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 static int update_block_group(struct btrfs_trans_handle *trans,
  64                               struct btrfs_root *root, u64 bytenr,
  65                               u64 num_bytes, int alloc);
  66 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  67                                 struct btrfs_root *root,
  68                                 struct btrfs_delayed_ref_node *node, u64 parent,
  69                                 u64 root_objectid, u64 owner_objectid,
  70                                 u64 owner_offset, int refs_to_drop,
  71                                 struct btrfs_delayed_extent_op *extra_op);
  72 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  73                                     struct extent_buffer *leaf,
  74                                     struct btrfs_extent_item *ei);
  75 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  76                                       struct btrfs_root *root,
  77                                       u64 parent, u64 root_objectid,
  78                                       u64 flags, u64 owner, u64 offset,
  79                                       struct btrfs_key *ins, int ref_mod);
  80 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  81                                      struct btrfs_root *root,
  82                                      u64 parent, u64 root_objectid,
  83                                      u64 flags, struct btrfs_disk_key *key,
  84                                      int level, struct btrfs_key *ins);
  85 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
  86                           struct btrfs_root *extent_root, u64 flags,
  87                           int force);
  88 static int find_next_key(struct btrfs_path *path, int level,
  89                          struct btrfs_key *key);
  90 static void dump_space_info(struct btrfs_fs_info *fs_info,
  91                             struct btrfs_space_info *info, u64 bytes,
  92                             int dump_block_groups);
  93 static int btrfs_add_reserved_bytes(struct btrfs_block_group_cache *cache,
  94                                     u64 ram_bytes, u64 num_bytes, int delalloc);
  95 static int btrfs_free_reserved_bytes(struct btrfs_block_group_cache *cache,
  96                                      u64 num_bytes, int delalloc);
  97 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
  98                                u64 num_bytes);
  99 int btrfs_pin_extent(struct btrfs_root *root,
 100                      u64 bytenr, u64 num_bytes, int reserved);
 101 static int __reserve_metadata_bytes(struct btrfs_root *root,
 102                                     struct btrfs_space_info *space_info,
 103                                     u64 orig_bytes,
 104                                     enum btrfs_reserve_flush_enum flush);
 105 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
 106                                      struct btrfs_space_info *space_info,
 107                                      u64 num_bytes);
 108 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
 109                                      struct btrfs_space_info *space_info,
 110                                      u64 num_bytes);
 111
 112 static noinline int
 113 block_group_cache_done(struct btrfs_block_group_cache *cache)
 114 {
 115         smp_mb();
 116         return cache->cached == BTRFS_CACHE_FINISHED ||
 117                 cache->cached == BTRFS_CACHE_ERROR;
 118 }
 119
 120 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
 121 {
 122         return (cache->flags & bits) == bits;
 123 }
 124
 125 void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 126 {
 127         atomic_inc(&cache->count);
 128 }
 129
 130 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 131 {
 132         if (atomic_dec_and_test(&cache->count)) {
 133                 WARN_ON(cache->pinned > 0);
 134                 WARN_ON(cache->reserved > 0);
 135                 kfree(cache->free_space_ctl);
 136                 kfree(cache);
 137         }
 138 }
 139
 140 /*
 141  * this adds the block group to the fs_info rb tree for the block group
 142  * cache
 143  */
 144 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 145                                 struct btrfs_block_group_cache *block_group)
 146 {
 147         struct rb_node **p;
 148         struct rb_node *parent = NULL;
 149         struct btrfs_block_group_cache *cache;
 150
 151         spin_lock(&info->block_group_cache_lock);
 152         p = &info->block_group_cache_tree.rb_node;
 153
 154         while (*p) {
 155                 parent = *p;
 156                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 157                                  cache_node);
 158                 if (block_group->key.objectid < cache->key.objectid) {
 159                         p = &(*p)->rb_left;
 160                 } else if (block_group->key.objectid > cache->key.objectid) {
 161                         p = &(*p)->rb_right;
 162                 } else {
 163                         spin_unlock(&info->block_group_cache_lock);
 164                         return -EEXIST;
 165                 }
 166         }
 167
 168         rb_link_node(&block_group->cache_node, parent, p);
 169         rb_insert_color(&block_group->cache_node,
 170                         &info->block_group_cache_tree);
 171
 172         if (info->first_logical_byte > block_group->key.objectid)
 173                 info->first_logical_byte = block_group->key.objectid;
 174
 175         spin_unlock(&info->block_group_cache_lock);
 176
 177         return 0;
 178 }
 179
 180 /*
 181  * This will return the block group at or after bytenr if contains is 0, else
 182  * it will return the block group that contains the bytenr
 183  */
 184 static struct btrfs_block_group_cache *
 185 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 186                               int contains)
 187 {
 188         struct btrfs_block_group_cache *cache, *ret = NULL;
 189         struct rb_node *n;
 190         u64 end, start;
 191
 192         spin_lock(&info->block_group_cache_lock);
 193         n = info->block_group_cache_tree.rb_node;
 194
 195         while (n) {
 196                 cache = rb_entry(n, struct btrfs_block_group_cache,
 197                                  cache_node);
 198                 end = cache->key.objectid + cache->key.offset - 1;
 199                 start = cache->key.objectid;
 200
 201                 if (bytenr < start) {
 202                         if (!contains && (!ret || start < ret->key.objectid))
 203                                 ret = cache;
 204                         n = n->rb_left;
 205                 } else if (bytenr > start) {
 206                         if (contains && bytenr <= end) {
 207                                 ret = cache;
 208                                 break;
 209                         }
 210                         n = n->rb_right;
 211                 } else {
 212                         ret = cache;
 213                         break;
 214                 }
 215         }
 216         if (ret) {
 217                 btrfs_get_block_group(ret);
 218                 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
 219                         info->first_logical_byte = ret->key.objectid;
 220         }
 221         spin_unlock(&info->block_group_cache_lock);
 222
 223         return ret;
 224 }
 225
 226 static int add_excluded_extent(struct btrfs_root *root,
 227                                u64 start, u64 num_bytes)
 228 {
 229         u64 end = start + num_bytes - 1;
 230         set_extent_bits(&root->fs_info->freed_extents[0],
 231                         start, end, EXTENT_UPTODATE);
 232         set_extent_bits(&root->fs_info->freed_extents[1],
 233                         start, end, EXTENT_UPTODATE);
 234         return 0;
 235 }
 236
 237 static void free_excluded_extents(struct btrfs_root *root,
 238                                   struct btrfs_block_group_cache *cache)
 239 {
 240         u64 start, end;
 241
 242         start = cache->key.objectid;
 243         end = start + cache->key.offset - 1;
 244
 245         clear_extent_bits(&root->fs_info->freed_extents[0],
 246                           start, end, EXTENT_UPTODATE);
 247         clear_extent_bits(&root->fs_info->freed_extents[1],
 248                           start, end, EXTENT_UPTODATE);
 249 }
 250
 251 static int exclude_super_stripes(struct btrfs_root *root,
 252                                  struct btrfs_block_group_cache *cache)
 253 {
 254         u64 bytenr;
 255         u64 *logical;
 256         int stripe_len;
 257         int i, nr, ret;
 258
 259         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 260                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 261                 cache->bytes_super += stripe_len;
 262                 ret = add_excluded_extent(root, cache->key.objectid,
 263                                           stripe_len);
 264                 if (ret)
 265                         return ret;
 266         }
 267
 268         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 269                 bytenr = btrfs_sb_offset(i);
 270                 ret = btrfs_rmap_block(root->fs_info, cache->key.objectid,
 271                                        bytenr, 0, &logical, &nr, &stripe_len);
 272                 if (ret)
 273                         return ret;
 274
 275                 while (nr--) {
 276                         u64 start, len;
 277
 278                         if (logical[nr] > cache->key.objectid +
 279                             cache->key.offset)
 280                                 continue;
 281
 282                         if (logical[nr] + stripe_len <= cache->key.objectid)
 283                                 continue;
 284
 285                         start = logical[nr];
 286                         if (start < cache->key.objectid) {
 287                                 start = cache->key.objectid;
 288                                 len = (logical[nr] + stripe_len) - start;
 289                         } else {
 290                                 len = min_t(u64, stripe_len,
 291                                             cache->key.objectid +
 292                                             cache->key.offset - start);
 293                         }
 294
 295                         cache->bytes_super += len;
 296                         ret = add_excluded_extent(root, start, len);
 297                         if (ret) {
 298                                 kfree(logical);
 299                                 return ret;
 300                         }
 301                 }
 302
 303                 kfree(logical);
 304         }
 305         return 0;
 306 }
 307
 308 static struct btrfs_caching_control *
 309 get_caching_control(struct btrfs_block_group_cache *cache)
 310 {
 311         struct btrfs_caching_control *ctl;
 312
 313         spin_lock(&cache->lock);
 314         if (!cache->caching_ctl) {
 315                 spin_unlock(&cache->lock);
 316                 return NULL;
 317         }
 318
 319         ctl = cache->caching_ctl;
 320         atomic_inc(&ctl->count);
 321         spin_unlock(&cache->lock);
 322         return ctl;
 323 }
 324
 325 static void put_caching_control(struct btrfs_caching_control *ctl)
 326 {
 327         if (atomic_dec_and_test(&ctl->count))
 328                 kfree(ctl);
 329 }
 330
 331 #ifdef CONFIG_BTRFS_DEBUG
 332 static void fragment_free_space(struct btrfs_root *root,
 333                                 struct btrfs_block_group_cache *block_group)
 334 {
 335         u64 start = block_group->key.objectid;
 336         u64 len = block_group->key.offset;
 337         u64 chunk = block_group->flags & BTRFS_BLOCK_GROUP_METADATA ?
 338                 root->nodesize : root->sectorsize;
 339         u64 step = chunk << 1;
 340
 341         while (len > chunk) {
 342                 btrfs_remove_free_space(block_group, start, chunk);
 343                 start += step;
 344                 if (len < step)
 345                         len = 0;
 346                 else
 347                         len -= step;
 348         }
 349 }
 350 #endif
 351
 352 /*
 353  * this is only called by cache_block_group, since we could have freed extents
 354  * we need to check the pinned_extents for any extents that can't be used yet
 355  * since their free space will be released as soon as the transaction commits.
 356  */
 357 u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 358                        struct btrfs_fs_info *info, u64 start, u64 end)
 359 {
 360         u64 extent_start, extent_end, size, total_added = 0;
 361         int ret;
 362
 363         while (start < end) {
 364                 ret = find_first_extent_bit(info->pinned_extents, start,
 365                                             &extent_start, &extent_end,
 366                                             EXTENT_DIRTY | EXTENT_UPTODATE,
 367                                             NULL);
 368                 if (ret)
 369                         break;
 370
 371                 if (extent_start <= start) {
 372                         start = extent_end + 1;
 373                 } else if (extent_start > start && extent_start < end) {
 374                         size = extent_start - start;
 375                         total_added += size;
 376                         ret = btrfs_add_free_space(block_group, start,
 377                                                    size);
 378                         BUG_ON(ret); /* -ENOMEM or logic error */
 379                         start = extent_end + 1;
 380                 } else {
 381                         break;
 382                 }
 383         }
 384
 385         if (start < end) {
 386                 size = end - start;
 387                 total_added += size;
 388                 ret = btrfs_add_free_space(block_group, start, size);
 389                 BUG_ON(ret); /* -ENOMEM or logic error */
 390         }
 391
 392         return total_added;
 393 }
 394
 395 static int load_extent_tree_free(struct btrfs_caching_control *caching_ctl)
 396 {
 397         struct btrfs_block_group_cache *block_group;
 398         struct btrfs_fs_info *fs_info;
 399         struct btrfs_root *extent_root;
 400         struct btrfs_path *path;
 401         struct extent_buffer *leaf;
 402         struct btrfs_key key;
 403         u64 total_found = 0;
 404         u64 last = 0;
 405         u32 nritems;
 406         int ret;
 407         bool wakeup = true;
 408
 409         block_group = caching_ctl->block_group;
 410         fs_info = block_group->fs_info;
 411         extent_root = fs_info->extent_root;
 412
 413         path = btrfs_alloc_path();
 414         if (!path)
 415                 return -ENOMEM;
 416
 417         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 418
 419 #ifdef CONFIG_BTRFS_DEBUG
 420         /*
 421          * If we're fragmenting we don't want to make anybody think we can
 422          * allocate from this block group until we've had a chance to fragment
 423          * the free space.
 424          */
 425         if (btrfs_should_fragment_free_space(extent_root, block_group))
 426                 wakeup = false;
 427 #endif
 428         /*
 429          * We don't want to deadlock with somebody trying to allocate a new
 430          * extent for the extent root while also trying to search the extent
 431          * root to add free space.  So we skip locking and search the commit
 432          * root, since its read-only
 433          */
 434         path->skip_locking = 1;
 435         path->search_commit_root = 1;
 436         path->reada = READA_FORWARD;
 437
 438         key.objectid = last;
 439         key.offset = 0;
 440         key.type = BTRFS_EXTENT_ITEM_KEY;
 441
 442 next:
 443         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 444         if (ret < 0)
 445                 goto out;
 446
 447         leaf = path->nodes[0];
 448         nritems = btrfs_header_nritems(leaf);
 449
 450         while (1) {
 451                 if (btrfs_fs_closing(fs_info) > 1) {
 452                         last = (u64)-1;
 453                         break;
 454                 }
 455
 456                 if (path->slots[0] < nritems) {
 457                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 458                 } else {
 459                         ret = find_next_key(path, 0, &key);
 460                         if (ret)
 461                                 break;
 462
 463                         if (need_resched() ||
 464                             rwsem_is_contended(&fs_info->commit_root_sem)) {
 465                                 if (wakeup)
 466                                         caching_ctl->progress = last;
 467                                 btrfs_release_path(path);
 468                                 up_read(&fs_info->commit_root_sem);
 469                                 mutex_unlock(&caching_ctl->mutex);
 470                                 cond_resched();
 471                                 mutex_lock(&caching_ctl->mutex);
 472                                 down_read(&fs_info->commit_root_sem);
 473                                 goto next;
 474                         }
 475
 476                         ret = btrfs_next_leaf(extent_root, path);
 477                         if (ret < 0)
 478                                 goto out;
 479                         if (ret)
 480                                 break;
 481                         leaf = path->nodes[0];
 482                         nritems = btrfs_header_nritems(leaf);
 483                         continue;
 484                 }
 485
 486                 if (key.objectid < last) {
 487                         key.objectid = last;
 488                         key.offset = 0;
 489                         key.type = BTRFS_EXTENT_ITEM_KEY;
 490
 491                         if (wakeup)
 492                                 caching_ctl->progress = last;
 493                         btrfs_release_path(path);
 494                         goto next;
 495                 }
 496
 497                 if (key.objectid < block_group->key.objectid) {
 498                         path->slots[0]++;
 499                         continue;
 500                 }
 501
 502                 if (key.objectid >= block_group->key.objectid +
 503                     block_group->key.offset)
 504                         break;
 505
 506                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 507                     key.type == BTRFS_METADATA_ITEM_KEY) {
 508                         total_found += add_new_free_space(block_group,
 509                                                           fs_info, last,
 510                                                           key.objectid);
 511                         if (key.type == BTRFS_METADATA_ITEM_KEY)
 512                                 last = key.objectid +
 513                                         fs_info->tree_root->nodesize;
 514                         else
 515                                 last = key.objectid + key.offset;
 516
 517                         if (total_found > CACHING_CTL_WAKE_UP) {
 518                                 total_found = 0;
 519                                 if (wakeup)
 520                                         wake_up(&caching_ctl->wait);
 521                         }
 522                 }
 523                 path->slots[0]++;
 524         }
 525         ret = 0;
 526
 527         total_found += add_new_free_space(block_group, fs_info, last,
 528                                           block_group->key.objectid +
 529                                           block_group->key.offset);
 530         caching_ctl->progress = (u64)-1;
 531
 532 out:
 533         btrfs_free_path(path);
 534         return ret;
 535 }
 536
 537 static noinline void caching_thread(struct btrfs_work *work)
 538 {
 539         struct btrfs_block_group_cache *block_group;
 540         struct btrfs_fs_info *fs_info;
 541         struct btrfs_caching_control *caching_ctl;
 542         struct btrfs_root *extent_root;
 543         int ret;
 544
 545         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 546         block_group = caching_ctl->block_group;
 547         fs_info = block_group->fs_info;
 548         extent_root = fs_info->extent_root;
 549
 550         mutex_lock(&caching_ctl->mutex);
 551         down_read(&fs_info->commit_root_sem);
 552
 553         if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
 554                 ret = load_free_space_tree(caching_ctl);
 555         else
 556                 ret = load_extent_tree_free(caching_ctl);
 557
 558         spin_lock(&block_group->lock);
 559         block_group->caching_ctl = NULL;
 560         block_group->cached = ret ? BTRFS_CACHE_ERROR : BTRFS_CACHE_FINISHED;
 561         spin_unlock(&block_group->lock);
 562
 563 #ifdef CONFIG_BTRFS_DEBUG
 564         if (btrfs_should_fragment_free_space(extent_root, block_group)) {
 565                 u64 bytes_used;
 566
 567                 spin_lock(&block_group->space_info->lock);
 568                 spin_lock(&block_group->lock);
 569                 bytes_used = block_group->key.offset -
 570                         btrfs_block_group_used(&block_group->item);
 571                 block_group->space_info->bytes_used += bytes_used >> 1;
 572                 spin_unlock(&block_group->lock);
 573                 spin_unlock(&block_group->space_info->lock);
 574                 fragment_free_space(extent_root, block_group);
 575         }
 576 #endif
 577
 578         caching_ctl->progress = (u64)-1;
 579
 580         up_read(&fs_info->commit_root_sem);
 581         free_excluded_extents(fs_info->extent_root, block_group);
 582         mutex_unlock(&caching_ctl->mutex);
 583
 584         wake_up(&caching_ctl->wait);
 585
 586         put_caching_control(caching_ctl);
 587         btrfs_put_block_group(block_group);
 588 }
 589
 590 static int cache_block_group(struct btrfs_block_group_cache *cache,
 591                              int load_cache_only)
 592 {
 593         DEFINE_WAIT(wait);
 594         struct btrfs_fs_info *fs_info = cache->fs_info;
 595         struct btrfs_caching_control *caching_ctl;
 596         int ret = 0;
 597
 598         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 599         if (!caching_ctl)
 600                 return -ENOMEM;
 601
 602         INIT_LIST_HEAD(&caching_ctl->list);
 603         mutex_init(&caching_ctl->mutex);
 604         init_waitqueue_head(&caching_ctl->wait);
 605         caching_ctl->block_group = cache;
 606         caching_ctl->progress = cache->key.objectid;
 607         atomic_set(&caching_ctl->count, 1);
 608         btrfs_init_work(&caching_ctl->work, btrfs_cache_helper,
 609                         caching_thread, NULL, NULL);
 610
 611         spin_lock(&cache->lock);
 612         /*
 613          * This should be a rare occasion, but this could happen I think in the
 614          * case where one thread starts to load the space cache info, and then
 615          * some other thread starts a transaction commit which tries to do an
 616          * allocation while the other thread is still loading the space cache
 617          * info.  The previous loop should have kept us from choosing this block
 618          * group, but if we've moved to the state where we will wait on caching
 619          * block groups we need to first check if we're doing a fast load here,
 620          * so we can wait for it to finish, otherwise we could end up allocating
 621          * from a block group who's cache gets evicted for one reason or
 622          * another.
 623          */
 624         while (cache->cached == BTRFS_CACHE_FAST) {
 625                 struct btrfs_caching_control *ctl;
 626
 627                 ctl = cache->caching_ctl;
 628                 atomic_inc(&ctl->count);
 629                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 630                 spin_unlock(&cache->lock);
 631
 632                 schedule();
 633
 634                 finish_wait(&ctl->wait, &wait);
 635                 put_caching_control(ctl);
 636                 spin_lock(&cache->lock);
 637         }
 638
 639         if (cache->cached != BTRFS_CACHE_NO) {
 640                 spin_unlock(&cache->lock);
 641                 kfree(caching_ctl);
 642                 return 0;
 643         }
 644         WARN_ON(cache->caching_ctl);
 645         cache->caching_ctl = caching_ctl;
 646         cache->cached = BTRFS_CACHE_FAST;
 647         spin_unlock(&cache->lock);
 648
 649         if (fs_info->mount_opt & BTRFS_MOUNT_SPACE_CACHE) {
 650                 mutex_lock(&caching_ctl->mutex);
 651                 ret = load_free_space_cache(fs_info, cache);
 652
 653                 spin_lock(&cache->lock);
 654                 if (ret == 1) {
 655                         cache->caching_ctl = NULL;
 656                         cache->cached = BTRFS_CACHE_FINISHED;
 657                         cache->last_byte_to_unpin = (u64)-1;
 658                         caching_ctl->progress = (u64)-1;
 659                 } else {
 660                         if (load_cache_only) {
 661                                 cache->caching_ctl = NULL;
 662                                 cache->cached = BTRFS_CACHE_NO;
 663                         } else {
 664                                 cache->cached = BTRFS_CACHE_STARTED;
 665                                 cache->has_caching_ctl = 1;
 666                         }
 667                 }
 668                 spin_unlock(&cache->lock);
 669 #ifdef CONFIG_BTRFS_DEBUG
 670                 if (ret == 1 &&
 671                     btrfs_should_fragment_free_space(fs_info->extent_root,
 672                                                      cache)) {
 673                         u64 bytes_used;
 674
 675                         spin_lock(&cache->space_info->lock);
 676                         spin_lock(&cache->lock);
 677                         bytes_used = cache->key.offset -
 678                                 btrfs_block_group_used(&cache->item);
 679                         cache->space_info->bytes_used += bytes_used >> 1;
 680                         spin_unlock(&cache->lock);
 681                         spin_unlock(&cache->space_info->lock);
 682                         fragment_free_space(fs_info->extent_root, cache);
 683                 }
 684 #endif
 685                 mutex_unlock(&caching_ctl->mutex);
 686
 687                 wake_up(&caching_ctl->wait);
 688                 if (ret == 1) {
 689                         put_caching_control(caching_ctl);
 690                         free_excluded_extents(fs_info->extent_root, cache);
 691                         return 0;
 692                 }
 693         } else {
 694                 /*
 695                  * We're either using the free space tree or no caching at all.
 696                  * Set cached to the appropriate value and wakeup any waiters.
 697                  */
 698                 spin_lock(&cache->lock);
 699                 if (load_cache_only) {
 700                         cache->caching_ctl = NULL;
 701                         cache->cached = BTRFS_CACHE_NO;
 702                 } else {
 703                         cache->cached = BTRFS_CACHE_STARTED;
 704                         cache->has_caching_ctl = 1;
 705                 }
 706                 spin_unlock(&cache->lock);
 707                 wake_up(&caching_ctl->wait);
 708         }
 709
 710         if (load_cache_only) {
 711                 put_caching_control(caching_ctl);
 712                 return 0;
 713         }
 714
 715         down_write(&fs_info->commit_root_sem);
 716         atomic_inc(&caching_ctl->count);
 717         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 718         up_write(&fs_info->commit_root_sem);
 719
 720         btrfs_get_block_group(cache);
 721
 722         btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
 723
 724         return ret;
 725 }
 726
 727 /*
 728  * return the block group that starts at or after bytenr
 729  */
 730 static struct btrfs_block_group_cache *
 731 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 732 {
 733         return block_group_cache_tree_search(info, bytenr, 0);
 734 }
 735
 736 /*
 737  * return the block group that contains the given bytenr
 738  */
 739 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 740                                                  struct btrfs_fs_info *info,
 741                                                  u64 bytenr)
 742 {
 743         return block_group_cache_tree_search(info, bytenr, 1);
 744 }
 745
 746 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 747                                                   u64 flags)
 748 {
 749         struct list_head *head = &info->space_info;
 750         struct btrfs_space_info *found;
 751
 752         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 753
 754         rcu_read_lock();
 755         list_for_each_entry_rcu(found, head, list) {
 756                 if (found->flags & flags) {
 757                         rcu_read_unlock();
 758                         return found;
 759                 }
 760         }
 761         rcu_read_unlock();
 762         return NULL;
 763 }
 764
 765 /*
 766  * after adding space to the filesystem, we need to clear the full flags
 767  * on all the space infos.
 768  */
 769 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 770 {
 771         struct list_head *head = &info->space_info;
 772         struct btrfs_space_info *found;
 773
 774         rcu_read_lock();
 775         list_for_each_entry_rcu(found, head, list)
 776                 found->full = 0;
 777         rcu_read_unlock();
 778 }
 779
 780 /* simple helper to search for an existing data extent at a given offset */
 781 int btrfs_lookup_data_extent(struct btrfs_root *root, u64 start, u64 len)
 782 {
 783         int ret;
 784         struct btrfs_key key;
 785         struct btrfs_path *path;
 786
 787         path = btrfs_alloc_path();
 788         if (!path)
 789                 return -ENOMEM;
 790
 791         key.objectid = start;
 792         key.offset = len;
 793         key.type = BTRFS_EXTENT_ITEM_KEY;
 794         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
 795                                 0, 0);
 796         btrfs_free_path(path);
 797         return ret;
 798 }
 799
 800 /*
 801  * helper function to lookup reference count and flags of a tree block.
 802  *
 803  * the head node for delayed ref is used to store the sum of all the
 804  * reference count modifications queued up in the rbtree. the head
 805  * node may also store the extent flags to set. This way you can check
 806  * to see what the reference count and extent flags would be if all of
 807  * the delayed refs are not processed.
 808  */
 809 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 810                              struct btrfs_root *root, u64 bytenr,
 811                              u64 offset, int metadata, u64 *refs, u64 *flags)
 812 {
 813         struct btrfs_delayed_ref_head *head;
 814         struct btrfs_delayed_ref_root *delayed_refs;
 815         struct btrfs_path *path;
 816         struct btrfs_extent_item *ei;
 817         struct extent_buffer *leaf;
 818         struct btrfs_key key;
 819         u32 item_size;
 820         u64 num_refs;
 821         u64 extent_flags;
 822         int ret;
 823
 824         /*
 825          * If we don't have skinny metadata, don't bother doing anything
 826          * different
 827          */
 828         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA)) {
 829                 offset = root->nodesize;
 830                 metadata = 0;
 831         }
 832
 833         path = btrfs_alloc_path();
 834         if (!path)
 835                 return -ENOMEM;
 836
 837         if (!trans) {
 838                 path->skip_locking = 1;
 839                 path->search_commit_root = 1;
 840         }
 841
 842 search_again:
 843         key.objectid = bytenr;
 844         key.offset = offset;
 845         if (metadata)
 846                 key.type = BTRFS_METADATA_ITEM_KEY;
 847         else
 848                 key.type = BTRFS_EXTENT_ITEM_KEY;
 849
 850         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
 851                                 &key, path, 0, 0);
 852         if (ret < 0)
 853                 goto out_free;
 854
 855         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 856                 if (path->slots[0]) {
 857                         path->slots[0]--;
 858                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 859                                               path->slots[0]);
 860                         if (key.objectid == bytenr &&
 861                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 862                             key.offset == root->nodesize)
 863                                 ret = 0;
 864                 }
 865         }
 866
 867         if (ret == 0) {
 868                 leaf = path->nodes[0];
 869                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 870                 if (item_size >= sizeof(*ei)) {
 871                         ei = btrfs_item_ptr(leaf, path->slots[0],
 872                                             struct btrfs_extent_item);
 873                         num_refs = btrfs_extent_refs(leaf, ei);
 874                         extent_flags = btrfs_extent_flags(leaf, ei);
 875                 } else {
 876 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 877                         struct btrfs_extent_item_v0 *ei0;
 878                         BUG_ON(item_size != sizeof(*ei0));
 879                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 880                                              struct btrfs_extent_item_v0);
 881                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
 882                         /* FIXME: this isn't correct for data */
 883                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 884 #else
 885                         BUG();
 886 #endif
 887                 }
 888                 BUG_ON(num_refs == 0);
 889         } else {
 890                 num_refs = 0;
 891                 extent_flags = 0;
 892                 ret = 0;
 893         }
 894
 895         if (!trans)
 896                 goto out;
 897
 898         delayed_refs = &trans->transaction->delayed_refs;
 899         spin_lock(&delayed_refs->lock);
 900         head = btrfs_find_delayed_ref_head(trans, bytenr);
 901         if (head) {
 902                 if (!mutex_trylock(&head->mutex)) {
 903                         atomic_inc(&head->node.refs);
 904                         spin_unlock(&delayed_refs->lock);
 905
 906                         btrfs_release_path(path);
 907
 908                         /*
 909                          * Mutex was contended, block until it's released and try
 910                          * again
 911                          */
 912                         mutex_lock(&head->mutex);
 913                         mutex_unlock(&head->mutex);
 914                         btrfs_put_delayed_ref(&head->node);
 915                         goto search_again;
 916                 }
 917                 spin_lock(&head->lock);
 918                 if (head->extent_op && head->extent_op->update_flags)
 919                         extent_flags |= head->extent_op->flags_to_set;
 920                 else
 921                         BUG_ON(num_refs == 0);
 922
 923                 num_refs += head->node.ref_mod;
 924                 spin_unlock(&head->lock);
 925                 mutex_unlock(&head->mutex);
 926         }
 927         spin_unlock(&delayed_refs->lock);
 928 out:
 929         WARN_ON(num_refs == 0);
 930         if (refs)
 931                 *refs = num_refs;
 932         if (flags)
 933                 *flags = extent_flags;
 934 out_free:
 935         btrfs_free_path(path);
 936         return ret;
 937 }
 938
 939 /*
 940  * Back reference rules.  Back refs have three main goals:
 941  *
 942  * 1) differentiate between all holders of references to an extent so that
 943  *    when a reference is dropped we can make sure it was a valid reference
 944  *    before freeing the extent.
 945  *
 946  * 2) Provide enough information to quickly find the holders of an extent
 947  *    if we notice a given block is corrupted or bad.
 948  *
 949  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 950  *    maintenance.  This is actually the same as #2, but with a slightly
 951  *    different use case.
 952  *
 953  * There are two kinds of back refs. The implicit back refs is optimized
 954  * for pointers in non-shared tree blocks. For a given pointer in a block,
 955  * back refs of this kind provide information about the block's owner tree
 956  * and the pointer's key. These information allow us to find the block by
 957  * b-tree searching. The full back refs is for pointers in tree blocks not
 958  * referenced by their owner trees. The location of tree block is recorded
 959  * in the back refs. Actually the full back refs is generic, and can be
 960  * used in all cases the implicit back refs is used. The major shortcoming
 961  * of the full back refs is its overhead. Every time a tree block gets
 962  * COWed, we have to update back refs entry for all pointers in it.
 963  *
 964  * For a newly allocated tree block, we use implicit back refs for
 965  * pointers in it. This means most tree related operations only involve
 966  * implicit back refs. For a tree block created in old transaction, the
 967  * only way to drop a reference to it is COW it. So we can detect the
 968  * event that tree block loses its owner tree's reference and do the
 969  * back refs conversion.
 970  *
 971  * When a tree block is COWed through a tree, there are four cases:
 972  *
 973  * The reference count of the block is one and the tree is the block's
 974  * owner tree. Nothing to do in this case.
 975  *
 976  * The reference count of the block is one and the tree is not the
 977  * block's owner tree. In this case, full back refs is used for pointers
 978  * in the block. Remove these full back refs, add implicit back refs for
 979  * every pointers in the new block.
 980  *
 981  * The reference count of the block is greater than one and the tree is
 982  * the block's owner tree. In this case, implicit back refs is used for
 983  * pointers in the block. Add full back refs for every pointers in the
 984  * block, increase lower level extents' reference counts. The original
 985  * implicit back refs are entailed to the new block.
 986  *
 987  * The reference count of the block is greater than one and the tree is
 988  * not the block's owner tree. Add implicit back refs for every pointer in
 989  * the new block, increase lower level extents' reference count.
 990  *
 991  * Back Reference Key composing:
 992  *
 993  * The key objectid corresponds to the first byte in the extent,
 994  * The key type is used to differentiate between types of back refs.
 995  * There are different meanings of the key offset for different types
 996  * of back refs.
 997  *
 998  * File extents can be referenced by:
 999  *
1000  * - multiple snapshots, subvolumes, or different generations in one subvol
1001  * - different files inside a single subvolume
1002  * - different offsets inside a file (bookend extents in file.c)
1003  *
1004  * The extent ref structure for the implicit back refs has fields for:
1005  *
1006  * - Objectid of the subvolume root
1007  * - objectid of the file holding the reference
1008  * - original offset in the file
1009  * - how many bookend extents
1010  *
1011  * The key offset for the implicit back refs is hash of the first
1012  * three fields.
1013  *
1014  * The extent ref structure for the full back refs has field for:
1015  *
1016  * - number of pointers in the tree leaf
1017  *
1018  * The key offset for the implicit back refs is the first byte of
1019  * the tree leaf
1020  *
1021  * When a file extent is allocated, The implicit back refs is used.
1022  * the fields are filled in:
1023  *
1024  *     (root_key.objectid, inode objectid, offset in file, 1)
1025  *
1026  * When a file extent is removed file truncation, we find the
1027  * corresponding implicit back refs and check the following fields:
1028  *
1029  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
1030  *
1031  * Btree extents can be referenced by:
1032  *
1033  * - Different subvolumes
1034  *
1035  * Both the implicit back refs and the full back refs for tree blocks
1036  * only consist of key. The key offset for the implicit back refs is
1037  * objectid of block's owner tree. The key offset for the full back refs
1038  * is the first byte of parent block.
1039  *
1040  * When implicit back refs is used, information about the lowest key and
1041  * level of the tree block are required. These information are stored in
1042  * tree block info structure.
1043  */
1044
1045 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1046 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
1047                                   struct btrfs_root *root,
1048                                   struct btrfs_path *path,
1049                                   u64 owner, u32 extra_size)
1050 {
1051         struct btrfs_extent_item *item;
1052         struct btrfs_extent_item_v0 *ei0;
1053         struct btrfs_extent_ref_v0 *ref0;
1054         struct btrfs_tree_block_info *bi;
1055         struct extent_buffer *leaf;
1056         struct btrfs_key key;
1057         struct btrfs_key found_key;
1058         u32 new_size = sizeof(*item);
1059         u64 refs;
1060         int ret;
1061
1062         leaf = path->nodes[0];
1063         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
1064
1065         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1066         ei0 = btrfs_item_ptr(leaf, path->slots[0],
1067                              struct btrfs_extent_item_v0);
1068         refs = btrfs_extent_refs_v0(leaf, ei0);
1069
1070         if (owner == (u64)-1) {
1071                 while (1) {
1072                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1073                                 ret = btrfs_next_leaf(root, path);
1074                                 if (ret < 0)
1075                                         return ret;
1076                                 BUG_ON(ret > 0); /* Corruption */
1077                                 leaf = path->nodes[0];
1078                         }
1079                         btrfs_item_key_to_cpu(leaf, &found_key,
1080                                               path->slots[0]);
1081                         BUG_ON(key.objectid != found_key.objectid);
1082                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
1083                                 path->slots[0]++;
1084                                 continue;
1085                         }
1086                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1087                                               struct btrfs_extent_ref_v0);
1088                         owner = btrfs_ref_objectid_v0(leaf, ref0);
1089                         break;
1090                 }
1091         }
1092         btrfs_release_path(path);
1093
1094         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1095                 new_size += sizeof(*bi);
1096
1097         new_size -= sizeof(*ei0);
1098         ret = btrfs_search_slot(trans, root, &key, path,
1099                                 new_size + extra_size, 1);
1100         if (ret < 0)
1101                 return ret;
1102         BUG_ON(ret); /* Corruption */
1103
1104         btrfs_extend_item(root, path, new_size);
1105
1106         leaf = path->nodes[0];
1107         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1108         btrfs_set_extent_refs(leaf, item, refs);
1109         /* FIXME: get real generation */
1110         btrfs_set_extent_generation(leaf, item, 0);
1111         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1112                 btrfs_set_extent_flags(leaf, item,
1113                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1114                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1115                 bi = (struct btrfs_tree_block_info *)(item + 1);
1116                 /* FIXME: get first key of the block */
1117                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
1118                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1119         } else {
1120                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1121         }
1122         btrfs_mark_buffer_dirty(leaf);
1123         return 0;
1124 }
1125 #endif
1126
1127 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1128 {
1129         u32 high_crc = ~(u32)0;
1130         u32 low_crc = ~(u32)0;
1131         __le64 lenum;
1132
1133         lenum = cpu_to_le64(root_objectid);
1134         high_crc = btrfs_crc32c(high_crc, &lenum, sizeof(lenum));
1135         lenum = cpu_to_le64(owner);
1136         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
1137         lenum = cpu_to_le64(offset);
1138         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
1139
1140         return ((u64)high_crc << 31) ^ (u64)low_crc;
1141 }
1142
1143 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1144                                      struct btrfs_extent_data_ref *ref)
1145 {
1146         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1147                                     btrfs_extent_data_ref_objectid(leaf, ref),
1148                                     btrfs_extent_data_ref_offset(leaf, ref));
1149 }
1150
1151 static int match_extent_data_ref(struct extent_buffer *leaf,
1152                                  struct btrfs_extent_data_ref *ref,
1153                                  u64 root_objectid, u64 owner, u64 offset)
1154 {
1155         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1156             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1157             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1158                 return 0;
1159         return 1;
1160 }
1161
1162 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1163                                            struct btrfs_root *root,
1164                                            struct btrfs_path *path,
1165                                            u64 bytenr, u64 parent,
1166                                            u64 root_objectid,
1167                                            u64 owner, u64 offset)
1168 {
1169         struct btrfs_key key;
1170         struct btrfs_extent_data_ref *ref;
1171         struct extent_buffer *leaf;
1172         u32 nritems;
1173         int ret;
1174         int recow;
1175         int err = -ENOENT;
1176
1177         key.objectid = bytenr;
1178         if (parent) {
1179                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1180                 key.offset = parent;
1181         } else {
1182                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1183                 key.offset = hash_extent_data_ref(root_objectid,
1184                                                   owner, offset);
1185         }
1186 again:
1187         recow = 0;
1188         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1189         if (ret < 0) {
1190                 err = ret;
1191                 goto fail;
1192         }
1193
1194         if (parent) {
1195                 if (!ret)
1196                         return 0;
1197 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1198                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1199                 btrfs_release_path(path);
1200                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1201                 if (ret < 0) {
1202                         err = ret;
1203                         goto fail;
1204                 }
1205                 if (!ret)
1206                         return 0;
1207 #endif
1208                 goto fail;
1209         }
1210
1211         leaf = path->nodes[0];
1212         nritems = btrfs_header_nritems(leaf);
1213         while (1) {
1214                 if (path->slots[0] >= nritems) {
1215                         ret = btrfs_next_leaf(root, path);
1216                         if (ret < 0)
1217                                 err = ret;
1218                         if (ret)
1219                                 goto fail;
1220
1221                         leaf = path->nodes[0];
1222                         nritems = btrfs_header_nritems(leaf);
1223                         recow = 1;
1224                 }
1225
1226                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1227                 if (key.objectid != bytenr ||
1228                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1229                         goto fail;
1230
1231                 ref = btrfs_item_ptr(leaf, path->slots[0],
1232                                      struct btrfs_extent_data_ref);
1233
1234                 if (match_extent_data_ref(leaf, ref, root_objectid,
1235                                           owner, offset)) {
1236                         if (recow) {
1237                                 btrfs_release_path(path);
1238                                 goto again;
1239                         }
1240                         err = 0;
1241                         break;
1242                 }
1243                 path->slots[0]++;
1244         }
1245 fail:
1246         return err;
1247 }
1248
1249 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1250                                            struct btrfs_root *root,
1251                                            struct btrfs_path *path,
1252                                            u64 bytenr, u64 parent,
1253                                            u64 root_objectid, u64 owner,
1254                                            u64 offset, int refs_to_add)
1255 {
1256         struct btrfs_key key;
1257         struct extent_buffer *leaf;
1258         u32 size;
1259         u32 num_refs;
1260         int ret;
1261
1262         key.objectid = bytenr;
1263         if (parent) {
1264                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1265                 key.offset = parent;
1266                 size = sizeof(struct btrfs_shared_data_ref);
1267         } else {
1268                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1269                 key.offset = hash_extent_data_ref(root_objectid,
1270                                                   owner, offset);
1271                 size = sizeof(struct btrfs_extent_data_ref);
1272         }
1273
1274         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1275         if (ret && ret != -EEXIST)
1276                 goto fail;
1277
1278         leaf = path->nodes[0];
1279         if (parent) {
1280                 struct btrfs_shared_data_ref *ref;
1281                 ref = btrfs_item_ptr(leaf, path->slots[0],
1282                                      struct btrfs_shared_data_ref);
1283                 if (ret == 0) {
1284                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1285                 } else {
1286                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1287                         num_refs += refs_to_add;
1288                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1289                 }
1290         } else {
1291                 struct btrfs_extent_data_ref *ref;
1292                 while (ret == -EEXIST) {
1293                         ref = btrfs_item_ptr(leaf, path->slots[0],
1294                                              struct btrfs_extent_data_ref);
1295                         if (match_extent_data_ref(leaf, ref, root_objectid,
1296                                                   owner, offset))
1297                                 break;
1298                         btrfs_release_path(path);
1299                         key.offset++;
1300                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1301                                                       size);
1302                         if (ret && ret != -EEXIST)
1303                                 goto fail;
1304
1305                         leaf = path->nodes[0];
1306                 }
1307                 ref = btrfs_item_ptr(leaf, path->slots[0],
1308                                      struct btrfs_extent_data_ref);
1309                 if (ret == 0) {
1310                         btrfs_set_extent_data_ref_root(leaf, ref,
1311                                                        root_objectid);
1312                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1313                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1314                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1315                 } else {
1316                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1317                         num_refs += refs_to_add;
1318                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1319                 }
1320         }
1321         btrfs_mark_buffer_dirty(leaf);
1322         ret = 0;
1323 fail:
1324         btrfs_release_path(path);
1325         return ret;
1326 }
1327
1328 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1329                                            struct btrfs_root *root,
1330                                            struct btrfs_path *path,
1331                                            int refs_to_drop, int *last_ref)
1332 {
1333         struct btrfs_key key;
1334         struct btrfs_extent_data_ref *ref1 = NULL;
1335         struct btrfs_shared_data_ref *ref2 = NULL;
1336         struct extent_buffer *leaf;
1337         u32 num_refs = 0;
1338         int ret = 0;
1339
1340         leaf = path->nodes[0];
1341         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1342
1343         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1344                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1345                                       struct btrfs_extent_data_ref);
1346                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1347         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1348                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1349                                       struct btrfs_shared_data_ref);
1350                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1351 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1352         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1353                 struct btrfs_extent_ref_v0 *ref0;
1354                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1355                                       struct btrfs_extent_ref_v0);
1356                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1357 #endif
1358         } else {
1359                 BUG();
1360         }
1361
1362         BUG_ON(num_refs < refs_to_drop);
1363         num_refs -= refs_to_drop;
1364
1365         if (num_refs == 0) {
1366                 ret = btrfs_del_item(trans, root, path);
1367                 *last_ref = 1;
1368         } else {
1369                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1370                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1371                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1372                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1373 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1374                 else {
1375                         struct btrfs_extent_ref_v0 *ref0;
1376                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1377                                         struct btrfs_extent_ref_v0);
1378                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1379                 }
1380 #endif
1381                 btrfs_mark_buffer_dirty(leaf);
1382         }
1383         return ret;
1384 }
1385
1386 static noinline u32 extent_data_ref_count(struct btrfs_path *path,
1387                                           struct btrfs_extent_inline_ref *iref)
1388 {
1389         struct btrfs_key key;
1390         struct extent_buffer *leaf;
1391         struct btrfs_extent_data_ref *ref1;
1392         struct btrfs_shared_data_ref *ref2;
1393         u32 num_refs = 0;
1394
1395         leaf = path->nodes[0];
1396         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1397         if (iref) {
1398                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1399                     BTRFS_EXTENT_DATA_REF_KEY) {
1400                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1401                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1402                 } else {
1403                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1404                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1405                 }
1406         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1407                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1408                                       struct btrfs_extent_data_ref);
1409                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1410         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1411                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1412                                       struct btrfs_shared_data_ref);
1413                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1414 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1415         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1416                 struct btrfs_extent_ref_v0 *ref0;
1417                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1418                                       struct btrfs_extent_ref_v0);
1419                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1420 #endif
1421         } else {
1422                 WARN_ON(1);
1423         }
1424         return num_refs;
1425 }
1426
1427 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1428                                           struct btrfs_root *root,
1429                                           struct btrfs_path *path,
1430                                           u64 bytenr, u64 parent,
1431                                           u64 root_objectid)
1432 {
1433         struct btrfs_key key;
1434         int ret;
1435
1436         key.objectid = bytenr;
1437         if (parent) {
1438                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1439                 key.offset = parent;
1440         } else {
1441                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1442                 key.offset = root_objectid;
1443         }
1444
1445         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1446         if (ret > 0)
1447                 ret = -ENOENT;
1448 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1449         if (ret == -ENOENT && parent) {
1450                 btrfs_release_path(path);
1451                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1452                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1453                 if (ret > 0)
1454                         ret = -ENOENT;
1455         }
1456 #endif
1457         return ret;
1458 }
1459
1460 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1461                                           struct btrfs_root *root,
1462                                           struct btrfs_path *path,
1463                                           u64 bytenr, u64 parent,
1464                                           u64 root_objectid)
1465 {
1466         struct btrfs_key key;
1467         int ret;
1468
1469         key.objectid = bytenr;
1470         if (parent) {
1471                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1472                 key.offset = parent;
1473         } else {
1474                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1475                 key.offset = root_objectid;
1476         }
1477
1478         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1479         btrfs_release_path(path);
1480         return ret;
1481 }
1482
1483 static inline int extent_ref_type(u64 parent, u64 owner)
1484 {
1485         int type;
1486         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1487                 if (parent > 0)
1488                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1489                 else
1490                         type = BTRFS_TREE_BLOCK_REF_KEY;
1491         } else {
1492                 if (parent > 0)
1493                         type = BTRFS_SHARED_DATA_REF_KEY;
1494                 else
1495                         type = BTRFS_EXTENT_DATA_REF_KEY;
1496         }
1497         return type;
1498 }
1499
1500 static int find_next_key(struct btrfs_path *path, int level,
1501                          struct btrfs_key *key)
1502
1503 {
1504         for (; level < BTRFS_MAX_LEVEL; level++) {
1505                 if (!path->nodes[level])
1506                         break;
1507                 if (path->slots[level] + 1 >=
1508                     btrfs_header_nritems(path->nodes[level]))
1509                         continue;
1510                 if (level == 0)
1511                         btrfs_item_key_to_cpu(path->nodes[level], key,
1512                                               path->slots[level] + 1);
1513                 else
1514                         btrfs_node_key_to_cpu(path->nodes[level], key,
1515                                               path->slots[level] + 1);
1516                 return 0;
1517         }
1518         return 1;
1519 }
1520
1521 /*
1522  * look for inline back ref. if back ref is found, *ref_ret is set
1523  * to the address of inline back ref, and 0 is returned.
1524  *
1525  * if back ref isn't found, *ref_ret is set to the address where it
1526  * should be inserted, and -ENOENT is returned.
1527  *
1528  * if insert is true and there are too many inline back refs, the path
1529  * points to the extent item, and -EAGAIN is returned.
1530  *
1531  * NOTE: inline back refs are ordered in the same way that back ref
1532  *       items in the tree are ordered.
1533  */
1534 static noinline_for_stack
1535 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1536                                  struct btrfs_root *root,
1537                                  struct btrfs_path *path,
1538                                  struct btrfs_extent_inline_ref **ref_ret,
1539                                  u64 bytenr, u64 num_bytes,
1540                                  u64 parent, u64 root_objectid,
1541                                  u64 owner, u64 offset, int insert)
1542 {
1543         struct btrfs_key key;
1544         struct extent_buffer *leaf;
1545         struct btrfs_extent_item *ei;
1546         struct btrfs_extent_inline_ref *iref;
1547         u64 flags;
1548         u64 item_size;
1549         unsigned long ptr;
1550         unsigned long end;
1551         int extra_size;
1552         int type;
1553         int want;
1554         int ret;
1555         int err = 0;
1556         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
1557                                                  SKINNY_METADATA);
1558
1559         key.objectid = bytenr;
1560         key.type = BTRFS_EXTENT_ITEM_KEY;
1561         key.offset = num_bytes;
1562
1563         want = extent_ref_type(parent, owner);
1564         if (insert) {
1565                 extra_size = btrfs_extent_inline_ref_size(want);
1566                 path->keep_locks = 1;
1567         } else
1568                 extra_size = -1;
1569
1570         /*
1571          * Owner is our parent level, so we can just add one to get the level
1572          * for the block we are interested in.
1573          */
1574         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
1575                 key.type = BTRFS_METADATA_ITEM_KEY;
1576                 key.offset = owner;
1577         }
1578
1579 again:
1580         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1581         if (ret < 0) {
1582                 err = ret;
1583                 goto out;
1584         }
1585
1586         /*
1587          * We may be a newly converted file system which still has the old fat
1588          * extent entries for metadata, so try and see if we have one of those.
1589          */
1590         if (ret > 0 && skinny_metadata) {
1591                 skinny_metadata = false;
1592                 if (path->slots[0]) {
1593                         path->slots[0]--;
1594                         btrfs_item_key_to_cpu(path->nodes[0], &key,
1595                                               path->slots[0]);
1596                         if (key.objectid == bytenr &&
1597                             key.type == BTRFS_EXTENT_ITEM_KEY &&
1598                             key.offset == num_bytes)
1599                                 ret = 0;
1600                 }
1601                 if (ret) {
1602                         key.objectid = bytenr;
1603                         key.type = BTRFS_EXTENT_ITEM_KEY;
1604                         key.offset = num_bytes;
1605                         btrfs_release_path(path);
1606                         goto again;
1607                 }
1608         }
1609
1610         if (ret && !insert) {
1611                 err = -ENOENT;
1612                 goto out;
1613         } else if (WARN_ON(ret)) {
1614                 err = -EIO;
1615                 goto out;
1616         }
1617
1618         leaf = path->nodes[0];
1619         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1620 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1621         if (item_size < sizeof(*ei)) {
1622                 if (!insert) {
1623                         err = -ENOENT;
1624                         goto out;
1625                 }
1626                 ret = convert_extent_item_v0(trans, root, path, owner,
1627                                              extra_size);
1628                 if (ret < 0) {
1629                         err = ret;
1630                         goto out;
1631                 }
1632                 leaf = path->nodes[0];
1633                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1634         }
1635 #endif
1636         BUG_ON(item_size < sizeof(*ei));
1637
1638         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1639         flags = btrfs_extent_flags(leaf, ei);
1640
1641         ptr = (unsigned long)(ei + 1);
1642         end = (unsigned long)ei + item_size;
1643
1644         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
1645                 ptr += sizeof(struct btrfs_tree_block_info);
1646                 BUG_ON(ptr > end);
1647         }
1648
1649         err = -ENOENT;
1650         while (1) {
1651                 if (ptr >= end) {
1652                         WARN_ON(ptr > end);
1653                         break;
1654                 }
1655                 iref = (struct btrfs_extent_inline_ref *)ptr;
1656                 type = btrfs_extent_inline_ref_type(leaf, iref);
1657                 if (want < type)
1658                         break;
1659                 if (want > type) {
1660                         ptr += btrfs_extent_inline_ref_size(type);
1661                         continue;
1662                 }
1663
1664                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1665                         struct btrfs_extent_data_ref *dref;
1666                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1667                         if (match_extent_data_ref(leaf, dref, root_objectid,
1668                                                   owner, offset)) {
1669                                 err = 0;
1670                                 break;
1671                         }
1672                         if (hash_extent_data_ref_item(leaf, dref) <
1673                             hash_extent_data_ref(root_objectid, owner, offset))
1674                                 break;
1675                 } else {
1676                         u64 ref_offset;
1677                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1678                         if (parent > 0) {
1679                                 if (parent == ref_offset) {
1680                                         err = 0;
1681                                         break;
1682                                 }
1683                                 if (ref_offset < parent)
1684                                         break;
1685                         } else {
1686                                 if (root_objectid == ref_offset) {
1687                                         err = 0;
1688                                         break;
1689                                 }
1690                                 if (ref_offset < root_objectid)
1691                                         break;
1692                         }
1693                 }
1694                 ptr += btrfs_extent_inline_ref_size(type);
1695         }
1696         if (err == -ENOENT && insert) {
1697                 if (item_size + extra_size >=
1698                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1699                         err = -EAGAIN;
1700                         goto out;
1701                 }
1702                 /*
1703                  * To add new inline back ref, we have to make sure
1704                  * there is no corresponding back ref item.
1705                  * For simplicity, we just do not add new inline back
1706                  * ref if there is any kind of item for this block
1707                  */
1708                 if (find_next_key(path, 0, &key) == 0 &&
1709                     key.objectid == bytenr &&
1710                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1711                         err = -EAGAIN;
1712                         goto out;
1713                 }
1714         }
1715         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1716 out:
1717         if (insert) {
1718                 path->keep_locks = 0;
1719                 btrfs_unlock_up_safe(path, 1);
1720         }
1721         return err;
1722 }
1723
1724 /*
1725  * helper to add new inline back ref
1726  */
1727 static noinline_for_stack
1728 void setup_inline_extent_backref(struct btrfs_root *root,
1729                                  struct btrfs_path *path,
1730                                  struct btrfs_extent_inline_ref *iref,
1731                                  u64 parent, u64 root_objectid,
1732                                  u64 owner, u64 offset, int refs_to_add,
1733                                  struct btrfs_delayed_extent_op *extent_op)
1734 {
1735         struct extent_buffer *leaf;
1736         struct btrfs_extent_item *ei;
1737         unsigned long ptr;
1738         unsigned long end;
1739         unsigned long item_offset;
1740         u64 refs;
1741         int size;
1742         int type;
1743
1744         leaf = path->nodes[0];
1745         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1746         item_offset = (unsigned long)iref - (unsigned long)ei;
1747
1748         type = extent_ref_type(parent, owner);
1749         size = btrfs_extent_inline_ref_size(type);
1750
1751         btrfs_extend_item(root, path, size);
1752
1753         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1754         refs = btrfs_extent_refs(leaf, ei);
1755         refs += refs_to_add;
1756         btrfs_set_extent_refs(leaf, ei, refs);
1757         if (extent_op)
1758                 __run_delayed_extent_op(extent_op, leaf, ei);
1759
1760         ptr = (unsigned long)ei + item_offset;
1761         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1762         if (ptr < end - size)
1763                 memmove_extent_buffer(leaf, ptr + size, ptr,
1764                                       end - size - ptr);
1765
1766         iref = (struct btrfs_extent_inline_ref *)ptr;
1767         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1768         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1769                 struct btrfs_extent_data_ref *dref;
1770                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1771                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1772                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1773                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1774                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1775         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1776                 struct btrfs_shared_data_ref *sref;
1777                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1778                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1779                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1780         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1781                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1782         } else {
1783                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1784         }
1785         btrfs_mark_buffer_dirty(leaf);
1786 }
1787
1788 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1789                                  struct btrfs_root *root,
1790                                  struct btrfs_path *path,
1791                                  struct btrfs_extent_inline_ref **ref_ret,
1792                                  u64 bytenr, u64 num_bytes, u64 parent,
1793                                  u64 root_objectid, u64 owner, u64 offset)
1794 {
1795         int ret;
1796
1797         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1798                                            bytenr, num_bytes, parent,
1799                                            root_objectid, owner, offset, 0);
1800         if (ret != -ENOENT)
1801                 return ret;
1802
1803         btrfs_release_path(path);
1804         *ref_ret = NULL;
1805
1806         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1807                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1808                                             root_objectid);
1809         } else {
1810                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1811                                              root_objectid, owner, offset);
1812         }
1813         return ret;
1814 }
1815
1816 /*
1817  * helper to update/remove inline back ref
1818  */
1819 static noinline_for_stack
1820 void update_inline_extent_backref(struct btrfs_root *root,
1821                                   struct btrfs_path *path,
1822                                   struct btrfs_extent_inline_ref *iref,
1823                                   int refs_to_mod,
1824                                   struct btrfs_delayed_extent_op *extent_op,
1825                                   int *last_ref)
1826 {
1827         struct extent_buffer *leaf;
1828         struct btrfs_extent_item *ei;
1829         struct btrfs_extent_data_ref *dref = NULL;
1830         struct btrfs_shared_data_ref *sref = NULL;
1831         unsigned long ptr;
1832         unsigned long end;
1833         u32 item_size;
1834         int size;
1835         int type;
1836         u64 refs;
1837
1838         leaf = path->nodes[0];
1839         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1840         refs = btrfs_extent_refs(leaf, ei);
1841         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1842         refs += refs_to_mod;
1843         btrfs_set_extent_refs(leaf, ei, refs);
1844         if (extent_op)
1845                 __run_delayed_extent_op(extent_op, leaf, ei);
1846
1847         type = btrfs_extent_inline_ref_type(leaf, iref);
1848
1849         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1850                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1851                 refs = btrfs_extent_data_ref_count(leaf, dref);
1852         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1853                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1854                 refs = btrfs_shared_data_ref_count(leaf, sref);
1855         } else {
1856                 refs = 1;
1857                 BUG_ON(refs_to_mod != -1);
1858         }
1859
1860         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1861         refs += refs_to_mod;
1862
1863         if (refs > 0) {
1864                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1865                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1866                 else
1867                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1868         } else {
1869                 *last_ref = 1;
1870                 size =  btrfs_extent_inline_ref_size(type);
1871                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1872                 ptr = (unsigned long)iref;
1873                 end = (unsigned long)ei + item_size;
1874                 if (ptr + size < end)
1875                         memmove_extent_buffer(leaf, ptr, ptr + size,
1876                                               end - ptr - size);
1877                 item_size -= size;
1878                 btrfs_truncate_item(root, path, item_size, 1);
1879         }
1880         btrfs_mark_buffer_dirty(leaf);
1881 }
1882
1883 static noinline_for_stack
1884 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1885                                  struct btrfs_root *root,
1886                                  struct btrfs_path *path,
1887                                  u64 bytenr, u64 num_bytes, u64 parent,
1888                                  u64 root_objectid, u64 owner,
1889                                  u64 offset, int refs_to_add,
1890                                  struct btrfs_delayed_extent_op *extent_op)
1891 {
1892         struct btrfs_extent_inline_ref *iref;
1893         int ret;
1894
1895         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1896                                            bytenr, num_bytes, parent,
1897                                            root_objectid, owner, offset, 1);
1898         if (ret == 0) {
1899                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1900                 update_inline_extent_backref(root, path, iref,
1901                                              refs_to_add, extent_op, NULL);
1902         } else if (ret == -ENOENT) {
1903                 setup_inline_extent_backref(root, path, iref, parent,
1904                                             root_objectid, owner, offset,
1905                                             refs_to_add, extent_op);
1906                 ret = 0;
1907         }
1908         return ret;
1909 }
1910
1911 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1912                                  struct btrfs_root *root,
1913                                  struct btrfs_path *path,
1914                                  u64 bytenr, u64 parent, u64 root_objectid,
1915                                  u64 owner, u64 offset, int refs_to_add)
1916 {
1917         int ret;
1918         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1919                 BUG_ON(refs_to_add != 1);
1920                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1921                                             parent, root_objectid);
1922         } else {
1923                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1924                                              parent, root_objectid,
1925                                              owner, offset, refs_to_add);
1926         }
1927         return ret;
1928 }
1929
1930 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1931                                  struct btrfs_root *root,
1932                                  struct btrfs_path *path,
1933                                  struct btrfs_extent_inline_ref *iref,
1934                                  int refs_to_drop, int is_data, int *last_ref)
1935 {
1936         int ret = 0;
1937
1938         BUG_ON(!is_data && refs_to_drop != 1);
1939         if (iref) {
1940                 update_inline_extent_backref(root, path, iref,
1941                                              -refs_to_drop, NULL, last_ref);
1942         } else if (is_data) {
1943                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop,
1944                                              last_ref);
1945         } else {
1946                 *last_ref = 1;
1947                 ret = btrfs_del_item(trans, root, path);
1948         }
1949         return ret;
1950 }
1951
1952 #define in_range(b, first, len)        ((b) >= (first) && (b) < (first) + (len))
1953 static int btrfs_issue_discard(struct block_device *bdev, u64 start, u64 len,
1954                                u64 *discarded_bytes)
1955 {
1956         int j, ret = 0;
1957         u64 bytes_left, end;
1958         u64 aligned_start = ALIGN(start, 1 << 9);
1959
1960         if (WARN_ON(start != aligned_start)) {
1961                 len -= aligned_start - start;
1962                 len = round_down(len, 1 << 9);
1963                 start = aligned_start;
1964         }
1965
1966         *discarded_bytes = 0;
1967
1968         if (!len)
1969                 return 0;
1970
1971         end = start + len;
1972         bytes_left = len;
1973
1974         /* Skip any superblocks on this device. */
1975         for (j = 0; j < BTRFS_SUPER_MIRROR_MAX; j++) {
1976                 u64 sb_start = btrfs_sb_offset(j);
1977                 u64 sb_end = sb_start + BTRFS_SUPER_INFO_SIZE;
1978                 u64 size = sb_start - start;
1979
1980                 if (!in_range(sb_start, start, bytes_left) &&
1981                     !in_range(sb_end, start, bytes_left) &&
1982                     !in_range(start, sb_start, BTRFS_SUPER_INFO_SIZE))
1983                         continue;
1984
1985                 /*
1986                  * Superblock spans beginning of range.  Adjust start and
1987                  * try again.
1988                  */
1989                 if (sb_start <= start) {
1990                         start += sb_end - start;
1991                         if (start > end) {
1992                                 bytes_left = 0;
1993                                 break;
1994                         }
1995                         bytes_left = end - start;
1996                         continue;
1997                 }
1998
1999                 if (size) {
2000                         ret = blkdev_issue_discard(bdev, start >> 9, size >> 9,
2001                                                    GFP_NOFS, 0);
2002                         if (!ret)
2003                                 *discarded_bytes += size;
2004                         else if (ret != -EOPNOTSUPP)
2005                                 return ret;
2006                 }
2007
2008                 start = sb_end;
2009                 if (start > end) {
2010                         bytes_left = 0;
2011                         break;
2012                 }
2013                 bytes_left = end - start;
2014         }
2015
2016         if (bytes_left) {
2017                 ret = blkdev_issue_discard(bdev, start >> 9, bytes_left >> 9,
2018                                            GFP_NOFS, 0);
2019                 if (!ret)
2020                         *discarded_bytes += bytes_left;
2021         }
2022         return ret;
2023 }
2024
2025 int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
2026                          u64 num_bytes, u64 *actual_bytes)
2027 {
2028         int ret;
2029         u64 discarded_bytes = 0;
2030         struct btrfs_bio *bbio = NULL;
2031
2032
2033         /*
2034          * Avoid races with device replace and make sure our bbio has devices
2035          * associated to its stripes that don't go away while we are discarding.
2036          */
2037         btrfs_bio_counter_inc_blocked(root->fs_info);
2038         /* Tell the block device(s) that the sectors can be discarded */
2039         ret = btrfs_map_block(root->fs_info, REQ_OP_DISCARD,
2040                               bytenr, &num_bytes, &bbio, 0);
2041         /* Error condition is -ENOMEM */
2042         if (!ret) {
2043                 struct btrfs_bio_stripe *stripe = bbio->stripes;
2044                 int i;
2045
2046
2047                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
2048                         u64 bytes;
2049                         if (!stripe->dev->can_discard)
2050                                 continue;
2051
2052                         ret = btrfs_issue_discard(stripe->dev->bdev,
2053                                                   stripe->physical,
2054                                                   stripe->length,
2055                                                   &bytes);
2056                         if (!ret)
2057                                 discarded_bytes += bytes;
2058                         else if (ret != -EOPNOTSUPP)
2059                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
2060
2061                         /*
2062                          * Just in case we get back EOPNOTSUPP for some reason,
2063                          * just ignore the return value so we don't screw up
2064                          * people calling discard_extent.
2065                          */
2066                         ret = 0;
2067                 }
2068                 btrfs_put_bbio(bbio);
2069         }
2070         btrfs_bio_counter_dec(root->fs_info);
2071
2072         if (actual_bytes)
2073                 *actual_bytes = discarded_bytes;
2074
2075
2076         if (ret == -EOPNOTSUPP)
2077                 ret = 0;
2078         return ret;
2079 }
2080
2081 /* Can return -ENOMEM */
2082 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2083                          struct btrfs_root *root,
2084                          u64 bytenr, u64 num_bytes, u64 parent,
2085                          u64 root_objectid, u64 owner, u64 offset)
2086 {
2087         int ret;
2088         struct btrfs_fs_info *fs_info = root->fs_info;
2089
2090         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
2091                root_objectid == BTRFS_TREE_LOG_OBJECTID);
2092
2093         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
2094                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
2095                                         num_bytes,
2096                                         parent, root_objectid, (int)owner,
2097                                         BTRFS_ADD_DELAYED_REF, NULL);
2098         } else {
2099                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
2100                                         num_bytes, parent, root_objectid,
2101                                         owner, offset, 0,
2102                                         BTRFS_ADD_DELAYED_REF, NULL);
2103         }
2104         return ret;
2105 }
2106
2107 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2108                                   struct btrfs_root *root,
2109                                   struct btrfs_delayed_ref_node *node,
2110                                   u64 parent, u64 root_objectid,
2111                                   u64 owner, u64 offset, int refs_to_add,
2112                                   struct btrfs_delayed_extent_op *extent_op)
2113 {
2114         struct btrfs_fs_info *fs_info = root->fs_info;
2115         struct btrfs_path *path;
2116         struct extent_buffer *leaf;
2117         struct btrfs_extent_item *item;
2118         struct btrfs_key key;
2119         u64 bytenr = node->bytenr;
2120         u64 num_bytes = node->num_bytes;
2121         u64 refs;
2122         int ret;
2123
2124         path = btrfs_alloc_path();
2125         if (!path)
2126                 return -ENOMEM;
2127
2128         path->reada = READA_FORWARD;
2129         path->leave_spinning = 1;
2130         /* this will setup the path even if it fails to insert the back ref */
2131         ret = insert_inline_extent_backref(trans, fs_info->extent_root, path,
2132                                            bytenr, num_bytes, parent,
2133                                            root_objectid, owner, offset,
2134                                            refs_to_add, extent_op);
2135         if ((ret < 0 && ret != -EAGAIN) || !ret)
2136                 goto out;
2137
2138         /*
2139          * Ok we had -EAGAIN which means we didn't have space to insert and
2140          * inline extent ref, so just update the reference count and add a
2141          * normal backref.
2142          */
2143         leaf = path->nodes[0];
2144         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2145         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2146         refs = btrfs_extent_refs(leaf, item);
2147         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
2148         if (extent_op)
2149                 __run_delayed_extent_op(extent_op, leaf, item);
2150
2151         btrfs_mark_buffer_dirty(leaf);
2152         btrfs_release_path(path);
2153
2154         path->reada = READA_FORWARD;
2155         path->leave_spinning = 1;
2156         /* now insert the actual backref */
2157         ret = insert_extent_backref(trans, root->fs_info->extent_root,
2158                                     path, bytenr, parent, root_objectid,
2159                                     owner, offset, refs_to_add);
2160         if (ret)
2161                 btrfs_abort_transaction(trans, ret);
2162 out:
2163         btrfs_free_path(path);
2164         return ret;
2165 }
2166
2167 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
2168                                 struct btrfs_root *root,
2169                                 struct btrfs_delayed_ref_node *node,
2170                                 struct btrfs_delayed_extent_op *extent_op,
2171                                 int insert_reserved)
2172 {
2173         int ret = 0;
2174         struct btrfs_delayed_data_ref *ref;
2175         struct btrfs_key ins;
2176         u64 parent = 0;
2177         u64 ref_root = 0;
2178         u64 flags = 0;
2179
2180         ins.objectid = node->bytenr;
2181         ins.offset = node->num_bytes;
2182         ins.type = BTRFS_EXTENT_ITEM_KEY;
2183
2184         ref = btrfs_delayed_node_to_data_ref(node);
2185         trace_run_delayed_data_ref(root->fs_info, node, ref, node->action);
2186
2187         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2188                 parent = ref->parent;
2189         ref_root = ref->root;
2190
2191         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2192                 if (extent_op)
2193                         flags |= extent_op->flags_to_set;
2194                 ret = alloc_reserved_file_extent(trans, root,
2195                                                  parent, ref_root, flags,
2196                                                  ref->objectid, ref->offset,
2197                                                  &ins, node->ref_mod);
2198         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2199                 ret = __btrfs_inc_extent_ref(trans, root, node, parent,
2200                                              ref_root, ref->objectid,
2201                                              ref->offset, node->ref_mod,
2202                                              extent_op);
2203         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2204                 ret = __btrfs_free_extent(trans, root, node, parent,
2205                                           ref_root, ref->objectid,
2206                                           ref->offset, node->ref_mod,
2207                                           extent_op);
2208         } else {
2209                 BUG();
2210         }
2211         return ret;
2212 }
2213
2214 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2215                                     struct extent_buffer *leaf,
2216                                     struct btrfs_extent_item *ei)
2217 {
2218         u64 flags = btrfs_extent_flags(leaf, ei);
2219         if (extent_op->update_flags) {
2220                 flags |= extent_op->flags_to_set;
2221                 btrfs_set_extent_flags(leaf, ei, flags);
2222         }
2223
2224         if (extent_op->update_key) {
2225                 struct btrfs_tree_block_info *bi;
2226                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2227                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2228                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2229         }
2230 }
2231
2232 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2233                                  struct btrfs_root *root,
2234                                  struct btrfs_delayed_ref_node *node,
2235                                  struct btrfs_delayed_extent_op *extent_op)
2236 {
2237         struct btrfs_key key;
2238         struct btrfs_path *path;
2239         struct btrfs_extent_item *ei;
2240         struct extent_buffer *leaf;
2241         u32 item_size;
2242         int ret;
2243         int err = 0;
2244         int metadata = !extent_op->is_data;
2245
2246         if (trans->aborted)
2247                 return 0;
2248
2249         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2250                 metadata = 0;
2251
2252         path = btrfs_alloc_path();
2253         if (!path)
2254                 return -ENOMEM;
2255
2256         key.objectid = node->bytenr;
2257
2258         if (metadata) {
2259                 key.type = BTRFS_METADATA_ITEM_KEY;
2260                 key.offset = extent_op->level;
2261         } else {
2262                 key.type = BTRFS_EXTENT_ITEM_KEY;
2263                 key.offset = node->num_bytes;
2264         }
2265
2266 again:
2267         path->reada = READA_FORWARD;
2268         path->leave_spinning = 1;
2269         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
2270                                 path, 0, 1);
2271         if (ret < 0) {
2272                 err = ret;
2273                 goto out;
2274         }
2275         if (ret > 0) {
2276                 if (metadata) {
2277                         if (path->slots[0] > 0) {
2278                                 path->slots[0]--;
2279                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
2280                                                       path->slots[0]);
2281                                 if (key.objectid == node->bytenr &&
2282                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
2283                                     key.offset == node->num_bytes)
2284                                         ret = 0;
2285                         }
2286                         if (ret > 0) {
2287                                 btrfs_release_path(path);
2288                                 metadata = 0;
2289
2290                                 key.objectid = node->bytenr;
2291                                 key.offset = node->num_bytes;
2292                                 key.type = BTRFS_EXTENT_ITEM_KEY;
2293                                 goto again;
2294                         }
2295                 } else {
2296                         err = -EIO;
2297                         goto out;
2298                 }
2299         }
2300
2301         leaf = path->nodes[0];
2302         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2303 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2304         if (item_size < sizeof(*ei)) {
2305                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2306                                              path, (u64)-1, 0);
2307                 if (ret < 0) {
2308                         err = ret;
2309                         goto out;
2310                 }
2311                 leaf = path->nodes[0];
2312                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2313         }
2314 #endif
2315         BUG_ON(item_size < sizeof(*ei));
2316         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2317         __run_delayed_extent_op(extent_op, leaf, ei);
2318
2319         btrfs_mark_buffer_dirty(leaf);
2320 out:
2321         btrfs_free_path(path);
2322         return err;
2323 }
2324
2325 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2326                                 struct btrfs_root *root,
2327                                 struct btrfs_delayed_ref_node *node,
2328                                 struct btrfs_delayed_extent_op *extent_op,
2329                                 int insert_reserved)
2330 {
2331         int ret = 0;
2332         struct btrfs_delayed_tree_ref *ref;
2333         struct btrfs_key ins;
2334         u64 parent = 0;
2335         u64 ref_root = 0;
2336         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
2337                                                  SKINNY_METADATA);
2338
2339         ref = btrfs_delayed_node_to_tree_ref(node);
2340         trace_run_delayed_tree_ref(root->fs_info, node, ref, node->action);
2341
2342         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2343                 parent = ref->parent;
2344         ref_root = ref->root;
2345
2346         ins.objectid = node->bytenr;
2347         if (skinny_metadata) {
2348                 ins.offset = ref->level;
2349                 ins.type = BTRFS_METADATA_ITEM_KEY;
2350         } else {
2351                 ins.offset = node->num_bytes;
2352                 ins.type = BTRFS_EXTENT_ITEM_KEY;
2353         }
2354
2355         if (node->ref_mod != 1) {
2356                 btrfs_err(root->fs_info,
2357         "btree block(%llu) has %d references rather than 1: action %d ref_root %llu parent %llu",
2358                           node->bytenr, node->ref_mod, node->action, ref_root,
2359                           parent);
2360                 return -EIO;
2361         }
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(root->fs_info, node, head,
2411                                            node->action);
2412
2413                 if (insert_reserved) {
2414                         btrfs_pin_extent(root, node->bytenr,
2415                                          node->num_bytes, 1);
2416                         if (head->is_data) {
2417                                 ret = btrfs_del_csums(trans, root,
2418                                                       node->bytenr,
2419                                                       node->num_bytes);
2420                         }
2421                 }
2422
2423                 /* Also free its reserved qgroup space */
2424                 btrfs_qgroup_free_delayed_ref(root->fs_info,
2425                                               head->qgroup_ref_root,
2426                                               head->qgroup_reserved);
2427                 return ret;
2428         }
2429
2430         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2431             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2432                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2433                                            insert_reserved);
2434         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2435                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2436                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2437                                            insert_reserved);
2438         else
2439                 BUG();
2440         return ret;
2441 }
2442
2443 static inline struct btrfs_delayed_ref_node *
2444 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2445 {
2446         struct btrfs_delayed_ref_node *ref;
2447
2448         if (list_empty(&head->ref_list))
2449                 return NULL;
2450
2451         /*
2452          * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
2453          * This is to prevent a ref count from going down to zero, which deletes
2454          * the extent item from the extent tree, when there still are references
2455          * to add, which would fail because they would not find the extent item.
2456          */
2457         list_for_each_entry(ref, &head->ref_list, list) {
2458                 if (ref->action == BTRFS_ADD_DELAYED_REF)
2459                         return ref;
2460         }
2461
2462         return list_entry(head->ref_list.next, struct btrfs_delayed_ref_node,
2463                           list);
2464 }
2465
2466 /*
2467  * Returns 0 on success or if called with an already aborted transaction.
2468  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2469  */
2470 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2471                                              struct btrfs_root *root,
2472                                              unsigned long nr)
2473 {
2474         struct btrfs_delayed_ref_root *delayed_refs;
2475         struct btrfs_delayed_ref_node *ref;
2476         struct btrfs_delayed_ref_head *locked_ref = NULL;
2477         struct btrfs_delayed_extent_op *extent_op;
2478         struct btrfs_fs_info *fs_info = root->fs_info;
2479         ktime_t start = ktime_get();
2480         int ret;
2481         unsigned long count = 0;
2482         unsigned long actual_count = 0;
2483         int must_insert_reserved = 0;
2484
2485         delayed_refs = &trans->transaction->delayed_refs;
2486         while (1) {
2487                 if (!locked_ref) {
2488                         if (count >= nr)
2489                                 break;
2490
2491                         spin_lock(&delayed_refs->lock);
2492                         locked_ref = btrfs_select_ref_head(trans);
2493                         if (!locked_ref) {
2494                                 spin_unlock(&delayed_refs->lock);
2495                                 break;
2496                         }
2497
2498                         /* grab the lock that says we are going to process
2499                          * all the refs for this head */
2500                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2501                         spin_unlock(&delayed_refs->lock);
2502                         /*
2503                          * we may have dropped the spin lock to get the head
2504                          * mutex lock, and that might have given someone else
2505                          * time to free the head.  If that's true, it has been
2506                          * removed from our list and we can move on.
2507                          */
2508                         if (ret == -EAGAIN) {
2509                                 locked_ref = NULL;
2510                                 count++;
2511                                 continue;
2512                         }
2513                 }
2514
2515                 /*
2516                  * We need to try and merge add/drops of the same ref since we
2517                  * can run into issues with relocate dropping the implicit ref
2518                  * and then it being added back again before the drop can
2519                  * finish.  If we merged anything we need to re-loop so we can
2520                  * get a good ref.
2521                  * Or we can get node references of the same type that weren't
2522                  * merged when created due to bumps in the tree mod seq, and
2523                  * we need to merge them to prevent adding an inline extent
2524                  * backref before dropping it (triggering a BUG_ON at
2525                  * insert_inline_extent_backref()).
2526                  */
2527                 spin_lock(&locked_ref->lock);
2528                 btrfs_merge_delayed_refs(trans, fs_info, delayed_refs,
2529                                          locked_ref);
2530
2531                 /*
2532                  * locked_ref is the head node, so we have to go one
2533                  * node back for any delayed ref updates
2534                  */
2535                 ref = select_delayed_ref(locked_ref);
2536
2537                 if (ref && ref->seq &&
2538                     btrfs_check_delayed_seq(fs_info, delayed_refs, ref->seq)) {
2539                         spin_unlock(&locked_ref->lock);
2540                         btrfs_delayed_ref_unlock(locked_ref);
2541                         spin_lock(&delayed_refs->lock);
2542                         locked_ref->processing = 0;
2543                         delayed_refs->num_heads_ready++;
2544                         spin_unlock(&delayed_refs->lock);
2545                         locked_ref = NULL;
2546                         cond_resched();
2547                         count++;
2548                         continue;
2549                 }
2550
2551                 /*
2552                  * record the must insert reserved flag before we
2553                  * drop the spin lock.
2554                  */
2555                 must_insert_reserved = locked_ref->must_insert_reserved;
2556                 locked_ref->must_insert_reserved = 0;
2557
2558                 extent_op = locked_ref->extent_op;
2559                 locked_ref->extent_op = NULL;
2560
2561                 if (!ref) {
2562
2563
2564                         /* All delayed refs have been processed, Go ahead
2565                          * and send the head node to run_one_delayed_ref,
2566                          * so that any accounting fixes can happen
2567                          */
2568                         ref = &locked_ref->node;
2569
2570                         if (extent_op && must_insert_reserved) {
2571                                 btrfs_free_delayed_extent_op(extent_op);
2572                                 extent_op = NULL;
2573                         }
2574
2575                         if (extent_op) {
2576                                 spin_unlock(&locked_ref->lock);
2577                                 ret = run_delayed_extent_op(trans, root,
2578                                                             ref, extent_op);
2579                                 btrfs_free_delayed_extent_op(extent_op);
2580
2581                                 if (ret) {
2582                                         /*
2583                                          * Need to reset must_insert_reserved if
2584                                          * there was an error so the abort stuff
2585                                          * can cleanup the reserved space
2586                                          * properly.
2587                                          */
2588                                         if (must_insert_reserved)
2589                                                 locked_ref->must_insert_reserved = 1;
2590                                         locked_ref->processing = 0;
2591                                         btrfs_debug(fs_info,
2592                                                     "run_delayed_extent_op returned %d",
2593                                                     ret);
2594                                         btrfs_delayed_ref_unlock(locked_ref);
2595                                         return ret;
2596                                 }
2597                                 continue;
2598                         }
2599
2600                         /*
2601                          * Need to drop our head ref lock and re-acquire the
2602                          * delayed ref lock and then re-check to make sure
2603                          * nobody got added.
2604                          */
2605                         spin_unlock(&locked_ref->lock);
2606                         spin_lock(&delayed_refs->lock);
2607                         spin_lock(&locked_ref->lock);
2608                         if (!list_empty(&locked_ref->ref_list) ||
2609                             locked_ref->extent_op) {
2610                                 spin_unlock(&locked_ref->lock);
2611                                 spin_unlock(&delayed_refs->lock);
2612                                 continue;
2613                         }
2614                         ref->in_tree = 0;
2615                         delayed_refs->num_heads--;
2616                         rb_erase(&locked_ref->href_node,
2617                                  &delayed_refs->href_root);
2618                         spin_unlock(&delayed_refs->lock);
2619                 } else {
2620                         actual_count++;
2621                         ref->in_tree = 0;
2622                         list_del(&ref->list);
2623                 }
2624                 atomic_dec(&delayed_refs->num_entries);
2625
2626                 if (!btrfs_delayed_ref_is_head(ref)) {
2627                         /*
2628                          * when we play the delayed ref, also correct the
2629                          * ref_mod on head
2630                          */
2631                         switch (ref->action) {
2632                         case BTRFS_ADD_DELAYED_REF:
2633                         case BTRFS_ADD_DELAYED_EXTENT:
2634                                 locked_ref->node.ref_mod -= ref->ref_mod;
2635                                 break;
2636                         case BTRFS_DROP_DELAYED_REF:
2637                                 locked_ref->node.ref_mod += ref->ref_mod;
2638                                 break;
2639                         default:
2640                                 WARN_ON(1);
2641                         }
2642                 }
2643                 spin_unlock(&locked_ref->lock);
2644
2645                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2646                                           must_insert_reserved);
2647
2648                 btrfs_free_delayed_extent_op(extent_op);
2649                 if (ret) {
2650                         locked_ref->processing = 0;
2651                         btrfs_delayed_ref_unlock(locked_ref);
2652                         btrfs_put_delayed_ref(ref);
2653                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d",
2654                                     ret);
2655                         return ret;
2656                 }
2657
2658                 /*
2659                  * If this node is a head, that means all the refs in this head
2660                  * have been dealt with, and we will pick the next head to deal
2661                  * with, so we must unlock the head and drop it from the cluster
2662                  * list before we release it.
2663                  */
2664                 if (btrfs_delayed_ref_is_head(ref)) {
2665                         if (locked_ref->is_data &&
2666                             locked_ref->total_ref_mod < 0) {
2667                                 spin_lock(&delayed_refs->lock);
2668                                 delayed_refs->pending_csums -= ref->num_bytes;
2669                                 spin_unlock(&delayed_refs->lock);
2670                         }
2671                         btrfs_delayed_ref_unlock(locked_ref);
2672                         locked_ref = NULL;
2673                 }
2674                 btrfs_put_delayed_ref(ref);
2675                 count++;
2676                 cond_resched();
2677         }
2678
2679         /*
2680          * We don't want to include ref heads since we can have empty ref heads
2681          * and those will drastically skew our runtime down since we just do
2682          * accounting, no actual extent tree updates.
2683          */
2684         if (actual_count > 0) {
2685                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2686                 u64 avg;
2687
2688                 /*
2689                  * We weigh the current average higher than our current runtime
2690                  * to avoid large swings in the average.
2691                  */
2692                 spin_lock(&delayed_refs->lock);
2693                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2694                 fs_info->avg_delayed_ref_runtime = avg >> 2;    /* div by 4 */
2695                 spin_unlock(&delayed_refs->lock);
2696         }
2697         return 0;
2698 }
2699
2700 #ifdef SCRAMBLE_DELAYED_REFS
2701 /*
2702  * Normally delayed refs get processed in ascending bytenr order. This
2703  * correlates in most cases to the order added. To expose dependencies on this
2704  * order, we start to process the tree in the middle instead of the beginning
2705  */
2706 static u64 find_middle(struct rb_root *root)
2707 {
2708         struct rb_node *n = root->rb_node;
2709         struct btrfs_delayed_ref_node *entry;
2710         int alt = 1;
2711         u64 middle;
2712         u64 first = 0, last = 0;
2713
2714         n = rb_first(root);
2715         if (n) {
2716                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2717                 first = entry->bytenr;
2718         }
2719         n = rb_last(root);
2720         if (n) {
2721                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2722                 last = entry->bytenr;
2723         }
2724         n = root->rb_node;
2725
2726         while (n) {
2727                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2728                 WARN_ON(!entry->in_tree);
2729
2730                 middle = entry->bytenr;
2731
2732                 if (alt)
2733                         n = n->rb_left;
2734                 else
2735                         n = n->rb_right;
2736
2737                 alt = 1 - alt;
2738         }
2739         return middle;
2740 }
2741 #endif
2742
2743 static inline u64 heads_to_leaves(struct btrfs_root *root, u64 heads)
2744 {
2745         u64 num_bytes;
2746
2747         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2748                              sizeof(struct btrfs_extent_inline_ref));
2749         if (!btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2750                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2751
2752         /*
2753          * We don't ever fill up leaves all the way so multiply by 2 just to be
2754          * closer to what we're really going to want to use.
2755          */
2756         return div_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(root));
2757 }
2758
2759 /*
2760  * Takes the number of bytes to be csumm'ed and figures out how many leaves it
2761  * would require to store the csums for that many bytes.
2762  */
2763 u64 btrfs_csum_bytes_to_leaves(struct btrfs_root *root, u64 csum_bytes)
2764 {
2765         u64 csum_size;
2766         u64 num_csums_per_leaf;
2767         u64 num_csums;
2768
2769         csum_size = BTRFS_MAX_ITEM_SIZE(root);
2770         num_csums_per_leaf = div64_u64(csum_size,
2771                         (u64)btrfs_super_csum_size(root->fs_info->super_copy));
2772         num_csums = div64_u64(csum_bytes, root->sectorsize);
2773         num_csums += num_csums_per_leaf - 1;
2774         num_csums = div64_u64(num_csums, num_csums_per_leaf);
2775         return num_csums;
2776 }
2777
2778 int btrfs_check_space_for_delayed_refs(struct btrfs_trans_handle *trans,
2779                                        struct btrfs_root *root)
2780 {
2781         struct btrfs_block_rsv *global_rsv;
2782         u64 num_heads = trans->transaction->delayed_refs.num_heads_ready;
2783         u64 csum_bytes = trans->transaction->delayed_refs.pending_csums;
2784         u64 num_dirty_bgs = trans->transaction->num_dirty_bgs;
2785         u64 num_bytes, num_dirty_bgs_bytes;
2786         int ret = 0;
2787
2788         num_bytes = btrfs_calc_trans_metadata_size(root, 1);
2789         num_heads = heads_to_leaves(root, num_heads);
2790         if (num_heads > 1)
2791                 num_bytes += (num_heads - 1) * root->nodesize;
2792         num_bytes <<= 1;
2793         num_bytes += btrfs_csum_bytes_to_leaves(root, csum_bytes) * root->nodesize;
2794         num_dirty_bgs_bytes = btrfs_calc_trans_metadata_size(root,
2795                                                              num_dirty_bgs);
2796         global_rsv = &root->fs_info->global_block_rsv;
2797
2798         /*
2799          * If we can't allocate any more chunks lets make sure we have _lots_ of
2800          * wiggle room since running delayed refs can create more delayed refs.
2801          */
2802         if (global_rsv->space_info->full) {
2803                 num_dirty_bgs_bytes <<= 1;
2804                 num_bytes <<= 1;
2805         }
2806
2807         spin_lock(&global_rsv->lock);
2808         if (global_rsv->reserved <= num_bytes + num_dirty_bgs_bytes)
2809                 ret = 1;
2810         spin_unlock(&global_rsv->lock);
2811         return ret;
2812 }
2813
2814 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans,
2815                                        struct btrfs_root *root)
2816 {
2817         struct btrfs_fs_info *fs_info = root->fs_info;
2818         u64 num_entries =
2819                 atomic_read(&trans->transaction->delayed_refs.num_entries);
2820         u64 avg_runtime;
2821         u64 val;
2822
2823         smp_mb();
2824         avg_runtime = fs_info->avg_delayed_ref_runtime;
2825         val = num_entries * avg_runtime;
2826         if (num_entries * avg_runtime >= NSEC_PER_SEC)
2827                 return 1;
2828         if (val >= NSEC_PER_SEC / 2)
2829                 return 2;
2830
2831         return btrfs_check_space_for_delayed_refs(trans, root);
2832 }
2833
2834 struct async_delayed_refs {
2835         struct btrfs_root *root;
2836         u64 transid;
2837         int count;
2838         int error;
2839         int sync;
2840         struct completion wait;
2841         struct btrfs_work work;
2842 };
2843
2844 static void delayed_ref_async_start(struct btrfs_work *work)
2845 {
2846         struct async_delayed_refs *async;
2847         struct btrfs_trans_handle *trans;
2848         int ret;
2849
2850         async = container_of(work, struct async_delayed_refs, work);
2851
2852         /* if the commit is already started, we don't need to wait here */
2853         if (btrfs_transaction_blocked(async->root->fs_info))
2854                 goto done;
2855
2856         trans = btrfs_join_transaction(async->root);
2857         if (IS_ERR(trans)) {
2858                 async->error = PTR_ERR(trans);
2859                 goto done;
2860         }
2861
2862         /*
2863          * trans->sync means that when we call end_transaction, we won't
2864          * wait on delayed refs
2865          */
2866         trans->sync = true;
2867
2868         /* Don't bother flushing if we got into a different transaction */
2869         if (trans->transid > async->transid)
2870                 goto end;
2871
2872         ret = btrfs_run_delayed_refs(trans, async->root, async->count);
2873         if (ret)
2874                 async->error = ret;
2875 end:
2876         ret = btrfs_end_transaction(trans, async->root);
2877         if (ret && !async->error)
2878                 async->error = ret;
2879 done:
2880         if (async->sync)
2881                 complete(&async->wait);
2882         else
2883                 kfree(async);
2884 }
2885
2886 int btrfs_async_run_delayed_refs(struct btrfs_root *root,
2887                                  unsigned long count, u64 transid, int wait)
2888 {
2889         struct async_delayed_refs *async;
2890         int ret;
2891
2892         async = kmalloc(sizeof(*async), GFP_NOFS);
2893         if (!async)
2894                 return -ENOMEM;
2895
2896         async->root = root->fs_info->tree_root;
2897         async->count = count;
2898         async->error = 0;
2899         async->transid = transid;
2900         if (wait)
2901                 async->sync = 1;
2902         else
2903                 async->sync = 0;
2904         init_completion(&async->wait);
2905
2906         btrfs_init_work(&async->work, btrfs_extent_refs_helper,
2907                         delayed_ref_async_start, NULL, NULL);
2908
2909         btrfs_queue_work(root->fs_info->extent_workers, &async->work);
2910
2911         if (wait) {
2912                 wait_for_completion(&async->wait);
2913                 ret = async->error;
2914                 kfree(async);
2915                 return ret;
2916         }
2917         return 0;
2918 }
2919
2920 /*
2921  * this starts processing the delayed reference count updates and
2922  * extent insertions we have queued up so far.  count can be
2923  * 0, which means to process everything in the tree at the start
2924  * of the run (but not newly added entries), or it can be some target
2925  * number you'd like to process.
2926  *
2927  * Returns 0 on success or if called with an aborted transaction
2928  * Returns <0 on error and aborts the transaction
2929  */
2930 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2931                            struct btrfs_root *root, unsigned long count)
2932 {
2933         struct rb_node *node;
2934         struct btrfs_delayed_ref_root *delayed_refs;
2935         struct btrfs_delayed_ref_head *head;
2936         int ret;
2937         int run_all = count == (unsigned long)-1;
2938         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
2939
2940         /* We'll clean this up in btrfs_cleanup_transaction */
2941         if (trans->aborted)
2942                 return 0;
2943
2944         if (test_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &root->fs_info->flags))
2945                 return 0;
2946
2947         if (root == root->fs_info->extent_root)
2948                 root = root->fs_info->tree_root;
2949
2950         delayed_refs = &trans->transaction->delayed_refs;
2951         if (count == 0)
2952                 count = atomic_read(&delayed_refs->num_entries) * 2;
2953
2954 again:
2955 #ifdef SCRAMBLE_DELAYED_REFS
2956         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2957 #endif
2958         trans->can_flush_pending_bgs = false;
2959         ret = __btrfs_run_delayed_refs(trans, root, count);
2960         if (ret < 0) {
2961                 btrfs_abort_transaction(trans, ret);
2962                 return ret;
2963         }
2964
2965         if (run_all) {
2966                 if (!list_empty(&trans->new_bgs))
2967                         btrfs_create_pending_block_groups(trans, root);
2968
2969                 spin_lock(&delayed_refs->lock);
2970                 node = rb_first(&delayed_refs->href_root);
2971                 if (!node) {
2972                         spin_unlock(&delayed_refs->lock);
2973                         goto out;
2974                 }
2975                 count = (unsigned long)-1;
2976
2977                 while (node) {
2978                         head = rb_entry(node, struct btrfs_delayed_ref_head,
2979                                         href_node);
2980                         if (btrfs_delayed_ref_is_head(&head->node)) {
2981                                 struct btrfs_delayed_ref_node *ref;
2982
2983                                 ref = &head->node;
2984                                 atomic_inc(&ref->refs);
2985
2986                                 spin_unlock(&delayed_refs->lock);
2987                                 /*
2988                                  * Mutex was contended, block until it's
2989                                  * released and try again
2990                                  */
2991                                 mutex_lock(&head->mutex);
2992                                 mutex_unlock(&head->mutex);
2993
2994                                 btrfs_put_delayed_ref(ref);
2995                                 cond_resched();
2996                                 goto again;
2997                         } else {
2998                                 WARN_ON(1);
2999                         }
3000                         node = rb_next(node);
3001                 }
3002                 spin_unlock(&delayed_refs->lock);
3003                 cond_resched();
3004                 goto again;
3005         }
3006 out:
3007         assert_qgroups_uptodate(trans);
3008         trans->can_flush_pending_bgs = can_flush_pending_bgs;
3009         return 0;
3010 }
3011
3012 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
3013                                 struct btrfs_root *root,
3014                                 u64 bytenr, u64 num_bytes, u64 flags,
3015                                 int level, int is_data)
3016 {
3017         struct btrfs_delayed_extent_op *extent_op;
3018         int ret;
3019
3020         extent_op = btrfs_alloc_delayed_extent_op();
3021         if (!extent_op)
3022                 return -ENOMEM;
3023
3024         extent_op->flags_to_set = flags;
3025         extent_op->update_flags = true;
3026         extent_op->update_key = false;
3027         extent_op->is_data = is_data ? true : false;
3028         extent_op->level = level;
3029
3030         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
3031                                           num_bytes, extent_op);
3032         if (ret)
3033                 btrfs_free_delayed_extent_op(extent_op);
3034         return ret;
3035 }
3036
3037 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
3038                                       struct btrfs_root *root,
3039                                       struct btrfs_path *path,
3040                                       u64 objectid, u64 offset, u64 bytenr)
3041 {
3042         struct btrfs_delayed_ref_head *head;
3043         struct btrfs_delayed_ref_node *ref;
3044         struct btrfs_delayed_data_ref *data_ref;
3045         struct btrfs_delayed_ref_root *delayed_refs;
3046         int ret = 0;
3047
3048         delayed_refs = &trans->transaction->delayed_refs;
3049         spin_lock(&delayed_refs->lock);
3050         head = btrfs_find_delayed_ref_head(trans, bytenr);
3051         if (!head) {
3052                 spin_unlock(&delayed_refs->lock);
3053                 return 0;
3054         }
3055
3056         if (!mutex_trylock(&head->mutex)) {
3057                 atomic_inc(&head->node.refs);
3058                 spin_unlock(&delayed_refs->lock);
3059
3060                 btrfs_release_path(path);
3061
3062                 /*
3063                  * Mutex was contended, block until it's released and let
3064                  * caller try again
3065                  */
3066                 mutex_lock(&head->mutex);
3067                 mutex_unlock(&head->mutex);
3068                 btrfs_put_delayed_ref(&head->node);
3069                 return -EAGAIN;
3070         }
3071         spin_unlock(&delayed_refs->lock);
3072
3073         spin_lock(&head->lock);
3074         list_for_each_entry(ref, &head->ref_list, list) {
3075                 /* If it's a shared ref we know a cross reference exists */
3076                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
3077                         ret = 1;
3078                         break;
3079                 }
3080
3081                 data_ref = btrfs_delayed_node_to_data_ref(ref);
3082
3083                 /*
3084                  * If our ref doesn't match the one we're currently looking at
3085                  * then we have a cross reference.
3086                  */
3087                 if (data_ref->root != root->root_key.objectid ||
3088                     data_ref->objectid != objectid ||
3089                     data_ref->offset != offset) {
3090                         ret = 1;
3091                         break;
3092                 }
3093         }
3094         spin_unlock(&head->lock);
3095         mutex_unlock(&head->mutex);
3096         return ret;
3097 }
3098
3099 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
3100                                         struct btrfs_root *root,
3101                                         struct btrfs_path *path,
3102                                         u64 objectid, u64 offset, u64 bytenr)
3103 {
3104         struct btrfs_root *extent_root = root->fs_info->extent_root;
3105         struct extent_buffer *leaf;
3106         struct btrfs_extent_data_ref *ref;
3107         struct btrfs_extent_inline_ref *iref;
3108         struct btrfs_extent_item *ei;
3109         struct btrfs_key key;
3110         u32 item_size;
3111         int ret;
3112
3113         key.objectid = bytenr;
3114         key.offset = (u64)-1;
3115         key.type = BTRFS_EXTENT_ITEM_KEY;
3116
3117         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
3118         if (ret < 0)
3119                 goto out;
3120         BUG_ON(ret == 0); /* Corruption */
3121
3122         ret = -ENOENT;
3123         if (path->slots[0] == 0)
3124                 goto out;
3125
3126         path->slots[0]--;
3127         leaf = path->nodes[0];
3128         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3129
3130         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
3131                 goto out;
3132
3133         ret = 1;
3134         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
3135 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3136         if (item_size < sizeof(*ei)) {
3137                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3138                 goto out;
3139         }
3140 #endif
3141         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
3142
3143         if (item_size != sizeof(*ei) +
3144             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
3145                 goto out;
3146
3147         if (btrfs_extent_generation(leaf, ei) <=
3148             btrfs_root_last_snapshot(&root->root_item))
3149                 goto out;
3150
3151         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
3152         if (btrfs_extent_inline_ref_type(leaf, iref) !=
3153             BTRFS_EXTENT_DATA_REF_KEY)
3154                 goto out;
3155
3156         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
3157         if (btrfs_extent_refs(leaf, ei) !=
3158             btrfs_extent_data_ref_count(leaf, ref) ||
3159             btrfs_extent_data_ref_root(leaf, ref) !=
3160             root->root_key.objectid ||
3161             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
3162             btrfs_extent_data_ref_offset(leaf, ref) != offset)
3163                 goto out;
3164
3165         ret = 0;
3166 out:
3167         return ret;
3168 }
3169
3170 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
3171                           struct btrfs_root *root,
3172                           u64 objectid, u64 offset, u64 bytenr)
3173 {
3174         struct btrfs_path *path;
3175         int ret;
3176         int ret2;
3177
3178         path = btrfs_alloc_path();
3179         if (!path)
3180                 return -ENOENT;
3181
3182         do {
3183                 ret = check_committed_ref(trans, root, path, objectid,
3184                                           offset, bytenr);
3185                 if (ret && ret != -ENOENT)
3186                         goto out;
3187
3188                 ret2 = check_delayed_ref(trans, root, path, objectid,
3189                                          offset, bytenr);
3190         } while (ret2 == -EAGAIN);
3191
3192         if (ret2 && ret2 != -ENOENT) {
3193                 ret = ret2;
3194                 goto out;
3195         }
3196
3197         if (ret != -ENOENT || ret2 != -ENOENT)
3198                 ret = 0;
3199 out:
3200         btrfs_free_path(path);
3201         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
3202                 WARN_ON(ret > 0);
3203         return ret;
3204 }
3205
3206 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
3207                            struct btrfs_root *root,
3208                            struct extent_buffer *buf,
3209                            int full_backref, int inc)
3210 {
3211         u64 bytenr;
3212         u64 num_bytes;
3213         u64 parent;
3214         u64 ref_root;
3215         u32 nritems;
3216         struct btrfs_key key;
3217         struct btrfs_file_extent_item *fi;
3218         int i;
3219         int level;
3220         int ret = 0;
3221         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
3222                             u64, u64, u64, u64, u64, u64);
3223
3224
3225         if (btrfs_is_testing(root->fs_info))
3226                 return 0;
3227
3228         ref_root = btrfs_header_owner(buf);
3229         nritems = btrfs_header_nritems(buf);
3230         level = btrfs_header_level(buf);
3231
3232         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state) && level == 0)
3233                 return 0;
3234
3235         if (inc)
3236                 process_func = btrfs_inc_extent_ref;
3237         else
3238                 process_func = btrfs_free_extent;
3239
3240         if (full_backref)
3241                 parent = buf->start;
3242         else
3243                 parent = 0;
3244
3245         for (i = 0; i < nritems; i++) {
3246                 if (level == 0) {
3247                         btrfs_item_key_to_cpu(buf, &key, i);
3248                         if (key.type != BTRFS_EXTENT_DATA_KEY)
3249                                 continue;
3250                         fi = btrfs_item_ptr(buf, i,
3251                                             struct btrfs_file_extent_item);
3252                         if (btrfs_file_extent_type(buf, fi) ==
3253                             BTRFS_FILE_EXTENT_INLINE)
3254                                 continue;
3255                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3256                         if (bytenr == 0)
3257                                 continue;
3258
3259                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3260                         key.offset -= btrfs_file_extent_offset(buf, fi);
3261                         ret = process_func(trans, root, bytenr, num_bytes,
3262                                            parent, ref_root, key.objectid,
3263                                            key.offset);
3264                         if (ret)
3265                                 goto fail;
3266                 } else {
3267                         bytenr = btrfs_node_blockptr(buf, i);
3268                         num_bytes = root->nodesize;
3269                         ret = process_func(trans, root, bytenr, num_bytes,
3270                                            parent, ref_root, level - 1, 0);
3271                         if (ret)
3272                                 goto fail;
3273                 }
3274         }
3275         return 0;
3276 fail:
3277         return ret;
3278 }
3279
3280 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3281                   struct extent_buffer *buf, int full_backref)
3282 {
3283         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
3284 }
3285
3286 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3287                   struct extent_buffer *buf, int full_backref)
3288 {
3289         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
3290 }
3291
3292 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3293                                  struct btrfs_root *root,
3294                                  struct btrfs_path *path,
3295                                  struct btrfs_block_group_cache *cache)
3296 {
3297         int ret;
3298         struct btrfs_root *extent_root = root->fs_info->extent_root;
3299         unsigned long bi;
3300         struct extent_buffer *leaf;
3301
3302         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3303         if (ret) {
3304                 if (ret > 0)
3305                         ret = -ENOENT;
3306                 goto fail;
3307         }
3308
3309         leaf = path->nodes[0];
3310         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3311         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3312         btrfs_mark_buffer_dirty(leaf);
3313 fail:
3314         btrfs_release_path(path);
3315         return ret;
3316
3317 }
3318
3319 static struct btrfs_block_group_cache *
3320 next_block_group(struct btrfs_root *root,
3321                  struct btrfs_block_group_cache *cache)
3322 {
3323         struct rb_node *node;
3324
3325         spin_lock(&root->fs_info->block_group_cache_lock);
3326
3327         /* If our block group was removed, we need a full search. */
3328         if (RB_EMPTY_NODE(&cache->cache_node)) {
3329                 const u64 next_bytenr = cache->key.objectid + cache->key.offset;
3330
3331                 spin_unlock(&root->fs_info->block_group_cache_lock);
3332                 btrfs_put_block_group(cache);
3333                 cache = btrfs_lookup_first_block_group(root->fs_info,
3334                                                        next_bytenr);
3335                 return cache;
3336         }
3337         node = rb_next(&cache->cache_node);
3338         btrfs_put_block_group(cache);
3339         if (node) {
3340                 cache = rb_entry(node, struct btrfs_block_group_cache,
3341                                  cache_node);
3342                 btrfs_get_block_group(cache);
3343         } else
3344                 cache = NULL;
3345         spin_unlock(&root->fs_info->block_group_cache_lock);
3346         return cache;
3347 }
3348
3349 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3350                             struct btrfs_trans_handle *trans,
3351                             struct btrfs_path *path)
3352 {
3353         struct btrfs_root *root = block_group->fs_info->tree_root;
3354         struct inode *inode = NULL;
3355         u64 alloc_hint = 0;
3356         int dcs = BTRFS_DC_ERROR;
3357         u64 num_pages = 0;
3358         int retries = 0;
3359         int ret = 0;
3360
3361         /*
3362          * If this block group is smaller than 100 megs don't bother caching the
3363          * block group.
3364          */
3365         if (block_group->key.offset < (100 * SZ_1M)) {
3366                 spin_lock(&block_group->lock);
3367                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3368                 spin_unlock(&block_group->lock);
3369                 return 0;
3370         }
3371
3372         if (trans->aborted)
3373                 return 0;
3374 again:
3375         inode = lookup_free_space_inode(root, block_group, path);
3376         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3377                 ret = PTR_ERR(inode);
3378                 btrfs_release_path(path);
3379                 goto out;
3380         }
3381
3382         if (IS_ERR(inode)) {
3383                 BUG_ON(retries);
3384                 retries++;
3385
3386                 if (block_group->ro)
3387                         goto out_free;
3388
3389                 ret = create_free_space_inode(root, trans, block_group, path);
3390                 if (ret)
3391                         goto out_free;
3392                 goto again;
3393         }
3394
3395         /* We've already setup this transaction, go ahead and exit */
3396         if (block_group->cache_generation == trans->transid &&
3397             i_size_read(inode)) {
3398                 dcs = BTRFS_DC_SETUP;
3399                 goto out_put;
3400         }
3401
3402         /*
3403          * We want to set the generation to 0, that way if anything goes wrong
3404          * from here on out we know not to trust this cache when we load up next
3405          * time.
3406          */
3407         BTRFS_I(inode)->generation = 0;
3408         ret = btrfs_update_inode(trans, root, inode);
3409         if (ret) {
3410                 /*
3411                  * So theoretically we could recover from this, simply set the
3412                  * super cache generation to 0 so we know to invalidate the
3413                  * cache, but then we'd have to keep track of the block groups
3414                  * that fail this way so we know we _have_ to reset this cache
3415                  * before the next commit or risk reading stale cache.  So to
3416                  * limit our exposure to horrible edge cases lets just abort the
3417                  * transaction, this only happens in really bad situations
3418                  * anyway.
3419                  */
3420                 btrfs_abort_transaction(trans, ret);
3421                 goto out_put;
3422         }
3423         WARN_ON(ret);
3424
3425         if (i_size_read(inode) > 0) {
3426                 ret = btrfs_check_trunc_cache_free_space(root,
3427                                         &root->fs_info->global_block_rsv);
3428                 if (ret)
3429                         goto out_put;
3430
3431                 ret = btrfs_truncate_free_space_cache(root, trans, NULL, inode);
3432                 if (ret)
3433                         goto out_put;
3434         }
3435
3436         spin_lock(&block_group->lock);
3437         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3438             !btrfs_test_opt(root->fs_info, SPACE_CACHE)) {
3439                 /*
3440                  * don't bother trying to write stuff out _if_
3441                  * a) we're not cached,
3442                  * b) we're with nospace_cache mount option.
3443                  */
3444                 dcs = BTRFS_DC_WRITTEN;
3445                 spin_unlock(&block_group->lock);
3446                 goto out_put;
3447         }
3448         spin_unlock(&block_group->lock);
3449
3450         /*
3451          * We hit an ENOSPC when setting up the cache in this transaction, just
3452          * skip doing the setup, we've already cleared the cache so we're safe.
3453          */
3454         if (test_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags)) {
3455                 ret = -ENOSPC;
3456                 goto out_put;
3457         }
3458
3459         /*
3460          * Try to preallocate enough space based on how big the block group is.
3461          * Keep in mind this has to include any pinned space which could end up
3462          * taking up quite a bit since it's not folded into the other space
3463          * cache.
3464          */
3465         num_pages = div_u64(block_group->key.offset, SZ_256M);
3466         if (!num_pages)
3467                 num_pages = 1;
3468
3469         num_pages *= 16;
3470         num_pages *= PAGE_SIZE;
3471
3472         ret = btrfs_check_data_free_space(inode, 0, num_pages);
3473         if (ret)
3474                 goto out_put;
3475
3476         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3477                                               num_pages, num_pages,
3478                                               &alloc_hint);
3479         /*
3480          * Our cache requires contiguous chunks so that we don't modify a bunch
3481          * of metadata or split extents when writing the cache out, which means
3482          * we can enospc if we are heavily fragmented in addition to just normal
3483          * out of space conditions.  So if we hit this just skip setting up any
3484          * other block groups for this transaction, maybe we'll unpin enough
3485          * space the next time around.
3486          */
3487         if (!ret)
3488                 dcs = BTRFS_DC_SETUP;
3489         else if (ret == -ENOSPC)
3490                 set_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags);
3491
3492 out_put:
3493         iput(inode);
3494 out_free:
3495         btrfs_release_path(path);
3496 out:
3497         spin_lock(&block_group->lock);
3498         if (!ret && dcs == BTRFS_DC_SETUP)
3499                 block_group->cache_generation = trans->transid;
3500         block_group->disk_cache_state = dcs;
3501         spin_unlock(&block_group->lock);
3502
3503         return ret;
3504 }
3505
3506 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans,
3507                             struct btrfs_root *root)
3508 {
3509         struct btrfs_block_group_cache *cache, *tmp;
3510         struct btrfs_transaction *cur_trans = trans->transaction;
3511         struct btrfs_path *path;
3512
3513         if (list_empty(&cur_trans->dirty_bgs) ||
3514             !btrfs_test_opt(root->fs_info, SPACE_CACHE))
3515                 return 0;
3516
3517         path = btrfs_alloc_path();
3518         if (!path)
3519                 return -ENOMEM;
3520
3521         /* Could add new block groups, use _safe just in case */
3522         list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3523                                  dirty_list) {
3524                 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3525                         cache_save_setup(cache, trans, path);
3526         }
3527
3528         btrfs_free_path(path);
3529         return 0;
3530 }
3531
3532 /*
3533  * transaction commit does final block group cache writeback during a
3534  * critical section where nothing is allowed to change the FS.  This is
3535  * required in order for the cache to actually match the block group,
3536  * but can introduce a lot of latency into the commit.
3537  *
3538  * So, btrfs_start_dirty_block_groups is here to kick off block group
3539  * cache IO.  There's a chance we'll have to redo some of it if the
3540  * block group changes again during the commit, but it greatly reduces
3541  * the commit latency by getting rid of the easy block groups while
3542  * we're still allowing others to join the commit.
3543  */
3544 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans,
3545                                    struct btrfs_root *root)
3546 {
3547         struct btrfs_block_group_cache *cache;
3548         struct btrfs_transaction *cur_trans = trans->transaction;
3549         int ret = 0;
3550         int should_put;
3551         struct btrfs_path *path = NULL;
3552         LIST_HEAD(dirty);
3553         struct list_head *io = &cur_trans->io_bgs;
3554         int num_started = 0;
3555         int loops = 0;
3556
3557         spin_lock(&cur_trans->dirty_bgs_lock);
3558         if (list_empty(&cur_trans->dirty_bgs)) {
3559                 spin_unlock(&cur_trans->dirty_bgs_lock);
3560                 return 0;
3561         }
3562         list_splice_init(&cur_trans->dirty_bgs, &dirty);
3563         spin_unlock(&cur_trans->dirty_bgs_lock);
3564
3565 again:
3566         /*
3567          * make sure all the block groups on our dirty list actually
3568          * exist
3569          */
3570         btrfs_create_pending_block_groups(trans, root);
3571
3572         if (!path) {
3573                 path = btrfs_alloc_path();
3574                 if (!path)
3575                         return -ENOMEM;
3576         }
3577
3578         /*
3579          * cache_write_mutex is here only to save us from balance or automatic
3580          * removal of empty block groups deleting this block group while we are
3581          * writing out the cache
3582          */
3583         mutex_lock(&trans->transaction->cache_write_mutex);
3584         while (!list_empty(&dirty)) {
3585                 cache = list_first_entry(&dirty,
3586                                          struct btrfs_block_group_cache,
3587                                          dirty_list);
3588                 /*
3589                  * this can happen if something re-dirties a block
3590                  * group that is already under IO.  Just wait for it to
3591                  * finish and then do it all again
3592                  */
3593                 if (!list_empty(&cache->io_list)) {
3594                         list_del_init(&cache->io_list);
3595                         btrfs_wait_cache_io(root, trans, cache,
3596                                             &cache->io_ctl, path,
3597                                             cache->key.objectid);
3598                         btrfs_put_block_group(cache);
3599                 }
3600
3601
3602                 /*
3603                  * btrfs_wait_cache_io uses the cache->dirty_list to decide
3604                  * if it should update the cache_state.  Don't delete
3605                  * until after we wait.
3606                  *
3607                  * Since we're not running in the commit critical section
3608                  * we need the dirty_bgs_lock to protect from update_block_group
3609                  */
3610                 spin_lock(&cur_trans->dirty_bgs_lock);
3611                 list_del_init(&cache->dirty_list);
3612                 spin_unlock(&cur_trans->dirty_bgs_lock);
3613
3614                 should_put = 1;
3615
3616                 cache_save_setup(cache, trans, path);
3617
3618                 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
3619                         cache->io_ctl.inode = NULL;
3620                         ret = btrfs_write_out_cache(root, trans, cache, path);
3621                         if (ret == 0 && cache->io_ctl.inode) {
3622                                 num_started++;
3623                                 should_put = 0;
3624
3625                                 /*
3626                                  * the cache_write_mutex is protecting
3627                                  * the io_list
3628                                  */
3629                                 list_add_tail(&cache->io_list, io);
3630                         } else {
3631                                 /*
3632                                  * if we failed to write the cache, the
3633                                  * generation will be bad and life goes on
3634                                  */
3635                                 ret = 0;
3636                         }
3637                 }
3638                 if (!ret) {
3639                         ret = write_one_cache_group(trans, root, path, cache);
3640                         /*
3641                          * Our block group might still be attached to the list
3642                          * of new block groups in the transaction handle of some
3643                          * other task (struct btrfs_trans_handle->new_bgs). This
3644                          * means its block group item isn't yet in the extent
3645                          * tree. If this happens ignore the error, as we will
3646                          * try again later in the critical section of the
3647                          * transaction commit.
3648                          */
3649                         if (ret == -ENOENT) {
3650                                 ret = 0;
3651                                 spin_lock(&cur_trans->dirty_bgs_lock);
3652                                 if (list_empty(&cache->dirty_list)) {
3653                                         list_add_tail(&cache->dirty_list,
3654                                                       &cur_trans->dirty_bgs);
3655                                         btrfs_get_block_group(cache);
3656                                 }
3657                                 spin_unlock(&cur_trans->dirty_bgs_lock);
3658                         } else if (ret) {
3659                                 btrfs_abort_transaction(trans, ret);
3660                         }
3661                 }
3662
3663                 /* if its not on the io list, we need to put the block group */
3664                 if (should_put)
3665                         btrfs_put_block_group(cache);
3666
3667                 if (ret)
3668                         break;
3669
3670                 /*
3671                  * Avoid blocking other tasks for too long. It might even save
3672                  * us from writing caches for block groups that are going to be
3673                  * removed.
3674                  */
3675                 mutex_unlock(&trans->transaction->cache_write_mutex);
3676                 mutex_lock(&trans->transaction->cache_write_mutex);
3677         }
3678         mutex_unlock(&trans->transaction->cache_write_mutex);
3679
3680         /*
3681          * go through delayed refs for all the stuff we've just kicked off
3682          * and then loop back (just once)
3683          */
3684         ret = btrfs_run_delayed_refs(trans, root, 0);
3685         if (!ret && loops == 0) {
3686                 loops++;
3687                 spin_lock(&cur_trans->dirty_bgs_lock);
3688                 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3689                 /*
3690                  * dirty_bgs_lock protects us from concurrent block group
3691                  * deletes too (not just cache_write_mutex).
3692                  */
3693                 if (!list_empty(&dirty)) {
3694                         spin_unlock(&cur_trans->dirty_bgs_lock);
3695                         goto again;
3696                 }
3697                 spin_unlock(&cur_trans->dirty_bgs_lock);
3698         } else if (ret < 0) {
3699                 btrfs_cleanup_dirty_bgs(cur_trans, root);
3700         }
3701
3702         btrfs_free_path(path);
3703         return ret;
3704 }
3705
3706 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3707                                    struct btrfs_root *root)
3708 {
3709         struct btrfs_block_group_cache *cache;
3710         struct btrfs_transaction *cur_trans = trans->transaction;
3711         int ret = 0;
3712         int should_put;
3713         struct btrfs_path *path;
3714         struct list_head *io = &cur_trans->io_bgs;
3715         int num_started = 0;
3716
3717         path = btrfs_alloc_path();
3718         if (!path)
3719                 return -ENOMEM;
3720
3721         /*
3722          * Even though we are in the critical section of the transaction commit,
3723          * we can still have concurrent tasks adding elements to this
3724          * transaction's list of dirty block groups. These tasks correspond to
3725          * endio free space workers started when writeback finishes for a
3726          * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
3727          * allocate new block groups as a result of COWing nodes of the root
3728          * tree when updating the free space inode. The writeback for the space
3729          * caches is triggered by an earlier call to
3730          * btrfs_start_dirty_block_groups() and iterations of the following
3731          * loop.
3732          * Also we want to do the cache_save_setup first and then run the
3733          * delayed refs to make sure we have the best chance at doing this all
3734          * in one shot.
3735          */
3736         spin_lock(&cur_trans->dirty_bgs_lock);
3737         while (!list_empty(&cur_trans->dirty_bgs)) {
3738                 cache = list_first_entry(&cur_trans->dirty_bgs,
3739                                          struct btrfs_block_group_cache,
3740                                          dirty_list);
3741
3742                 /*
3743                  * this can happen if cache_save_setup re-dirties a block
3744                  * group that is already under IO.  Just wait for it to
3745                  * finish and then do it all again
3746                  */
3747                 if (!list_empty(&cache->io_list)) {
3748                         spin_unlock(&cur_trans->dirty_bgs_lock);
3749                         list_del_init(&cache->io_list);
3750                         btrfs_wait_cache_io(root, trans, cache,
3751                                             &cache->io_ctl, path,
3752                                             cache->key.objectid);
3753                         btrfs_put_block_group(cache);
3754                         spin_lock(&cur_trans->dirty_bgs_lock);
3755                 }
3756
3757                 /*
3758                  * don't remove from the dirty list until after we've waited
3759                  * on any pending IO
3760                  */
3761                 list_del_init(&cache->dirty_list);
3762                 spin_unlock(&cur_trans->dirty_bgs_lock);
3763                 should_put = 1;
3764
3765                 cache_save_setup(cache, trans, path);
3766
3767                 if (!ret)
3768                         ret = btrfs_run_delayed_refs(trans, root, (unsigned long) -1);
3769
3770                 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3771                         cache->io_ctl.inode = NULL;
3772                         ret = btrfs_write_out_cache(root, trans, cache, path);
3773                         if (ret == 0 && cache->io_ctl.inode) {
3774                                 num_started++;
3775                                 should_put = 0;
3776                                 list_add_tail(&cache->io_list, io);
3777                         } else {
3778                                 /*
3779                                  * if we failed to write the cache, the
3780                                  * generation will be bad and life goes on
3781                                  */
3782                                 ret = 0;
3783                         }
3784                 }
3785                 if (!ret) {
3786                         ret = write_one_cache_group(trans, root, path, cache);
3787                         /*
3788                          * One of the free space endio workers might have
3789                          * created a new block group while updating a free space
3790                          * cache's inode (at inode.c:btrfs_finish_ordered_io())
3791                          * and hasn't released its transaction handle yet, in
3792                          * which case the new block group is still attached to
3793                          * its transaction handle and its creation has not
3794                          * finished yet (no block group item in the extent tree
3795                          * yet, etc). If this is the case, wait for all free
3796                          * space endio workers to finish and retry. This is a
3797                          * a very rare case so no need for a more efficient and
3798                          * complex approach.
3799                          */
3800                         if (ret == -ENOENT) {
3801                                 wait_event(cur_trans->writer_wait,
3802                                    atomic_read(&cur_trans->num_writers) == 1);
3803                                 ret = write_one_cache_group(trans, root, path,
3804                                                             cache);
3805                         }
3806                         if (ret)
3807                                 btrfs_abort_transaction(trans, ret);
3808                 }
3809
3810                 /* if its not on the io list, we need to put the block group */
3811                 if (should_put)
3812                         btrfs_put_block_group(cache);
3813                 spin_lock(&cur_trans->dirty_bgs_lock);
3814         }
3815         spin_unlock(&cur_trans->dirty_bgs_lock);
3816
3817         while (!list_empty(io)) {
3818                 cache = list_first_entry(io, struct btrfs_block_group_cache,
3819                                          io_list);
3820                 list_del_init(&cache->io_list);
3821                 btrfs_wait_cache_io(root, trans, cache,
3822                                     &cache->io_ctl, path, cache->key.objectid);
3823                 btrfs_put_block_group(cache);
3824         }
3825
3826         btrfs_free_path(path);
3827         return ret;
3828 }
3829
3830 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3831 {
3832         struct btrfs_block_group_cache *block_group;
3833         int readonly = 0;
3834
3835         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3836         if (!block_group || block_group->ro)
3837                 readonly = 1;
3838         if (block_group)
3839                 btrfs_put_block_group(block_group);
3840         return readonly;
3841 }
3842
3843 bool btrfs_inc_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3844 {
3845         struct btrfs_block_group_cache *bg;
3846         bool ret = true;
3847
3848         bg = btrfs_lookup_block_group(fs_info, bytenr);
3849         if (!bg)
3850                 return false;
3851
3852         spin_lock(&bg->lock);
3853         if (bg->ro)
3854                 ret = false;
3855         else
3856                 atomic_inc(&bg->nocow_writers);
3857         spin_unlock(&bg->lock);
3858
3859         /* no put on block group, done by btrfs_dec_nocow_writers */
3860         if (!ret)
3861                 btrfs_put_block_group(bg);
3862
3863         return ret;
3864
3865 }
3866
3867 void btrfs_dec_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3868 {
3869         struct btrfs_block_group_cache *bg;
3870
3871         bg = btrfs_lookup_block_group(fs_info, bytenr);
3872         ASSERT(bg);
3873         if (atomic_dec_and_test(&bg->nocow_writers))
3874                 wake_up_atomic_t(&bg->nocow_writers);
3875         /*
3876          * Once for our lookup and once for the lookup done by a previous call
3877          * to btrfs_inc_nocow_writers()
3878          */
3879         btrfs_put_block_group(bg);
3880         btrfs_put_block_group(bg);
3881 }
3882
3883 static int btrfs_wait_nocow_writers_atomic_t(atomic_t *a)
3884 {
3885         schedule();
3886         return 0;
3887 }
3888
3889 void btrfs_wait_nocow_writers(struct btrfs_block_group_cache *bg)
3890 {
3891         wait_on_atomic_t(&bg->nocow_writers,
3892                          btrfs_wait_nocow_writers_atomic_t,
3893                          TASK_UNINTERRUPTIBLE);
3894 }
3895
3896 static const char *alloc_name(u64 flags)
3897 {
3898         switch (flags) {
3899         case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA:
3900                 return "mixed";
3901         case BTRFS_BLOCK_GROUP_METADATA:
3902                 return "metadata";
3903         case BTRFS_BLOCK_GROUP_DATA:
3904                 return "data";
3905         case BTRFS_BLOCK_GROUP_SYSTEM:
3906                 return "system";
3907         default:
3908                 WARN_ON(1);
3909                 return "invalid-combination";
3910         };
3911 }
3912
3913 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3914                              u64 total_bytes, u64 bytes_used,
3915                              u64 bytes_readonly,
3916                              struct btrfs_space_info **space_info)
3917 {
3918         struct btrfs_space_info *found;
3919         int i;
3920         int factor;
3921         int ret;
3922
3923         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3924                      BTRFS_BLOCK_GROUP_RAID10))
3925                 factor = 2;
3926         else
3927                 factor = 1;
3928
3929         found = __find_space_info(info, flags);
3930         if (found) {
3931                 spin_lock(&found->lock);
3932                 found->total_bytes += total_bytes;
3933                 found->disk_total += total_bytes * factor;
3934                 found->bytes_used += bytes_used;
3935                 found->disk_used += bytes_used * factor;
3936                 found->bytes_readonly += bytes_readonly;
3937                 if (total_bytes > 0)
3938                         found->full = 0;
3939                 space_info_add_new_bytes(info, found, total_bytes -
3940                                          bytes_used - bytes_readonly);
3941                 spin_unlock(&found->lock);
3942                 *space_info = found;
3943                 return 0;
3944         }
3945         found = kzalloc(sizeof(*found), GFP_NOFS);
3946         if (!found)
3947                 return -ENOMEM;
3948
3949         ret = percpu_counter_init(&found->total_bytes_pinned, 0, GFP_KERNEL);
3950         if (ret) {
3951                 kfree(found);
3952                 return ret;
3953         }
3954
3955         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3956                 INIT_LIST_HEAD(&found->block_groups[i]);
3957         init_rwsem(&found->groups_sem);
3958         spin_lock_init(&found->lock);
3959         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3960         found->total_bytes = total_bytes;
3961         found->disk_total = total_bytes * factor;
3962         found->bytes_used = bytes_used;
3963         found->disk_used = bytes_used * factor;
3964         found->bytes_pinned = 0;
3965         found->bytes_reserved = 0;
3966         found->bytes_readonly = bytes_readonly;
3967         found->bytes_may_use = 0;
3968         found->full = 0;
3969         found->max_extent_size = 0;
3970         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3971         found->chunk_alloc = 0;
3972         found->flush = 0;
3973         init_waitqueue_head(&found->wait);
3974         INIT_LIST_HEAD(&found->ro_bgs);
3975         INIT_LIST_HEAD(&found->tickets);
3976         INIT_LIST_HEAD(&found->priority_tickets);
3977
3978         ret = kobject_init_and_add(&found->kobj, &space_info_ktype,
3979                                     info->space_info_kobj, "%s",
3980                                     alloc_name(found->flags));
3981         if (ret) {
3982                 kfree(found);
3983                 return ret;
3984         }
3985
3986         *space_info = found;
3987         list_add_rcu(&found->list, &info->space_info);
3988         if (flags & BTRFS_BLOCK_GROUP_DATA)
3989                 info->data_sinfo = found;
3990
3991         return ret;
3992 }
3993
3994 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3995 {
3996         u64 extra_flags = chunk_to_extended(flags) &
3997                                 BTRFS_EXTENDED_PROFILE_MASK;
3998
3999         write_seqlock(&fs_info->profiles_lock);
4000         if (flags & BTRFS_BLOCK_GROUP_DATA)
4001                 fs_info->avail_data_alloc_bits |= extra_flags;
4002         if (flags & BTRFS_BLOCK_GROUP_METADATA)
4003                 fs_info->avail_metadata_alloc_bits |= extra_flags;
4004         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4005                 fs_info->avail_system_alloc_bits |= extra_flags;
4006         write_sequnlock(&fs_info->profiles_lock);
4007 }
4008
4009 /*
4010  * returns target flags in extended format or 0 if restripe for this
4011  * chunk_type is not in progress
4012  *
4013  * should be called with either volume_mutex or balance_lock held
4014  */
4015 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
4016 {
4017         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
4018         u64 target = 0;
4019
4020         if (!bctl)
4021                 return 0;
4022
4023         if (flags & BTRFS_BLOCK_GROUP_DATA &&
4024             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4025                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
4026         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
4027                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4028                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
4029         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
4030                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4031                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
4032         }
4033
4034         return target;
4035 }
4036
4037 /*
4038  * @flags: available profiles in extended format (see ctree.h)
4039  *
4040  * Returns reduced profile in chunk format.  If profile changing is in
4041  * progress (either running or paused) picks the target profile (if it's
4042  * already available), otherwise falls back to plain reducing.
4043  */
4044 static u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
4045 {
4046         u64 num_devices = root->fs_info->fs_devices->rw_devices;
4047         u64 target;
4048         u64 raid_type;
4049         u64 allowed = 0;
4050
4051         /*
4052          * see if restripe for this chunk_type is in progress, if so
4053          * try to reduce to the target profile
4054          */
4055         spin_lock(&root->fs_info->balance_lock);
4056         target = get_restripe_target(root->fs_info, flags);
4057         if (target) {
4058                 /* pick target profile only if it's already available */
4059                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
4060                         spin_unlock(&root->fs_info->balance_lock);
4061                         return extended_to_chunk(target);
4062                 }
4063         }
4064         spin_unlock(&root->fs_info->balance_lock);
4065
4066         /* First, mask out the RAID levels which aren't possible */
4067         for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
4068                 if (num_devices >= btrfs_raid_array[raid_type].devs_min)
4069                         allowed |= btrfs_raid_group[raid_type];
4070         }
4071         allowed &= flags;
4072
4073         if (allowed & BTRFS_BLOCK_GROUP_RAID6)
4074                 allowed = BTRFS_BLOCK_GROUP_RAID6;
4075         else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
4076                 allowed = BTRFS_BLOCK_GROUP_RAID5;
4077         else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
4078                 allowed = BTRFS_BLOCK_GROUP_RAID10;
4079         else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
4080                 allowed = BTRFS_BLOCK_GROUP_RAID1;
4081         else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
4082                 allowed = BTRFS_BLOCK_GROUP_RAID0;
4083
4084         flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
4085
4086         return extended_to_chunk(flags | allowed);
4087 }
4088
4089 static u64 get_alloc_profile(struct btrfs_root *root, u64 orig_flags)
4090 {
4091         unsigned seq;
4092         u64 flags;
4093
4094         do {
4095                 flags = orig_flags;
4096                 seq = read_seqbegin(&root->fs_info->profiles_lock);
4097
4098                 if (flags & BTRFS_BLOCK_GROUP_DATA)
4099                         flags |= root->fs_info->avail_data_alloc_bits;
4100                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4101                         flags |= root->fs_info->avail_system_alloc_bits;
4102                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
4103                         flags |= root->fs_info->avail_metadata_alloc_bits;
4104         } while (read_seqretry(&root->fs_info->profiles_lock, seq));
4105
4106         return btrfs_reduce_alloc_profile(root, flags);
4107 }
4108
4109 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
4110 {
4111         u64 flags;
4112         u64 ret;
4113
4114         if (data)
4115                 flags = BTRFS_BLOCK_GROUP_DATA;
4116         else if (root == root->fs_info->chunk_root)
4117                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
4118         else
4119                 flags = BTRFS_BLOCK_GROUP_METADATA;
4120
4121         ret = get_alloc_profile(root, flags);
4122         return ret;
4123 }
4124
4125 int btrfs_alloc_data_chunk_ondemand(struct inode *inode, u64 bytes)
4126 {
4127         struct btrfs_space_info *data_sinfo;
4128         struct btrfs_root *root = BTRFS_I(inode)->root;
4129         struct btrfs_fs_info *fs_info = root->fs_info;
4130         u64 used;
4131         int ret = 0;
4132         int need_commit = 2;
4133         int have_pinned_space;
4134
4135         /* make sure bytes are sectorsize aligned */
4136         bytes = ALIGN(bytes, root->sectorsize);
4137
4138         if (btrfs_is_free_space_inode(inode)) {
4139                 need_commit = 0;
4140                 ASSERT(current->journal_info);
4141         }
4142
4143         data_sinfo = fs_info->data_sinfo;
4144         if (!data_sinfo)
4145                 goto alloc;
4146
4147 again:
4148         /* make sure we have enough space to handle the data first */
4149         spin_lock(&data_sinfo->lock);
4150         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
4151                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
4152                 data_sinfo->bytes_may_use;
4153
4154         if (used + bytes > data_sinfo->total_bytes) {
4155                 struct btrfs_trans_handle *trans;
4156
4157                 /*
4158                  * if we don't have enough free bytes in this space then we need
4159                  * to alloc a new chunk.
4160                  */
4161                 if (!data_sinfo->full) {
4162                         u64 alloc_target;
4163
4164                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
4165                         spin_unlock(&data_sinfo->lock);
4166 alloc:
4167                         alloc_target = btrfs_get_alloc_profile(root, 1);
4168                         /*
4169                          * It is ugly that we don't call nolock join
4170                          * transaction for the free space inode case here.
4171                          * But it is safe because we only do the data space
4172                          * reservation for the free space cache in the
4173                          * transaction context, the common join transaction
4174                          * just increase the counter of the current transaction
4175                          * handler, doesn't try to acquire the trans_lock of
4176                          * the fs.
4177                          */
4178                         trans = btrfs_join_transaction(root);
4179                         if (IS_ERR(trans))
4180                                 return PTR_ERR(trans);
4181
4182                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
4183                                              alloc_target,
4184                                              CHUNK_ALLOC_NO_FORCE);
4185                         btrfs_end_transaction(trans, root);
4186                         if (ret < 0) {
4187                                 if (ret != -ENOSPC)
4188                                         return ret;
4189                                 else {
4190                                         have_pinned_space = 1;
4191                                         goto commit_trans;
4192                                 }
4193                         }
4194
4195                         if (!data_sinfo)
4196                                 data_sinfo = fs_info->data_sinfo;
4197
4198                         goto again;
4199                 }
4200
4201                 /*
4202                  * If we don't have enough pinned space to deal with this
4203                  * allocation, and no removed chunk in current transaction,
4204                  * don't bother committing the transaction.
4205                  */
4206                 have_pinned_space = percpu_counter_compare(
4207                         &data_sinfo->total_bytes_pinned,
4208                         used + bytes - data_sinfo->total_bytes);
4209                 spin_unlock(&data_sinfo->lock);
4210
4211                 /* commit the current transaction and try again */
4212 commit_trans:
4213                 if (need_commit &&
4214                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
4215                         need_commit--;
4216
4217                         if (need_commit > 0) {
4218                                 btrfs_start_delalloc_roots(fs_info, 0, -1);
4219                                 btrfs_wait_ordered_roots(fs_info, -1, 0, (u64)-1);
4220                         }
4221
4222                         trans = btrfs_join_transaction(root);
4223                         if (IS_ERR(trans))
4224                                 return PTR_ERR(trans);
4225                         if (have_pinned_space >= 0 ||
4226                             test_bit(BTRFS_TRANS_HAVE_FREE_BGS,
4227                                      &trans->transaction->flags) ||
4228                             need_commit > 0) {
4229                                 ret = btrfs_commit_transaction(trans, root);
4230                                 if (ret)
4231                                         return ret;
4232                                 /*
4233                                  * The cleaner kthread might still be doing iput
4234                                  * operations. Wait for it to finish so that
4235                                  * more space is released.
4236                                  */
4237                                 mutex_lock(&root->fs_info->cleaner_delayed_iput_mutex);
4238                                 mutex_unlock(&root->fs_info->cleaner_delayed_iput_mutex);
4239                                 goto again;
4240                         } else {
4241                                 btrfs_end_transaction(trans, root);
4242                         }
4243                 }
4244
4245                 trace_btrfs_space_reservation(root->fs_info,
4246                                               "space_info:enospc",
4247                                               data_sinfo->flags, bytes, 1);
4248                 return -ENOSPC;
4249         }
4250         data_sinfo->bytes_may_use += bytes;
4251         trace_btrfs_space_reservation(root->fs_info, "space_info",
4252                                       data_sinfo->flags, bytes, 1);
4253         spin_unlock(&data_sinfo->lock);
4254
4255         return ret;
4256 }
4257
4258 /*
4259  * New check_data_free_space() with ability for precious data reservation
4260  * Will replace old btrfs_check_data_free_space(), but for patch split,
4261  * add a new function first and then replace it.
4262  */
4263 int btrfs_check_data_free_space(struct inode *inode, u64 start, u64 len)
4264 {
4265         struct btrfs_root *root = BTRFS_I(inode)->root;
4266         int ret;
4267
4268         /* align the range */
4269         len = round_up(start + len, root->sectorsize) -
4270               round_down(start, root->sectorsize);
4271         start = round_down(start, root->sectorsize);
4272
4273         ret = btrfs_alloc_data_chunk_ondemand(inode, len);
4274         if (ret < 0)
4275                 return ret;
4276
4277         /* Use new btrfs_qgroup_reserve_data to reserve precious data space. */
4278         ret = btrfs_qgroup_reserve_data(inode, start, len);
4279         if (ret)
4280                 btrfs_free_reserved_data_space_noquota(inode, start, len);
4281         return ret;
4282 }
4283
4284 /*
4285  * Called if we need to clear a data reservation for this inode
4286  * Normally in a error case.
4287  *
4288  * This one will *NOT* use accurate qgroup reserved space API, just for case
4289  * which we can't sleep and is sure it won't affect qgroup reserved space.
4290  * Like clear_bit_hook().
4291  */
4292 void btrfs_free_reserved_data_space_noquota(struct inode *inode, u64 start,
4293                                             u64 len)
4294 {
4295         struct btrfs_root *root = BTRFS_I(inode)->root;
4296         struct btrfs_space_info *data_sinfo;
4297
4298         /* Make sure the range is aligned to sectorsize */
4299         len = round_up(start + len, root->sectorsize) -
4300               round_down(start, root->sectorsize);
4301         start = round_down(start, root->sectorsize);
4302
4303         data_sinfo = root->fs_info->data_sinfo;
4304         spin_lock(&data_sinfo->lock);
4305         if (WARN_ON(data_sinfo->bytes_may_use < len))
4306                 data_sinfo->bytes_may_use = 0;
4307         else
4308                 data_sinfo->bytes_may_use -= len;
4309         trace_btrfs_space_reservation(root->fs_info, "space_info",
4310                                       data_sinfo->flags, len, 0);
4311         spin_unlock(&data_sinfo->lock);
4312 }
4313
4314 /*
4315  * Called if we need to clear a data reservation for this inode
4316  * Normally in a error case.
4317  *
4318  * This one will handle the per-inode data rsv map for accurate reserved
4319  * space framework.
4320  */
4321 void btrfs_free_reserved_data_space(struct inode *inode, u64 start, u64 len)
4322 {
4323         btrfs_free_reserved_data_space_noquota(inode, start, len);
4324         btrfs_qgroup_free_data(inode, start, len);
4325 }
4326
4327 static void force_metadata_allocation(struct btrfs_fs_info *info)
4328 {
4329         struct list_head *head = &info->space_info;
4330         struct btrfs_space_info *found;
4331
4332         rcu_read_lock();
4333         list_for_each_entry_rcu(found, head, list) {
4334                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
4335                         found->force_alloc = CHUNK_ALLOC_FORCE;
4336         }
4337         rcu_read_unlock();
4338 }
4339
4340 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
4341 {
4342         return (global->size << 1);
4343 }
4344
4345 static int should_alloc_chunk(struct btrfs_root *root,
4346                               struct btrfs_space_info *sinfo, int force)
4347 {
4348         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4349         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
4350         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
4351         u64 thresh;
4352
4353         if (force == CHUNK_ALLOC_FORCE)
4354                 return 1;
4355
4356         /*
4357          * We need to take into account the global rsv because for all intents
4358          * and purposes it's used space.  Don't worry about locking the
4359          * global_rsv, it doesn't change except when the transaction commits.
4360          */
4361         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
4362                 num_allocated += calc_global_rsv_need_space(global_rsv);
4363
4364         /*
4365          * in limited mode, we want to have some free space up to
4366          * about 1% of the FS size.
4367          */
4368         if (force == CHUNK_ALLOC_LIMITED) {
4369                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
4370                 thresh = max_t(u64, SZ_64M, div_factor_fine(thresh, 1));
4371
4372                 if (num_bytes - num_allocated < thresh)
4373                         return 1;
4374         }
4375
4376         if (num_allocated + SZ_2M < div_factor(num_bytes, 8))
4377                 return 0;
4378         return 1;
4379 }
4380
4381 static u64 get_profile_num_devs(struct btrfs_root *root, u64 type)
4382 {
4383         u64 num_dev;
4384
4385         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
4386                     BTRFS_BLOCK_GROUP_RAID0 |
4387                     BTRFS_BLOCK_GROUP_RAID5 |
4388                     BTRFS_BLOCK_GROUP_RAID6))
4389                 num_dev = root->fs_info->fs_devices->rw_devices;
4390         else if (type & BTRFS_BLOCK_GROUP_RAID1)
4391                 num_dev = 2;
4392         else
4393                 num_dev = 1;    /* DUP or single */
4394
4395         return num_dev;
4396 }
4397
4398 /*
4399  * If @is_allocation is true, reserve space in the system space info necessary
4400  * for allocating a chunk, otherwise if it's false, reserve space necessary for
4401  * removing a chunk.
4402  */
4403 void check_system_chunk(struct btrfs_trans_handle *trans,
4404                         struct btrfs_root *root,
4405                         u64 type)
4406 {
4407         struct btrfs_space_info *info;
4408         u64 left;
4409         u64 thresh;
4410         int ret = 0;
4411         u64 num_devs;
4412
4413         /*
4414          * Needed because we can end up allocating a system chunk and for an
4415          * atomic and race free space reservation in the chunk block reserve.
4416          */
4417         ASSERT(mutex_is_locked(&root->fs_info->chunk_mutex));
4418
4419         info = __find_space_info(root->fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4420         spin_lock(&info->lock);
4421         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
4422                 info->bytes_reserved - info->bytes_readonly -
4423                 info->bytes_may_use;
4424         spin_unlock(&info->lock);
4425
4426         num_devs = get_profile_num_devs(root, type);
4427
4428         /* num_devs device items to update and 1 chunk item to add or remove */
4429         thresh = btrfs_calc_trunc_metadata_size(root, num_devs) +
4430                 btrfs_calc_trans_metadata_size(root, 1);
4431
4432         if (left < thresh && btrfs_test_opt(root->fs_info, ENOSPC_DEBUG)) {
4433                 btrfs_info(root->fs_info, "left=%llu, need=%llu, flags=%llu",
4434                         left, thresh, type);
4435                 dump_space_info(root->fs_info, info, 0, 0);
4436         }
4437
4438         if (left < thresh) {
4439                 u64 flags;
4440
4441                 flags = btrfs_get_alloc_profile(root->fs_info->chunk_root, 0);
4442                 /*
4443                  * Ignore failure to create system chunk. We might end up not
4444                  * needing it, as we might not need to COW all nodes/leafs from
4445                  * the paths we visit in the chunk tree (they were already COWed
4446                  * or created in the current transaction for example).
4447                  */
4448                 ret = btrfs_alloc_chunk(trans, root, flags);
4449         }
4450
4451         if (!ret) {
4452                 ret = btrfs_block_rsv_add(root->fs_info->chunk_root,
4453                                           &root->fs_info->chunk_block_rsv,
4454                                           thresh, BTRFS_RESERVE_NO_FLUSH);
4455                 if (!ret)
4456                         trans->chunk_bytes_reserved += thresh;
4457         }
4458 }
4459
4460 /*
4461  * If force is CHUNK_ALLOC_FORCE:
4462  *    - return 1 if it successfully allocates a chunk,
4463  *    - return errors including -ENOSPC otherwise.
4464  * If force is NOT CHUNK_ALLOC_FORCE:
4465  *    - return 0 if it doesn't need to allocate a new chunk,
4466  *    - return 1 if it successfully allocates a chunk,
4467  *    - return errors including -ENOSPC otherwise.
4468  */
4469 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
4470                           struct btrfs_root *extent_root, u64 flags, int force)
4471 {
4472         struct btrfs_space_info *space_info;
4473         struct btrfs_fs_info *fs_info = extent_root->fs_info;
4474         int wait_for_alloc = 0;
4475         int ret = 0;
4476
4477         /* Don't re-enter if we're already allocating a chunk */
4478         if (trans->allocating_chunk)
4479                 return -ENOSPC;
4480
4481         space_info = __find_space_info(extent_root->fs_info, flags);
4482         if (!space_info) {
4483                 ret = update_space_info(extent_root->fs_info, flags,
4484                                         0, 0, 0, &space_info);
4485                 BUG_ON(ret); /* -ENOMEM */
4486         }
4487         BUG_ON(!space_info); /* Logic error */
4488
4489 again:
4490         spin_lock(&space_info->lock);
4491         if (force < space_info->force_alloc)
4492                 force = space_info->force_alloc;
4493         if (space_info->full) {
4494                 if (should_alloc_chunk(extent_root, space_info, force))
4495                         ret = -ENOSPC;
4496                 else
4497                         ret = 0;
4498                 spin_unlock(&space_info->lock);
4499                 return ret;
4500         }
4501
4502         if (!should_alloc_chunk(extent_root, space_info, force)) {
4503                 spin_unlock(&space_info->lock);
4504                 return 0;
4505         } else if (space_info->chunk_alloc) {
4506                 wait_for_alloc = 1;
4507         } else {
4508                 space_info->chunk_alloc = 1;
4509         }
4510
4511         spin_unlock(&space_info->lock);
4512
4513         mutex_lock(&fs_info->chunk_mutex);
4514
4515         /*
4516          * The chunk_mutex is held throughout the entirety of a chunk
4517          * allocation, so once we've acquired the chunk_mutex we know that the
4518          * other guy is done and we need to recheck and see if we should
4519          * allocate.
4520          */
4521         if (wait_for_alloc) {
4522                 mutex_unlock(&fs_info->chunk_mutex);
4523                 wait_for_alloc = 0;
4524                 goto again;
4525         }
4526
4527         trans->allocating_chunk = true;
4528
4529         /*
4530          * If we have mixed data/metadata chunks we want to make sure we keep
4531          * allocating mixed chunks instead of individual chunks.
4532          */
4533         if (btrfs_mixed_space_info(space_info))
4534                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4535
4536         /*
4537          * if we're doing a data chunk, go ahead and make sure that
4538          * we keep a reasonable number of metadata chunks allocated in the
4539          * FS as well.
4540          */
4541         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4542                 fs_info->data_chunk_allocations++;
4543                 if (!(fs_info->data_chunk_allocations %
4544                       fs_info->metadata_ratio))
4545                         force_metadata_allocation(fs_info);
4546         }
4547
4548         /*
4549          * Check if we have enough space in SYSTEM chunk because we may need
4550          * to update devices.
4551          */
4552         check_system_chunk(trans, extent_root, flags);
4553
4554         ret = btrfs_alloc_chunk(trans, extent_root, flags);
4555         trans->allocating_chunk = false;
4556
4557         spin_lock(&space_info->lock);
4558         if (ret < 0 && ret != -ENOSPC)
4559                 goto out;
4560         if (ret)
4561                 space_info->full = 1;
4562         else
4563                 ret = 1;
4564
4565         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4566 out:
4567         space_info->chunk_alloc = 0;
4568         spin_unlock(&space_info->lock);
4569         mutex_unlock(&fs_info->chunk_mutex);
4570         /*
4571          * When we allocate a new chunk we reserve space in the chunk block
4572          * reserve to make sure we can COW nodes/leafs in the chunk tree or
4573          * add new nodes/leafs to it if we end up needing to do it when
4574          * inserting the chunk item and updating device items as part of the
4575          * second phase of chunk allocation, performed by
4576          * btrfs_finish_chunk_alloc(). So make sure we don't accumulate a
4577          * large number of new block groups to create in our transaction
4578          * handle's new_bgs list to avoid exhausting the chunk block reserve
4579          * in extreme cases - like having a single transaction create many new
4580          * block groups when starting to write out the free space caches of all
4581          * the block groups that were made dirty during the lifetime of the
4582          * transaction.
4583          */
4584         if (trans->can_flush_pending_bgs &&
4585             trans->chunk_bytes_reserved >= (u64)SZ_2M) {
4586                 btrfs_create_pending_block_groups(trans, extent_root);
4587                 btrfs_trans_release_chunk_metadata(trans);
4588         }
4589         return ret;
4590 }
4591
4592 static int can_overcommit(struct btrfs_root *root,
4593                           struct btrfs_space_info *space_info, u64 bytes,
4594                           enum btrfs_reserve_flush_enum flush)
4595 {
4596         struct btrfs_block_rsv *global_rsv;
4597         u64 profile;
4598         u64 space_size;
4599         u64 avail;
4600         u64 used;
4601
4602         /* Don't overcommit when in mixed mode. */
4603         if (space_info->flags & BTRFS_BLOCK_GROUP_DATA)
4604                 return 0;
4605
4606         BUG_ON(root->fs_info == NULL);
4607         global_rsv = &root->fs_info->global_block_rsv;
4608         profile = btrfs_get_alloc_profile(root, 0);
4609         used = space_info->bytes_used + space_info->bytes_reserved +
4610                 space_info->bytes_pinned + space_info->bytes_readonly;
4611
4612         /*
4613          * We only want to allow over committing if we have lots of actual space
4614          * free, but if we don't have enough space to handle the global reserve
4615          * space then we could end up having a real enospc problem when trying
4616          * to allocate a chunk or some other such important allocation.
4617          */
4618         spin_lock(&global_rsv->lock);
4619         space_size = calc_global_rsv_need_space(global_rsv);
4620         spin_unlock(&global_rsv->lock);
4621         if (used + space_size >= space_info->total_bytes)
4622                 return 0;
4623
4624         used += space_info->bytes_may_use;
4625
4626         spin_lock(&root->fs_info->free_chunk_lock);
4627         avail = root->fs_info->free_chunk_space;
4628         spin_unlock(&root->fs_info->free_chunk_lock);
4629
4630         /*
4631          * If we have dup, raid1 or raid10 then only half of the free
4632          * space is actually useable.  For raid56, the space info used
4633          * doesn't include the parity drive, so we don't have to
4634          * change the math
4635          */
4636         if (profile & (BTRFS_BLOCK_GROUP_DUP |
4637                        BTRFS_BLOCK_GROUP_RAID1 |
4638                        BTRFS_BLOCK_GROUP_RAID10))
4639                 avail >>= 1;
4640
4641         /*
4642          * If we aren't flushing all things, let us overcommit up to
4643          * 1/2th of the space. If we can flush, don't let us overcommit
4644          * too much, let it overcommit up to 1/8 of the space.
4645          */
4646         if (flush == BTRFS_RESERVE_FLUSH_ALL)
4647                 avail >>= 3;
4648         else
4649                 avail >>= 1;
4650
4651         if (used + bytes < space_info->total_bytes + avail)
4652                 return 1;
4653         return 0;
4654 }
4655
4656 static void btrfs_writeback_inodes_sb_nr(struct btrfs_root *root,
4657                                          unsigned long nr_pages, int nr_items)
4658 {
4659         struct super_block *sb = root->fs_info->sb;
4660
4661         if (down_read_trylock(&sb->s_umount)) {
4662                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
4663                 up_read(&sb->s_umount);
4664         } else {
4665                 /*
4666                  * We needn't worry the filesystem going from r/w to r/o though
4667                  * we don't acquire ->s_umount mutex, because the filesystem
4668                  * should guarantee the delalloc inodes list be empty after
4669                  * the filesystem is readonly(all dirty pages are written to
4670                  * the disk).
4671                  */
4672                 btrfs_start_delalloc_roots(root->fs_info, 0, nr_items);
4673                 if (!current->journal_info)
4674                         btrfs_wait_ordered_roots(root->fs_info, nr_items,
4675                                                  0, (u64)-1);
4676         }
4677 }
4678
4679 static inline int calc_reclaim_items_nr(struct btrfs_root *root, u64 to_reclaim)
4680 {
4681         u64 bytes;
4682         int nr;
4683
4684         bytes = btrfs_calc_trans_metadata_size(root, 1);
4685         nr = (int)div64_u64(to_reclaim, bytes);
4686         if (!nr)
4687                 nr = 1;
4688         return nr;
4689 }
4690
4691 #define EXTENT_SIZE_PER_ITEM    SZ_256K
4692
4693 /*
4694  * shrink metadata reservation for delalloc
4695  */
4696 static void shrink_delalloc(struct btrfs_root *root, u64 to_reclaim, u64 orig,
4697                             bool wait_ordered)
4698 {
4699         struct btrfs_block_rsv *block_rsv;
4700         struct btrfs_space_info *space_info;
4701         struct btrfs_trans_handle *trans;
4702         u64 delalloc_bytes;
4703         u64 max_reclaim;
4704         long time_left;
4705         unsigned long nr_pages;
4706         int loops;
4707         int items;
4708         enum btrfs_reserve_flush_enum flush;
4709
4710         /* Calc the number of the pages we need flush for space reservation */
4711         items = calc_reclaim_items_nr(root, to_reclaim);
4712         to_reclaim = (u64)items * EXTENT_SIZE_PER_ITEM;
4713
4714         trans = (struct btrfs_trans_handle *)current->journal_info;
4715         block_rsv = &root->fs_info->delalloc_block_rsv;
4716         space_info = block_rsv->space_info;
4717
4718         delalloc_bytes = percpu_counter_sum_positive(
4719                                                 &root->fs_info->delalloc_bytes);
4720         if (delalloc_bytes == 0) {
4721                 if (trans)
4722                         return;
4723                 if (wait_ordered)
4724                         btrfs_wait_ordered_roots(root->fs_info, items,
4725                                                  0, (u64)-1);
4726                 return;
4727         }
4728
4729         loops = 0;
4730         while (delalloc_bytes && loops < 3) {
4731                 max_reclaim = min(delalloc_bytes, to_reclaim);
4732                 nr_pages = max_reclaim >> PAGE_SHIFT;
4733                 btrfs_writeback_inodes_sb_nr(root, nr_pages, items);
4734                 /*
4735                  * We need to wait for the async pages to actually start before
4736                  * we do anything.
4737                  */
4738                 max_reclaim = atomic_read(&root->fs_info->async_delalloc_pages);
4739                 if (!max_reclaim)
4740                         goto skip_async;
4741
4742                 if (max_reclaim <= nr_pages)
4743                         max_reclaim = 0;
4744                 else
4745                         max_reclaim -= nr_pages;
4746
4747                 wait_event(root->fs_info->async_submit_wait,
4748                            atomic_read(&root->fs_info->async_delalloc_pages) <=
4749                            (int)max_reclaim);
4750 skip_async:
4751                 if (!trans)
4752                         flush = BTRFS_RESERVE_FLUSH_ALL;
4753                 else
4754                         flush = BTRFS_RESERVE_NO_FLUSH;
4755                 spin_lock(&space_info->lock);
4756                 if (can_overcommit(root, space_info, orig, flush)) {
4757                         spin_unlock(&space_info->lock);
4758                         break;
4759                 }
4760                 if (list_empty(&space_info->tickets) &&
4761                     list_empty(&space_info->priority_tickets)) {
4762                         spin_unlock(&space_info->lock);
4763                         break;
4764                 }
4765                 spin_unlock(&space_info->lock);
4766
4767                 loops++;
4768                 if (wait_ordered && !trans) {
4769                         btrfs_wait_ordered_roots(root->fs_info, items,
4770                                                  0, (u64)-1);
4771                 } else {
4772                         time_left = schedule_timeout_killable(1);
4773                         if (time_left)
4774                                 break;
4775                 }
4776                 delalloc_bytes = percpu_counter_sum_positive(
4777                                                 &root->fs_info->delalloc_bytes);
4778         }
4779 }
4780
4781 /**
4782  * maybe_commit_transaction - possibly commit the transaction if its ok to
4783  * @root - the root we're allocating for
4784  * @bytes - the number of bytes we want to reserve
4785  * @force - force the commit
4786  *
4787  * This will check to make sure that committing the transaction will actually
4788  * get us somewhere and then commit the transaction if it does.  Otherwise it
4789  * will return -ENOSPC.
4790  */
4791 static int may_commit_transaction(struct btrfs_root *root,
4792                                   struct btrfs_space_info *space_info,
4793                                   u64 bytes, int force)
4794 {
4795         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
4796         struct btrfs_trans_handle *trans;
4797
4798         trans = (struct btrfs_trans_handle *)current->journal_info;
4799         if (trans)
4800                 return -EAGAIN;
4801
4802         if (force)
4803                 goto commit;
4804
4805         /* See if there is enough pinned space to make this reservation */
4806         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4807                                    bytes) >= 0)
4808                 goto commit;
4809
4810         /*
4811          * See if there is some space in the delayed insertion reservation for
4812          * this reservation.
4813          */
4814         if (space_info != delayed_rsv->space_info)
4815                 return -ENOSPC;
4816
4817         spin_lock(&delayed_rsv->lock);
4818         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4819                                    bytes - delayed_rsv->size) >= 0) {
4820                 spin_unlock(&delayed_rsv->lock);
4821                 return -ENOSPC;
4822         }
4823         spin_unlock(&delayed_rsv->lock);
4824
4825 commit:
4826         trans = btrfs_join_transaction(root);
4827         if (IS_ERR(trans))
4828                 return -ENOSPC;
4829
4830         return btrfs_commit_transaction(trans, root);
4831 }
4832
4833 struct reserve_ticket {
4834         u64 bytes;
4835         int error;
4836         struct list_head list;
4837         wait_queue_head_t wait;
4838 };
4839
4840 static int flush_space(struct btrfs_root *root,
4841                        struct btrfs_space_info *space_info, u64 num_bytes,
4842                        u64 orig_bytes, int state)
4843 {
4844         struct btrfs_trans_handle *trans;
4845         int nr;
4846         int ret = 0;
4847
4848         switch (state) {
4849         case FLUSH_DELAYED_ITEMS_NR:
4850         case FLUSH_DELAYED_ITEMS:
4851                 if (state == FLUSH_DELAYED_ITEMS_NR)
4852                         nr = calc_reclaim_items_nr(root, num_bytes) * 2;
4853                 else
4854                         nr = -1;
4855
4856                 trans = btrfs_join_transaction(root);
4857                 if (IS_ERR(trans)) {
4858                         ret = PTR_ERR(trans);
4859                         break;
4860                 }
4861                 ret = btrfs_run_delayed_items_nr(trans, root, nr);
4862                 btrfs_end_transaction(trans, root);
4863                 break;
4864         case FLUSH_DELALLOC:
4865         case FLUSH_DELALLOC_WAIT:
4866                 shrink_delalloc(root, num_bytes * 2, orig_bytes,
4867                                 state == FLUSH_DELALLOC_WAIT);
4868                 break;
4869         case ALLOC_CHUNK:
4870                 trans = btrfs_join_transaction(root);
4871                 if (IS_ERR(trans)) {
4872                         ret = PTR_ERR(trans);
4873                         break;
4874                 }
4875                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
4876                                      btrfs_get_alloc_profile(root, 0),
4877                                      CHUNK_ALLOC_NO_FORCE);
4878                 btrfs_end_transaction(trans, root);
4879                 if (ret > 0 || ret == -ENOSPC)
4880                         ret = 0;
4881                 break;
4882         case COMMIT_TRANS:
4883                 ret = may_commit_transaction(root, space_info, orig_bytes, 0);
4884                 break;
4885         default:
4886                 ret = -ENOSPC;
4887                 break;
4888         }
4889
4890         trace_btrfs_flush_space(root->fs_info, space_info->flags, num_bytes,
4891                                 orig_bytes, state, ret);
4892         return ret;
4893 }
4894
4895 static inline u64
4896 btrfs_calc_reclaim_metadata_size(struct btrfs_root *root,
4897                                  struct btrfs_space_info *space_info)
4898 {
4899         struct reserve_ticket *ticket;
4900         u64 used;
4901         u64 expected;
4902         u64 to_reclaim = 0;
4903
4904         list_for_each_entry(ticket, &space_info->tickets, list)
4905                 to_reclaim += ticket->bytes;
4906         list_for_each_entry(ticket, &space_info->priority_tickets, list)
4907                 to_reclaim += ticket->bytes;
4908         if (to_reclaim)
4909                 return to_reclaim;
4910
4911         to_reclaim = min_t(u64, num_online_cpus() * SZ_1M, SZ_16M);
4912         if (can_overcommit(root, space_info, to_reclaim,
4913                            BTRFS_RESERVE_FLUSH_ALL))
4914                 return 0;
4915
4916         used = space_info->bytes_used + space_info->bytes_reserved +
4917                space_info->bytes_pinned + space_info->bytes_readonly +
4918                space_info->bytes_may_use;
4919         if (can_overcommit(root, space_info, SZ_1M, BTRFS_RESERVE_FLUSH_ALL))
4920                 expected = div_factor_fine(space_info->total_bytes, 95);
4921         else
4922                 expected = div_factor_fine(space_info->total_bytes, 90);
4923
4924         if (used > expected)
4925                 to_reclaim = used - expected;
4926         else
4927                 to_reclaim = 0;
4928         to_reclaim = min(to_reclaim, space_info->bytes_may_use +
4929                                      space_info->bytes_reserved);
4930         return to_reclaim;
4931 }
4932
4933 static inline int need_do_async_reclaim(struct btrfs_space_info *space_info,
4934                                         struct btrfs_root *root, u64 used)
4935 {
4936         u64 thresh = div_factor_fine(space_info->total_bytes, 98);
4937
4938         /* If we're just plain full then async reclaim just slows us down. */
4939         if ((space_info->bytes_used + space_info->bytes_reserved) >= thresh)
4940                 return 0;
4941
4942         if (!btrfs_calc_reclaim_metadata_size(root, space_info))
4943                 return 0;
4944
4945         return (used >= thresh && !btrfs_fs_closing(root->fs_info) &&
4946                 !test_bit(BTRFS_FS_STATE_REMOUNTING,
4947                           &root->fs_info->fs_state));
4948 }
4949
4950 static void wake_all_tickets(struct list_head *head)
4951 {
4952         struct reserve_ticket *ticket;
4953
4954         while (!list_empty(head)) {
4955                 ticket = list_first_entry(head, struct reserve_ticket, list);
4956                 list_del_init(&ticket->list);
4957                 ticket->error = -ENOSPC;
4958                 wake_up(&ticket->wait);
4959         }
4960 }
4961
4962 /*
4963  * This is for normal flushers, we can wait all goddamned day if we want to.  We
4964  * will loop and continuously try to flush as long as we are making progress.
4965  * We count progress as clearing off tickets each time we have to loop.
4966  */
4967 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
4968 {
4969         struct btrfs_fs_info *fs_info;
4970         struct btrfs_space_info *space_info;
4971         u64 to_reclaim;
4972         int flush_state;
4973         int commit_cycles = 0;
4974         u64 last_tickets_id;
4975
4976         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
4977         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4978
4979         spin_lock(&space_info->lock);
4980         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info->fs_root,
4981                                                       space_info);
4982         if (!to_reclaim) {
4983                 space_info->flush = 0;
4984                 spin_unlock(&space_info->lock);
4985                 return;
4986         }
4987         last_tickets_id = space_info->tickets_id;
4988         spin_unlock(&space_info->lock);
4989
4990         flush_state = FLUSH_DELAYED_ITEMS_NR;
4991         do {
4992                 struct reserve_ticket *ticket;
4993                 int ret;
4994
4995                 ret = flush_space(fs_info->fs_root, space_info, to_reclaim,
4996                             to_reclaim, flush_state);
4997                 spin_lock(&space_info->lock);
4998                 if (list_empty(&space_info->tickets)) {
4999                         space_info->flush = 0;
5000                         spin_unlock(&space_info->lock);
5001                         return;
5002                 }
5003                 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info->fs_root,
5004                                                               space_info);
5005                 ticket = list_first_entry(&space_info->tickets,
5006                                           struct reserve_ticket, list);
5007                 if (last_tickets_id == space_info->tickets_id) {
5008                         flush_state++;
5009                 } else {
5010                         last_tickets_id = space_info->tickets_id;
5011                         flush_state = FLUSH_DELAYED_ITEMS_NR;
5012                         if (commit_cycles)
5013                                 commit_cycles--;
5014                 }
5015
5016                 if (flush_state > COMMIT_TRANS) {
5017                         commit_cycles++;
5018                         if (commit_cycles > 2) {
5019                                 wake_all_tickets(&space_info->tickets);
5020                                 space_info->flush = 0;
5021                         } else {
5022                                 flush_state = FLUSH_DELAYED_ITEMS_NR;
5023                         }
5024                 }
5025                 spin_unlock(&space_info->lock);
5026         } while (flush_state <= COMMIT_TRANS);
5027 }
5028
5029 void btrfs_init_async_reclaim_work(struct work_struct *work)
5030 {
5031         INIT_WORK(work, btrfs_async_reclaim_metadata_space);
5032 }
5033
5034 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info,
5035                                             struct btrfs_space_info *space_info,
5036                                             struct reserve_ticket *ticket)
5037 {
5038         u64 to_reclaim;
5039         int flush_state = FLUSH_DELAYED_ITEMS_NR;
5040
5041         spin_lock(&space_info->lock);
5042         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info->fs_root,
5043                                                       space_info);
5044         if (!to_reclaim) {
5045                 spin_unlock(&space_info->lock);
5046                 return;
5047         }
5048         spin_unlock(&space_info->lock);
5049
5050         do {
5051                 flush_space(fs_info->fs_root, space_info, to_reclaim,
5052                             to_reclaim, flush_state);
5053                 flush_state++;
5054                 spin_lock(&space_info->lock);
5055                 if (ticket->bytes == 0) {
5056                         spin_unlock(&space_info->lock);
5057                         return;
5058                 }
5059                 spin_unlock(&space_info->lock);
5060
5061                 /*
5062                  * Priority flushers can't wait on delalloc without
5063                  * deadlocking.
5064                  */
5065                 if (flush_state == FLUSH_DELALLOC ||
5066                     flush_state == FLUSH_DELALLOC_WAIT)
5067                         flush_state = ALLOC_CHUNK;
5068         } while (flush_state < COMMIT_TRANS);
5069 }
5070
5071 static int wait_reserve_ticket(struct btrfs_fs_info *fs_info,
5072                                struct btrfs_space_info *space_info,
5073                                struct reserve_ticket *ticket, u64 orig_bytes)
5074
5075 {
5076         DEFINE_WAIT(wait);
5077         int ret = 0;
5078
5079         spin_lock(&space_info->lock);
5080         while (ticket->bytes > 0 && ticket->error == 0) {
5081                 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE);
5082                 if (ret) {
5083                         ret = -EINTR;
5084                         break;
5085                 }
5086                 spin_unlock(&space_info->lock);
5087
5088                 schedule();
5089
5090                 finish_wait(&ticket->wait, &wait);
5091                 spin_lock(&space_info->lock);
5092         }
5093         if (!ret)
5094                 ret = ticket->error;
5095         if (!list_empty(&ticket->list))
5096                 list_del_init(&ticket->list);
5097         if (ticket->bytes && ticket->bytes < orig_bytes) {
5098                 u64 num_bytes = orig_bytes - ticket->bytes;
5099                 space_info->bytes_may_use -= num_bytes;
5100                 trace_btrfs_space_reservation(fs_info, "space_info",
5101                                               space_info->flags, num_bytes, 0);
5102         }
5103         spin_unlock(&space_info->lock);
5104
5105         return ret;
5106 }
5107
5108 /**
5109  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5110  * @root - the root we're allocating for
5111  * @space_info - the space info we want to allocate from
5112  * @orig_bytes - the number of bytes we want
5113  * @flush - whether or not we can flush to make our reservation
5114  *
5115  * This will reserve orig_bytes number of bytes from the space info associated
5116  * with the block_rsv.  If there is not enough space it will make an attempt to
5117  * flush out space to make room.  It will do this by flushing delalloc if
5118  * possible or committing the transaction.  If flush is 0 then no attempts to
5119  * regain reservations will be made and this will fail if there is not enough
5120  * space already.
5121  */
5122 static int __reserve_metadata_bytes(struct btrfs_root *root,
5123                                     struct btrfs_space_info *space_info,
5124                                     u64 orig_bytes,
5125                                     enum btrfs_reserve_flush_enum flush)
5126 {
5127         struct reserve_ticket ticket;
5128         u64 used;
5129         int ret = 0;
5130
5131         ASSERT(orig_bytes);
5132         ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL);
5133
5134         spin_lock(&space_info->lock);
5135         ret = -ENOSPC;
5136         used = space_info->bytes_used + space_info->bytes_reserved +
5137                 space_info->bytes_pinned + space_info->bytes_readonly +
5138                 space_info->bytes_may_use;
5139
5140         /*
5141          * If we have enough space then hooray, make our reservation and carry
5142          * on.  If not see if we can overcommit, and if we can, hooray carry on.
5143          * If not things get more complicated.
5144          */
5145         if (used + orig_bytes <= space_info->total_bytes) {
5146                 space_info->bytes_may_use += orig_bytes;
5147                 trace_btrfs_space_reservation(root->fs_info, "space_info",
5148                                               space_info->flags, orig_bytes,
5149                                               1);
5150                 ret = 0;
5151         } else if (can_overcommit(root, space_info, orig_bytes, flush)) {
5152                 space_info->bytes_may_use += orig_bytes;
5153                 trace_btrfs_space_reservation(root->fs_info, "space_info",
5154                                               space_info->flags, orig_bytes,
5155                                               1);
5156                 ret = 0;
5157         }
5158
5159         /*
5160          * If we couldn't make a reservation then setup our reservation ticket
5161          * and kick the async worker if it's not already running.
5162          *
5163          * If we are a priority flusher then we just need to add our ticket to
5164          * the list and we will do our own flushing further down.
5165          */
5166         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
5167                 ticket.bytes = orig_bytes;
5168                 ticket.error = 0;
5169                 init_waitqueue_head(&ticket.wait);
5170                 if (flush == BTRFS_RESERVE_FLUSH_ALL) {
5171                         list_add_tail(&ticket.list, &space_info->tickets);
5172                         if (!space_info->flush) {
5173                                 space_info->flush = 1;
5174                                 trace_btrfs_trigger_flush(root->fs_info,
5175                                                           space_info->flags,
5176                                                           orig_bytes, flush,
5177                                                           "enospc");
5178                                 queue_work(system_unbound_wq,
5179                                            &root->fs_info->async_reclaim_work);
5180                         }
5181                 } else {
5182                         list_add_tail(&ticket.list,
5183                                       &space_info->priority_tickets);
5184                 }
5185         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
5186                 used += orig_bytes;
5187                 /*
5188                  * We will do the space reservation dance during log replay,
5189                  * which means we won't have fs_info->fs_root set, so don't do
5190                  * the async reclaim as we will panic.
5191                  */
5192                 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &root->fs_info->flags) &&
5193                     need_do_async_reclaim(space_info, root, used) &&
5194                     !work_busy(&root->fs_info->async_reclaim_work)) {
5195                         trace_btrfs_trigger_flush(root->fs_info,
5196                                                   space_info->flags,
5197                                                   orig_bytes, flush,
5198                                                   "preempt");
5199                         queue_work(system_unbound_wq,
5200                                    &root->fs_info->async_reclaim_work);
5201                 }
5202         }
5203         spin_unlock(&space_info->lock);
5204         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
5205                 return ret;
5206
5207         if (flush == BTRFS_RESERVE_FLUSH_ALL)
5208                 return wait_reserve_ticket(root->fs_info, space_info, &ticket,
5209                                            orig_bytes);
5210
5211         ret = 0;
5212         priority_reclaim_metadata_space(root->fs_info, space_info, &ticket);
5213         spin_lock(&space_info->lock);
5214         if (ticket.bytes) {
5215                 if (ticket.bytes < orig_bytes) {
5216                         u64 num_bytes = orig_bytes - ticket.bytes;
5217                         space_info->bytes_may_use -= num_bytes;
5218                         trace_btrfs_space_reservation(root->fs_info,
5219                                         "space_info", space_info->flags,
5220                                         num_bytes, 0);
5221
5222                 }
5223                 list_del_init(&ticket.list);
5224                 ret = -ENOSPC;
5225         }
5226         spin_unlock(&space_info->lock);
5227         ASSERT(list_empty(&ticket.list));
5228         return ret;
5229 }
5230
5231 /**
5232  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5233  * @root - the root we're allocating for
5234  * @block_rsv - the block_rsv we're allocating for
5235  * @orig_bytes - the number of bytes we want
5236  * @flush - whether or not we can flush to make our reservation
5237  *
5238  * This will reserve orgi_bytes number of bytes from the space info associated
5239  * with the block_rsv.  If there is not enough space it will make an attempt to
5240  * flush out space to make room.  It will do this by flushing delalloc if
5241  * possible or committing the transaction.  If flush is 0 then no attempts to
5242  * regain reservations will be made and this will fail if there is not enough
5243  * space already.
5244  */
5245 static int reserve_metadata_bytes(struct btrfs_root *root,
5246                                   struct btrfs_block_rsv *block_rsv,
5247                                   u64 orig_bytes,
5248                                   enum btrfs_reserve_flush_enum flush)
5249 {
5250         int ret;
5251
5252         ret = __reserve_metadata_bytes(root, block_rsv->space_info, orig_bytes,
5253                                        flush);
5254         if (ret == -ENOSPC &&
5255             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
5256                 struct btrfs_block_rsv *global_rsv =
5257                         &root->fs_info->global_block_rsv;
5258
5259                 if (block_rsv != global_rsv &&
5260                     !block_rsv_use_bytes(global_rsv, orig_bytes))
5261                         ret = 0;
5262         }
5263         if (ret == -ENOSPC)
5264                 trace_btrfs_space_reservation(root->fs_info,
5265                                               "space_info:enospc",
5266                                               block_rsv->space_info->flags,
5267                                               orig_bytes, 1);
5268         return ret;
5269 }
5270
5271 static struct btrfs_block_rsv *get_block_rsv(
5272                                         const struct btrfs_trans_handle *trans,
5273                                         const struct btrfs_root *root)
5274 {
5275         struct btrfs_block_rsv *block_rsv = NULL;
5276
5277         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
5278             (root == root->fs_info->csum_root && trans->adding_csums) ||
5279              (root == root->fs_info->uuid_root))
5280                 block_rsv = trans->block_rsv;
5281
5282         if (!block_rsv)
5283                 block_rsv = root->block_rsv;
5284
5285         if (!block_rsv)
5286                 block_rsv = &root->fs_info->empty_block_rsv;
5287
5288         return block_rsv;
5289 }
5290
5291 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
5292                                u64 num_bytes)
5293 {
5294         int ret = -ENOSPC;
5295         spin_lock(&block_rsv->lock);
5296         if (block_rsv->reserved >= num_bytes) {
5297                 block_rsv->reserved -= num_bytes;
5298                 if (block_rsv->reserved < block_rsv->size)
5299                         block_rsv->full = 0;
5300                 ret = 0;
5301         }
5302         spin_unlock(&block_rsv->lock);
5303         return ret;
5304 }
5305
5306 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
5307                                 u64 num_bytes, int update_size)
5308 {
5309         spin_lock(&block_rsv->lock);
5310         block_rsv->reserved += num_bytes;
5311         if (update_size)
5312                 block_rsv->size += num_bytes;
5313         else if (block_rsv->reserved >= block_rsv->size)
5314                 block_rsv->full = 1;
5315         spin_unlock(&block_rsv->lock);
5316 }
5317
5318 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
5319                              struct btrfs_block_rsv *dest, u64 num_bytes,
5320                              int min_factor)
5321 {
5322         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5323         u64 min_bytes;
5324
5325         if (global_rsv->space_info != dest->space_info)
5326                 return -ENOSPC;
5327
5328         spin_lock(&global_rsv->lock);
5329         min_bytes = div_factor(global_rsv->size, min_factor);
5330         if (global_rsv->reserved < min_bytes + num_bytes) {
5331                 spin_unlock(&global_rsv->lock);
5332                 return -ENOSPC;
5333         }
5334         global_rsv->reserved -= num_bytes;
5335         if (global_rsv->reserved < global_rsv->size)
5336                 global_rsv->full = 0;
5337         spin_unlock(&global_rsv->lock);
5338
5339         block_rsv_add_bytes(dest, num_bytes, 1);
5340         return 0;
5341 }
5342
5343 /*
5344  * This is for space we already have accounted in space_info->bytes_may_use, so
5345  * basically when we're returning space from block_rsv's.
5346  */
5347 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
5348                                      struct btrfs_space_info *space_info,
5349                                      u64 num_bytes)
5350 {
5351         struct reserve_ticket *ticket;
5352         struct list_head *head;
5353         u64 used;
5354         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH;
5355         bool check_overcommit = false;
5356
5357         spin_lock(&space_info->lock);
5358         head = &space_info->priority_tickets;
5359
5360         /*
5361          * If we are over our limit then we need to check and see if we can
5362          * overcommit, and if we can't then we just need to free up our space
5363          * and not satisfy any requests.
5364          */
5365         used = space_info->bytes_used + space_info->bytes_reserved +
5366                 space_info->bytes_pinned + space_info->bytes_readonly +
5367                 space_info->bytes_may_use;
5368         if (used - num_bytes >= space_info->total_bytes)
5369                 check_overcommit = true;
5370 again:
5371         while (!list_empty(head) && num_bytes) {
5372                 ticket = list_first_entry(head, struct reserve_ticket,
5373                                           list);
5374                 /*
5375                  * We use 0 bytes because this space is already reserved, so
5376                  * adding the ticket space would be a double count.
5377                  */
5378                 if (check_overcommit &&
5379                     !can_overcommit(fs_info->extent_root, space_info, 0,
5380                                     flush))
5381                         break;
5382                 if (num_bytes >= ticket->bytes) {
5383                         list_del_init(&ticket->list);
5384                         num_bytes -= ticket->bytes;
5385                         ticket->bytes = 0;
5386                         space_info->tickets_id++;
5387                         wake_up(&ticket->wait);
5388                 } else {
5389                         ticket->bytes -= num_bytes;
5390                         num_bytes = 0;
5391                 }
5392         }
5393
5394         if (num_bytes && head == &space_info->priority_tickets) {
5395                 head = &space_info->tickets;
5396                 flush = BTRFS_RESERVE_FLUSH_ALL;
5397                 goto again;
5398         }
5399         space_info->bytes_may_use -= num_bytes;
5400         trace_btrfs_space_reservation(fs_info, "space_info",
5401                                       space_info->flags, num_bytes, 0);
5402         spin_unlock(&space_info->lock);
5403 }
5404
5405 /*
5406  * This is for newly allocated space that isn't accounted in
5407  * space_info->bytes_may_use yet.  So if we allocate a chunk or unpin an extent
5408  * we use this helper.
5409  */
5410 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
5411                                      struct btrfs_space_info *space_info,
5412                                      u64 num_bytes)
5413 {
5414         struct reserve_ticket *ticket;
5415         struct list_head *head = &space_info->priority_tickets;
5416
5417 again:
5418         while (!list_empty(head) && num_bytes) {
5419                 ticket = list_first_entry(head, struct reserve_ticket,
5420                                           list);
5421                 if (num_bytes >= ticket->bytes) {
5422                         trace_btrfs_space_reservation(fs_info, "space_info",
5423                                                       space_info->flags,
5424                                                       ticket->bytes, 1);
5425                         list_del_init(&ticket->list);
5426                         num_bytes -= ticket->bytes;
5427                         space_info->bytes_may_use += ticket->bytes;
5428                         ticket->bytes = 0;
5429                         space_info->tickets_id++;
5430                         wake_up(&ticket->wait);
5431                 } else {
5432                         trace_btrfs_space_reservation(fs_info, "space_info",
5433                                                       space_info->flags,
5434                                                       num_bytes, 1);
5435                         space_info->bytes_may_use += num_bytes;
5436                         ticket->bytes -= num_bytes;
5437                         num_bytes = 0;
5438                 }
5439         }
5440
5441         if (num_bytes && head == &space_info->priority_tickets) {
5442                 head = &space_info->tickets;
5443                 goto again;
5444         }
5445 }
5446
5447 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
5448                                     struct btrfs_block_rsv *block_rsv,
5449                                     struct btrfs_block_rsv *dest, u64 num_bytes)
5450 {
5451         struct btrfs_space_info *space_info = block_rsv->space_info;
5452
5453         spin_lock(&block_rsv->lock);
5454         if (num_bytes == (u64)-1)
5455                 num_bytes = block_rsv->size;
5456         block_rsv->size -= num_bytes;
5457         if (block_rsv->reserved >= block_rsv->size) {
5458                 num_bytes = block_rsv->reserved - block_rsv->size;
5459                 block_rsv->reserved = block_rsv->size;
5460                 block_rsv->full = 1;
5461         } else {
5462                 num_bytes = 0;
5463         }
5464         spin_unlock(&block_rsv->lock);
5465
5466         if (num_bytes > 0) {
5467                 if (dest) {
5468                         spin_lock(&dest->lock);
5469                         if (!dest->full) {
5470                                 u64 bytes_to_add;
5471
5472                                 bytes_to_add = dest->size - dest->reserved;
5473                                 bytes_to_add = min(num_bytes, bytes_to_add);
5474                                 dest->reserved += bytes_to_add;
5475                                 if (dest->reserved >= dest->size)
5476                                         dest->full = 1;
5477                                 num_bytes -= bytes_to_add;
5478                         }
5479                         spin_unlock(&dest->lock);
5480                 }
5481                 if (num_bytes)
5482                         space_info_add_old_bytes(fs_info, space_info,
5483                                                  num_bytes);
5484         }
5485 }
5486
5487 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src,
5488                             struct btrfs_block_rsv *dst, u64 num_bytes,
5489                             int update_size)
5490 {
5491         int ret;
5492
5493         ret = block_rsv_use_bytes(src, num_bytes);
5494         if (ret)
5495                 return ret;
5496
5497         block_rsv_add_bytes(dst, num_bytes, update_size);
5498         return 0;
5499 }
5500
5501 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
5502 {
5503         memset(rsv, 0, sizeof(*rsv));
5504         spin_lock_init(&rsv->lock);
5505         rsv->type = type;
5506 }
5507
5508 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root,
5509                                               unsigned short type)
5510 {
5511         struct btrfs_block_rsv *block_rsv;
5512         struct btrfs_fs_info *fs_info = root->fs_info;
5513
5514         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
5515         if (!block_rsv)
5516                 return NULL;
5517
5518         btrfs_init_block_rsv(block_rsv, type);
5519         block_rsv->space_info = __find_space_info(fs_info,
5520                                                   BTRFS_BLOCK_GROUP_METADATA);
5521         return block_rsv;
5522 }
5523
5524 void btrfs_free_block_rsv(struct btrfs_root *root,
5525                           struct btrfs_block_rsv *rsv)
5526 {
5527         if (!rsv)
5528                 return;
5529         btrfs_block_rsv_release(root, rsv, (u64)-1);
5530         kfree(rsv);
5531 }
5532
5533 void __btrfs_free_block_rsv(struct btrfs_block_rsv *rsv)
5534 {
5535         kfree(rsv);
5536 }
5537
5538 int btrfs_block_rsv_add(struct btrfs_root *root,
5539                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
5540                         enum btrfs_reserve_flush_enum flush)
5541 {
5542         int ret;
5543
5544         if (num_bytes == 0)
5545                 return 0;
5546
5547         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5548         if (!ret) {
5549                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
5550                 return 0;
5551         }
5552
5553         return ret;
5554 }
5555
5556 int btrfs_block_rsv_check(struct btrfs_root *root,
5557                           struct btrfs_block_rsv *block_rsv, int min_factor)
5558 {
5559         u64 num_bytes = 0;
5560         int ret = -ENOSPC;
5561
5562         if (!block_rsv)
5563                 return 0;
5564
5565         spin_lock(&block_rsv->lock);
5566         num_bytes = div_factor(block_rsv->size, min_factor);
5567         if (block_rsv->reserved >= num_bytes)
5568                 ret = 0;
5569         spin_unlock(&block_rsv->lock);
5570
5571         return ret;
5572 }
5573
5574 int btrfs_block_rsv_refill(struct btrfs_root *root,
5575                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
5576                            enum btrfs_reserve_flush_enum flush)
5577 {
5578         u64 num_bytes = 0;
5579         int ret = -ENOSPC;
5580
5581         if (!block_rsv)
5582                 return 0;
5583
5584         spin_lock(&block_rsv->lock);
5585         num_bytes = min_reserved;
5586         if (block_rsv->reserved >= num_bytes)
5587                 ret = 0;
5588         else
5589                 num_bytes -= block_rsv->reserved;
5590         spin_unlock(&block_rsv->lock);
5591
5592         if (!ret)
5593                 return 0;
5594
5595         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5596         if (!ret) {
5597                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
5598                 return 0;
5599         }
5600
5601         return ret;
5602 }
5603
5604 void btrfs_block_rsv_release(struct btrfs_root *root,
5605                              struct btrfs_block_rsv *block_rsv,
5606                              u64 num_bytes)
5607 {
5608         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
5609         if (global_rsv == block_rsv ||
5610             block_rsv->space_info != global_rsv->space_info)
5611                 global_rsv = NULL;
5612         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
5613                                 num_bytes);
5614 }
5615
5616 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
5617 {
5618         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
5619         struct btrfs_space_info *sinfo = block_rsv->space_info;
5620         u64 num_bytes;
5621
5622         /*
5623          * The global block rsv is based on the size of the extent tree, the
5624          * checksum tree and the root tree.  If the fs is empty we want to set
5625          * it to a minimal amount for safety.
5626          */
5627         num_bytes = btrfs_root_used(&fs_info->extent_root->root_item) +
5628                 btrfs_root_used(&fs_info->csum_root->root_item) +
5629                 btrfs_root_used(&fs_info->tree_root->root_item);
5630         num_bytes = max_t(u64, num_bytes, SZ_16M);
5631
5632         spin_lock(&sinfo->lock);
5633         spin_lock(&block_rsv->lock);
5634
5635         block_rsv->size = min_t(u64, num_bytes, SZ_512M);
5636
5637         if (block_rsv->reserved < block_rsv->size) {
5638                 num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
5639                         sinfo->bytes_reserved + sinfo->bytes_readonly +
5640                         sinfo->bytes_may_use;
5641                 if (sinfo->total_bytes > num_bytes) {
5642                         num_bytes = sinfo->total_bytes - num_bytes;
5643                         num_bytes = min(num_bytes,
5644                                         block_rsv->size - block_rsv->reserved);
5645                         block_rsv->reserved += num_bytes;
5646                         sinfo->bytes_may_use += num_bytes;
5647                         trace_btrfs_space_reservation(fs_info, "space_info",
5648                                                       sinfo->flags, num_bytes,
5649                                                       1);
5650                 }
5651         } else if (block_rsv->reserved > block_rsv->size) {
5652                 num_bytes = block_rsv->reserved - block_rsv->size;
5653                 sinfo->bytes_may_use -= num_bytes;
5654                 trace_btrfs_space_reservation(fs_info, "space_info",
5655                                       sinfo->flags, num_bytes, 0);
5656                 block_rsv->reserved = block_rsv->size;
5657         }
5658
5659         if (block_rsv->reserved == block_rsv->size)
5660                 block_rsv->full = 1;
5661         else
5662                 block_rsv->full = 0;
5663
5664         spin_unlock(&block_rsv->lock);
5665         spin_unlock(&sinfo->lock);
5666 }
5667
5668 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
5669 {
5670         struct btrfs_space_info *space_info;
5671
5672         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
5673         fs_info->chunk_block_rsv.space_info = space_info;
5674
5675         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5676         fs_info->global_block_rsv.space_info = space_info;
5677         fs_info->delalloc_block_rsv.space_info = space_info;
5678         fs_info->trans_block_rsv.space_info = space_info;
5679         fs_info->empty_block_rsv.space_info = space_info;
5680         fs_info->delayed_block_rsv.space_info = space_info;
5681
5682         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
5683         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
5684         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
5685         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
5686         if (fs_info->quota_root)
5687                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
5688         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
5689
5690         update_global_block_rsv(fs_info);
5691 }
5692
5693 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
5694 {
5695         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
5696                                 (u64)-1);
5697         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
5698         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
5699         WARN_ON(fs_info->trans_block_rsv.size > 0);
5700         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
5701         WARN_ON(fs_info->chunk_block_rsv.size > 0);
5702         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
5703         WARN_ON(fs_info->delayed_block_rsv.size > 0);
5704         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
5705 }
5706
5707 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
5708                                   struct btrfs_root *root)
5709 {
5710         if (!trans->block_rsv)
5711                 return;
5712
5713         if (!trans->bytes_reserved)
5714                 return;
5715
5716         trace_btrfs_space_reservation(root->fs_info, "transaction",
5717                                       trans->transid, trans->bytes_reserved, 0);
5718         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
5719         trans->bytes_reserved = 0;
5720 }
5721
5722 /*
5723  * To be called after all the new block groups attached to the transaction
5724  * handle have been created (btrfs_create_pending_block_groups()).
5725  */
5726 void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans)
5727 {
5728         struct btrfs_fs_info *fs_info = trans->fs_info;
5729
5730         if (!trans->chunk_bytes_reserved)
5731                 return;
5732
5733         WARN_ON_ONCE(!list_empty(&trans->new_bgs));
5734
5735         block_rsv_release_bytes(fs_info, &fs_info->chunk_block_rsv, NULL,
5736                                 trans->chunk_bytes_reserved);
5737         trans->chunk_bytes_reserved = 0;
5738 }
5739
5740 /* Can only return 0 or -ENOSPC */
5741 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
5742                                   struct inode *inode)
5743 {
5744         struct btrfs_root *root = BTRFS_I(inode)->root;
5745         /*
5746          * We always use trans->block_rsv here as we will have reserved space
5747          * for our orphan when starting the transaction, using get_block_rsv()
5748          * here will sometimes make us choose the wrong block rsv as we could be
5749          * doing a reloc inode for a non refcounted root.
5750          */
5751         struct btrfs_block_rsv *src_rsv = trans->block_rsv;
5752         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
5753
5754         /*
5755          * We need to hold space in order to delete our orphan item once we've
5756          * added it, so this takes the reservation so we can release it later
5757          * when we are truly done with the orphan item.
5758          */
5759         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
5760         trace_btrfs_space_reservation(root->fs_info, "orphan",
5761                                       btrfs_ino(inode), num_bytes, 1);
5762         return btrfs_block_rsv_migrate(src_rsv, dst_rsv, num_bytes, 1);
5763 }
5764
5765 void btrfs_orphan_release_metadata(struct inode *inode)
5766 {
5767         struct btrfs_root *root = BTRFS_I(inode)->root;
5768         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
5769         trace_btrfs_space_reservation(root->fs_info, "orphan",
5770                                       btrfs_ino(inode), num_bytes, 0);
5771         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
5772 }
5773
5774 /*
5775  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
5776  * root: the root of the parent directory
5777  * rsv: block reservation
5778  * items: the number of items that we need do reservation
5779  * qgroup_reserved: used to return the reserved size in qgroup
5780  *
5781  * This function is used to reserve the space for snapshot/subvolume
5782  * creation and deletion. Those operations are different with the
5783  * common file/directory operations, they change two fs/file trees
5784  * and root tree, the number of items that the qgroup reserves is
5785  * different with the free space reservation. So we can not use
5786  * the space reservation mechanism in start_transaction().
5787  */
5788 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
5789                                      struct btrfs_block_rsv *rsv,
5790                                      int items,
5791                                      u64 *qgroup_reserved,
5792                                      bool use_global_rsv)
5793 {
5794         u64 num_bytes;
5795         int ret;
5796         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
5797
5798         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &root->fs_info->flags)) {
5799                 /* One for parent inode, two for dir entries */
5800                 num_bytes = 3 * root->nodesize;
5801                 ret = btrfs_qgroup_reserve_meta(root, num_bytes);
5802                 if (ret)
5803                         return ret;
5804         } else {
5805                 num_bytes = 0;
5806         }
5807
5808         *qgroup_reserved = num_bytes;
5809
5810         num_bytes = btrfs_calc_trans_metadata_size(root, items);
5811         rsv->space_info = __find_space_info(root->fs_info,
5812                                             BTRFS_BLOCK_GROUP_METADATA);
5813         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
5814                                   BTRFS_RESERVE_FLUSH_ALL);
5815
5816         if (ret == -ENOSPC && use_global_rsv)
5817                 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes, 1);
5818
5819         if (ret && *qgroup_reserved)
5820                 btrfs_qgroup_free_meta(root, *qgroup_reserved);
5821
5822         return ret;
5823 }
5824
5825 void btrfs_subvolume_release_metadata(struct btrfs_root *root,
5826                                       struct btrfs_block_rsv *rsv,
5827                                       u64 qgroup_reserved)
5828 {
5829         btrfs_block_rsv_release(root, rsv, (u64)-1);
5830 }
5831
5832 /**
5833  * drop_outstanding_extent - drop an outstanding extent
5834  * @inode: the inode we're dropping the extent for
5835  * @num_bytes: the number of bytes we're releasing.
5836  *
5837  * This is called when we are freeing up an outstanding extent, either called
5838  * after an error or after an extent is written.  This will return the number of
5839  * reserved extents that need to be freed.  This must be called with
5840  * BTRFS_I(inode)->lock held.
5841  */
5842 static unsigned drop_outstanding_extent(struct inode *inode, u64 num_bytes)
5843 {
5844         unsigned drop_inode_space = 0;
5845         unsigned dropped_extents = 0;
5846         unsigned num_extents = 0;
5847
5848         num_extents = (unsigned)div64_u64(num_bytes +
5849                                           BTRFS_MAX_EXTENT_SIZE - 1,
5850                                           BTRFS_MAX_EXTENT_SIZE);
5851         ASSERT(num_extents);
5852         ASSERT(BTRFS_I(inode)->outstanding_extents >= num_extents);
5853         BTRFS_I(inode)->outstanding_extents -= num_extents;
5854
5855         if (BTRFS_I(inode)->outstanding_extents == 0 &&
5856             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5857                                &BTRFS_I(inode)->runtime_flags))
5858                 drop_inode_space = 1;
5859
5860         /*
5861          * If we have more or the same amount of outstanding extents than we have
5862          * reserved then we need to leave the reserved extents count alone.
5863          */
5864         if (BTRFS_I(inode)->outstanding_extents >=
5865             BTRFS_I(inode)->reserved_extents)
5866                 return drop_inode_space;
5867
5868         dropped_extents = BTRFS_I(inode)->reserved_extents -
5869                 BTRFS_I(inode)->outstanding_extents;
5870         BTRFS_I(inode)->reserved_extents -= dropped_extents;
5871         return dropped_extents + drop_inode_space;
5872 }
5873
5874 /**
5875  * calc_csum_metadata_size - return the amount of metadata space that must be
5876  *      reserved/freed for the given bytes.
5877  * @inode: the inode we're manipulating
5878  * @num_bytes: the number of bytes in question
5879  * @reserve: 1 if we are reserving space, 0 if we are freeing space
5880  *
5881  * This adjusts the number of csum_bytes in the inode and then returns the
5882  * correct amount of metadata that must either be reserved or freed.  We
5883  * calculate how many checksums we can fit into one leaf and then divide the
5884  * number of bytes that will need to be checksumed by this value to figure out
5885  * how many checksums will be required.  If we are adding bytes then the number
5886  * may go up and we will return the number of additional bytes that must be
5887  * reserved.  If it is going down we will return the number of bytes that must
5888  * be freed.
5889  *
5890  * This must be called with BTRFS_I(inode)->lock held.
5891  */
5892 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
5893                                    int reserve)
5894 {
5895         struct btrfs_root *root = BTRFS_I(inode)->root;
5896         u64 old_csums, num_csums;
5897
5898         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
5899             BTRFS_I(inode)->csum_bytes == 0)
5900                 return 0;
5901
5902         old_csums = btrfs_csum_bytes_to_leaves(root, BTRFS_I(inode)->csum_bytes);
5903         if (reserve)
5904                 BTRFS_I(inode)->csum_bytes += num_bytes;
5905         else
5906                 BTRFS_I(inode)->csum_bytes -= num_bytes;
5907         num_csums = btrfs_csum_bytes_to_leaves(root, BTRFS_I(inode)->csum_bytes);
5908
5909         /* No change, no need to reserve more */
5910         if (old_csums == num_csums)
5911                 return 0;
5912
5913         if (reserve)
5914                 return btrfs_calc_trans_metadata_size(root,
5915                                                       num_csums - old_csums);
5916
5917         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
5918 }
5919
5920 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
5921 {
5922         struct btrfs_root *root = BTRFS_I(inode)->root;
5923         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
5924         u64 to_reserve = 0;
5925         u64 csum_bytes;
5926         unsigned nr_extents = 0;
5927         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
5928         int ret = 0;
5929         bool delalloc_lock = true;
5930         u64 to_free = 0;
5931         unsigned dropped;
5932         bool release_extra = false;
5933
5934         /* If we are a free space inode we need to not flush since we will be in
5935          * the middle of a transaction commit.  We also don't need the delalloc
5936          * mutex since we won't race with anybody.  We need this mostly to make
5937          * lockdep shut its filthy mouth.
5938          *
5939          * If we have a transaction open (can happen if we call truncate_block
5940          * from truncate), then we need FLUSH_LIMIT so we don't deadlock.
5941          */
5942         if (btrfs_is_free_space_inode(inode)) {
5943                 flush = BTRFS_RESERVE_NO_FLUSH;
5944                 delalloc_lock = false;
5945         } else if (current->journal_info) {
5946                 flush = BTRFS_RESERVE_FLUSH_LIMIT;
5947         }
5948
5949         if (flush != BTRFS_RESERVE_NO_FLUSH &&
5950             btrfs_transaction_in_commit(root->fs_info))
5951                 schedule_timeout(1);
5952
5953         if (delalloc_lock)
5954                 mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
5955
5956         num_bytes = ALIGN(num_bytes, root->sectorsize);
5957
5958         spin_lock(&BTRFS_I(inode)->lock);
5959         nr_extents = (unsigned)div64_u64(num_bytes +
5960                                          BTRFS_MAX_EXTENT_SIZE - 1,
5961                                          BTRFS_MAX_EXTENT_SIZE);
5962         BTRFS_I(inode)->outstanding_extents += nr_extents;
5963
5964         nr_extents = 0;
5965         if (BTRFS_I(inode)->outstanding_extents >
5966             BTRFS_I(inode)->reserved_extents)
5967                 nr_extents += BTRFS_I(inode)->outstanding_extents -
5968                         BTRFS_I(inode)->reserved_extents;
5969
5970         /* We always want to reserve a slot for updating the inode. */
5971         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents + 1);
5972         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
5973         csum_bytes = BTRFS_I(inode)->csum_bytes;
5974         spin_unlock(&BTRFS_I(inode)->lock);
5975
5976         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &root->fs_info->flags)) {
5977                 ret = btrfs_qgroup_reserve_meta(root,
5978                                 nr_extents * root->nodesize);
5979                 if (ret)
5980                         goto out_fail;
5981         }
5982
5983         ret = btrfs_block_rsv_add(root, block_rsv, to_reserve, flush);
5984         if (unlikely(ret)) {
5985                 btrfs_qgroup_free_meta(root, nr_extents * root->nodesize);
5986                 goto out_fail;
5987         }
5988
5989         spin_lock(&BTRFS_I(inode)->lock);
5990         if (test_and_set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5991                              &BTRFS_I(inode)->runtime_flags)) {
5992                 to_reserve -= btrfs_calc_trans_metadata_size(root, 1);
5993                 release_extra = true;
5994         }
5995         BTRFS_I(inode)->reserved_extents += nr_extents;
5996         spin_unlock(&BTRFS_I(inode)->lock);
5997
5998         if (delalloc_lock)
5999                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
6000
6001         if (to_reserve)
6002                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
6003                                               btrfs_ino(inode), to_reserve, 1);
6004         if (release_extra)
6005                 btrfs_block_rsv_release(root, block_rsv,
6006                                         btrfs_calc_trans_metadata_size(root,
6007                                                                        1));
6008         return 0;
6009
6010 out_fail:
6011         spin_lock(&BTRFS_I(inode)->lock);
6012         dropped = drop_outstanding_extent(inode, num_bytes);
6013         /*
6014          * If the inodes csum_bytes is the same as the original
6015          * csum_bytes then we know we haven't raced with any free()ers
6016          * so we can just reduce our inodes csum bytes and carry on.
6017          */
6018         if (BTRFS_I(inode)->csum_bytes == csum_bytes) {
6019                 calc_csum_metadata_size(inode, num_bytes, 0);
6020         } else {
6021                 u64 orig_csum_bytes = BTRFS_I(inode)->csum_bytes;
6022                 u64 bytes;
6023
6024                 /*
6025                  * This is tricky, but first we need to figure out how much we
6026                  * freed from any free-ers that occurred during this
6027                  * reservation, so we reset ->csum_bytes to the csum_bytes
6028                  * before we dropped our lock, and then call the free for the
6029                  * number of bytes that were freed while we were trying our
6030                  * reservation.
6031                  */
6032                 bytes = csum_bytes - BTRFS_I(inode)->csum_bytes;
6033                 BTRFS_I(inode)->csum_bytes = csum_bytes;
6034                 to_free = calc_csum_metadata_size(inode, bytes, 0);
6035
6036
6037                 /*
6038                  * Now we need to see how much we would have freed had we not
6039                  * been making this reservation and our ->csum_bytes were not
6040                  * artificially inflated.
6041                  */
6042                 BTRFS_I(inode)->csum_bytes = csum_bytes - num_bytes;
6043                 bytes = csum_bytes - orig_csum_bytes;
6044                 bytes = calc_csum_metadata_size(inode, bytes, 0);
6045
6046                 /*
6047                  * Now reset ->csum_bytes to what it should be.  If bytes is
6048                  * more than to_free then we would have freed more space had we
6049                  * not had an artificially high ->csum_bytes, so we need to free
6050                  * the remainder.  If bytes is the same or less then we don't
6051                  * need to do anything, the other free-ers did the correct
6052                  * thing.
6053                  */
6054                 BTRFS_I(inode)->csum_bytes = orig_csum_bytes - num_bytes;
6055                 if (bytes > to_free)
6056                         to_free = bytes - to_free;
6057                 else
6058                         to_free = 0;
6059         }
6060         spin_unlock(&BTRFS_I(inode)->lock);
6061         if (dropped)
6062                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
6063
6064         if (to_free) {
6065                 btrfs_block_rsv_release(root, block_rsv, to_free);
6066                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
6067                                               btrfs_ino(inode), to_free, 0);
6068         }
6069         if (delalloc_lock)
6070                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
6071         return ret;
6072 }
6073
6074 /**
6075  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
6076  * @inode: the inode to release the reservation for
6077  * @num_bytes: the number of bytes we're releasing
6078  *
6079  * This will release the metadata reservation for an inode.  This can be called
6080  * once we complete IO for a given set of bytes to release their metadata
6081  * reservations.
6082  */
6083 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
6084 {
6085         struct btrfs_root *root = BTRFS_I(inode)->root;
6086         u64 to_free = 0;
6087         unsigned dropped;
6088
6089         num_bytes = ALIGN(num_bytes, root->sectorsize);
6090         spin_lock(&BTRFS_I(inode)->lock);
6091         dropped = drop_outstanding_extent(inode, num_bytes);
6092
6093         if (num_bytes)
6094                 to_free = calc_csum_metadata_size(inode, num_bytes, 0);
6095         spin_unlock(&BTRFS_I(inode)->lock);
6096         if (dropped > 0)
6097                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
6098
6099         if (btrfs_is_testing(root->fs_info))
6100                 return;
6101
6102         trace_btrfs_space_reservation(root->fs_info, "delalloc",
6103                                       btrfs_ino(inode), to_free, 0);
6104
6105         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
6106                                 to_free);
6107 }
6108
6109 /**
6110  * btrfs_delalloc_reserve_space - reserve data and metadata space for
6111  * delalloc
6112  * @inode: inode we're writing to
6113  * @start: start range we are writing to
6114  * @len: how long the range we are writing to
6115  *
6116  * TODO: This function will finally replace old btrfs_delalloc_reserve_space()
6117  *
6118  * This will do the following things
6119  *
6120  * o reserve space in data space info for num bytes
6121  *   and reserve precious corresponding qgroup space
6122  *   (Done in check_data_free_space)
6123  *
6124  * o reserve space for metadata space, based on the number of outstanding
6125  *   extents and how much csums will be needed
6126  *   also reserve metadata space in a per root over-reserve method.
6127  * o add to the inodes->delalloc_bytes
6128  * o add it to the fs_info's delalloc inodes list.
6129  *   (Above 3 all done in delalloc_reserve_metadata)
6130  *
6131  * Return 0 for success
6132  * Return <0 for error(-ENOSPC or -EQUOT)
6133  */
6134 int btrfs_delalloc_reserve_space(struct inode *inode, u64 start, u64 len)
6135 {
6136         int ret;
6137
6138         ret = btrfs_check_data_free_space(inode, start, len);
6139         if (ret < 0)
6140                 return ret;
6141         ret = btrfs_delalloc_reserve_metadata(inode, len);
6142         if (ret < 0)
6143                 btrfs_free_reserved_data_space(inode, start, len);
6144         return ret;
6145 }
6146
6147 /**
6148  * btrfs_delalloc_release_space - release data and metadata space for delalloc
6149  * @inode: inode we're releasing space for
6150  * @start: start position of the space already reserved
6151  * @len: the len of the space already reserved
6152  *
6153  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
6154  * called in the case that we don't need the metadata AND data reservations
6155  * anymore.  So if there is an error or we insert an inline extent.
6156  *
6157  * This function will release the metadata space that was not used and will
6158  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
6159  * list if there are no delalloc bytes left.
6160  * Also it will handle the qgroup reserved space.
6161  */
6162 void btrfs_delalloc_release_space(struct inode *inode, u64 start, u64 len)
6163 {
6164         btrfs_delalloc_release_metadata(inode, len);
6165         btrfs_free_reserved_data_space(inode, start, len);
6166 }
6167
6168 static int update_block_group(struct btrfs_trans_handle *trans,
6169                               struct btrfs_root *root, u64 bytenr,
6170                               u64 num_bytes, int alloc)
6171 {
6172         struct btrfs_block_group_cache *cache = NULL;
6173         struct btrfs_fs_info *info = root->fs_info;
6174         u64 total = num_bytes;
6175         u64 old_val;
6176         u64 byte_in_group;
6177         int factor;
6178
6179         /* block accounting for super block */
6180         spin_lock(&info->delalloc_root_lock);
6181         old_val = btrfs_super_bytes_used(info->super_copy);
6182         if (alloc)
6183                 old_val += num_bytes;
6184         else
6185                 old_val -= num_bytes;
6186         btrfs_set_super_bytes_used(info->super_copy, old_val);
6187         spin_unlock(&info->delalloc_root_lock);
6188
6189         while (total) {
6190                 cache = btrfs_lookup_block_group(info, bytenr);
6191                 if (!cache)
6192                         return -ENOENT;
6193                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
6194                                     BTRFS_BLOCK_GROUP_RAID1 |
6195                                     BTRFS_BLOCK_GROUP_RAID10))
6196                         factor = 2;
6197                 else
6198                         factor = 1;
6199                 /*
6200                  * If this block group has free space cache written out, we
6201                  * need to make sure to load it if we are removing space.  This
6202                  * is because we need the unpinning stage to actually add the
6203                  * space back to the block group, otherwise we will leak space.
6204                  */
6205                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
6206                         cache_block_group(cache, 1);
6207
6208                 byte_in_group = bytenr - cache->key.objectid;
6209                 WARN_ON(byte_in_group > cache->key.offset);
6210
6211                 spin_lock(&cache->space_info->lock);
6212                 spin_lock(&cache->lock);
6213
6214                 if (btrfs_test_opt(root->fs_info, SPACE_CACHE) &&
6215                     cache->disk_cache_state < BTRFS_DC_CLEAR)
6216                         cache->disk_cache_state = BTRFS_DC_CLEAR;
6217
6218                 old_val = btrfs_block_group_used(&cache->item);
6219                 num_bytes = min(total, cache->key.offset - byte_in_group);
6220                 if (alloc) {
6221                         old_val += num_bytes;
6222                         btrfs_set_block_group_used(&cache->item, old_val);
6223                         cache->reserved -= num_bytes;
6224                         cache->space_info->bytes_reserved -= num_bytes;
6225                         cache->space_info->bytes_used += num_bytes;
6226                         cache->space_info->disk_used += num_bytes * factor;
6227                         spin_unlock(&cache->lock);
6228                         spin_unlock(&cache->space_info->lock);
6229                 } else {
6230                         old_val -= num_bytes;
6231                         btrfs_set_block_group_used(&cache->item, old_val);
6232                         cache->pinned += num_bytes;
6233                         cache->space_info->bytes_pinned += num_bytes;
6234                         cache->space_info->bytes_used -= num_bytes;
6235                         cache->space_info->disk_used -= num_bytes * factor;
6236                         spin_unlock(&cache->lock);
6237                         spin_unlock(&cache->space_info->lock);
6238
6239                         trace_btrfs_space_reservation(root->fs_info, "pinned",
6240                                                       cache->space_info->flags,
6241                                                       num_bytes, 1);
6242                         set_extent_dirty(info->pinned_extents,
6243                                          bytenr, bytenr + num_bytes - 1,
6244                                          GFP_NOFS | __GFP_NOFAIL);
6245                 }
6246
6247                 spin_lock(&trans->transaction->dirty_bgs_lock);
6248                 if (list_empty(&cache->dirty_list)) {
6249                         list_add_tail(&cache->dirty_list,
6250                                       &trans->transaction->dirty_bgs);
6251                                 trans->transaction->num_dirty_bgs++;
6252                         btrfs_get_block_group(cache);
6253                 }
6254                 spin_unlock(&trans->transaction->dirty_bgs_lock);
6255
6256                 /*
6257                  * No longer have used bytes in this block group, queue it for
6258                  * deletion. We do this after adding the block group to the
6259                  * dirty list to avoid races between cleaner kthread and space
6260                  * cache writeout.
6261                  */
6262                 if (!alloc && old_val == 0) {
6263                         spin_lock(&info->unused_bgs_lock);
6264                         if (list_empty(&cache->bg_list)) {
6265                                 btrfs_get_block_group(cache);
6266                                 list_add_tail(&cache->bg_list,
6267                                               &info->unused_bgs);
6268                         }
6269                         spin_unlock(&info->unused_bgs_lock);
6270                 }
6271
6272                 btrfs_put_block_group(cache);
6273                 total -= num_bytes;
6274                 bytenr += num_bytes;
6275         }
6276         return 0;
6277 }
6278
6279 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
6280 {
6281         struct btrfs_block_group_cache *cache;
6282         u64 bytenr;
6283
6284         spin_lock(&root->fs_info->block_group_cache_lock);
6285         bytenr = root->fs_info->first_logical_byte;
6286         spin_unlock(&root->fs_info->block_group_cache_lock);
6287
6288         if (bytenr < (u64)-1)
6289                 return bytenr;
6290
6291         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
6292         if (!cache)
6293                 return 0;
6294
6295         bytenr = cache->key.objectid;
6296         btrfs_put_block_group(cache);
6297
6298         return bytenr;
6299 }
6300
6301 static int pin_down_extent(struct btrfs_root *root,
6302                            struct btrfs_block_group_cache *cache,
6303                            u64 bytenr, u64 num_bytes, int reserved)
6304 {
6305         spin_lock(&cache->space_info->lock);
6306         spin_lock(&cache->lock);
6307         cache->pinned += num_bytes;
6308         cache->space_info->bytes_pinned += num_bytes;
6309         if (reserved) {
6310                 cache->reserved -= num_bytes;
6311                 cache->space_info->bytes_reserved -= num_bytes;
6312         }
6313         spin_unlock(&cache->lock);
6314         spin_unlock(&cache->space_info->lock);
6315
6316         trace_btrfs_space_reservation(root->fs_info, "pinned",
6317                                       cache->space_info->flags, num_bytes, 1);
6318         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
6319                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
6320         return 0;
6321 }
6322
6323 /*
6324  * this function must be called within transaction
6325  */
6326 int btrfs_pin_extent(struct btrfs_root *root,
6327                      u64 bytenr, u64 num_bytes, int reserved)
6328 {
6329         struct btrfs_block_group_cache *cache;
6330
6331         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
6332         BUG_ON(!cache); /* Logic error */
6333
6334         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
6335
6336         btrfs_put_block_group(cache);
6337         return 0;
6338 }
6339
6340 /*
6341  * this function must be called within transaction
6342  */
6343 int btrfs_pin_extent_for_log_replay(struct btrfs_root *root,
6344                                     u64 bytenr, u64 num_bytes)
6345 {
6346         struct btrfs_block_group_cache *cache;
6347         int ret;
6348
6349         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
6350         if (!cache)
6351                 return -EINVAL;
6352
6353         /*
6354          * pull in the free space cache (if any) so that our pin
6355          * removes the free space from the cache.  We have load_only set
6356          * to one because the slow code to read in the free extents does check
6357          * the pinned extents.
6358          */
6359         cache_block_group(cache, 1);
6360
6361         pin_down_extent(root, cache, bytenr, num_bytes, 0);
6362
6363         /* remove us from the free space cache (if we're there at all) */
6364         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
6365         btrfs_put_block_group(cache);
6366         return ret;
6367 }
6368
6369 static int __exclude_logged_extent(struct btrfs_root *root, u64 start, u64 num_bytes)
6370 {
6371         int ret;
6372         struct btrfs_block_group_cache *block_group;
6373         struct btrfs_caching_control *caching_ctl;
6374
6375         block_group = btrfs_lookup_block_group(root->fs_info, start);
6376         if (!block_group)
6377                 return -EINVAL;
6378
6379         cache_block_group(block_group, 0);
6380         caching_ctl = get_caching_control(block_group);
6381
6382         if (!caching_ctl) {
6383                 /* Logic error */
6384                 BUG_ON(!block_group_cache_done(block_group));
6385                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6386         } else {
6387                 mutex_lock(&caching_ctl->mutex);
6388
6389                 if (start >= caching_ctl->progress) {
6390                         ret = add_excluded_extent(root, start, num_bytes);
6391                 } else if (start + num_bytes <= caching_ctl->progress) {
6392                         ret = btrfs_remove_free_space(block_group,
6393                                                       start, num_bytes);
6394                 } else {
6395                         num_bytes = caching_ctl->progress - start;
6396                         ret = btrfs_remove_free_space(block_group,
6397                                                       start, num_bytes);
6398                         if (ret)
6399                                 goto out_lock;
6400
6401                         num_bytes = (start + num_bytes) -
6402                                 caching_ctl->progress;
6403                         start = caching_ctl->progress;
6404                         ret = add_excluded_extent(root, start, num_bytes);
6405                 }
6406 out_lock:
6407                 mutex_unlock(&caching_ctl->mutex);
6408                 put_caching_control(caching_ctl);
6409         }
6410         btrfs_put_block_group(block_group);
6411         return ret;
6412 }
6413
6414 int btrfs_exclude_logged_extents(struct btrfs_root *log,
6415                                  struct extent_buffer *eb)
6416 {
6417         struct btrfs_file_extent_item *item;
6418         struct btrfs_key key;
6419         int found_type;
6420         int i;
6421
6422         if (!btrfs_fs_incompat(log->fs_info, MIXED_GROUPS))
6423                 return 0;
6424
6425         for (i = 0; i < btrfs_header_nritems(eb); i++) {
6426                 btrfs_item_key_to_cpu(eb, &key, i);
6427                 if (key.type != BTRFS_EXTENT_DATA_KEY)
6428                         continue;
6429                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
6430                 found_type = btrfs_file_extent_type(eb, item);
6431                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
6432                         continue;
6433                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
6434                         continue;
6435                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
6436                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
6437                 __exclude_logged_extent(log, key.objectid, key.offset);
6438         }
6439
6440         return 0;
6441 }
6442
6443 static void
6444 btrfs_inc_block_group_reservations(struct btrfs_block_group_cache *bg)
6445 {
6446         atomic_inc(&bg->reservations);
6447 }
6448
6449 void btrfs_dec_block_group_reservations(struct btrfs_fs_info *fs_info,
6450                                         const u64 start)
6451 {
6452         struct btrfs_block_group_cache *bg;
6453
6454         bg = btrfs_lookup_block_group(fs_info, start);
6455         ASSERT(bg);
6456         if (atomic_dec_and_test(&bg->reservations))
6457                 wake_up_atomic_t(&bg->reservations);
6458         btrfs_put_block_group(bg);
6459 }
6460
6461 static int btrfs_wait_bg_reservations_atomic_t(atomic_t *a)
6462 {
6463         schedule();
6464         return 0;
6465 }
6466
6467 void btrfs_wait_block_group_reservations(struct btrfs_block_group_cache *bg)
6468 {
6469         struct btrfs_space_info *space_info = bg->space_info;
6470
6471         ASSERT(bg->ro);
6472
6473         if (!(bg->flags & BTRFS_BLOCK_GROUP_DATA))
6474                 return;
6475
6476         /*
6477          * Our block group is read only but before we set it to read only,
6478          * some task might have had allocated an extent from it already, but it
6479          * has not yet created a respective ordered extent (and added it to a
6480          * root's list of ordered extents).
6481          * Therefore wait for any task currently allocating extents, since the
6482          * block group's reservations counter is incremented while a read lock
6483          * on the groups' semaphore is held and decremented after releasing
6484          * the read access on that semaphore and creating the ordered extent.
6485          */
6486         down_write(&space_info->groups_sem);
6487         up_write(&space_info->groups_sem);
6488
6489         wait_on_atomic_t(&bg->reservations,
6490                          btrfs_wait_bg_reservations_atomic_t,
6491                          TASK_UNINTERRUPTIBLE);
6492 }
6493
6494 /**
6495  * btrfs_add_reserved_bytes - update the block_group and space info counters
6496  * @cache:      The cache we are manipulating
6497  * @ram_bytes:  The number of bytes of file content, and will be same to
6498  *              @num_bytes except for the compress path.
6499  * @num_bytes:  The number of bytes in question
6500  * @delalloc:   The blocks are allocated for the delalloc write
6501  *
6502  * This is called by the allocator when it reserves space. Metadata
6503  * reservations should be called with RESERVE_ALLOC so we do the proper
6504  * ENOSPC accounting.  For data we handle the reservation through clearing the
6505  * delalloc bits in the io_tree.  We have to do this since we could end up
6506  * allocating less disk space for the amount of data we have reserved in the
6507  * case of compression.
6508  *
6509  * If this is a reservation and the block group has become read only we cannot
6510  * make the reservation and return -EAGAIN, otherwise this function always
6511  * succeeds.
6512  */
6513 static int btrfs_add_reserved_bytes(struct btrfs_block_group_cache *cache,
6514                                     u64 ram_bytes, u64 num_bytes, int delalloc)
6515 {
6516         struct btrfs_space_info *space_info = cache->space_info;
6517         int ret = 0;
6518
6519         spin_lock(&space_info->lock);
6520         spin_lock(&cache->lock);
6521         if (cache->ro) {
6522                 ret = -EAGAIN;
6523         } else {
6524                 cache->reserved += num_bytes;
6525                 space_info->bytes_reserved += num_bytes;
6526
6527                 trace_btrfs_space_reservation(cache->fs_info,
6528                                 "space_info", space_info->flags,
6529                                 ram_bytes, 0);
6530                 space_info->bytes_may_use -= ram_bytes;
6531                 if (delalloc)
6532                         cache->delalloc_bytes += num_bytes;
6533         }
6534         spin_unlock(&cache->lock);
6535         spin_unlock(&space_info->lock);
6536         return ret;
6537 }
6538
6539 /**
6540  * btrfs_free_reserved_bytes - update the block_group and space info counters
6541  * @cache:      The cache we are manipulating
6542  * @num_bytes:  The number of bytes in question
6543  * @delalloc:   The blocks are allocated for the delalloc write
6544  *
6545  * This is called by somebody who is freeing space that was never actually used
6546  * on disk.  For example if you reserve some space for a new leaf in transaction
6547  * A and before transaction A commits you free that leaf, you call this with
6548  * reserve set to 0 in order to clear the reservation.
6549  */
6550
6551 static int btrfs_free_reserved_bytes(struct btrfs_block_group_cache *cache,
6552                                      u64 num_bytes, int delalloc)
6553 {
6554         struct btrfs_space_info *space_info = cache->space_info;
6555         int ret = 0;
6556
6557         spin_lock(&space_info->lock);
6558         spin_lock(&cache->lock);
6559         if (cache->ro)
6560                 space_info->bytes_readonly += num_bytes;
6561         cache->reserved -= num_bytes;
6562         space_info->bytes_reserved -= num_bytes;
6563
6564         if (delalloc)
6565                 cache->delalloc_bytes -= num_bytes;
6566         spin_unlock(&cache->lock);
6567         spin_unlock(&space_info->lock);
6568         return ret;
6569 }
6570 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
6571                                 struct btrfs_root *root)
6572 {
6573         struct btrfs_fs_info *fs_info = root->fs_info;
6574         struct btrfs_caching_control *next;
6575         struct btrfs_caching_control *caching_ctl;
6576         struct btrfs_block_group_cache *cache;
6577
6578         down_write(&fs_info->commit_root_sem);
6579
6580         list_for_each_entry_safe(caching_ctl, next,
6581                                  &fs_info->caching_block_groups, list) {
6582                 cache = caching_ctl->block_group;
6583                 if (block_group_cache_done(cache)) {
6584                         cache->last_byte_to_unpin = (u64)-1;
6585                         list_del_init(&caching_ctl->list);
6586                         put_caching_control(caching_ctl);
6587                 } else {
6588                         cache->last_byte_to_unpin = caching_ctl->progress;
6589                 }
6590         }
6591
6592         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6593                 fs_info->pinned_extents = &fs_info->freed_extents[1];
6594         else
6595                 fs_info->pinned_extents = &fs_info->freed_extents[0];
6596
6597         up_write(&fs_info->commit_root_sem);
6598
6599         update_global_block_rsv(fs_info);
6600 }
6601
6602 /*
6603  * Returns the free cluster for the given space info and sets empty_cluster to
6604  * what it should be based on the mount options.
6605  */
6606 static struct btrfs_free_cluster *
6607 fetch_cluster_info(struct btrfs_root *root, struct btrfs_space_info *space_info,
6608                    u64 *empty_cluster)
6609 {
6610         struct btrfs_free_cluster *ret = NULL;
6611         bool ssd = btrfs_test_opt(root->fs_info, SSD);
6612
6613         *empty_cluster = 0;
6614         if (btrfs_mixed_space_info(space_info))
6615                 return ret;
6616
6617         if (ssd)
6618                 *empty_cluster = SZ_2M;
6619         if (space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
6620                 ret = &root->fs_info->meta_alloc_cluster;
6621                 if (!ssd)
6622                         *empty_cluster = SZ_64K;
6623         } else if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) && ssd) {
6624                 ret = &root->fs_info->data_alloc_cluster;
6625         }
6626
6627         return ret;
6628 }
6629
6630 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end,
6631                               const bool return_free_space)
6632 {
6633         struct btrfs_fs_info *fs_info = root->fs_info;
6634         struct btrfs_block_group_cache *cache = NULL;
6635         struct btrfs_space_info *space_info;
6636         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
6637         struct btrfs_free_cluster *cluster = NULL;
6638         u64 len;
6639         u64 total_unpinned = 0;
6640         u64 empty_cluster = 0;
6641         bool readonly;
6642
6643         while (start <= end) {
6644                 readonly = false;
6645                 if (!cache ||
6646                     start >= cache->key.objectid + cache->key.offset) {
6647                         if (cache)
6648                                 btrfs_put_block_group(cache);
6649                         total_unpinned = 0;
6650                         cache = btrfs_lookup_block_group(fs_info, start);
6651                         BUG_ON(!cache); /* Logic error */
6652
6653                         cluster = fetch_cluster_info(root,
6654                                                      cache->space_info,
6655                                                      &empty_cluster);
6656                         empty_cluster <<= 1;
6657                 }
6658
6659                 len = cache->key.objectid + cache->key.offset - start;
6660                 len = min(len, end + 1 - start);
6661
6662                 if (start < cache->last_byte_to_unpin) {
6663                         len = min(len, cache->last_byte_to_unpin - start);
6664                         if (return_free_space)
6665                                 btrfs_add_free_space(cache, start, len);
6666                 }
6667
6668                 start += len;
6669                 total_unpinned += len;
6670                 space_info = cache->space_info;
6671
6672                 /*
6673                  * If this space cluster has been marked as fragmented and we've
6674                  * unpinned enough in this block group to potentially allow a
6675                  * cluster to be created inside of it go ahead and clear the
6676                  * fragmented check.
6677                  */
6678                 if (cluster && cluster->fragmented &&
6679                     total_unpinned > empty_cluster) {
6680                         spin_lock(&cluster->lock);
6681                         cluster->fragmented = 0;
6682                         spin_unlock(&cluster->lock);
6683                 }
6684
6685                 spin_lock(&space_info->lock);
6686                 spin_lock(&cache->lock);
6687                 cache->pinned -= len;
6688                 space_info->bytes_pinned -= len;
6689
6690                 trace_btrfs_space_reservation(fs_info, "pinned",
6691                                               space_info->flags, len, 0);
6692                 space_info->max_extent_size = 0;
6693                 percpu_counter_add(&space_info->total_bytes_pinned, -len);
6694                 if (cache->ro) {
6695                         space_info->bytes_readonly += len;
6696                         readonly = true;
6697                 }
6698                 spin_unlock(&cache->lock);
6699                 if (!readonly && return_free_space &&
6700                     global_rsv->space_info == space_info) {
6701                         u64 to_add = len;
6702                         WARN_ON(!return_free_space);
6703                         spin_lock(&global_rsv->lock);
6704                         if (!global_rsv->full) {
6705                                 to_add = min(len, global_rsv->size -
6706                                              global_rsv->reserved);
6707                                 global_rsv->reserved += to_add;
6708                                 space_info->bytes_may_use += to_add;
6709                                 if (global_rsv->reserved >= global_rsv->size)
6710                                         global_rsv->full = 1;
6711                                 trace_btrfs_space_reservation(fs_info,
6712                                                               "space_info",
6713                                                               space_info->flags,
6714                                                               to_add, 1);
6715                                 len -= to_add;
6716                         }
6717                         spin_unlock(&global_rsv->lock);
6718                         /* Add to any tickets we may have */
6719                         if (len)
6720                                 space_info_add_new_bytes(fs_info, space_info,
6721                                                          len);
6722                 }
6723                 spin_unlock(&space_info->lock);
6724         }
6725
6726         if (cache)
6727                 btrfs_put_block_group(cache);
6728         return 0;
6729 }
6730
6731 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
6732                                struct btrfs_root *root)
6733 {
6734         struct btrfs_fs_info *fs_info = root->fs_info;
6735         struct btrfs_block_group_cache *block_group, *tmp;
6736         struct list_head *deleted_bgs;
6737         struct extent_io_tree *unpin;
6738         u64 start;
6739         u64 end;
6740         int ret;
6741
6742         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6743                 unpin = &fs_info->freed_extents[1];
6744         else
6745                 unpin = &fs_info->freed_extents[0];
6746
6747         while (!trans->aborted) {
6748                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
6749                 ret = find_first_extent_bit(unpin, 0, &start, &end,
6750                                             EXTENT_DIRTY, NULL);
6751                 if (ret) {
6752                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6753                         break;
6754                 }
6755
6756                 if (btrfs_test_opt(root->fs_info, DISCARD))
6757                         ret = btrfs_discard_extent(root, start,
6758                                                    end + 1 - start, NULL);
6759
6760                 clear_extent_dirty(unpin, start, end);
6761                 unpin_extent_range(root, start, end, true);
6762                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6763                 cond_resched();
6764         }
6765
6766         /*
6767          * Transaction is finished.  We don't need the lock anymore.  We
6768          * do need to clean up the block groups in case of a transaction
6769          * abort.
6770          */
6771         deleted_bgs = &trans->transaction->deleted_bgs;
6772         list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
6773                 u64 trimmed = 0;
6774
6775                 ret = -EROFS;
6776                 if (!trans->aborted)
6777                         ret = btrfs_discard_extent(root,
6778                                                    block_group->key.objectid,
6779                                                    block_group->key.offset,
6780                                                    &trimmed);
6781
6782                 list_del_init(&block_group->bg_list);
6783                 btrfs_put_block_group_trimming(block_group);
6784                 btrfs_put_block_group(block_group);
6785
6786                 if (ret) {
6787                         const char *errstr = btrfs_decode_error(ret);
6788                         btrfs_warn(fs_info,
6789                                    "Discard failed while removing blockgroup: errno=%d %s\n",
6790                                    ret, errstr);
6791                 }
6792         }
6793
6794         return 0;
6795 }
6796
6797 static void add_pinned_bytes(struct btrfs_fs_info *fs_info, u64 num_bytes,
6798                              u64 owner, u64 root_objectid)
6799 {
6800         struct btrfs_space_info *space_info;
6801         u64 flags;
6802
6803         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
6804                 if (root_objectid == BTRFS_CHUNK_TREE_OBJECTID)
6805                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
6806                 else
6807                         flags = BTRFS_BLOCK_GROUP_METADATA;
6808         } else {
6809                 flags = BTRFS_BLOCK_GROUP_DATA;
6810         }
6811
6812         space_info = __find_space_info(fs_info, flags);
6813         BUG_ON(!space_info); /* Logic bug */
6814         percpu_counter_add(&space_info->total_bytes_pinned, num_bytes);
6815 }
6816
6817
6818 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
6819                                 struct btrfs_root *root,
6820                                 struct btrfs_delayed_ref_node *node, u64 parent,
6821                                 u64 root_objectid, u64 owner_objectid,
6822                                 u64 owner_offset, int refs_to_drop,
6823                                 struct btrfs_delayed_extent_op *extent_op)
6824 {
6825         struct btrfs_key key;
6826         struct btrfs_path *path;
6827         struct btrfs_fs_info *info = root->fs_info;
6828         struct btrfs_root *extent_root = info->extent_root;
6829         struct extent_buffer *leaf;
6830         struct btrfs_extent_item *ei;
6831         struct btrfs_extent_inline_ref *iref;
6832         int ret;
6833         int is_data;
6834         int extent_slot = 0;
6835         int found_extent = 0;
6836         int num_to_del = 1;
6837         u32 item_size;
6838         u64 refs;
6839         u64 bytenr = node->bytenr;
6840         u64 num_bytes = node->num_bytes;
6841         int last_ref = 0;
6842         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
6843                                                  SKINNY_METADATA);
6844
6845         path = btrfs_alloc_path();
6846         if (!path)
6847                 return -ENOMEM;
6848
6849         path->reada = READA_FORWARD;
6850         path->leave_spinning = 1;
6851
6852         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
6853         BUG_ON(!is_data && refs_to_drop != 1);
6854
6855         if (is_data)
6856                 skinny_metadata = 0;
6857
6858         ret = lookup_extent_backref(trans, extent_root, path, &iref,
6859                                     bytenr, num_bytes, parent,
6860                                     root_objectid, owner_objectid,
6861                                     owner_offset);
6862         if (ret == 0) {
6863                 extent_slot = path->slots[0];
6864                 while (extent_slot >= 0) {
6865                         btrfs_item_key_to_cpu(path->nodes[0], &key,
6866                                               extent_slot);
6867                         if (key.objectid != bytenr)
6868                                 break;
6869                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
6870                             key.offset == num_bytes) {
6871                                 found_extent = 1;
6872                                 break;
6873                         }
6874                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
6875                             key.offset == owner_objectid) {
6876                                 found_extent = 1;
6877                                 break;
6878                         }
6879                         if (path->slots[0] - extent_slot > 5)
6880                                 break;
6881                         extent_slot--;
6882                 }
6883 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6884                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
6885                 if (found_extent && item_size < sizeof(*ei))
6886                         found_extent = 0;
6887 #endif
6888                 if (!found_extent) {
6889                         BUG_ON(iref);
6890                         ret = remove_extent_backref(trans, extent_root, path,
6891                                                     NULL, refs_to_drop,
6892                                                     is_data, &last_ref);
6893                         if (ret) {
6894                                 btrfs_abort_transaction(trans, ret);
6895                                 goto out;
6896                         }
6897                         btrfs_release_path(path);
6898                         path->leave_spinning = 1;
6899
6900                         key.objectid = bytenr;
6901                         key.type = BTRFS_EXTENT_ITEM_KEY;
6902                         key.offset = num_bytes;
6903
6904                         if (!is_data && skinny_metadata) {
6905                                 key.type = BTRFS_METADATA_ITEM_KEY;
6906                                 key.offset = owner_objectid;
6907                         }
6908
6909                         ret = btrfs_search_slot(trans, extent_root,
6910                                                 &key, path, -1, 1);
6911                         if (ret > 0 && skinny_metadata && path->slots[0]) {
6912                                 /*
6913                                  * Couldn't find our skinny metadata item,
6914                                  * see if we have ye olde extent item.
6915                                  */
6916                                 path->slots[0]--;
6917                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
6918                                                       path->slots[0]);
6919                                 if (key.objectid == bytenr &&
6920                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
6921                                     key.offset == num_bytes)
6922                                         ret = 0;
6923                         }
6924
6925                         if (ret > 0 && skinny_metadata) {
6926                                 skinny_metadata = false;
6927                                 key.objectid = bytenr;
6928                                 key.type = BTRFS_EXTENT_ITEM_KEY;
6929                                 key.offset = num_bytes;
6930                                 btrfs_release_path(path);
6931                                 ret = btrfs_search_slot(trans, extent_root,
6932                                                         &key, path, -1, 1);
6933                         }
6934
6935                         if (ret) {
6936                                 btrfs_err(info,
6937                                           "umm, got %d back from search, was looking for %llu",
6938                                           ret, bytenr);
6939                                 if (ret > 0)
6940                                         btrfs_print_leaf(extent_root,
6941                                                          path->nodes[0]);
6942                         }
6943                         if (ret < 0) {
6944                                 btrfs_abort_transaction(trans, ret);
6945                                 goto out;
6946                         }
6947                         extent_slot = path->slots[0];
6948                 }
6949         } else if (WARN_ON(ret == -ENOENT)) {
6950                 btrfs_print_leaf(extent_root, path->nodes[0]);
6951                 btrfs_err(info,
6952                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
6953                         bytenr, parent, root_objectid, owner_objectid,
6954                         owner_offset);
6955                 btrfs_abort_transaction(trans, ret);
6956                 goto out;
6957         } else {
6958                 btrfs_abort_transaction(trans, ret);
6959                 goto out;
6960         }
6961
6962         leaf = path->nodes[0];
6963         item_size = btrfs_item_size_nr(leaf, extent_slot);
6964 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6965         if (item_size < sizeof(*ei)) {
6966                 BUG_ON(found_extent || extent_slot != path->slots[0]);
6967                 ret = convert_extent_item_v0(trans, extent_root, path,
6968                                              owner_objectid, 0);
6969                 if (ret < 0) {
6970                         btrfs_abort_transaction(trans, ret);
6971                         goto out;
6972                 }
6973
6974                 btrfs_release_path(path);
6975                 path->leave_spinning = 1;
6976
6977                 key.objectid = bytenr;
6978                 key.type = BTRFS_EXTENT_ITEM_KEY;
6979                 key.offset = num_bytes;
6980
6981                 ret = btrfs_search_slot(trans, extent_root, &key, path,
6982                                         -1, 1);
6983                 if (ret) {
6984                         btrfs_err(info,
6985                                   "umm, got %d back from search, was looking for %llu",
6986                                 ret, bytenr);
6987                         btrfs_print_leaf(extent_root, path->nodes[0]);
6988                 }
6989                 if (ret < 0) {
6990                         btrfs_abort_transaction(trans, ret);
6991                         goto out;
6992                 }
6993
6994                 extent_slot = path->slots[0];
6995                 leaf = path->nodes[0];
6996                 item_size = btrfs_item_size_nr(leaf, extent_slot);
6997         }
6998 #endif
6999         BUG_ON(item_size < sizeof(*ei));
7000         ei = btrfs_item_ptr(leaf, extent_slot,
7001                             struct btrfs_extent_item);
7002         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
7003             key.type == BTRFS_EXTENT_ITEM_KEY) {
7004                 struct btrfs_tree_block_info *bi;
7005                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
7006                 bi = (struct btrfs_tree_block_info *)(ei + 1);
7007                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
7008         }
7009
7010         refs = btrfs_extent_refs(leaf, ei);
7011         if (refs < refs_to_drop) {
7012                 btrfs_err(info,
7013                           "trying to drop %d refs but we only have %Lu for bytenr %Lu",
7014                           refs_to_drop, refs, bytenr);
7015                 ret = -EINVAL;
7016                 btrfs_abort_transaction(trans, ret);
7017                 goto out;
7018         }
7019         refs -= refs_to_drop;
7020
7021         if (refs > 0) {
7022                 if (extent_op)
7023                         __run_delayed_extent_op(extent_op, leaf, ei);
7024                 /*
7025                  * In the case of inline back ref, reference count will
7026                  * be updated by remove_extent_backref
7027                  */
7028                 if (iref) {
7029                         BUG_ON(!found_extent);
7030                 } else {
7031                         btrfs_set_extent_refs(leaf, ei, refs);
7032                         btrfs_mark_buffer_dirty(leaf);
7033                 }
7034                 if (found_extent) {
7035                         ret = remove_extent_backref(trans, extent_root, path,
7036                                                     iref, refs_to_drop,
7037                                                     is_data, &last_ref);
7038                         if (ret) {
7039                                 btrfs_abort_transaction(trans, ret);
7040                                 goto out;
7041                         }
7042                 }
7043                 add_pinned_bytes(root->fs_info, -num_bytes, owner_objectid,
7044                                  root_objectid);
7045         } else {
7046                 if (found_extent) {
7047                         BUG_ON(is_data && refs_to_drop !=
7048                                extent_data_ref_count(path, iref));
7049                         if (iref) {
7050                                 BUG_ON(path->slots[0] != extent_slot);
7051                         } else {
7052                                 BUG_ON(path->slots[0] != extent_slot + 1);
7053                                 path->slots[0] = extent_slot;
7054                                 num_to_del = 2;
7055                         }
7056                 }
7057
7058                 last_ref = 1;
7059                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
7060                                       num_to_del);
7061                 if (ret) {
7062                         btrfs_abort_transaction(trans, ret);
7063                         goto out;
7064                 }
7065                 btrfs_release_path(path);
7066
7067                 if (is_data) {
7068                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
7069                         if (ret) {
7070                                 btrfs_abort_transaction(trans, ret);
7071                                 goto out;
7072                         }
7073                 }
7074
7075                 ret = add_to_free_space_tree(trans, root->fs_info, bytenr,
7076                                              num_bytes);
7077                 if (ret) {
7078                         btrfs_abort_transaction(trans, ret);
7079                         goto out;
7080                 }
7081
7082                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
7083                 if (ret) {
7084                         btrfs_abort_transaction(trans, ret);
7085                         goto out;
7086                 }
7087         }
7088         btrfs_release_path(path);
7089
7090 out:
7091         btrfs_free_path(path);
7092         return ret;
7093 }
7094
7095 /*
7096  * when we free an block, it is possible (and likely) that we free the last
7097  * delayed ref for that extent as well.  This searches the delayed ref tree for
7098  * a given extent, and if there are no other delayed refs to be processed, it
7099  * removes it from the tree.
7100  */
7101 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
7102                                       struct btrfs_root *root, u64 bytenr)
7103 {
7104         struct btrfs_delayed_ref_head *head;
7105         struct btrfs_delayed_ref_root *delayed_refs;
7106         int ret = 0;
7107
7108         delayed_refs = &trans->transaction->delayed_refs;
7109         spin_lock(&delayed_refs->lock);
7110         head = btrfs_find_delayed_ref_head(trans, bytenr);
7111         if (!head)
7112                 goto out_delayed_unlock;
7113
7114         spin_lock(&head->lock);
7115         if (!list_empty(&head->ref_list))
7116                 goto out;
7117
7118         if (head->extent_op) {
7119                 if (!head->must_insert_reserved)
7120                         goto out;
7121                 btrfs_free_delayed_extent_op(head->extent_op);
7122                 head->extent_op = NULL;
7123         }
7124
7125         /*
7126          * waiting for the lock here would deadlock.  If someone else has it
7127          * locked they are already in the process of dropping it anyway
7128          */
7129         if (!mutex_trylock(&head->mutex))
7130                 goto out;
7131
7132         /*
7133          * at this point we have a head with no other entries.  Go
7134          * ahead and process it.
7135          */
7136         head->node.in_tree = 0;
7137         rb_erase(&head->href_node, &delayed_refs->href_root);
7138
7139         atomic_dec(&delayed_refs->num_entries);
7140
7141         /*
7142          * we don't take a ref on the node because we're removing it from the
7143          * tree, so we just steal the ref the tree was holding.
7144          */
7145         delayed_refs->num_heads--;
7146         if (head->processing == 0)
7147                 delayed_refs->num_heads_ready--;
7148         head->processing = 0;
7149         spin_unlock(&head->lock);
7150         spin_unlock(&delayed_refs->lock);
7151
7152         BUG_ON(head->extent_op);
7153         if (head->must_insert_reserved)
7154                 ret = 1;
7155
7156         mutex_unlock(&head->mutex);
7157         btrfs_put_delayed_ref(&head->node);
7158         return ret;
7159 out:
7160         spin_unlock(&head->lock);
7161
7162 out_delayed_unlock:
7163         spin_unlock(&delayed_refs->lock);
7164         return 0;
7165 }
7166
7167 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
7168                            struct btrfs_root *root,
7169                            struct extent_buffer *buf,
7170                            u64 parent, int last_ref)
7171 {
7172         int pin = 1;
7173         int ret;
7174
7175         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7176                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
7177                                         buf->start, buf->len,
7178                                         parent, root->root_key.objectid,
7179                                         btrfs_header_level(buf),
7180                                         BTRFS_DROP_DELAYED_REF, NULL);
7181                 BUG_ON(ret); /* -ENOMEM */
7182         }
7183
7184         if (!last_ref)
7185                 return;
7186
7187         if (btrfs_header_generation(buf) == trans->transid) {
7188                 struct btrfs_block_group_cache *cache;
7189
7190                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7191                         ret = check_ref_cleanup(trans, root, buf->start);
7192                         if (!ret)
7193                                 goto out;
7194                 }
7195
7196                 cache = btrfs_lookup_block_group(root->fs_info, buf->start);
7197
7198                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
7199                         pin_down_extent(root, cache, buf->start, buf->len, 1);
7200                         btrfs_put_block_group(cache);
7201                         goto out;
7202                 }
7203
7204                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
7205
7206                 btrfs_add_free_space(cache, buf->start, buf->len);
7207                 btrfs_free_reserved_bytes(cache, buf->len, 0);
7208                 btrfs_put_block_group(cache);
7209                 trace_btrfs_reserved_extent_free(root, buf->start, buf->len);
7210                 pin = 0;
7211         }
7212 out:
7213         if (pin)
7214                 add_pinned_bytes(root->fs_info, buf->len,
7215                                  btrfs_header_level(buf),
7216                                  root->root_key.objectid);
7217
7218         /*
7219          * Deleting the buffer, clear the corrupt flag since it doesn't matter
7220          * anymore.
7221          */
7222         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
7223 }
7224
7225 /* Can return -ENOMEM */
7226 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
7227                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
7228                       u64 owner, u64 offset)
7229 {
7230         int ret;
7231         struct btrfs_fs_info *fs_info = root->fs_info;
7232
7233         if (btrfs_is_testing(fs_info))
7234                 return 0;
7235
7236         add_pinned_bytes(root->fs_info, num_bytes, owner, root_objectid);
7237
7238         /*
7239          * tree log blocks never actually go into the extent allocation
7240          * tree, just update pinning info and exit early.
7241          */
7242         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
7243                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
7244                 /* unlocks the pinned mutex */
7245                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
7246                 ret = 0;
7247         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
7248                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
7249                                         num_bytes,
7250                                         parent, root_objectid, (int)owner,
7251                                         BTRFS_DROP_DELAYED_REF, NULL);
7252         } else {
7253                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
7254                                                 num_bytes,
7255                                                 parent, root_objectid, owner,
7256                                                 offset, 0,
7257                                                 BTRFS_DROP_DELAYED_REF, NULL);
7258         }
7259         return ret;
7260 }
7261
7262 /*
7263  * when we wait for progress in the block group caching, its because
7264  * our allocation attempt failed at least once.  So, we must sleep
7265  * and let some progress happen before we try again.
7266  *
7267  * This function will sleep at least once waiting for new free space to
7268  * show up, and then it will check the block group free space numbers
7269  * for our min num_bytes.  Another option is to have it go ahead
7270  * and look in the rbtree for a free extent of a given size, but this
7271  * is a good start.
7272  *
7273  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
7274  * any of the information in this block group.
7275  */
7276 static noinline void
7277 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
7278                                 u64 num_bytes)
7279 {
7280         struct btrfs_caching_control *caching_ctl;
7281
7282         caching_ctl = get_caching_control(cache);
7283         if (!caching_ctl)
7284                 return;
7285
7286         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
7287                    (cache->free_space_ctl->free_space >= num_bytes));
7288
7289         put_caching_control(caching_ctl);
7290 }
7291
7292 static noinline int
7293 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
7294 {
7295         struct btrfs_caching_control *caching_ctl;
7296         int ret = 0;
7297
7298         caching_ctl = get_caching_control(cache);
7299         if (!caching_ctl)
7300                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
7301
7302         wait_event(caching_ctl->wait, block_group_cache_done(cache));
7303         if (cache->cached == BTRFS_CACHE_ERROR)
7304                 ret = -EIO;
7305         put_caching_control(caching_ctl);
7306         return ret;
7307 }
7308
7309 int __get_raid_index(u64 flags)
7310 {
7311         if (flags & BTRFS_BLOCK_GROUP_RAID10)
7312                 return BTRFS_RAID_RAID10;
7313         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
7314                 return BTRFS_RAID_RAID1;
7315         else if (flags & BTRFS_BLOCK_GROUP_DUP)
7316                 return BTRFS_RAID_DUP;
7317         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
7318                 return BTRFS_RAID_RAID0;
7319         else if (flags & BTRFS_BLOCK_GROUP_RAID5)
7320                 return BTRFS_RAID_RAID5;
7321         else if (flags & BTRFS_BLOCK_GROUP_RAID6)
7322                 return BTRFS_RAID_RAID6;
7323
7324         return BTRFS_RAID_SINGLE; /* BTRFS_BLOCK_GROUP_SINGLE */
7325 }
7326
7327 int get_block_group_index(struct btrfs_block_group_cache *cache)
7328 {
7329         return __get_raid_index(cache->flags);
7330 }
7331
7332 static const char *btrfs_raid_type_names[BTRFS_NR_RAID_TYPES] = {
7333         [BTRFS_RAID_RAID10]     = "raid10",
7334         [BTRFS_RAID_RAID1]      = "raid1",
7335         [BTRFS_RAID_DUP]        = "dup",
7336         [BTRFS_RAID_RAID0]      = "raid0",
7337         [BTRFS_RAID_SINGLE]     = "single",
7338         [BTRFS_RAID_RAID5]      = "raid5",
7339         [BTRFS_RAID_RAID6]      = "raid6",
7340 };
7341
7342 static const char *get_raid_name(enum btrfs_raid_types type)
7343 {
7344         if (type >= BTRFS_NR_RAID_TYPES)
7345                 return NULL;
7346
7347         return btrfs_raid_type_names[type];
7348 }
7349
7350 enum btrfs_loop_type {
7351         LOOP_CACHING_NOWAIT = 0,
7352         LOOP_CACHING_WAIT = 1,
7353         LOOP_ALLOC_CHUNK = 2,
7354         LOOP_NO_EMPTY_SIZE = 3,
7355 };
7356
7357 static inline void
7358 btrfs_lock_block_group(struct btrfs_block_group_cache *cache,
7359                        int delalloc)
7360 {
7361         if (delalloc)
7362                 down_read(&cache->data_rwsem);
7363 }
7364
7365 static inline void
7366 btrfs_grab_block_group(struct btrfs_block_group_cache *cache,
7367                        int delalloc)
7368 {
7369         btrfs_get_block_group(cache);
7370         if (delalloc)
7371                 down_read(&cache->data_rwsem);
7372 }
7373
7374 static struct btrfs_block_group_cache *
7375 btrfs_lock_cluster(struct btrfs_block_group_cache *block_group,
7376                    struct btrfs_free_cluster *cluster,
7377                    int delalloc)
7378 {
7379         struct btrfs_block_group_cache *used_bg = NULL;
7380
7381         spin_lock(&cluster->refill_lock);
7382         while (1) {
7383                 used_bg = cluster->block_group;
7384                 if (!used_bg)
7385                         return NULL;
7386
7387                 if (used_bg == block_group)
7388                         return used_bg;
7389
7390                 btrfs_get_block_group(used_bg);
7391
7392                 if (!delalloc)
7393                         return used_bg;
7394
7395                 if (down_read_trylock(&used_bg->data_rwsem))
7396                         return used_bg;
7397
7398                 spin_unlock(&cluster->refill_lock);
7399
7400                 down_read(&used_bg->data_rwsem);
7401
7402                 spin_lock(&cluster->refill_lock);
7403                 if (used_bg == cluster->block_group)
7404                         return used_bg;
7405
7406                 up_read(&used_bg->data_rwsem);
7407                 btrfs_put_block_group(used_bg);
7408         }
7409 }
7410
7411 static inline void
7412 btrfs_release_block_group(struct btrfs_block_group_cache *cache,
7413                          int delalloc)
7414 {
7415         if (delalloc)
7416                 up_read(&cache->data_rwsem);
7417         btrfs_put_block_group(cache);
7418 }
7419
7420 /*
7421  * walks the btree of allocated extents and find a hole of a given size.
7422  * The key ins is changed to record the hole:
7423  * ins->objectid == start position
7424  * ins->flags = BTRFS_EXTENT_ITEM_KEY
7425  * ins->offset == the size of the hole.
7426  * Any available blocks before search_start are skipped.
7427  *
7428  * If there is no suitable free space, we will record the max size of
7429  * the free space extent currently.
7430  */
7431 static noinline int find_free_extent(struct btrfs_root *orig_root,
7432                                 u64 ram_bytes, u64 num_bytes, u64 empty_size,
7433                                 u64 hint_byte, struct btrfs_key *ins,
7434                                 u64 flags, int delalloc)
7435 {
7436         int ret = 0;
7437         struct btrfs_root *root = orig_root->fs_info->extent_root;
7438         struct btrfs_free_cluster *last_ptr = NULL;
7439         struct btrfs_block_group_cache *block_group = NULL;
7440         u64 search_start = 0;
7441         u64 max_extent_size = 0;
7442         u64 empty_cluster = 0;
7443         struct btrfs_space_info *space_info;
7444         int loop = 0;
7445         int index = __get_raid_index(flags);
7446         bool failed_cluster_refill = false;
7447         bool failed_alloc = false;
7448         bool use_cluster = true;
7449         bool have_caching_bg = false;
7450         bool orig_have_caching_bg = false;
7451         bool full_search = false;
7452
7453         WARN_ON(num_bytes < root->sectorsize);
7454         ins->type = BTRFS_EXTENT_ITEM_KEY;
7455         ins->objectid = 0;
7456         ins->offset = 0;
7457
7458         trace_find_free_extent(orig_root, num_bytes, empty_size, flags);
7459
7460         space_info = __find_space_info(root->fs_info, flags);
7461         if (!space_info) {
7462                 btrfs_err(root->fs_info, "No space info for %llu", flags);
7463                 return -ENOSPC;
7464         }
7465
7466         /*
7467          * If our free space is heavily fragmented we may not be able to make
7468          * big contiguous allocations, so instead of doing the expensive search
7469          * for free space, simply return ENOSPC with our max_extent_size so we
7470          * can go ahead and search for a more manageable chunk.
7471          *
7472          * If our max_extent_size is large enough for our allocation simply
7473          * disable clustering since we will likely not be able to find enough
7474          * space to create a cluster and induce latency trying.
7475          */
7476         if (unlikely(space_info->max_extent_size)) {
7477                 spin_lock(&space_info->lock);
7478                 if (space_info->max_extent_size &&
7479                     num_bytes > space_info->max_extent_size) {
7480                         ins->offset = space_info->max_extent_size;
7481                         spin_unlock(&space_info->lock);
7482                         return -ENOSPC;
7483                 } else if (space_info->max_extent_size) {
7484                         use_cluster = false;
7485                 }
7486                 spin_unlock(&space_info->lock);
7487         }
7488
7489         last_ptr = fetch_cluster_info(orig_root, space_info, &empty_cluster);
7490         if (last_ptr) {
7491                 spin_lock(&last_ptr->lock);
7492                 if (last_ptr->block_group)
7493                         hint_byte = last_ptr->window_start;
7494                 if (last_ptr->fragmented) {
7495                         /*
7496                          * We still set window_start so we can keep track of the
7497                          * last place we found an allocation to try and save
7498                          * some time.
7499                          */
7500                         hint_byte = last_ptr->window_start;
7501                         use_cluster = false;
7502                 }
7503                 spin_unlock(&last_ptr->lock);
7504         }
7505
7506         search_start = max(search_start, first_logical_byte(root, 0));
7507         search_start = max(search_start, hint_byte);
7508         if (search_start == hint_byte) {
7509                 block_group = btrfs_lookup_block_group(root->fs_info,
7510                                                        search_start);
7511                 /*
7512                  * we don't want to use the block group if it doesn't match our
7513                  * allocation bits, or if its not cached.
7514                  *
7515                  * However if we are re-searching with an ideal block group
7516                  * picked out then we don't care that the block group is cached.
7517                  */
7518                 if (block_group && block_group_bits(block_group, flags) &&
7519                     block_group->cached != BTRFS_CACHE_NO) {
7520                         down_read(&space_info->groups_sem);
7521                         if (list_empty(&block_group->list) ||
7522                             block_group->ro) {
7523                                 /*
7524                                  * someone is removing this block group,
7525                                  * we can't jump into the have_block_group
7526                                  * target because our list pointers are not
7527                                  * valid
7528                                  */
7529                                 btrfs_put_block_group(block_group);
7530                                 up_read(&space_info->groups_sem);
7531                         } else {
7532                                 index = get_block_group_index(block_group);
7533                                 btrfs_lock_block_group(block_group, delalloc);
7534                                 goto have_block_group;
7535                         }
7536                 } else if (block_group) {
7537                         btrfs_put_block_group(block_group);
7538                 }
7539         }
7540 search:
7541         have_caching_bg = false;
7542         if (index == 0 || index == __get_raid_index(flags))
7543                 full_search = true;
7544         down_read(&space_info->groups_sem);
7545         list_for_each_entry(block_group, &space_info->block_groups[index],
7546                             list) {
7547                 u64 offset;
7548                 int cached;
7549
7550                 btrfs_grab_block_group(block_group, delalloc);
7551                 search_start = block_group->key.objectid;
7552
7553                 /*
7554                  * this can happen if we end up cycling through all the
7555                  * raid types, but we want to make sure we only allocate
7556                  * for the proper type.
7557                  */
7558                 if (!block_group_bits(block_group, flags)) {
7559                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
7560                                 BTRFS_BLOCK_GROUP_RAID1 |
7561                                 BTRFS_BLOCK_GROUP_RAID5 |
7562                                 BTRFS_BLOCK_GROUP_RAID6 |
7563                                 BTRFS_BLOCK_GROUP_RAID10;
7564
7565                         /*
7566                          * if they asked for extra copies and this block group
7567                          * doesn't provide them, bail.  This does allow us to
7568                          * fill raid0 from raid1.
7569                          */
7570                         if ((flags & extra) && !(block_group->flags & extra))
7571                                 goto loop;
7572                 }
7573
7574 have_block_group:
7575                 cached = block_group_cache_done(block_group);
7576                 if (unlikely(!cached)) {
7577                         have_caching_bg = true;
7578                         ret = cache_block_group(block_group, 0);
7579                         BUG_ON(ret < 0);
7580                         ret = 0;
7581                 }
7582
7583                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
7584                         goto loop;
7585                 if (unlikely(block_group->ro))
7586                         goto loop;
7587
7588                 /*
7589                  * Ok we want to try and use the cluster allocator, so
7590                  * lets look there
7591                  */
7592                 if (last_ptr && use_cluster) {
7593                         struct btrfs_block_group_cache *used_block_group;
7594                         unsigned long aligned_cluster;
7595                         /*
7596                          * the refill lock keeps out other
7597                          * people trying to start a new cluster
7598                          */
7599                         used_block_group = btrfs_lock_cluster(block_group,
7600                                                               last_ptr,
7601                                                               delalloc);
7602                         if (!used_block_group)
7603                                 goto refill_cluster;
7604
7605                         if (used_block_group != block_group &&
7606                             (used_block_group->ro ||
7607                              !block_group_bits(used_block_group, flags)))
7608                                 goto release_cluster;
7609
7610                         offset = btrfs_alloc_from_cluster(used_block_group,
7611                                                 last_ptr,
7612                                                 num_bytes,
7613                                                 used_block_group->key.objectid,
7614                                                 &max_extent_size);
7615                         if (offset) {
7616                                 /* we have a block, we're done */
7617                                 spin_unlock(&last_ptr->refill_lock);
7618                                 trace_btrfs_reserve_extent_cluster(root,
7619                                                 used_block_group,
7620                                                 search_start, num_bytes);
7621                                 if (used_block_group != block_group) {
7622                                         btrfs_release_block_group(block_group,
7623                                                                   delalloc);
7624                                         block_group = used_block_group;
7625                                 }
7626                                 goto checks;
7627                         }
7628
7629                         WARN_ON(last_ptr->block_group != used_block_group);
7630 release_cluster:
7631                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
7632                          * set up a new clusters, so lets just skip it
7633                          * and let the allocator find whatever block
7634                          * it can find.  If we reach this point, we
7635                          * will have tried the cluster allocator
7636                          * plenty of times and not have found
7637                          * anything, so we are likely way too
7638                          * fragmented for the clustering stuff to find
7639                          * anything.
7640                          *
7641                          * However, if the cluster is taken from the
7642                          * current block group, release the cluster
7643                          * first, so that we stand a better chance of
7644                          * succeeding in the unclustered
7645                          * allocation.  */
7646                         if (loop >= LOOP_NO_EMPTY_SIZE &&
7647                             used_block_group != block_group) {
7648                                 spin_unlock(&last_ptr->refill_lock);
7649                                 btrfs_release_block_group(used_block_group,
7650                                                           delalloc);
7651                                 goto unclustered_alloc;
7652                         }
7653
7654                         /*
7655                          * this cluster didn't work out, free it and
7656                          * start over
7657                          */
7658                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7659
7660                         if (used_block_group != block_group)
7661                                 btrfs_release_block_group(used_block_group,
7662                                                           delalloc);
7663 refill_cluster:
7664                         if (loop >= LOOP_NO_EMPTY_SIZE) {
7665                                 spin_unlock(&last_ptr->refill_lock);
7666                                 goto unclustered_alloc;
7667                         }
7668
7669                         aligned_cluster = max_t(unsigned long,
7670                                                 empty_cluster + empty_size,
7671                                               block_group->full_stripe_len);
7672
7673                         /* allocate a cluster in this block group */
7674                         ret = btrfs_find_space_cluster(root, block_group,
7675                                                        last_ptr, search_start,
7676                                                        num_bytes,
7677                                                        aligned_cluster);
7678                         if (ret == 0) {
7679                                 /*
7680                                  * now pull our allocation out of this
7681                                  * cluster
7682                                  */
7683                                 offset = btrfs_alloc_from_cluster(block_group,
7684                                                         last_ptr,
7685                                                         num_bytes,
7686                                                         search_start,
7687                                                         &max_extent_size);
7688                                 if (offset) {
7689                                         /* we found one, proceed */
7690                                         spin_unlock(&last_ptr->refill_lock);
7691                                         trace_btrfs_reserve_extent_cluster(root,
7692                                                 block_group, search_start,
7693                                                 num_bytes);
7694                                         goto checks;
7695                                 }
7696                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
7697                                    && !failed_cluster_refill) {
7698                                 spin_unlock(&last_ptr->refill_lock);
7699
7700                                 failed_cluster_refill = true;
7701                                 wait_block_group_cache_progress(block_group,
7702                                        num_bytes + empty_cluster + empty_size);
7703                                 goto have_block_group;
7704                         }
7705
7706                         /*
7707                          * at this point we either didn't find a cluster
7708                          * or we weren't able to allocate a block from our
7709                          * cluster.  Free the cluster we've been trying
7710                          * to use, and go to the next block group
7711                          */
7712                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7713                         spin_unlock(&last_ptr->refill_lock);
7714                         goto loop;
7715                 }
7716
7717 unclustered_alloc:
7718                 /*
7719                  * We are doing an unclustered alloc, set the fragmented flag so
7720                  * we don't bother trying to setup a cluster again until we get
7721                  * more space.
7722                  */
7723                 if (unlikely(last_ptr)) {
7724                         spin_lock(&last_ptr->lock);
7725                         last_ptr->fragmented = 1;
7726                         spin_unlock(&last_ptr->lock);
7727                 }
7728                 spin_lock(&block_group->free_space_ctl->tree_lock);
7729                 if (cached &&
7730                     block_group->free_space_ctl->free_space <
7731                     num_bytes + empty_cluster + empty_size) {
7732                         if (block_group->free_space_ctl->free_space >
7733                             max_extent_size)
7734                                 max_extent_size =
7735                                         block_group->free_space_ctl->free_space;
7736                         spin_unlock(&block_group->free_space_ctl->tree_lock);
7737                         goto loop;
7738                 }
7739                 spin_unlock(&block_group->free_space_ctl->tree_lock);
7740
7741                 offset = btrfs_find_space_for_alloc(block_group, search_start,
7742                                                     num_bytes, empty_size,
7743                                                     &max_extent_size);
7744                 /*
7745                  * If we didn't find a chunk, and we haven't failed on this
7746                  * block group before, and this block group is in the middle of
7747                  * caching and we are ok with waiting, then go ahead and wait
7748                  * for progress to be made, and set failed_alloc to true.
7749                  *
7750                  * If failed_alloc is true then we've already waited on this
7751                  * block group once and should move on to the next block group.
7752                  */
7753                 if (!offset && !failed_alloc && !cached &&
7754                     loop > LOOP_CACHING_NOWAIT) {
7755                         wait_block_group_cache_progress(block_group,
7756                                                 num_bytes + empty_size);
7757                         failed_alloc = true;
7758                         goto have_block_group;
7759                 } else if (!offset) {
7760                         goto loop;
7761                 }
7762 checks:
7763                 search_start = ALIGN(offset, root->stripesize);
7764
7765                 /* move on to the next group */
7766                 if (search_start + num_bytes >
7767                     block_group->key.objectid + block_group->key.offset) {
7768                         btrfs_add_free_space(block_group, offset, num_bytes);
7769                         goto loop;
7770                 }
7771
7772                 if (offset < search_start)
7773                         btrfs_add_free_space(block_group, offset,
7774                                              search_start - offset);
7775                 BUG_ON(offset > search_start);
7776
7777                 ret = btrfs_add_reserved_bytes(block_group, ram_bytes,
7778                                 num_bytes, delalloc);
7779                 if (ret == -EAGAIN) {
7780                         btrfs_add_free_space(block_group, offset, num_bytes);
7781                         goto loop;
7782                 }
7783                 btrfs_inc_block_group_reservations(block_group);
7784
7785                 /* we are all good, lets return */
7786                 ins->objectid = search_start;
7787                 ins->offset = num_bytes;
7788
7789                 trace_btrfs_reserve_extent(orig_root, block_group,
7790                                            search_start, num_bytes);
7791                 btrfs_release_block_group(block_group, delalloc);
7792                 break;
7793 loop:
7794                 failed_cluster_refill = false;
7795                 failed_alloc = false;
7796                 BUG_ON(index != get_block_group_index(block_group));
7797                 btrfs_release_block_group(block_group, delalloc);
7798         }
7799         up_read(&space_info->groups_sem);
7800
7801         if ((loop == LOOP_CACHING_NOWAIT) && have_caching_bg
7802                 && !orig_have_caching_bg)
7803                 orig_have_caching_bg = true;
7804
7805         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
7806                 goto search;
7807
7808         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
7809                 goto search;
7810
7811         /*
7812          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
7813          *                      caching kthreads as we move along
7814          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
7815          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
7816          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
7817          *                      again
7818          */
7819         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
7820                 index = 0;
7821                 if (loop == LOOP_CACHING_NOWAIT) {
7822                         /*
7823                          * We want to skip the LOOP_CACHING_WAIT step if we
7824                          * don't have any uncached bgs and we've already done a
7825                          * full search through.
7826                          */
7827                         if (orig_have_caching_bg || !full_search)
7828                                 loop = LOOP_CACHING_WAIT;
7829                         else
7830                                 loop = LOOP_ALLOC_CHUNK;
7831                 } else {
7832                         loop++;
7833                 }
7834
7835                 if (loop == LOOP_ALLOC_CHUNK) {
7836                         struct btrfs_trans_handle *trans;
7837                         int exist = 0;
7838
7839                         trans = current->journal_info;
7840                         if (trans)
7841                                 exist = 1;
7842                         else
7843                                 trans = btrfs_join_transaction(root);
7844
7845                         if (IS_ERR(trans)) {
7846                                 ret = PTR_ERR(trans);
7847                                 goto out;
7848                         }
7849
7850                         ret = do_chunk_alloc(trans, root, flags,
7851                                              CHUNK_ALLOC_FORCE);
7852
7853                         /*
7854                          * If we can't allocate a new chunk we've already looped
7855                          * through at least once, move on to the NO_EMPTY_SIZE
7856                          * case.
7857                          */
7858                         if (ret == -ENOSPC)
7859                                 loop = LOOP_NO_EMPTY_SIZE;
7860
7861                         /*
7862                          * Do not bail out on ENOSPC since we
7863                          * can do more things.
7864                          */
7865                         if (ret < 0 && ret != -ENOSPC)
7866                                 btrfs_abort_transaction(trans, ret);
7867                         else
7868                                 ret = 0;
7869                         if (!exist)
7870                                 btrfs_end_transaction(trans, root);
7871                         if (ret)
7872                                 goto out;
7873                 }
7874
7875                 if (loop == LOOP_NO_EMPTY_SIZE) {
7876                         /*
7877                          * Don't loop again if we already have no empty_size and
7878                          * no empty_cluster.
7879                          */
7880                         if (empty_size == 0 &&
7881                             empty_cluster == 0) {
7882                                 ret = -ENOSPC;
7883                                 goto out;
7884                         }
7885                         empty_size = 0;
7886                         empty_cluster = 0;
7887                 }
7888
7889                 goto search;
7890         } else if (!ins->objectid) {
7891                 ret = -ENOSPC;
7892         } else if (ins->objectid) {
7893                 if (!use_cluster && last_ptr) {
7894                         spin_lock(&last_ptr->lock);
7895                         last_ptr->window_start = ins->objectid;
7896                         spin_unlock(&last_ptr->lock);
7897                 }
7898                 ret = 0;
7899         }
7900 out:
7901         if (ret == -ENOSPC) {
7902                 spin_lock(&space_info->lock);
7903                 space_info->max_extent_size = max_extent_size;
7904                 spin_unlock(&space_info->lock);
7905                 ins->offset = max_extent_size;
7906         }
7907         return ret;
7908 }
7909
7910 static void dump_space_info(struct btrfs_fs_info *fs_info,
7911                             struct btrfs_space_info *info, u64 bytes,
7912                             int dump_block_groups)
7913 {
7914         struct btrfs_block_group_cache *cache;
7915         int index = 0;
7916
7917         spin_lock(&info->lock);
7918         btrfs_info(fs_info, "space_info %llu has %llu free, is %sfull",
7919                    info->flags,
7920                    info->total_bytes - info->bytes_used - info->bytes_pinned -
7921                    info->bytes_reserved - info->bytes_readonly -
7922                    info->bytes_may_use, (info->full) ? "" : "not ");
7923         btrfs_info(fs_info,
7924                 "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu",
7925                 info->total_bytes, info->bytes_used, info->bytes_pinned,
7926                 info->bytes_reserved, info->bytes_may_use,
7927                 info->bytes_readonly);
7928         spin_unlock(&info->lock);
7929
7930         if (!dump_block_groups)
7931                 return;
7932
7933         down_read(&info->groups_sem);
7934 again:
7935         list_for_each_entry(cache, &info->block_groups[index], list) {
7936                 spin_lock(&cache->lock);
7937                 btrfs_info(fs_info,
7938                         "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s",
7939                         cache->key.objectid, cache->key.offset,
7940                         btrfs_block_group_used(&cache->item), cache->pinned,
7941                         cache->reserved, cache->ro ? "[readonly]" : "");
7942                 btrfs_dump_free_space(cache, bytes);
7943                 spin_unlock(&cache->lock);
7944         }
7945         if (++index < BTRFS_NR_RAID_TYPES)
7946                 goto again;
7947         up_read(&info->groups_sem);
7948 }
7949
7950 int btrfs_reserve_extent(struct btrfs_root *root, u64 ram_bytes,
7951                          u64 num_bytes, u64 min_alloc_size,
7952                          u64 empty_size, u64 hint_byte,
7953                          struct btrfs_key *ins, int is_data, int delalloc)
7954 {
7955         struct btrfs_fs_info *fs_info = root->fs_info;
7956         bool final_tried = num_bytes == min_alloc_size;
7957         u64 flags;
7958         int ret;
7959
7960         flags = btrfs_get_alloc_profile(root, is_data);
7961 again:
7962         WARN_ON(num_bytes < root->sectorsize);
7963         ret = find_free_extent(root, ram_bytes, num_bytes, empty_size,
7964                                hint_byte, ins, flags, delalloc);
7965         if (!ret && !is_data) {
7966                 btrfs_dec_block_group_reservations(fs_info, ins->objectid);
7967         } else if (ret == -ENOSPC) {
7968                 if (!final_tried && ins->offset) {
7969                         num_bytes = min(num_bytes >> 1, ins->offset);
7970                         num_bytes = round_down(num_bytes, root->sectorsize);
7971                         num_bytes = max(num_bytes, min_alloc_size);
7972                         ram_bytes = num_bytes;
7973                         if (num_bytes == min_alloc_size)
7974                                 final_tried = true;
7975                         goto again;
7976                 } else if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
7977                         struct btrfs_space_info *sinfo;
7978
7979                         sinfo = __find_space_info(fs_info, flags);
7980                         btrfs_err(root->fs_info,
7981                                   "allocation failed flags %llu, wanted %llu",
7982                                   flags, num_bytes);
7983                         if (sinfo)
7984                                 dump_space_info(fs_info, sinfo, num_bytes, 1);
7985                 }
7986         }
7987
7988         return ret;
7989 }
7990
7991 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
7992                                         u64 start, u64 len,
7993                                         int pin, int delalloc)
7994 {
7995         struct btrfs_block_group_cache *cache;
7996         int ret = 0;
7997
7998         cache = btrfs_lookup_block_group(root->fs_info, start);
7999         if (!cache) {
8000                 btrfs_err(root->fs_info, "Unable to find block group for %llu",
8001                         start);
8002                 return -ENOSPC;
8003         }
8004
8005         if (pin)
8006                 pin_down_extent(root, cache, start, len, 1);
8007         else {
8008                 if (btrfs_test_opt(root->fs_info, DISCARD))
8009                         ret = btrfs_discard_extent(root, start, len, NULL);
8010                 btrfs_add_free_space(cache, start, len);
8011                 btrfs_free_reserved_bytes(cache, len, delalloc);
8012                 trace_btrfs_reserved_extent_free(root, start, len);
8013         }
8014
8015         btrfs_put_block_group(cache);
8016         return ret;
8017 }
8018
8019 int btrfs_free_reserved_extent(struct btrfs_root *root,
8020                                u64 start, u64 len, int delalloc)
8021 {
8022         return __btrfs_free_reserved_extent(root, start, len, 0, delalloc);
8023 }
8024
8025 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
8026                                        u64 start, u64 len)
8027 {
8028         return __btrfs_free_reserved_extent(root, start, len, 1, 0);
8029 }
8030
8031 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8032                                       struct btrfs_root *root,
8033                                       u64 parent, u64 root_objectid,
8034                                       u64 flags, u64 owner, u64 offset,
8035                                       struct btrfs_key *ins, int ref_mod)
8036 {
8037         int ret;
8038         struct btrfs_fs_info *fs_info = root->fs_info;
8039         struct btrfs_extent_item *extent_item;
8040         struct btrfs_extent_inline_ref *iref;
8041         struct btrfs_path *path;
8042         struct extent_buffer *leaf;
8043         int type;
8044         u32 size;
8045
8046         if (parent > 0)
8047                 type = BTRFS_SHARED_DATA_REF_KEY;
8048         else
8049                 type = BTRFS_EXTENT_DATA_REF_KEY;
8050
8051         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
8052
8053         path = btrfs_alloc_path();
8054         if (!path)
8055                 return -ENOMEM;
8056
8057         path->leave_spinning = 1;
8058         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
8059                                       ins, size);
8060         if (ret) {
8061                 btrfs_free_path(path);
8062                 return ret;
8063         }
8064
8065         leaf = path->nodes[0];
8066         extent_item = btrfs_item_ptr(leaf, path->slots[0],
8067                                      struct btrfs_extent_item);
8068         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
8069         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8070         btrfs_set_extent_flags(leaf, extent_item,
8071                                flags | BTRFS_EXTENT_FLAG_DATA);
8072
8073         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8074         btrfs_set_extent_inline_ref_type(leaf, iref, type);
8075         if (parent > 0) {
8076                 struct btrfs_shared_data_ref *ref;
8077                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
8078                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
8079                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
8080         } else {
8081                 struct btrfs_extent_data_ref *ref;
8082                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
8083                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
8084                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
8085                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
8086                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
8087         }
8088
8089         btrfs_mark_buffer_dirty(path->nodes[0]);
8090         btrfs_free_path(path);
8091
8092         ret = remove_from_free_space_tree(trans, fs_info, ins->objectid,
8093                                           ins->offset);
8094         if (ret)
8095                 return ret;
8096
8097         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
8098         if (ret) { /* -ENOENT, logic error */
8099                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8100                         ins->objectid, ins->offset);
8101                 BUG();
8102         }
8103         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
8104         return ret;
8105 }
8106
8107 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
8108                                      struct btrfs_root *root,
8109                                      u64 parent, u64 root_objectid,
8110                                      u64 flags, struct btrfs_disk_key *key,
8111                                      int level, struct btrfs_key *ins)
8112 {
8113         int ret;
8114         struct btrfs_fs_info *fs_info = root->fs_info;
8115         struct btrfs_extent_item *extent_item;
8116         struct btrfs_tree_block_info *block_info;
8117         struct btrfs_extent_inline_ref *iref;
8118         struct btrfs_path *path;
8119         struct extent_buffer *leaf;
8120         u32 size = sizeof(*extent_item) + sizeof(*iref);
8121         u64 num_bytes = ins->offset;
8122         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
8123                                                  SKINNY_METADATA);
8124
8125         if (!skinny_metadata)
8126                 size += sizeof(*block_info);
8127
8128         path = btrfs_alloc_path();
8129         if (!path) {
8130                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
8131                                                    root->nodesize);
8132                 return -ENOMEM;
8133         }
8134
8135         path->leave_spinning = 1;
8136         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
8137                                       ins, size);
8138         if (ret) {
8139                 btrfs_free_path(path);
8140                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
8141                                                    root->nodesize);
8142                 return ret;
8143         }
8144
8145         leaf = path->nodes[0];
8146         extent_item = btrfs_item_ptr(leaf, path->slots[0],
8147                                      struct btrfs_extent_item);
8148         btrfs_set_extent_refs(leaf, extent_item, 1);
8149         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8150         btrfs_set_extent_flags(leaf, extent_item,
8151                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
8152
8153         if (skinny_metadata) {
8154                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8155                 num_bytes = root->nodesize;
8156         } else {
8157                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
8158                 btrfs_set_tree_block_key(leaf, block_info, key);
8159                 btrfs_set_tree_block_level(leaf, block_info, level);
8160                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
8161         }
8162
8163         if (parent > 0) {
8164                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
8165                 btrfs_set_extent_inline_ref_type(leaf, iref,
8166                                                  BTRFS_SHARED_BLOCK_REF_KEY);
8167                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
8168         } else {
8169                 btrfs_set_extent_inline_ref_type(leaf, iref,
8170                                                  BTRFS_TREE_BLOCK_REF_KEY);
8171                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
8172         }
8173
8174         btrfs_mark_buffer_dirty(leaf);
8175         btrfs_free_path(path);
8176
8177         ret = remove_from_free_space_tree(trans, fs_info, ins->objectid,
8178                                           num_bytes);
8179         if (ret)
8180                 return ret;
8181
8182         ret = update_block_group(trans, root, ins->objectid, root->nodesize,
8183                                  1);
8184         if (ret) { /* -ENOENT, logic error */
8185                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8186                         ins->objectid, ins->offset);
8187                 BUG();
8188         }
8189
8190         trace_btrfs_reserved_extent_alloc(root, ins->objectid, root->nodesize);
8191         return ret;
8192 }
8193
8194 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8195                                      struct btrfs_root *root,
8196                                      u64 root_objectid, u64 owner,
8197                                      u64 offset, u64 ram_bytes,
8198                                      struct btrfs_key *ins)
8199 {
8200         int ret;
8201
8202         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
8203
8204         ret = btrfs_add_delayed_data_ref(root->fs_info, trans, ins->objectid,
8205                                          ins->offset, 0,
8206                                          root_objectid, owner, offset,
8207                                          ram_bytes, BTRFS_ADD_DELAYED_EXTENT,
8208                                          NULL);
8209         return ret;
8210 }
8211
8212 /*
8213  * this is used by the tree logging recovery code.  It records that
8214  * an extent has been allocated and makes sure to clear the free
8215  * space cache bits as well
8216  */
8217 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
8218                                    struct btrfs_root *root,
8219                                    u64 root_objectid, u64 owner, u64 offset,
8220                                    struct btrfs_key *ins)
8221 {
8222         int ret;
8223         struct btrfs_block_group_cache *block_group;
8224         struct btrfs_space_info *space_info;
8225
8226         /*
8227          * Mixed block groups will exclude before processing the log so we only
8228          * need to do the exclude dance if this fs isn't mixed.
8229          */
8230         if (!btrfs_fs_incompat(root->fs_info, MIXED_GROUPS)) {
8231                 ret = __exclude_logged_extent(root, ins->objectid, ins->offset);
8232                 if (ret)
8233                         return ret;
8234         }
8235
8236         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
8237         if (!block_group)
8238                 return -EINVAL;
8239
8240         space_info = block_group->space_info;
8241         spin_lock(&space_info->lock);
8242         spin_lock(&block_group->lock);
8243         space_info->bytes_reserved += ins->offset;
8244         block_group->reserved += ins->offset;
8245         spin_unlock(&block_group->lock);
8246         spin_unlock(&space_info->lock);
8247
8248         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
8249                                          0, owner, offset, ins, 1);
8250         btrfs_put_block_group(block_group);
8251         return ret;
8252 }
8253
8254 static struct extent_buffer *
8255 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
8256                       u64 bytenr, int level)
8257 {
8258         struct extent_buffer *buf;
8259
8260         buf = btrfs_find_create_tree_block(root, bytenr);
8261         if (IS_ERR(buf))
8262                 return buf;
8263
8264         btrfs_set_header_generation(buf, trans->transid);
8265         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
8266         btrfs_tree_lock(buf);
8267         clean_tree_block(trans, root->fs_info, buf);
8268         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
8269
8270         btrfs_set_lock_blocking(buf);
8271         set_extent_buffer_uptodate(buf);
8272
8273         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
8274                 buf->log_index = root->log_transid % 2;
8275                 /*
8276                  * we allow two log transactions at a time, use different
8277                  * EXENT bit to differentiate dirty pages.
8278                  */
8279                 if (buf->log_index == 0)
8280                         set_extent_dirty(&root->dirty_log_pages, buf->start,
8281                                         buf->start + buf->len - 1, GFP_NOFS);
8282                 else
8283                         set_extent_new(&root->dirty_log_pages, buf->start,
8284                                         buf->start + buf->len - 1);
8285         } else {
8286                 buf->log_index = -1;
8287                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
8288                          buf->start + buf->len - 1, GFP_NOFS);
8289         }
8290         trans->dirty = true;
8291         /* this returns a buffer locked for blocking */
8292         return buf;
8293 }
8294
8295 static struct btrfs_block_rsv *
8296 use_block_rsv(struct btrfs_trans_handle *trans,
8297               struct btrfs_root *root, u32 blocksize)
8298 {
8299         struct btrfs_block_rsv *block_rsv;
8300         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
8301         int ret;
8302         bool global_updated = false;
8303
8304         block_rsv = get_block_rsv(trans, root);
8305
8306         if (unlikely(block_rsv->size == 0))
8307                 goto try_reserve;
8308 again:
8309         ret = block_rsv_use_bytes(block_rsv, blocksize);
8310         if (!ret)
8311                 return block_rsv;
8312
8313         if (block_rsv->failfast)
8314                 return ERR_PTR(ret);
8315
8316         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
8317                 global_updated = true;
8318                 update_global_block_rsv(root->fs_info);
8319                 goto again;
8320         }
8321
8322         if (btrfs_test_opt(root->fs_info, ENOSPC_DEBUG)) {
8323                 static DEFINE_RATELIMIT_STATE(_rs,
8324                                 DEFAULT_RATELIMIT_INTERVAL * 10,
8325                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
8326                 if (__ratelimit(&_rs))
8327                         WARN(1, KERN_DEBUG
8328                                 "BTRFS: block rsv returned %d\n", ret);
8329         }
8330 try_reserve:
8331         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
8332                                      BTRFS_RESERVE_NO_FLUSH);
8333         if (!ret)
8334                 return block_rsv;
8335         /*
8336          * If we couldn't reserve metadata bytes try and use some from
8337          * the global reserve if its space type is the same as the global
8338          * reservation.
8339          */
8340         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
8341             block_rsv->space_info == global_rsv->space_info) {
8342                 ret = block_rsv_use_bytes(global_rsv, blocksize);
8343                 if (!ret)
8344                         return global_rsv;
8345         }
8346         return ERR_PTR(ret);
8347 }
8348
8349 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
8350                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
8351 {
8352         block_rsv_add_bytes(block_rsv, blocksize, 0);
8353         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
8354 }
8355
8356 /*
8357  * finds a free extent and does all the dirty work required for allocation
8358  * returns the tree buffer or an ERR_PTR on error.
8359  */
8360 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
8361                                         struct btrfs_root *root,
8362                                         u64 parent, u64 root_objectid,
8363                                         struct btrfs_disk_key *key, int level,
8364                                         u64 hint, u64 empty_size)
8365 {
8366         struct btrfs_key ins;
8367         struct btrfs_block_rsv *block_rsv;
8368         struct extent_buffer *buf;
8369         struct btrfs_delayed_extent_op *extent_op;
8370         u64 flags = 0;
8371         int ret;
8372         u32 blocksize = root->nodesize;
8373         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
8374                                                  SKINNY_METADATA);
8375
8376 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
8377         if (btrfs_is_testing(root->fs_info)) {
8378                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
8379                                             level);
8380                 if (!IS_ERR(buf))
8381                         root->alloc_bytenr += blocksize;
8382                 return buf;
8383         }
8384 #endif
8385
8386         block_rsv = use_block_rsv(trans, root, blocksize);
8387         if (IS_ERR(block_rsv))
8388                 return ERR_CAST(block_rsv);
8389
8390         ret = btrfs_reserve_extent(root, blocksize, blocksize, blocksize,
8391                                    empty_size, hint, &ins, 0, 0);
8392         if (ret)
8393                 goto out_unuse;
8394
8395         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level);
8396         if (IS_ERR(buf)) {
8397                 ret = PTR_ERR(buf);
8398                 goto out_free_reserved;
8399         }
8400
8401         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
8402                 if (parent == 0)
8403                         parent = ins.objectid;
8404                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
8405         } else
8406                 BUG_ON(parent > 0);
8407
8408         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
8409                 extent_op = btrfs_alloc_delayed_extent_op();
8410                 if (!extent_op) {
8411                         ret = -ENOMEM;
8412                         goto out_free_buf;
8413                 }
8414                 if (key)
8415                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
8416                 else
8417                         memset(&extent_op->key, 0, sizeof(extent_op->key));
8418                 extent_op->flags_to_set = flags;
8419                 extent_op->update_key = skinny_metadata ? false : true;
8420                 extent_op->update_flags = true;
8421                 extent_op->is_data = false;
8422                 extent_op->level = level;
8423
8424                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
8425                                                  ins.objectid, ins.offset,
8426                                                  parent, root_objectid, level,
8427                                                  BTRFS_ADD_DELAYED_EXTENT,
8428                                                  extent_op);
8429                 if (ret)
8430                         goto out_free_delayed;
8431         }
8432         return buf;
8433
8434 out_free_delayed:
8435         btrfs_free_delayed_extent_op(extent_op);
8436 out_free_buf:
8437         free_extent_buffer(buf);
8438 out_free_reserved:
8439         btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 0);
8440 out_unuse:
8441         unuse_block_rsv(root->fs_info, block_rsv, blocksize);
8442         return ERR_PTR(ret);
8443 }
8444
8445 struct walk_control {
8446         u64 refs[BTRFS_MAX_LEVEL];
8447         u64 flags[BTRFS_MAX_LEVEL];
8448         struct btrfs_key update_progress;
8449         int stage;
8450         int level;
8451         int shared_level;
8452         int update_ref;
8453         int keep_locks;
8454         int reada_slot;
8455         int reada_count;
8456         int for_reloc;
8457 };
8458
8459 #define DROP_REFERENCE  1
8460 #define UPDATE_BACKREF  2
8461
8462 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
8463                                      struct btrfs_root *root,
8464                                      struct walk_control *wc,
8465                                      struct btrfs_path *path)
8466 {
8467         u64 bytenr;
8468         u64 generation;
8469         u64 refs;
8470         u64 flags;
8471         u32 nritems;
8472         struct btrfs_key key;
8473         struct extent_buffer *eb;
8474         int ret;
8475         int slot;
8476         int nread = 0;
8477
8478         if (path->slots[wc->level] < wc->reada_slot) {
8479                 wc->reada_count = wc->reada_count * 2 / 3;
8480                 wc->reada_count = max(wc->reada_count, 2);
8481         } else {
8482                 wc->reada_count = wc->reada_count * 3 / 2;
8483                 wc->reada_count = min_t(int, wc->reada_count,
8484                                         BTRFS_NODEPTRS_PER_BLOCK(root));
8485         }
8486
8487         eb = path->nodes[wc->level];
8488         nritems = btrfs_header_nritems(eb);
8489
8490         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
8491                 if (nread >= wc->reada_count)
8492                         break;
8493
8494                 cond_resched();
8495                 bytenr = btrfs_node_blockptr(eb, slot);
8496                 generation = btrfs_node_ptr_generation(eb, slot);
8497
8498                 if (slot == path->slots[wc->level])
8499                         goto reada;
8500
8501                 if (wc->stage == UPDATE_BACKREF &&
8502                     generation <= root->root_key.offset)
8503                         continue;
8504
8505                 /* We don't lock the tree block, it's OK to be racy here */
8506                 ret = btrfs_lookup_extent_info(trans, root, bytenr,
8507                                                wc->level - 1, 1, &refs,
8508                                                &flags);
8509                 /* We don't care about errors in readahead. */
8510                 if (ret < 0)
8511                         continue;
8512                 BUG_ON(refs == 0);
8513
8514                 if (wc->stage == DROP_REFERENCE) {
8515                         if (refs == 1)
8516                                 goto reada;
8517
8518                         if (wc->level == 1 &&
8519                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8520                                 continue;
8521                         if (!wc->update_ref ||
8522                             generation <= root->root_key.offset)
8523                                 continue;
8524                         btrfs_node_key_to_cpu(eb, &key, slot);
8525                         ret = btrfs_comp_cpu_keys(&key,
8526                                                   &wc->update_progress);
8527                         if (ret < 0)
8528                                 continue;
8529                 } else {
8530                         if (wc->level == 1 &&
8531                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8532                                 continue;
8533                 }
8534 reada:
8535                 readahead_tree_block(root, bytenr);
8536                 nread++;
8537         }
8538         wc->reada_slot = slot;
8539 }
8540
8541 static int account_leaf_items(struct btrfs_trans_handle *trans,
8542                               struct btrfs_root *root,
8543                               struct extent_buffer *eb)
8544 {
8545         int nr = btrfs_header_nritems(eb);
8546         int i, extent_type, ret;
8547         struct btrfs_key key;
8548         struct btrfs_file_extent_item *fi;
8549         u64 bytenr, num_bytes;
8550
8551         /* We can be called directly from walk_up_proc() */
8552         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &root->fs_info->flags))
8553                 return 0;
8554
8555         for (i = 0; i < nr; i++) {
8556                 btrfs_item_key_to_cpu(eb, &key, i);
8557
8558                 if (key.type != BTRFS_EXTENT_DATA_KEY)
8559                         continue;
8560
8561                 fi = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
8562                 /* filter out non qgroup-accountable extents  */
8563                 extent_type = btrfs_file_extent_type(eb, fi);
8564
8565                 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
8566                         continue;
8567
8568                 bytenr = btrfs_file_extent_disk_bytenr(eb, fi);
8569                 if (!bytenr)
8570                         continue;
8571
8572                 num_bytes = btrfs_file_extent_disk_num_bytes(eb, fi);
8573
8574                 ret = btrfs_qgroup_insert_dirty_extent(trans, root->fs_info,
8575                                 bytenr, num_bytes, GFP_NOFS);
8576                 if (ret)
8577                         return ret;
8578         }
8579         return 0;
8580 }
8581
8582 /*
8583  * Walk up the tree from the bottom, freeing leaves and any interior
8584  * nodes which have had all slots visited. If a node (leaf or
8585  * interior) is freed, the node above it will have it's slot
8586  * incremented. The root node will never be freed.
8587  *
8588  * At the end of this function, we should have a path which has all
8589  * slots incremented to the next position for a search. If we need to
8590  * read a new node it will be NULL and the node above it will have the
8591  * correct slot selected for a later read.
8592  *
8593  * If we increment the root nodes slot counter past the number of
8594  * elements, 1 is returned to signal completion of the search.
8595  */
8596 static int adjust_slots_upwards(struct btrfs_root *root,
8597                                 struct btrfs_path *path, int root_level)
8598 {
8599         int level = 0;
8600         int nr, slot;
8601         struct extent_buffer *eb;
8602
8603         if (root_level == 0)
8604                 return 1;
8605
8606         while (level <= root_level) {
8607                 eb = path->nodes[level];
8608                 nr = btrfs_header_nritems(eb);
8609                 path->slots[level]++;
8610                 slot = path->slots[level];
8611                 if (slot >= nr || level == 0) {
8612                         /*
8613                          * Don't free the root -  we will detect this
8614                          * condition after our loop and return a
8615                          * positive value for caller to stop walking the tree.
8616                          */
8617                         if (level != root_level) {
8618                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8619                                 path->locks[level] = 0;
8620
8621                                 free_extent_buffer(eb);
8622                                 path->nodes[level] = NULL;
8623                                 path->slots[level] = 0;
8624                         }
8625                 } else {
8626                         /*
8627                          * We have a valid slot to walk back down
8628                          * from. Stop here so caller can process these
8629                          * new nodes.
8630                          */
8631                         break;
8632                 }
8633
8634                 level++;
8635         }
8636
8637         eb = path->nodes[root_level];
8638         if (path->slots[root_level] >= btrfs_header_nritems(eb))
8639                 return 1;
8640
8641         return 0;
8642 }
8643
8644 /*
8645  * root_eb is the subtree root and is locked before this function is called.
8646  */
8647 static int account_shared_subtree(struct btrfs_trans_handle *trans,
8648                                   struct btrfs_root *root,
8649                                   struct extent_buffer *root_eb,
8650                                   u64 root_gen,
8651                                   int root_level)
8652 {
8653         int ret = 0;
8654         int level;
8655         struct extent_buffer *eb = root_eb;
8656         struct btrfs_path *path = NULL;
8657
8658         BUG_ON(root_level < 0 || root_level > BTRFS_MAX_LEVEL);
8659         BUG_ON(root_eb == NULL);
8660
8661         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &root->fs_info->flags))
8662                 return 0;
8663
8664         if (!extent_buffer_uptodate(root_eb)) {
8665                 ret = btrfs_read_buffer(root_eb, root_gen);
8666                 if (ret)
8667                         goto out;
8668         }
8669
8670         if (root_level == 0) {
8671                 ret = account_leaf_items(trans, root, root_eb);
8672                 goto out;
8673         }
8674
8675         path = btrfs_alloc_path();
8676         if (!path)
8677                 return -ENOMEM;
8678
8679         /*
8680          * Walk down the tree.  Missing extent blocks are filled in as
8681          * we go. Metadata is accounted every time we read a new
8682          * extent block.
8683          *
8684          * When we reach a leaf, we account for file extent items in it,
8685          * walk back up the tree (adjusting slot pointers as we go)
8686          * and restart the search process.
8687          */
8688         extent_buffer_get(root_eb); /* For path */
8689         path->nodes[root_level] = root_eb;
8690         path->slots[root_level] = 0;
8691         path->locks[root_level] = 0; /* so release_path doesn't try to unlock */
8692 walk_down:
8693         level = root_level;
8694         while (level >= 0) {
8695                 if (path->nodes[level] == NULL) {
8696                         int parent_slot;
8697                         u64 child_gen;
8698                         u64 child_bytenr;
8699
8700                         /* We need to get child blockptr/gen from
8701                          * parent before we can read it. */
8702                         eb = path->nodes[level + 1];
8703                         parent_slot = path->slots[level + 1];
8704                         child_bytenr = btrfs_node_blockptr(eb, parent_slot);
8705                         child_gen = btrfs_node_ptr_generation(eb, parent_slot);
8706
8707                         eb = read_tree_block(root, child_bytenr, child_gen);
8708                         if (IS_ERR(eb)) {
8709                                 ret = PTR_ERR(eb);
8710                                 goto out;
8711                         } else if (!extent_buffer_uptodate(eb)) {
8712                                 free_extent_buffer(eb);
8713                                 ret = -EIO;
8714                                 goto out;
8715                         }
8716
8717                         path->nodes[level] = eb;
8718                         path->slots[level] = 0;
8719
8720                         btrfs_tree_read_lock(eb);
8721                         btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
8722                         path->locks[level] = BTRFS_READ_LOCK_BLOCKING;
8723
8724                         ret = btrfs_qgroup_insert_dirty_extent(trans,
8725                                         root->fs_info, child_bytenr,
8726                                         root->nodesize, GFP_NOFS);
8727                         if (ret)
8728                                 goto out;
8729                 }
8730
8731                 if (level == 0) {
8732                         ret = account_leaf_items(trans, root, path->nodes[level]);
8733                         if (ret)
8734                                 goto out;
8735
8736                         /* Nonzero return here means we completed our search */
8737                         ret = adjust_slots_upwards(root, path, root_level);
8738                         if (ret)
8739                                 break;
8740
8741                         /* Restart search with new slots */
8742                         goto walk_down;
8743                 }
8744
8745                 level--;
8746         }
8747
8748         ret = 0;
8749 out:
8750         btrfs_free_path(path);
8751
8752         return ret;
8753 }
8754
8755 /*
8756  * helper to process tree block while walking down the tree.
8757  *
8758  * when wc->stage == UPDATE_BACKREF, this function updates
8759  * back refs for pointers in the block.
8760  *
8761  * NOTE: return value 1 means we should stop walking down.
8762  */
8763 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
8764                                    struct btrfs_root *root,
8765                                    struct btrfs_path *path,
8766                                    struct walk_control *wc, int lookup_info)
8767 {
8768         int level = wc->level;
8769         struct extent_buffer *eb = path->nodes[level];
8770         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8771         int ret;
8772
8773         if (wc->stage == UPDATE_BACKREF &&
8774             btrfs_header_owner(eb) != root->root_key.objectid)
8775                 return 1;
8776
8777         /*
8778          * when reference count of tree block is 1, it won't increase
8779          * again. once full backref flag is set, we never clear it.
8780          */
8781         if (lookup_info &&
8782             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
8783              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
8784                 BUG_ON(!path->locks[level]);
8785                 ret = btrfs_lookup_extent_info(trans, root,
8786                                                eb->start, level, 1,
8787                                                &wc->refs[level],
8788                                                &wc->flags[level]);
8789                 BUG_ON(ret == -ENOMEM);
8790                 if (ret)
8791                         return ret;
8792                 BUG_ON(wc->refs[level] == 0);
8793         }
8794
8795         if (wc->stage == DROP_REFERENCE) {
8796                 if (wc->refs[level] > 1)
8797                         return 1;
8798
8799                 if (path->locks[level] && !wc->keep_locks) {
8800                         btrfs_tree_unlock_rw(eb, path->locks[level]);
8801                         path->locks[level] = 0;
8802                 }
8803                 return 0;
8804         }
8805
8806         /* wc->stage == UPDATE_BACKREF */
8807         if (!(wc->flags[level] & flag)) {
8808                 BUG_ON(!path->locks[level]);
8809                 ret = btrfs_inc_ref(trans, root, eb, 1);
8810                 BUG_ON(ret); /* -ENOMEM */
8811                 ret = btrfs_dec_ref(trans, root, eb, 0);
8812                 BUG_ON(ret); /* -ENOMEM */
8813                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
8814                                                   eb->len, flag,
8815                                                   btrfs_header_level(eb), 0);
8816                 BUG_ON(ret); /* -ENOMEM */
8817                 wc->flags[level] |= flag;
8818         }
8819
8820         /*
8821          * the block is shared by multiple trees, so it's not good to
8822          * keep the tree lock
8823          */
8824         if (path->locks[level] && level > 0) {
8825                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8826                 path->locks[level] = 0;
8827         }
8828         return 0;
8829 }
8830
8831 /*
8832  * helper to process tree block pointer.
8833  *
8834  * when wc->stage == DROP_REFERENCE, this function checks
8835  * reference count of the block pointed to. if the block
8836  * is shared and we need update back refs for the subtree
8837  * rooted at the block, this function changes wc->stage to
8838  * UPDATE_BACKREF. if the block is shared and there is no
8839  * need to update back, this function drops the reference
8840  * to the block.
8841  *
8842  * NOTE: return value 1 means we should stop walking down.
8843  */
8844 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
8845                                  struct btrfs_root *root,
8846                                  struct btrfs_path *path,
8847                                  struct walk_control *wc, int *lookup_info)
8848 {
8849         u64 bytenr;
8850         u64 generation;
8851         u64 parent;
8852         u32 blocksize;
8853         struct btrfs_key key;
8854         struct extent_buffer *next;
8855         int level = wc->level;
8856         int reada = 0;
8857         int ret = 0;
8858         bool need_account = false;
8859
8860         generation = btrfs_node_ptr_generation(path->nodes[level],
8861                                                path->slots[level]);
8862         /*
8863          * if the lower level block was created before the snapshot
8864          * was created, we know there is no need to update back refs
8865          * for the subtree
8866          */
8867         if (wc->stage == UPDATE_BACKREF &&
8868             generation <= root->root_key.offset) {
8869                 *lookup_info = 1;
8870                 return 1;
8871         }
8872
8873         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
8874         blocksize = root->nodesize;
8875
8876         next = btrfs_find_tree_block(root->fs_info, bytenr);
8877         if (!next) {
8878                 next = btrfs_find_create_tree_block(root, bytenr);
8879                 if (IS_ERR(next))
8880                         return PTR_ERR(next);
8881
8882                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
8883                                                level - 1);
8884                 reada = 1;
8885         }
8886         btrfs_tree_lock(next);
8887         btrfs_set_lock_blocking(next);
8888
8889         ret = btrfs_lookup_extent_info(trans, root, bytenr, level - 1, 1,
8890                                        &wc->refs[level - 1],
8891                                        &wc->flags[level - 1]);
8892         if (ret < 0) {
8893                 btrfs_tree_unlock(next);
8894                 return ret;
8895         }
8896
8897         if (unlikely(wc->refs[level - 1] == 0)) {
8898                 btrfs_err(root->fs_info, "Missing references.");
8899                 BUG();
8900         }
8901         *lookup_info = 0;
8902
8903         if (wc->stage == DROP_REFERENCE) {
8904                 if (wc->refs[level - 1] > 1) {
8905                         need_account = true;
8906                         if (level == 1 &&
8907                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8908                                 goto skip;
8909
8910                         if (!wc->update_ref ||
8911                             generation <= root->root_key.offset)
8912                                 goto skip;
8913
8914                         btrfs_node_key_to_cpu(path->nodes[level], &key,
8915                                               path->slots[level]);
8916                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
8917                         if (ret < 0)
8918                                 goto skip;
8919
8920                         wc->stage = UPDATE_BACKREF;
8921                         wc->shared_level = level - 1;
8922                 }
8923         } else {
8924                 if (level == 1 &&
8925                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8926                         goto skip;
8927         }
8928
8929         if (!btrfs_buffer_uptodate(next, generation, 0)) {
8930                 btrfs_tree_unlock(next);
8931                 free_extent_buffer(next);
8932                 next = NULL;
8933                 *lookup_info = 1;
8934         }
8935
8936         if (!next) {
8937                 if (reada && level == 1)
8938                         reada_walk_down(trans, root, wc, path);
8939                 next = read_tree_block(root, bytenr, generation);
8940                 if (IS_ERR(next)) {
8941                         return PTR_ERR(next);
8942                 } else if (!extent_buffer_uptodate(next)) {
8943                         free_extent_buffer(next);
8944                         return -EIO;
8945                 }
8946                 btrfs_tree_lock(next);
8947                 btrfs_set_lock_blocking(next);
8948         }
8949
8950         level--;
8951         BUG_ON(level != btrfs_header_level(next));
8952         path->nodes[level] = next;
8953         path->slots[level] = 0;
8954         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8955         wc->level = level;
8956         if (wc->level == 1)
8957                 wc->reada_slot = 0;
8958         return 0;
8959 skip:
8960         wc->refs[level - 1] = 0;
8961         wc->flags[level - 1] = 0;
8962         if (wc->stage == DROP_REFERENCE) {
8963                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
8964                         parent = path->nodes[level]->start;
8965                 } else {
8966                         BUG_ON(root->root_key.objectid !=
8967                                btrfs_header_owner(path->nodes[level]));
8968                         parent = 0;
8969                 }
8970
8971                 if (need_account) {
8972                         ret = account_shared_subtree(trans, root, next,
8973                                                      generation, level - 1);
8974                         if (ret) {
8975                                 btrfs_err_rl(root->fs_info,
8976                                              "Error %d accounting shared subtree. Quota is out of sync, rescan required.",
8977                                              ret);
8978                         }
8979                 }
8980                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
8981                                 root->root_key.objectid, level - 1, 0);
8982                 BUG_ON(ret); /* -ENOMEM */
8983         }
8984         btrfs_tree_unlock(next);
8985         free_extent_buffer(next);
8986         *lookup_info = 1;
8987         return 1;
8988 }
8989
8990 /*
8991  * helper to process tree block while walking up the tree.
8992  *
8993  * when wc->stage == DROP_REFERENCE, this function drops
8994  * reference count on the block.
8995  *
8996  * when wc->stage == UPDATE_BACKREF, this function changes
8997  * wc->stage back to DROP_REFERENCE if we changed wc->stage
8998  * to UPDATE_BACKREF previously while processing the block.
8999  *
9000  * NOTE: return value 1 means we should stop walking up.
9001  */
9002 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
9003                                  struct btrfs_root *root,
9004                                  struct btrfs_path *path,
9005                                  struct walk_control *wc)
9006 {
9007         int ret;
9008         int level = wc->level;
9009         struct extent_buffer *eb = path->nodes[level];
9010         u64 parent = 0;
9011
9012         if (wc->stage == UPDATE_BACKREF) {
9013                 BUG_ON(wc->shared_level < level);
9014                 if (level < wc->shared_level)
9015                         goto out;
9016
9017                 ret = find_next_key(path, level + 1, &wc->update_progress);
9018                 if (ret > 0)
9019                         wc->update_ref = 0;
9020
9021                 wc->stage = DROP_REFERENCE;
9022                 wc->shared_level = -1;
9023                 path->slots[level] = 0;
9024
9025                 /*
9026                  * check reference count again if the block isn't locked.
9027                  * we should start walking down the tree again if reference
9028                  * count is one.
9029                  */
9030                 if (!path->locks[level]) {
9031                         BUG_ON(level == 0);
9032                         btrfs_tree_lock(eb);
9033                         btrfs_set_lock_blocking(eb);
9034                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9035
9036                         ret = btrfs_lookup_extent_info(trans, root,
9037                                                        eb->start, level, 1,
9038                                                        &wc->refs[level],
9039                                                        &wc->flags[level]);
9040                         if (ret < 0) {
9041                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
9042                                 path->locks[level] = 0;
9043                                 return ret;
9044                         }
9045                         BUG_ON(wc->refs[level] == 0);
9046                         if (wc->refs[level] == 1) {
9047                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
9048                                 path->locks[level] = 0;
9049                                 return 1;
9050                         }
9051                 }
9052         }
9053
9054         /* wc->stage == DROP_REFERENCE */
9055         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
9056
9057         if (wc->refs[level] == 1) {
9058                 if (level == 0) {
9059                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
9060                                 ret = btrfs_dec_ref(trans, root, eb, 1);
9061                         else
9062                                 ret = btrfs_dec_ref(trans, root, eb, 0);
9063                         BUG_ON(ret); /* -ENOMEM */
9064                         ret = account_leaf_items(trans, root, eb);
9065                         if (ret) {
9066                                 btrfs_err_rl(root->fs_info,
9067                                              "error %d accounting leaf items. Quota is out of sync, rescan required.",
9068                                              ret);
9069                         }
9070                 }
9071                 /* make block locked assertion in clean_tree_block happy */
9072                 if (!path->locks[level] &&
9073                     btrfs_header_generation(eb) == trans->transid) {
9074                         btrfs_tree_lock(eb);
9075                         btrfs_set_lock_blocking(eb);
9076                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9077                 }
9078                 clean_tree_block(trans, root->fs_info, eb);
9079         }
9080
9081         if (eb == root->node) {
9082                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
9083                         parent = eb->start;
9084                 else
9085                         BUG_ON(root->root_key.objectid !=
9086                                btrfs_header_owner(eb));
9087         } else {
9088                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
9089                         parent = path->nodes[level + 1]->start;
9090                 else
9091                         BUG_ON(root->root_key.objectid !=
9092                                btrfs_header_owner(path->nodes[level + 1]));
9093         }
9094
9095         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
9096 out:
9097         wc->refs[level] = 0;
9098         wc->flags[level] = 0;
9099         return 0;
9100 }
9101
9102 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
9103                                    struct btrfs_root *root,
9104                                    struct btrfs_path *path,
9105                                    struct walk_control *wc)
9106 {
9107         int level = wc->level;
9108         int lookup_info = 1;
9109         int ret;
9110
9111         while (level >= 0) {
9112                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
9113                 if (ret > 0)
9114                         break;
9115
9116                 if (level == 0)
9117                         break;
9118
9119                 if (path->slots[level] >=
9120                     btrfs_header_nritems(path->nodes[level]))
9121                         break;
9122
9123                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
9124                 if (ret > 0) {
9125                         path->slots[level]++;
9126                         continue;
9127                 } else if (ret < 0)
9128                         return ret;
9129                 level = wc->level;
9130         }
9131         return 0;
9132 }
9133
9134 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
9135                                  struct btrfs_root *root,
9136                                  struct btrfs_path *path,
9137                                  struct walk_control *wc, int max_level)
9138 {
9139         int level = wc->level;
9140         int ret;
9141
9142         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
9143         while (level < max_level && path->nodes[level]) {
9144                 wc->level = level;
9145                 if (path->slots[level] + 1 <
9146                     btrfs_header_nritems(path->nodes[level])) {
9147                         path->slots[level]++;
9148                         return 0;
9149                 } else {
9150                         ret = walk_up_proc(trans, root, path, wc);
9151                         if (ret > 0)
9152                                 return 0;
9153
9154                         if (path->locks[level]) {
9155                                 btrfs_tree_unlock_rw(path->nodes[level],
9156                                                      path->locks[level]);
9157                                 path->locks[level] = 0;
9158                         }
9159                         free_extent_buffer(path->nodes[level]);
9160                         path->nodes[level] = NULL;
9161                         level++;
9162                 }
9163         }
9164         return 1;
9165 }
9166
9167 /*
9168  * drop a subvolume tree.
9169  *
9170  * this function traverses the tree freeing any blocks that only
9171  * referenced by the tree.
9172  *
9173  * when a shared tree block is found. this function decreases its
9174  * reference count by one. if update_ref is true, this function
9175  * also make sure backrefs for the shared block and all lower level
9176  * blocks are properly updated.
9177  *
9178  * If called with for_reloc == 0, may exit early with -EAGAIN
9179  */
9180 int btrfs_drop_snapshot(struct btrfs_root *root,
9181                          struct btrfs_block_rsv *block_rsv, int update_ref,
9182                          int for_reloc)
9183 {
9184         struct btrfs_fs_info *fs_info = root->fs_info;
9185         struct btrfs_path *path;
9186         struct btrfs_trans_handle *trans;
9187         struct btrfs_root *tree_root = fs_info->tree_root;
9188         struct btrfs_root_item *root_item = &root->root_item;
9189         struct walk_control *wc;
9190         struct btrfs_key key;
9191         int err = 0;
9192         int ret;
9193         int level;
9194         bool root_dropped = false;
9195
9196         btrfs_debug(fs_info, "Drop subvolume %llu", root->objectid);
9197
9198         path = btrfs_alloc_path();
9199         if (!path) {
9200                 err = -ENOMEM;
9201                 goto out;
9202         }
9203
9204         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9205         if (!wc) {
9206                 btrfs_free_path(path);
9207                 err = -ENOMEM;
9208                 goto out;
9209         }
9210
9211         trans = btrfs_start_transaction(tree_root, 0);
9212         if (IS_ERR(trans)) {
9213                 err = PTR_ERR(trans);
9214                 goto out_free;
9215         }
9216
9217         if (block_rsv)
9218                 trans->block_rsv = block_rsv;
9219
9220         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
9221                 level = btrfs_header_level(root->node);
9222                 path->nodes[level] = btrfs_lock_root_node(root);
9223                 btrfs_set_lock_blocking(path->nodes[level]);
9224                 path->slots[level] = 0;
9225                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9226                 memset(&wc->update_progress, 0,
9227                        sizeof(wc->update_progress));
9228         } else {
9229                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
9230                 memcpy(&wc->update_progress, &key,
9231                        sizeof(wc->update_progress));
9232
9233                 level = root_item->drop_level;
9234                 BUG_ON(level == 0);
9235                 path->lowest_level = level;
9236                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
9237                 path->lowest_level = 0;
9238                 if (ret < 0) {
9239                         err = ret;
9240                         goto out_end_trans;
9241                 }
9242                 WARN_ON(ret > 0);
9243
9244                 /*
9245                  * unlock our path, this is safe because only this
9246                  * function is allowed to delete this snapshot
9247                  */
9248                 btrfs_unlock_up_safe(path, 0);
9249
9250                 level = btrfs_header_level(root->node);
9251                 while (1) {
9252                         btrfs_tree_lock(path->nodes[level]);
9253                         btrfs_set_lock_blocking(path->nodes[level]);
9254                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9255
9256                         ret = btrfs_lookup_extent_info(trans, root,
9257                                                 path->nodes[level]->start,
9258                                                 level, 1, &wc->refs[level],
9259                                                 &wc->flags[level]);
9260                         if (ret < 0) {
9261                                 err = ret;
9262                                 goto out_end_trans;
9263                         }
9264                         BUG_ON(wc->refs[level] == 0);
9265
9266                         if (level == root_item->drop_level)
9267                                 break;
9268
9269                         btrfs_tree_unlock(path->nodes[level]);
9270                         path->locks[level] = 0;
9271                         WARN_ON(wc->refs[level] != 1);
9272                         level--;
9273                 }
9274         }
9275
9276         wc->level = level;
9277         wc->shared_level = -1;
9278         wc->stage = DROP_REFERENCE;
9279         wc->update_ref = update_ref;
9280         wc->keep_locks = 0;
9281         wc->for_reloc = for_reloc;
9282         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
9283
9284         while (1) {
9285
9286                 ret = walk_down_tree(trans, root, path, wc);
9287                 if (ret < 0) {
9288                         err = ret;
9289                         break;
9290                 }
9291
9292                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
9293                 if (ret < 0) {
9294                         err = ret;
9295                         break;
9296                 }
9297
9298                 if (ret > 0) {
9299                         BUG_ON(wc->stage != DROP_REFERENCE);
9300                         break;
9301                 }
9302
9303                 if (wc->stage == DROP_REFERENCE) {
9304                         level = wc->level;
9305                         btrfs_node_key(path->nodes[level],
9306                                        &root_item->drop_progress,
9307                                        path->slots[level]);
9308                         root_item->drop_level = level;
9309                 }
9310
9311                 BUG_ON(wc->level == 0);
9312                 if (btrfs_should_end_transaction(trans, tree_root) ||
9313                     (!for_reloc && btrfs_need_cleaner_sleep(root))) {
9314                         ret = btrfs_update_root(trans, tree_root,
9315                                                 &root->root_key,
9316                                                 root_item);
9317                         if (ret) {
9318                                 btrfs_abort_transaction(trans, ret);
9319                                 err = ret;
9320                                 goto out_end_trans;
9321                         }
9322
9323                         btrfs_end_transaction_throttle(trans, tree_root);
9324                         if (!for_reloc && btrfs_need_cleaner_sleep(root)) {
9325                                 btrfs_debug(fs_info,
9326                                             "drop snapshot early exit");
9327                                 err = -EAGAIN;
9328                                 goto out_free;
9329                         }
9330
9331                         trans = btrfs_start_transaction(tree_root, 0);
9332                         if (IS_ERR(trans)) {
9333                                 err = PTR_ERR(trans);
9334                                 goto out_free;
9335                         }
9336                         if (block_rsv)
9337                                 trans->block_rsv = block_rsv;
9338                 }
9339         }
9340         btrfs_release_path(path);
9341         if (err)
9342                 goto out_end_trans;
9343
9344         ret = btrfs_del_root(trans, tree_root, &root->root_key);
9345         if (ret) {
9346                 btrfs_abort_transaction(trans, ret);
9347                 goto out_end_trans;
9348         }
9349
9350         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
9351                 ret = btrfs_find_root(tree_root, &root->root_key, path,
9352                                       NULL, NULL);
9353                 if (ret < 0) {
9354                         btrfs_abort_transaction(trans, ret);
9355                         err = ret;
9356                         goto out_end_trans;
9357                 } else if (ret > 0) {
9358                         /* if we fail to delete the orphan item this time
9359                          * around, it'll get picked up the next time.
9360                          *
9361                          * The most common failure here is just -ENOENT.
9362                          */
9363                         btrfs_del_orphan_item(trans, tree_root,
9364                                               root->root_key.objectid);
9365                 }
9366         }
9367
9368         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
9369                 btrfs_add_dropped_root(trans, root);
9370         } else {
9371                 free_extent_buffer(root->node);
9372                 free_extent_buffer(root->commit_root);
9373                 btrfs_put_fs_root(root);
9374         }
9375         root_dropped = true;
9376 out_end_trans:
9377         btrfs_end_transaction_throttle(trans, tree_root);
9378 out_free:
9379         kfree(wc);
9380         btrfs_free_path(path);
9381 out:
9382         /*
9383          * So if we need to stop dropping the snapshot for whatever reason we
9384          * need to make sure to add it back to the dead root list so that we
9385          * keep trying to do the work later.  This also cleans up roots if we
9386          * don't have it in the radix (like when we recover after a power fail
9387          * or unmount) so we don't leak memory.
9388          */
9389         if (!for_reloc && root_dropped == false)
9390                 btrfs_add_dead_root(root);
9391         if (err && err != -EAGAIN)
9392                 btrfs_handle_fs_error(fs_info, err, NULL);
9393         return err;
9394 }
9395
9396 /*
9397  * drop subtree rooted at tree block 'node'.
9398  *
9399  * NOTE: this function will unlock and release tree block 'node'
9400  * only used by relocation code
9401  */
9402 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
9403                         struct btrfs_root *root,
9404                         struct extent_buffer *node,
9405                         struct extent_buffer *parent)
9406 {
9407         struct btrfs_path *path;
9408         struct walk_control *wc;
9409         int level;
9410         int parent_level;
9411         int ret = 0;
9412         int wret;
9413
9414         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
9415
9416         path = btrfs_alloc_path();
9417         if (!path)
9418                 return -ENOMEM;
9419
9420         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9421         if (!wc) {
9422                 btrfs_free_path(path);
9423                 return -ENOMEM;
9424         }
9425
9426         btrfs_assert_tree_locked(parent);
9427         parent_level = btrfs_header_level(parent);
9428         extent_buffer_get(parent);
9429         path->nodes[parent_level] = parent;
9430         path->slots[parent_level] = btrfs_header_nritems(parent);
9431
9432         btrfs_assert_tree_locked(node);
9433         level = btrfs_header_level(node);
9434         path->nodes[level] = node;
9435         path->slots[level] = 0;
9436         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9437
9438         wc->refs[parent_level] = 1;
9439         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
9440         wc->level = level;
9441         wc->shared_level = -1;
9442         wc->stage = DROP_REFERENCE;
9443         wc->update_ref = 0;
9444         wc->keep_locks = 1;
9445         wc->for_reloc = 1;
9446         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
9447
9448         while (1) {
9449                 wret = walk_down_tree(trans, root, path, wc);
9450                 if (wret < 0) {
9451                         ret = wret;
9452                         break;
9453                 }
9454
9455                 wret = walk_up_tree(trans, root, path, wc, parent_level);
9456                 if (wret < 0)
9457                         ret = wret;
9458                 if (wret != 0)
9459                         break;
9460         }
9461
9462         kfree(wc);
9463         btrfs_free_path(path);
9464         return ret;
9465 }
9466
9467 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
9468 {
9469         u64 num_devices;
9470         u64 stripped;
9471
9472         /*
9473          * if restripe for this chunk_type is on pick target profile and
9474          * return, otherwise do the usual balance
9475          */
9476         stripped = get_restripe_target(root->fs_info, flags);
9477         if (stripped)
9478                 return extended_to_chunk(stripped);
9479
9480         num_devices = root->fs_info->fs_devices->rw_devices;
9481
9482         stripped = BTRFS_BLOCK_GROUP_RAID0 |
9483                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
9484                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
9485
9486         if (num_devices == 1) {
9487                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9488                 stripped = flags & ~stripped;
9489
9490                 /* turn raid0 into single device chunks */
9491                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
9492                         return stripped;
9493
9494                 /* turn mirroring into duplication */
9495                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
9496                              BTRFS_BLOCK_GROUP_RAID10))
9497                         return stripped | BTRFS_BLOCK_GROUP_DUP;
9498         } else {
9499                 /* they already had raid on here, just return */
9500                 if (flags & stripped)
9501                         return flags;
9502
9503                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9504                 stripped = flags & ~stripped;
9505
9506                 /* switch duplicated blocks with raid1 */
9507                 if (flags & BTRFS_BLOCK_GROUP_DUP)
9508                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
9509
9510                 /* this is drive concat, leave it alone */
9511         }
9512
9513         return flags;
9514 }
9515
9516 static int inc_block_group_ro(struct btrfs_block_group_cache *cache, int force)
9517 {
9518         struct btrfs_space_info *sinfo = cache->space_info;
9519         u64 num_bytes;
9520         u64 min_allocable_bytes;
9521         int ret = -ENOSPC;
9522
9523         /*
9524          * We need some metadata space and system metadata space for
9525          * allocating chunks in some corner cases until we force to set
9526          * it to be readonly.
9527          */
9528         if ((sinfo->flags &
9529              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
9530             !force)
9531                 min_allocable_bytes = SZ_1M;
9532         else
9533                 min_allocable_bytes = 0;
9534
9535         spin_lock(&sinfo->lock);
9536         spin_lock(&cache->lock);
9537
9538         if (cache->ro) {
9539                 cache->ro++;
9540                 ret = 0;
9541                 goto out;
9542         }
9543
9544         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
9545                     cache->bytes_super - btrfs_block_group_used(&cache->item);
9546
9547         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
9548             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
9549             min_allocable_bytes <= sinfo->total_bytes) {
9550                 sinfo->bytes_readonly += num_bytes;
9551                 cache->ro++;
9552                 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
9553                 ret = 0;
9554         }
9555 out:
9556         spin_unlock(&cache->lock);
9557         spin_unlock(&sinfo->lock);
9558         return ret;
9559 }
9560
9561 int btrfs_inc_block_group_ro(struct btrfs_root *root,
9562                              struct btrfs_block_group_cache *cache)
9563
9564 {
9565         struct btrfs_trans_handle *trans;
9566         u64 alloc_flags;
9567         int ret;
9568
9569 again:
9570         trans = btrfs_join_transaction(root);
9571         if (IS_ERR(trans))
9572                 return PTR_ERR(trans);
9573
9574         /*
9575          * we're not allowed to set block groups readonly after the dirty
9576          * block groups cache has started writing.  If it already started,
9577          * back off and let this transaction commit
9578          */
9579         mutex_lock(&root->fs_info->ro_block_group_mutex);
9580         if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
9581                 u64 transid = trans->transid;
9582
9583                 mutex_unlock(&root->fs_info->ro_block_group_mutex);
9584                 btrfs_end_transaction(trans, root);
9585
9586                 ret = btrfs_wait_for_commit(root, transid);
9587                 if (ret)
9588                         return ret;
9589                 goto again;
9590         }
9591
9592         /*
9593          * if we are changing raid levels, try to allocate a corresponding
9594          * block group with the new raid level.
9595          */
9596         alloc_flags = update_block_group_flags(root, cache->flags);
9597         if (alloc_flags != cache->flags) {
9598                 ret = do_chunk_alloc(trans, root, alloc_flags,
9599                                      CHUNK_ALLOC_FORCE);
9600                 /*
9601                  * ENOSPC is allowed here, we may have enough space
9602                  * already allocated at the new raid level to
9603                  * carry on
9604                  */
9605                 if (ret == -ENOSPC)
9606                         ret = 0;
9607                 if (ret < 0)
9608                         goto out;
9609         }
9610
9611         ret = inc_block_group_ro(cache, 0);
9612         if (!ret)
9613                 goto out;
9614         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
9615         ret = do_chunk_alloc(trans, root, alloc_flags,
9616                              CHUNK_ALLOC_FORCE);
9617         if (ret < 0)
9618                 goto out;
9619         ret = inc_block_group_ro(cache, 0);
9620 out:
9621         if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
9622                 alloc_flags = update_block_group_flags(root, cache->flags);
9623                 lock_chunks(root->fs_info->chunk_root);
9624                 check_system_chunk(trans, root, alloc_flags);
9625                 unlock_chunks(root->fs_info->chunk_root);
9626         }
9627         mutex_unlock(&root->fs_info->ro_block_group_mutex);
9628
9629         btrfs_end_transaction(trans, root);
9630         return ret;
9631 }
9632
9633 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
9634                             struct btrfs_root *root, u64 type)
9635 {
9636         u64 alloc_flags = get_alloc_profile(root, type);
9637         return do_chunk_alloc(trans, root, alloc_flags,
9638                               CHUNK_ALLOC_FORCE);
9639 }
9640
9641 /*
9642  * helper to account the unused space of all the readonly block group in the
9643  * space_info. takes mirrors into account.
9644  */
9645 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
9646 {
9647         struct btrfs_block_group_cache *block_group;
9648         u64 free_bytes = 0;
9649         int factor;
9650
9651         /* It's df, we don't care if it's racy */
9652         if (list_empty(&sinfo->ro_bgs))
9653                 return 0;
9654
9655         spin_lock(&sinfo->lock);
9656         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
9657                 spin_lock(&block_group->lock);
9658
9659                 if (!block_group->ro) {
9660                         spin_unlock(&block_group->lock);
9661                         continue;
9662                 }
9663
9664                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
9665                                           BTRFS_BLOCK_GROUP_RAID10 |
9666                                           BTRFS_BLOCK_GROUP_DUP))
9667                         factor = 2;
9668                 else
9669                         factor = 1;
9670
9671                 free_bytes += (block_group->key.offset -
9672                                btrfs_block_group_used(&block_group->item)) *
9673                                factor;
9674
9675                 spin_unlock(&block_group->lock);
9676         }
9677         spin_unlock(&sinfo->lock);
9678
9679         return free_bytes;
9680 }
9681
9682 void btrfs_dec_block_group_ro(struct btrfs_root *root,
9683                               struct btrfs_block_group_cache *cache)
9684 {
9685         struct btrfs_space_info *sinfo = cache->space_info;
9686         u64 num_bytes;
9687
9688         BUG_ON(!cache->ro);
9689
9690         spin_lock(&sinfo->lock);
9691         spin_lock(&cache->lock);
9692         if (!--cache->ro) {
9693                 num_bytes = cache->key.offset - cache->reserved -
9694                             cache->pinned - cache->bytes_super -
9695                             btrfs_block_group_used(&cache->item);
9696                 sinfo->bytes_readonly -= num_bytes;
9697                 list_del_init(&cache->ro_list);
9698         }
9699         spin_unlock(&cache->lock);
9700         spin_unlock(&sinfo->lock);
9701 }
9702
9703 /*
9704  * checks to see if its even possible to relocate this block group.
9705  *
9706  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
9707  * ok to go ahead and try.
9708  */
9709 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
9710 {
9711         struct btrfs_block_group_cache *block_group;
9712         struct btrfs_space_info *space_info;
9713         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
9714         struct btrfs_device *device;
9715         struct btrfs_trans_handle *trans;
9716         u64 min_free;
9717         u64 dev_min = 1;
9718         u64 dev_nr = 0;
9719         u64 target;
9720         int debug;
9721         int index;
9722         int full = 0;
9723         int ret = 0;
9724
9725         debug = btrfs_test_opt(root->fs_info, ENOSPC_DEBUG);
9726
9727         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
9728
9729         /* odd, couldn't find the block group, leave it alone */
9730         if (!block_group) {
9731                 if (debug)
9732                         btrfs_warn(root->fs_info,
9733                                    "can't find block group for bytenr %llu",
9734                                    bytenr);
9735                 return -1;
9736         }
9737
9738         min_free = btrfs_block_group_used(&block_group->item);
9739
9740         /* no bytes used, we're good */
9741         if (!min_free)
9742                 goto out;
9743
9744         space_info = block_group->space_info;
9745         spin_lock(&space_info->lock);
9746
9747         full = space_info->full;
9748
9749         /*
9750          * if this is the last block group we have in this space, we can't
9751          * relocate it unless we're able to allocate a new chunk below.
9752          *
9753          * Otherwise, we need to make sure we have room in the space to handle
9754          * all of the extents from this block group.  If we can, we're good
9755          */
9756         if ((space_info->total_bytes != block_group->key.offset) &&
9757             (space_info->bytes_used + space_info->bytes_reserved +
9758              space_info->bytes_pinned + space_info->bytes_readonly +
9759              min_free < space_info->total_bytes)) {
9760                 spin_unlock(&space_info->lock);
9761                 goto out;
9762         }
9763         spin_unlock(&space_info->lock);
9764
9765         /*
9766          * ok we don't have enough space, but maybe we have free space on our
9767          * devices to allocate new chunks for relocation, so loop through our
9768          * alloc devices and guess if we have enough space.  if this block
9769          * group is going to be restriped, run checks against the target
9770          * profile instead of the current one.
9771          */
9772         ret = -1;
9773
9774         /*
9775          * index:
9776          *      0: raid10
9777          *      1: raid1
9778          *      2: dup
9779          *      3: raid0
9780          *      4: single
9781          */
9782         target = get_restripe_target(root->fs_info, block_group->flags);
9783         if (target) {
9784                 index = __get_raid_index(extended_to_chunk(target));
9785         } else {
9786                 /*
9787                  * this is just a balance, so if we were marked as full
9788                  * we know there is no space for a new chunk
9789                  */
9790                 if (full) {
9791                         if (debug)
9792                                 btrfs_warn(root->fs_info,
9793                                         "no space to alloc new chunk for block group %llu",
9794                                         block_group->key.objectid);
9795                         goto out;
9796                 }
9797
9798                 index = get_block_group_index(block_group);
9799         }
9800
9801         if (index == BTRFS_RAID_RAID10) {
9802                 dev_min = 4;
9803                 /* Divide by 2 */
9804                 min_free >>= 1;
9805         } else if (index == BTRFS_RAID_RAID1) {
9806                 dev_min = 2;
9807         } else if (index == BTRFS_RAID_DUP) {
9808                 /* Multiply by 2 */
9809                 min_free <<= 1;
9810         } else if (index == BTRFS_RAID_RAID0) {
9811                 dev_min = fs_devices->rw_devices;
9812                 min_free = div64_u64(min_free, dev_min);
9813         }
9814
9815         /* We need to do this so that we can look at pending chunks */
9816         trans = btrfs_join_transaction(root);
9817         if (IS_ERR(trans)) {
9818                 ret = PTR_ERR(trans);
9819                 goto out;
9820         }
9821
9822         mutex_lock(&root->fs_info->chunk_mutex);
9823         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
9824                 u64 dev_offset;
9825
9826                 /*
9827                  * check to make sure we can actually find a chunk with enough
9828                  * space to fit our block group in.
9829                  */
9830                 if (device->total_bytes > device->bytes_used + min_free &&
9831                     !device->is_tgtdev_for_dev_replace) {
9832                         ret = find_free_dev_extent(trans, device, min_free,
9833                                                    &dev_offset, NULL);
9834                         if (!ret)
9835                                 dev_nr++;
9836
9837                         if (dev_nr >= dev_min)
9838                                 break;
9839
9840                         ret = -1;
9841                 }
9842         }
9843         if (debug && ret == -1)
9844                 btrfs_warn(root->fs_info,
9845                         "no space to allocate a new chunk for block group %llu",
9846                         block_group->key.objectid);
9847         mutex_unlock(&root->fs_info->chunk_mutex);
9848         btrfs_end_transaction(trans, root);
9849 out:
9850         btrfs_put_block_group(block_group);
9851         return ret;
9852 }
9853
9854 static int find_first_block_group(struct btrfs_root *root,
9855                 struct btrfs_path *path, struct btrfs_key *key)
9856 {
9857         int ret = 0;
9858         struct btrfs_key found_key;
9859         struct extent_buffer *leaf;
9860         int slot;
9861
9862         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
9863         if (ret < 0)
9864                 goto out;
9865
9866         while (1) {
9867                 slot = path->slots[0];
9868                 leaf = path->nodes[0];
9869                 if (slot >= btrfs_header_nritems(leaf)) {
9870                         ret = btrfs_next_leaf(root, path);
9871                         if (ret == 0)
9872                                 continue;
9873                         if (ret < 0)
9874                                 goto out;
9875                         break;
9876                 }
9877                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
9878
9879                 if (found_key.objectid >= key->objectid &&
9880                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
9881                         struct extent_map_tree *em_tree;
9882                         struct extent_map *em;
9883
9884                         em_tree = &root->fs_info->mapping_tree.map_tree;
9885                         read_lock(&em_tree->lock);
9886                         em = lookup_extent_mapping(em_tree, found_key.objectid,
9887                                                    found_key.offset);
9888                         read_unlock(&em_tree->lock);
9889                         if (!em) {
9890                                 btrfs_err(root->fs_info,
9891                         "logical %llu len %llu found bg but no related chunk",
9892                                           found_key.objectid, found_key.offset);
9893                                 ret = -ENOENT;
9894                         } else {
9895                                 ret = 0;
9896                         }
9897                         free_extent_map(em);
9898                         goto out;
9899                 }
9900                 path->slots[0]++;
9901         }
9902 out:
9903         return ret;
9904 }
9905
9906 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
9907 {
9908         struct btrfs_block_group_cache *block_group;
9909         u64 last = 0;
9910
9911         while (1) {
9912                 struct inode *inode;
9913
9914                 block_group = btrfs_lookup_first_block_group(info, last);
9915                 while (block_group) {
9916                         spin_lock(&block_group->lock);
9917                         if (block_group->iref)
9918                                 break;
9919                         spin_unlock(&block_group->lock);
9920                         block_group = next_block_group(info->tree_root,
9921                                                        block_group);
9922                 }
9923                 if (!block_group) {
9924                         if (last == 0)
9925                                 break;
9926                         last = 0;
9927                         continue;
9928                 }
9929
9930                 inode = block_group->inode;
9931                 block_group->iref = 0;
9932                 block_group->inode = NULL;
9933                 spin_unlock(&block_group->lock);
9934                 ASSERT(block_group->io_ctl.inode == NULL);
9935                 iput(inode);
9936                 last = block_group->key.objectid + block_group->key.offset;
9937                 btrfs_put_block_group(block_group);
9938         }
9939 }
9940
9941 int btrfs_free_block_groups(struct btrfs_fs_info *info)
9942 {
9943         struct btrfs_block_group_cache *block_group;
9944         struct btrfs_space_info *space_info;
9945         struct btrfs_caching_control *caching_ctl;
9946         struct rb_node *n;
9947
9948         down_write(&info->commit_root_sem);
9949         while (!list_empty(&info->caching_block_groups)) {
9950                 caching_ctl = list_entry(info->caching_block_groups.next,
9951                                          struct btrfs_caching_control, list);
9952                 list_del(&caching_ctl->list);
9953                 put_caching_control(caching_ctl);
9954         }
9955         up_write(&info->commit_root_sem);
9956
9957         spin_lock(&info->unused_bgs_lock);
9958         while (!list_empty(&info->unused_bgs)) {
9959                 block_group = list_first_entry(&info->unused_bgs,
9960                                                struct btrfs_block_group_cache,
9961                                                bg_list);
9962                 list_del_init(&block_group->bg_list);
9963                 btrfs_put_block_group(block_group);
9964         }
9965         spin_unlock(&info->unused_bgs_lock);
9966
9967         spin_lock(&info->block_group_cache_lock);
9968         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
9969                 block_group = rb_entry(n, struct btrfs_block_group_cache,
9970                                        cache_node);
9971                 rb_erase(&block_group->cache_node,
9972                          &info->block_group_cache_tree);
9973                 RB_CLEAR_NODE(&block_group->cache_node);
9974                 spin_unlock(&info->block_group_cache_lock);
9975
9976                 down_write(&block_group->space_info->groups_sem);
9977                 list_del(&block_group->list);
9978                 up_write(&block_group->space_info->groups_sem);
9979
9980                 if (block_group->cached == BTRFS_CACHE_STARTED)
9981                         wait_block_group_cache_done(block_group);
9982
9983                 /*
9984                  * We haven't cached this block group, which means we could
9985                  * possibly have excluded extents on this block group.
9986                  */
9987                 if (block_group->cached == BTRFS_CACHE_NO ||
9988                     block_group->cached == BTRFS_CACHE_ERROR)
9989                         free_excluded_extents(info->extent_root, block_group);
9990
9991                 btrfs_remove_free_space_cache(block_group);
9992                 ASSERT(list_empty(&block_group->dirty_list));
9993                 ASSERT(list_empty(&block_group->io_list));
9994                 ASSERT(list_empty(&block_group->bg_list));
9995                 ASSERT(atomic_read(&block_group->count) == 1);
9996                 btrfs_put_block_group(block_group);
9997
9998                 spin_lock(&info->block_group_cache_lock);
9999         }
10000         spin_unlock(&info->block_group_cache_lock);
10001
10002         /* now that all the block groups are freed, go through and
10003          * free all the space_info structs.  This is only called during
10004          * the final stages of unmount, and so we know nobody is
10005          * using them.  We call synchronize_rcu() once before we start,
10006          * just to be on the safe side.
10007          */
10008         synchronize_rcu();
10009
10010         release_global_block_rsv(info);
10011
10012         while (!list_empty(&info->space_info)) {
10013                 int i;
10014
10015                 space_info = list_entry(info->space_info.next,
10016                                         struct btrfs_space_info,
10017                                         list);
10018
10019                 /*
10020                  * Do not hide this behind enospc_debug, this is actually
10021                  * important and indicates a real bug if this happens.
10022                  */
10023                 if (WARN_ON(space_info->bytes_pinned > 0 ||
10024                             space_info->bytes_reserved > 0 ||
10025                             space_info->bytes_may_use > 0))
10026                         dump_space_info(info, space_info, 0, 0);
10027                 list_del(&space_info->list);
10028                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
10029                         struct kobject *kobj;
10030                         kobj = space_info->block_group_kobjs[i];
10031                         space_info->block_group_kobjs[i] = NULL;
10032                         if (kobj) {
10033                                 kobject_del(kobj);
10034                                 kobject_put(kobj);
10035                         }
10036                 }
10037                 kobject_del(&space_info->kobj);
10038                 kobject_put(&space_info->kobj);
10039         }
10040         return 0;
10041 }
10042
10043 static void __link_block_group(struct btrfs_space_info *space_info,
10044                                struct btrfs_block_group_cache *cache)
10045 {
10046         int index = get_block_group_index(cache);
10047         bool first = false;
10048
10049         down_write(&space_info->groups_sem);
10050         if (list_empty(&space_info->block_groups[index]))
10051                 first = true;
10052         list_add_tail(&cache->list, &space_info->block_groups[index]);
10053         up_write(&space_info->groups_sem);
10054
10055         if (first) {
10056                 struct raid_kobject *rkobj;
10057                 int ret;
10058
10059                 rkobj = kzalloc(sizeof(*rkobj), GFP_NOFS);
10060                 if (!rkobj)
10061                         goto out_err;
10062                 rkobj->raid_type = index;
10063                 kobject_init(&rkobj->kobj, &btrfs_raid_ktype);
10064                 ret = kobject_add(&rkobj->kobj, &space_info->kobj,
10065                                   "%s", get_raid_name(index));
10066                 if (ret) {
10067                         kobject_put(&rkobj->kobj);
10068                         goto out_err;
10069                 }
10070                 space_info->block_group_kobjs[index] = &rkobj->kobj;
10071         }
10072
10073         return;
10074 out_err:
10075         btrfs_warn(cache->fs_info,
10076                    "failed to add kobject for block cache, ignoring");
10077 }
10078
10079 static struct btrfs_block_group_cache *
10080 btrfs_create_block_group_cache(struct btrfs_root *root, u64 start, u64 size)
10081 {
10082         struct btrfs_block_group_cache *cache;
10083
10084         cache = kzalloc(sizeof(*cache), GFP_NOFS);
10085         if (!cache)
10086                 return NULL;
10087
10088         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
10089                                         GFP_NOFS);
10090         if (!cache->free_space_ctl) {
10091                 kfree(cache);
10092                 return NULL;
10093         }
10094
10095         cache->key.objectid = start;
10096         cache->key.offset = size;
10097         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
10098
10099         cache->sectorsize = root->sectorsize;
10100         cache->fs_info = root->fs_info;
10101         cache->full_stripe_len = btrfs_full_stripe_len(root,
10102                                                &root->fs_info->mapping_tree,
10103                                                start);
10104         set_free_space_tree_thresholds(cache);
10105
10106         atomic_set(&cache->count, 1);
10107         spin_lock_init(&cache->lock);
10108         init_rwsem(&cache->data_rwsem);
10109         INIT_LIST_HEAD(&cache->list);
10110         INIT_LIST_HEAD(&cache->cluster_list);
10111         INIT_LIST_HEAD(&cache->bg_list);
10112         INIT_LIST_HEAD(&cache->ro_list);
10113         INIT_LIST_HEAD(&cache->dirty_list);
10114         INIT_LIST_HEAD(&cache->io_list);
10115         btrfs_init_free_space_ctl(cache);
10116         atomic_set(&cache->trimming, 0);
10117         mutex_init(&cache->free_space_lock);
10118
10119         return cache;
10120 }
10121
10122 int btrfs_read_block_groups(struct btrfs_root *root)
10123 {
10124         struct btrfs_path *path;
10125         int ret;
10126         struct btrfs_block_group_cache *cache;
10127         struct btrfs_fs_info *info = root->fs_info;
10128         struct btrfs_space_info *space_info;
10129         struct btrfs_key key;
10130         struct btrfs_key found_key;
10131         struct extent_buffer *leaf;
10132         int need_clear = 0;
10133         u64 cache_gen;
10134         u64 feature;
10135         int mixed;
10136
10137         feature = btrfs_super_incompat_flags(info->super_copy);
10138         mixed = !!(feature & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS);
10139
10140         root = info->extent_root;
10141         key.objectid = 0;
10142         key.offset = 0;
10143         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
10144         path = btrfs_alloc_path();
10145         if (!path)
10146                 return -ENOMEM;
10147         path->reada = READA_FORWARD;
10148
10149         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
10150         if (btrfs_test_opt(root->fs_info, SPACE_CACHE) &&
10151             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
10152                 need_clear = 1;
10153         if (btrfs_test_opt(root->fs_info, CLEAR_CACHE))
10154                 need_clear = 1;
10155
10156         while (1) {
10157                 ret = find_first_block_group(root, path, &key);
10158                 if (ret > 0)
10159                         break;
10160                 if (ret != 0)
10161                         goto error;
10162
10163                 leaf = path->nodes[0];
10164                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
10165
10166                 cache = btrfs_create_block_group_cache(root, found_key.objectid,
10167                                                        found_key.offset);
10168                 if (!cache) {
10169                         ret = -ENOMEM;
10170                         goto error;
10171                 }
10172
10173                 if (need_clear) {
10174                         /*
10175                          * When we mount with old space cache, we need to
10176                          * set BTRFS_DC_CLEAR and set dirty flag.
10177                          *
10178                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
10179                          *    truncate the old free space cache inode and
10180                          *    setup a new one.
10181                          * b) Setting 'dirty flag' makes sure that we flush
10182                          *    the new space cache info onto disk.
10183                          */
10184                         if (btrfs_test_opt(root->fs_info, SPACE_CACHE))
10185                                 cache->disk_cache_state = BTRFS_DC_CLEAR;
10186                 }
10187
10188                 read_extent_buffer(leaf, &cache->item,
10189                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
10190                                    sizeof(cache->item));
10191                 cache->flags = btrfs_block_group_flags(&cache->item);
10192                 if (!mixed &&
10193                     ((cache->flags & BTRFS_BLOCK_GROUP_METADATA) &&
10194                     (cache->flags & BTRFS_BLOCK_GROUP_DATA))) {
10195                         btrfs_err(info,
10196 "bg %llu is a mixed block group but filesystem hasn't enabled mixed block groups",
10197                                   cache->key.objectid);
10198                         ret = -EINVAL;
10199                         goto error;
10200                 }
10201
10202                 key.objectid = found_key.objectid + found_key.offset;
10203                 btrfs_release_path(path);
10204
10205                 /*
10206                  * We need to exclude the super stripes now so that the space
10207                  * info has super bytes accounted for, otherwise we'll think
10208                  * we have more space than we actually do.
10209                  */
10210                 ret = exclude_super_stripes(root, cache);
10211                 if (ret) {
10212                         /*
10213                          * We may have excluded something, so call this just in
10214                          * case.
10215                          */
10216                         free_excluded_extents(root, cache);
10217                         btrfs_put_block_group(cache);
10218                         goto error;
10219                 }
10220
10221                 /*
10222                  * check for two cases, either we are full, and therefore
10223                  * don't need to bother with the caching work since we won't
10224                  * find any space, or we are empty, and we can just add all
10225                  * the space in and be done with it.  This saves us _alot_ of
10226                  * time, particularly in the full case.
10227                  */
10228                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
10229                         cache->last_byte_to_unpin = (u64)-1;
10230                         cache->cached = BTRFS_CACHE_FINISHED;
10231                         free_excluded_extents(root, cache);
10232                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10233                         cache->last_byte_to_unpin = (u64)-1;
10234                         cache->cached = BTRFS_CACHE_FINISHED;
10235                         add_new_free_space(cache, root->fs_info,
10236                                            found_key.objectid,
10237                                            found_key.objectid +
10238                                            found_key.offset);
10239                         free_excluded_extents(root, cache);
10240                 }
10241
10242                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
10243                 if (ret) {
10244                         btrfs_remove_free_space_cache(cache);
10245                         btrfs_put_block_group(cache);
10246                         goto error;
10247                 }
10248
10249                 trace_btrfs_add_block_group(root->fs_info, cache, 0);
10250                 ret = update_space_info(info, cache->flags, found_key.offset,
10251                                         btrfs_block_group_used(&cache->item),
10252                                         cache->bytes_super, &space_info);
10253                 if (ret) {
10254                         btrfs_remove_free_space_cache(cache);
10255                         spin_lock(&info->block_group_cache_lock);
10256                         rb_erase(&cache->cache_node,
10257                                  &info->block_group_cache_tree);
10258                         RB_CLEAR_NODE(&cache->cache_node);
10259                         spin_unlock(&info->block_group_cache_lock);
10260                         btrfs_put_block_group(cache);
10261                         goto error;
10262                 }
10263
10264                 cache->space_info = space_info;
10265
10266                 __link_block_group(space_info, cache);
10267
10268                 set_avail_alloc_bits(root->fs_info, cache->flags);
10269                 if (btrfs_chunk_readonly(root, cache->key.objectid)) {
10270                         inc_block_group_ro(cache, 1);
10271                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10272                         spin_lock(&info->unused_bgs_lock);
10273                         /* Should always be true but just in case. */
10274                         if (list_empty(&cache->bg_list)) {
10275                                 btrfs_get_block_group(cache);
10276                                 list_add_tail(&cache->bg_list,
10277                                               &info->unused_bgs);
10278                         }
10279                         spin_unlock(&info->unused_bgs_lock);
10280                 }
10281         }
10282
10283         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
10284                 if (!(get_alloc_profile(root, space_info->flags) &
10285                       (BTRFS_BLOCK_GROUP_RAID10 |
10286                        BTRFS_BLOCK_GROUP_RAID1 |
10287                        BTRFS_BLOCK_GROUP_RAID5 |
10288                        BTRFS_BLOCK_GROUP_RAID6 |
10289                        BTRFS_BLOCK_GROUP_DUP)))
10290                         continue;
10291                 /*
10292                  * avoid allocating from un-mirrored block group if there are
10293                  * mirrored block groups.
10294                  */
10295                 list_for_each_entry(cache,
10296                                 &space_info->block_groups[BTRFS_RAID_RAID0],
10297                                 list)
10298                         inc_block_group_ro(cache, 1);
10299                 list_for_each_entry(cache,
10300                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
10301                                 list)
10302                         inc_block_group_ro(cache, 1);
10303         }
10304
10305         init_global_block_rsv(info);
10306         ret = 0;
10307 error:
10308         btrfs_free_path(path);
10309         return ret;
10310 }
10311
10312 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans,
10313                                        struct btrfs_root *root)
10314 {
10315         struct btrfs_block_group_cache *block_group, *tmp;
10316         struct btrfs_root *extent_root = root->fs_info->extent_root;
10317         struct btrfs_block_group_item item;
10318         struct btrfs_key key;
10319         int ret = 0;
10320         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
10321
10322         trans->can_flush_pending_bgs = false;
10323         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) {
10324                 if (ret)
10325                         goto next;
10326
10327                 spin_lock(&block_group->lock);
10328                 memcpy(&item, &block_group->item, sizeof(item));
10329                 memcpy(&key, &block_group->key, sizeof(key));
10330                 spin_unlock(&block_group->lock);
10331
10332                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
10333                                         sizeof(item));
10334                 if (ret)
10335                         btrfs_abort_transaction(trans, ret);
10336                 ret = btrfs_finish_chunk_alloc(trans, extent_root,
10337                                                key.objectid, key.offset);
10338                 if (ret)
10339                         btrfs_abort_transaction(trans, ret);
10340                 add_block_group_free_space(trans, root->fs_info, block_group);
10341                 /* already aborted the transaction if it failed. */
10342 next:
10343                 list_del_init(&block_group->bg_list);
10344         }
10345         trans->can_flush_pending_bgs = can_flush_pending_bgs;
10346 }
10347
10348 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
10349                            struct btrfs_root *root, u64 bytes_used,
10350                            u64 type, u64 chunk_objectid, u64 chunk_offset,
10351                            u64 size)
10352 {
10353         int ret;
10354         struct btrfs_root *extent_root;
10355         struct btrfs_block_group_cache *cache;
10356         extent_root = root->fs_info->extent_root;
10357
10358         btrfs_set_log_full_commit(root->fs_info, trans);
10359
10360         cache = btrfs_create_block_group_cache(root, chunk_offset, size);
10361         if (!cache)
10362                 return -ENOMEM;
10363
10364         btrfs_set_block_group_used(&cache->item, bytes_used);
10365         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
10366         btrfs_set_block_group_flags(&cache->item, type);
10367
10368         cache->flags = type;
10369         cache->last_byte_to_unpin = (u64)-1;
10370         cache->cached = BTRFS_CACHE_FINISHED;
10371         cache->needs_free_space = 1;
10372         ret = exclude_super_stripes(root, cache);
10373         if (ret) {
10374                 /*
10375                  * We may have excluded something, so call this just in
10376                  * case.
10377                  */
10378                 free_excluded_extents(root, cache);
10379                 btrfs_put_block_group(cache);
10380                 return ret;
10381         }
10382
10383         add_new_free_space(cache, root->fs_info, chunk_offset,
10384                            chunk_offset + size);
10385
10386         free_excluded_extents(root, cache);
10387
10388 #ifdef CONFIG_BTRFS_DEBUG
10389         if (btrfs_should_fragment_free_space(root, cache)) {
10390                 u64 new_bytes_used = size - bytes_used;
10391
10392                 bytes_used += new_bytes_used >> 1;
10393                 fragment_free_space(root, cache);
10394         }
10395 #endif
10396         /*
10397          * Call to ensure the corresponding space_info object is created and
10398          * assigned to our block group, but don't update its counters just yet.
10399          * We want our bg to be added to the rbtree with its ->space_info set.
10400          */
10401         ret = update_space_info(root->fs_info, cache->flags, 0, 0, 0,
10402                                 &cache->space_info);
10403         if (ret) {
10404                 btrfs_remove_free_space_cache(cache);
10405                 btrfs_put_block_group(cache);
10406                 return ret;
10407         }
10408
10409         ret = btrfs_add_block_group_cache(root->fs_info, cache);
10410         if (ret) {
10411                 btrfs_remove_free_space_cache(cache);
10412                 btrfs_put_block_group(cache);
10413                 return ret;
10414         }
10415
10416         /*
10417          * Now that our block group has its ->space_info set and is inserted in
10418          * the rbtree, update the space info's counters.
10419          */
10420         trace_btrfs_add_block_group(root->fs_info, cache, 1);
10421         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
10422                                 cache->bytes_super, &cache->space_info);
10423         if (ret) {
10424                 btrfs_remove_free_space_cache(cache);
10425                 spin_lock(&root->fs_info->block_group_cache_lock);
10426                 rb_erase(&cache->cache_node,
10427                          &root->fs_info->block_group_cache_tree);
10428                 RB_CLEAR_NODE(&cache->cache_node);
10429                 spin_unlock(&root->fs_info->block_group_cache_lock);
10430                 btrfs_put_block_group(cache);
10431                 return ret;
10432         }
10433         update_global_block_rsv(root->fs_info);
10434
10435         __link_block_group(cache->space_info, cache);
10436
10437         list_add_tail(&cache->bg_list, &trans->new_bgs);
10438
10439         set_avail_alloc_bits(extent_root->fs_info, type);
10440         return 0;
10441 }
10442
10443 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
10444 {
10445         u64 extra_flags = chunk_to_extended(flags) &
10446                                 BTRFS_EXTENDED_PROFILE_MASK;
10447
10448         write_seqlock(&fs_info->profiles_lock);
10449         if (flags & BTRFS_BLOCK_GROUP_DATA)
10450                 fs_info->avail_data_alloc_bits &= ~extra_flags;
10451         if (flags & BTRFS_BLOCK_GROUP_METADATA)
10452                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
10453         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
10454                 fs_info->avail_system_alloc_bits &= ~extra_flags;
10455         write_sequnlock(&fs_info->profiles_lock);
10456 }
10457
10458 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
10459                              struct btrfs_root *root, u64 group_start,
10460                              struct extent_map *em)
10461 {
10462         struct btrfs_path *path;
10463         struct btrfs_block_group_cache *block_group;
10464         struct btrfs_free_cluster *cluster;
10465         struct btrfs_root *tree_root = root->fs_info->tree_root;
10466         struct btrfs_key key;
10467         struct inode *inode;
10468         struct kobject *kobj = NULL;
10469         int ret;
10470         int index;
10471         int factor;
10472         struct btrfs_caching_control *caching_ctl = NULL;
10473         bool remove_em;
10474
10475         root = root->fs_info->extent_root;
10476
10477         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
10478         BUG_ON(!block_group);
10479         BUG_ON(!block_group->ro);
10480
10481         /*
10482          * Free the reserved super bytes from this block group before
10483          * remove it.
10484          */
10485         free_excluded_extents(root, block_group);
10486
10487         memcpy(&key, &block_group->key, sizeof(key));
10488         index = get_block_group_index(block_group);
10489         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
10490                                   BTRFS_BLOCK_GROUP_RAID1 |
10491                                   BTRFS_BLOCK_GROUP_RAID10))
10492                 factor = 2;
10493         else
10494                 factor = 1;
10495
10496         /* make sure this block group isn't part of an allocation cluster */
10497         cluster = &root->fs_info->data_alloc_cluster;
10498         spin_lock(&cluster->refill_lock);
10499         btrfs_return_cluster_to_free_space(block_group, cluster);
10500         spin_unlock(&cluster->refill_lock);
10501
10502         /*
10503          * make sure this block group isn't part of a metadata
10504          * allocation cluster
10505          */
10506         cluster = &root->fs_info->meta_alloc_cluster;
10507         spin_lock(&cluster->refill_lock);
10508         btrfs_return_cluster_to_free_space(block_group, cluster);
10509         spin_unlock(&cluster->refill_lock);
10510
10511         path = btrfs_alloc_path();
10512         if (!path) {
10513                 ret = -ENOMEM;
10514                 goto out;
10515         }
10516
10517         /*
10518          * get the inode first so any iput calls done for the io_list
10519          * aren't the final iput (no unlinks allowed now)
10520          */
10521         inode = lookup_free_space_inode(tree_root, block_group, path);
10522
10523         mutex_lock(&trans->transaction->cache_write_mutex);
10524         /*
10525          * make sure our free spache cache IO is done before remove the
10526          * free space inode
10527          */
10528         spin_lock(&trans->transaction->dirty_bgs_lock);
10529         if (!list_empty(&block_group->io_list)) {
10530                 list_del_init(&block_group->io_list);
10531
10532                 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
10533
10534                 spin_unlock(&trans->transaction->dirty_bgs_lock);
10535                 btrfs_wait_cache_io(root, trans, block_group,
10536                                     &block_group->io_ctl, path,
10537                                     block_group->key.objectid);
10538                 btrfs_put_block_group(block_group);
10539                 spin_lock(&trans->transaction->dirty_bgs_lock);
10540         }
10541
10542         if (!list_empty(&block_group->dirty_list)) {
10543                 list_del_init(&block_group->dirty_list);
10544                 btrfs_put_block_group(block_group);
10545         }
10546         spin_unlock(&trans->transaction->dirty_bgs_lock);
10547         mutex_unlock(&trans->transaction->cache_write_mutex);
10548
10549         if (!IS_ERR(inode)) {
10550                 ret = btrfs_orphan_add(trans, inode);
10551                 if (ret) {
10552                         btrfs_add_delayed_iput(inode);
10553                         goto out;
10554                 }
10555                 clear_nlink(inode);
10556                 /* One for the block groups ref */
10557                 spin_lock(&block_group->lock);
10558                 if (block_group->iref) {
10559                         block_group->iref = 0;
10560                         block_group->inode = NULL;
10561                         spin_unlock(&block_group->lock);
10562                         iput(inode);
10563                 } else {
10564                         spin_unlock(&block_group->lock);
10565                 }
10566                 /* One for our lookup ref */
10567                 btrfs_add_delayed_iput(inode);
10568         }
10569
10570         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
10571         key.offset = block_group->key.objectid;
10572         key.type = 0;
10573
10574         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
10575         if (ret < 0)
10576                 goto out;
10577         if (ret > 0)
10578                 btrfs_release_path(path);
10579         if (ret == 0) {
10580                 ret = btrfs_del_item(trans, tree_root, path);
10581                 if (ret)
10582                         goto out;
10583                 btrfs_release_path(path);
10584         }
10585
10586         spin_lock(&root->fs_info->block_group_cache_lock);
10587         rb_erase(&block_group->cache_node,
10588                  &root->fs_info->block_group_cache_tree);
10589         RB_CLEAR_NODE(&block_group->cache_node);
10590
10591         if (root->fs_info->first_logical_byte == block_group->key.objectid)
10592                 root->fs_info->first_logical_byte = (u64)-1;
10593         spin_unlock(&root->fs_info->block_group_cache_lock);
10594
10595         down_write(&block_group->space_info->groups_sem);
10596         /*
10597          * we must use list_del_init so people can check to see if they
10598          * are still on the list after taking the semaphore
10599          */
10600         list_del_init(&block_group->list);
10601         if (list_empty(&block_group->space_info->block_groups[index])) {
10602                 kobj = block_group->space_info->block_group_kobjs[index];
10603                 block_group->space_info->block_group_kobjs[index] = NULL;
10604                 clear_avail_alloc_bits(root->fs_info, block_group->flags);
10605         }
10606         up_write(&block_group->space_info->groups_sem);
10607         if (kobj) {
10608                 kobject_del(kobj);
10609                 kobject_put(kobj);
10610         }
10611
10612         if (block_group->has_caching_ctl)
10613                 caching_ctl = get_caching_control(block_group);
10614         if (block_group->cached == BTRFS_CACHE_STARTED)
10615                 wait_block_group_cache_done(block_group);
10616         if (block_group->has_caching_ctl) {
10617                 down_write(&root->fs_info->commit_root_sem);
10618                 if (!caching_ctl) {
10619                         struct btrfs_caching_control *ctl;
10620
10621                         list_for_each_entry(ctl,
10622                                     &root->fs_info->caching_block_groups, list)
10623                                 if (ctl->block_group == block_group) {
10624                                         caching_ctl = ctl;
10625                                         atomic_inc(&caching_ctl->count);
10626                                         break;
10627                                 }
10628                 }
10629                 if (caching_ctl)
10630                         list_del_init(&caching_ctl->list);
10631                 up_write(&root->fs_info->commit_root_sem);
10632                 if (caching_ctl) {
10633                         /* Once for the caching bgs list and once for us. */
10634                         put_caching_control(caching_ctl);
10635                         put_caching_control(caching_ctl);
10636                 }
10637         }
10638
10639         spin_lock(&trans->transaction->dirty_bgs_lock);
10640         if (!list_empty(&block_group->dirty_list)) {
10641                 WARN_ON(1);
10642         }
10643         if (!list_empty(&block_group->io_list)) {
10644                 WARN_ON(1);
10645         }
10646         spin_unlock(&trans->transaction->dirty_bgs_lock);
10647         btrfs_remove_free_space_cache(block_group);
10648
10649         spin_lock(&block_group->space_info->lock);
10650         list_del_init(&block_group->ro_list);
10651
10652         if (btrfs_test_opt(root->fs_info, ENOSPC_DEBUG)) {
10653                 WARN_ON(block_group->space_info->total_bytes
10654                         < block_group->key.offset);
10655                 WARN_ON(block_group->space_info->bytes_readonly
10656                         < block_group->key.offset);
10657                 WARN_ON(block_group->space_info->disk_total
10658                         < block_group->key.offset * factor);
10659         }
10660         block_group->space_info->total_bytes -= block_group->key.offset;
10661         block_group->space_info->bytes_readonly -= block_group->key.offset;
10662         block_group->space_info->disk_total -= block_group->key.offset * factor;
10663
10664         spin_unlock(&block_group->space_info->lock);
10665
10666         memcpy(&key, &block_group->key, sizeof(key));
10667
10668         lock_chunks(root);
10669         if (!list_empty(&em->list)) {
10670                 /* We're in the transaction->pending_chunks list. */
10671                 free_extent_map(em);
10672         }
10673         spin_lock(&block_group->lock);
10674         block_group->removed = 1;
10675         /*
10676          * At this point trimming can't start on this block group, because we
10677          * removed the block group from the tree fs_info->block_group_cache_tree
10678          * so no one can't find it anymore and even if someone already got this
10679          * block group before we removed it from the rbtree, they have already
10680          * incremented block_group->trimming - if they didn't, they won't find
10681          * any free space entries because we already removed them all when we
10682          * called btrfs_remove_free_space_cache().
10683          *
10684          * And we must not remove the extent map from the fs_info->mapping_tree
10685          * to prevent the same logical address range and physical device space
10686          * ranges from being reused for a new block group. This is because our
10687          * fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
10688          * completely transactionless, so while it is trimming a range the
10689          * currently running transaction might finish and a new one start,
10690          * allowing for new block groups to be created that can reuse the same
10691          * physical device locations unless we take this special care.
10692          *
10693          * There may also be an implicit trim operation if the file system
10694          * is mounted with -odiscard. The same protections must remain
10695          * in place until the extents have been discarded completely when
10696          * the transaction commit has completed.
10697          */
10698         remove_em = (atomic_read(&block_group->trimming) == 0);
10699         /*
10700          * Make sure a trimmer task always sees the em in the pinned_chunks list
10701          * if it sees block_group->removed == 1 (needs to lock block_group->lock
10702          * before checking block_group->removed).
10703          */
10704         if (!remove_em) {
10705                 /*
10706                  * Our em might be in trans->transaction->pending_chunks which
10707                  * is protected by fs_info->chunk_mutex ([lock|unlock]_chunks),
10708                  * and so is the fs_info->pinned_chunks list.
10709                  *
10710                  * So at this point we must be holding the chunk_mutex to avoid
10711                  * any races with chunk allocation (more specifically at
10712                  * volumes.c:contains_pending_extent()), to ensure it always
10713                  * sees the em, either in the pending_chunks list or in the
10714                  * pinned_chunks list.
10715                  */
10716                 list_move_tail(&em->list, &root->fs_info->pinned_chunks);
10717         }
10718         spin_unlock(&block_group->lock);
10719
10720         if (remove_em) {
10721                 struct extent_map_tree *em_tree;
10722
10723                 em_tree = &root->fs_info->mapping_tree.map_tree;
10724                 write_lock(&em_tree->lock);
10725                 /*
10726                  * The em might be in the pending_chunks list, so make sure the
10727                  * chunk mutex is locked, since remove_extent_mapping() will
10728                  * delete us from that list.
10729                  */
10730                 remove_extent_mapping(em_tree, em);
10731                 write_unlock(&em_tree->lock);
10732                 /* once for the tree */
10733                 free_extent_map(em);
10734         }
10735
10736         unlock_chunks(root);
10737
10738         ret = remove_block_group_free_space(trans, root->fs_info, block_group);
10739         if (ret)
10740                 goto out;
10741
10742         btrfs_put_block_group(block_group);
10743         btrfs_put_block_group(block_group);
10744
10745         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
10746         if (ret > 0)
10747                 ret = -EIO;
10748         if (ret < 0)
10749                 goto out;
10750
10751         ret = btrfs_del_item(trans, root, path);
10752 out:
10753         btrfs_free_path(path);
10754         return ret;
10755 }
10756
10757 struct btrfs_trans_handle *
10758 btrfs_start_trans_remove_block_group(struct btrfs_fs_info *fs_info,
10759                                      const u64 chunk_offset)
10760 {
10761         struct extent_map_tree *em_tree = &fs_info->mapping_tree.map_tree;
10762         struct extent_map *em;
10763         struct map_lookup *map;
10764         unsigned int num_items;
10765
10766         read_lock(&em_tree->lock);
10767         em = lookup_extent_mapping(em_tree, chunk_offset, 1);
10768         read_unlock(&em_tree->lock);
10769         ASSERT(em && em->start == chunk_offset);
10770
10771         /*
10772          * We need to reserve 3 + N units from the metadata space info in order
10773          * to remove a block group (done at btrfs_remove_chunk() and at
10774          * btrfs_remove_block_group()), which are used for:
10775          *
10776          * 1 unit for adding the free space inode's orphan (located in the tree
10777          * of tree roots).
10778          * 1 unit for deleting the block group item (located in the extent
10779          * tree).
10780          * 1 unit for deleting the free space item (located in tree of tree
10781          * roots).
10782          * N units for deleting N device extent items corresponding to each
10783          * stripe (located in the device tree).
10784          *
10785          * In order to remove a block group we also need to reserve units in the
10786          * system space info in order to update the chunk tree (update one or
10787          * more device items and remove one chunk item), but this is done at
10788          * btrfs_remove_chunk() through a call to check_system_chunk().
10789          */
10790         map = em->map_lookup;
10791         num_items = 3 + map->num_stripes;
10792         free_extent_map(em);
10793
10794         return btrfs_start_transaction_fallback_global_rsv(fs_info->extent_root,
10795                                                            num_items, 1);
10796 }
10797
10798 /*
10799  * Process the unused_bgs list and remove any that don't have any allocated
10800  * space inside of them.
10801  */
10802 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
10803 {
10804         struct btrfs_block_group_cache *block_group;
10805         struct btrfs_space_info *space_info;
10806         struct btrfs_root *root = fs_info->extent_root;
10807         struct btrfs_trans_handle *trans;
10808         int ret = 0;
10809
10810         if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
10811                 return;
10812
10813         spin_lock(&fs_info->unused_bgs_lock);
10814         while (!list_empty(&fs_info->unused_bgs)) {
10815                 u64 start, end;
10816                 int trimming;
10817
10818                 block_group = list_first_entry(&fs_info->unused_bgs,
10819                                                struct btrfs_block_group_cache,
10820                                                bg_list);
10821                 list_del_init(&block_group->bg_list);
10822
10823                 space_info = block_group->space_info;
10824
10825                 if (ret || btrfs_mixed_space_info(space_info)) {
10826                         btrfs_put_block_group(block_group);
10827                         continue;
10828                 }
10829                 spin_unlock(&fs_info->unused_bgs_lock);
10830
10831                 mutex_lock(&fs_info->delete_unused_bgs_mutex);
10832
10833                 /* Don't want to race with allocators so take the groups_sem */
10834                 down_write(&space_info->groups_sem);
10835                 spin_lock(&block_group->lock);
10836                 if (block_group->reserved ||
10837                     btrfs_block_group_used(&block_group->item) ||
10838                     (block_group->ro && !block_group->removed) ||
10839                     list_is_singular(&block_group->list)) {
10840                         /*
10841                          * We want to bail if we made new allocations or have
10842                          * outstanding allocations in this block group.  We do
10843                          * the ro check in case balance is currently acting on
10844                          * this block group.
10845                          */
10846                         spin_unlock(&block_group->lock);
10847                         up_write(&space_info->groups_sem);
10848                         goto next;
10849                 }
10850                 spin_unlock(&block_group->lock);
10851
10852                 /* We don't want to force the issue, only flip if it's ok. */
10853                 ret = inc_block_group_ro(block_group, 0);
10854                 up_write(&space_info->groups_sem);
10855                 if (ret < 0) {
10856                         ret = 0;
10857                         goto next;
10858                 }
10859
10860                 /*
10861                  * Want to do this before we do anything else so we can recover
10862                  * properly if we fail to join the transaction.
10863                  */
10864                 trans = btrfs_start_trans_remove_block_group(fs_info,
10865                                                      block_group->key.objectid);
10866                 if (IS_ERR(trans)) {
10867                         btrfs_dec_block_group_ro(root, block_group);
10868                         ret = PTR_ERR(trans);
10869                         goto next;
10870                 }
10871
10872                 /*
10873                  * We could have pending pinned extents for this block group,
10874                  * just delete them, we don't care about them anymore.
10875                  */
10876                 start = block_group->key.objectid;
10877                 end = start + block_group->key.offset - 1;
10878                 /*
10879                  * Hold the unused_bg_unpin_mutex lock to avoid racing with
10880                  * btrfs_finish_extent_commit(). If we are at transaction N,
10881                  * another task might be running finish_extent_commit() for the
10882                  * previous transaction N - 1, and have seen a range belonging
10883                  * to the block group in freed_extents[] before we were able to
10884                  * clear the whole block group range from freed_extents[]. This
10885                  * means that task can lookup for the block group after we
10886                  * unpinned it from freed_extents[] and removed it, leading to
10887                  * a BUG_ON() at btrfs_unpin_extent_range().
10888                  */
10889                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
10890                 ret = clear_extent_bits(&fs_info->freed_extents[0], start, end,
10891                                   EXTENT_DIRTY);
10892                 if (ret) {
10893                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10894                         btrfs_dec_block_group_ro(root, block_group);
10895                         goto end_trans;
10896                 }
10897                 ret = clear_extent_bits(&fs_info->freed_extents[1], start, end,
10898                                   EXTENT_DIRTY);
10899                 if (ret) {
10900                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10901                         btrfs_dec_block_group_ro(root, block_group);
10902                         goto end_trans;
10903                 }
10904                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10905
10906                 /* Reset pinned so btrfs_put_block_group doesn't complain */
10907                 spin_lock(&space_info->lock);
10908                 spin_lock(&block_group->lock);
10909
10910                 space_info->bytes_pinned -= block_group->pinned;
10911                 space_info->bytes_readonly += block_group->pinned;
10912                 percpu_counter_add(&space_info->total_bytes_pinned,
10913                                    -block_group->pinned);
10914                 block_group->pinned = 0;
10915
10916                 spin_unlock(&block_group->lock);
10917                 spin_unlock(&space_info->lock);
10918
10919                 /* DISCARD can flip during remount */
10920                 trimming = btrfs_test_opt(root->fs_info, DISCARD);
10921
10922                 /* Implicit trim during transaction commit. */
10923                 if (trimming)
10924                         btrfs_get_block_group_trimming(block_group);
10925
10926                 /*
10927                  * Btrfs_remove_chunk will abort the transaction if things go
10928                  * horribly wrong.
10929                  */
10930                 ret = btrfs_remove_chunk(trans, root,
10931                                          block_group->key.objectid);
10932
10933                 if (ret) {
10934                         if (trimming)
10935                                 btrfs_put_block_group_trimming(block_group);
10936                         goto end_trans;
10937                 }
10938
10939                 /*
10940                  * If we're not mounted with -odiscard, we can just forget
10941                  * about this block group. Otherwise we'll need to wait
10942                  * until transaction commit to do the actual discard.
10943                  */
10944                 if (trimming) {
10945                         spin_lock(&fs_info->unused_bgs_lock);
10946                         /*
10947                          * A concurrent scrub might have added us to the list
10948                          * fs_info->unused_bgs, so use a list_move operation
10949                          * to add the block group to the deleted_bgs list.
10950                          */
10951                         list_move(&block_group->bg_list,
10952                                   &trans->transaction->deleted_bgs);
10953                         spin_unlock(&fs_info->unused_bgs_lock);
10954                         btrfs_get_block_group(block_group);
10955                 }
10956 end_trans:
10957                 btrfs_end_transaction(trans, root);
10958 next:
10959                 mutex_unlock(&fs_info->delete_unused_bgs_mutex);
10960                 btrfs_put_block_group(block_group);
10961                 spin_lock(&fs_info->unused_bgs_lock);
10962         }
10963         spin_unlock(&fs_info->unused_bgs_lock);
10964 }
10965
10966 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
10967 {
10968         struct btrfs_space_info *space_info;
10969         struct btrfs_super_block *disk_super;
10970         u64 features;
10971         u64 flags;
10972         int mixed = 0;
10973         int ret;
10974
10975         disk_super = fs_info->super_copy;
10976         if (!btrfs_super_root(disk_super))
10977                 return -EINVAL;
10978
10979         features = btrfs_super_incompat_flags(disk_super);
10980         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
10981                 mixed = 1;
10982
10983         flags = BTRFS_BLOCK_GROUP_SYSTEM;
10984         ret = update_space_info(fs_info, flags, 0, 0, 0, &space_info);
10985         if (ret)
10986                 goto out;
10987
10988         if (mixed) {
10989                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
10990                 ret = update_space_info(fs_info, flags, 0, 0, 0, &space_info);
10991         } else {
10992                 flags = BTRFS_BLOCK_GROUP_METADATA;
10993                 ret = update_space_info(fs_info, flags, 0, 0, 0, &space_info);
10994                 if (ret)
10995                         goto out;
10996
10997                 flags = BTRFS_BLOCK_GROUP_DATA;
10998                 ret = update_space_info(fs_info, flags, 0, 0, 0, &space_info);
10999         }
11000 out:
11001         return ret;
11002 }
11003
11004 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
11005 {
11006         return unpin_extent_range(root, start, end, false);
11007 }
11008
11009 /*
11010  * It used to be that old block groups would be left around forever.
11011  * Iterating over them would be enough to trim unused space.  Since we
11012  * now automatically remove them, we also need to iterate over unallocated
11013  * space.
11014  *
11015  * We don't want a transaction for this since the discard may take a
11016  * substantial amount of time.  We don't require that a transaction be
11017  * running, but we do need to take a running transaction into account
11018  * to ensure that we're not discarding chunks that were released in
11019  * the current transaction.
11020  *
11021  * Holding the chunks lock will prevent other threads from allocating
11022  * or releasing chunks, but it won't prevent a running transaction
11023  * from committing and releasing the memory that the pending chunks
11024  * list head uses.  For that, we need to take a reference to the
11025  * transaction.
11026  */
11027 static int btrfs_trim_free_extents(struct btrfs_device *device,
11028                                    u64 minlen, u64 *trimmed)
11029 {
11030         u64 start = 0, len = 0;
11031         int ret;
11032
11033         *trimmed = 0;
11034
11035         /* Not writeable = nothing to do. */
11036         if (!device->writeable)
11037                 return 0;
11038
11039         /* No free space = nothing to do. */
11040         if (device->total_bytes <= device->bytes_used)
11041                 return 0;
11042
11043         ret = 0;
11044
11045         while (1) {
11046                 struct btrfs_fs_info *fs_info = device->dev_root->fs_info;
11047                 struct btrfs_transaction *trans;
11048                 u64 bytes;
11049
11050                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
11051                 if (ret)
11052                         return ret;
11053
11054                 down_read(&fs_info->commit_root_sem);
11055
11056                 spin_lock(&fs_info->trans_lock);
11057                 trans = fs_info->running_transaction;
11058                 if (trans)
11059                         atomic_inc(&trans->use_count);
11060                 spin_unlock(&fs_info->trans_lock);
11061
11062                 ret = find_free_dev_extent_start(trans, device, minlen, start,
11063                                                  &start, &len);
11064                 if (trans)
11065                         btrfs_put_transaction(trans);
11066
11067                 if (ret) {
11068                         up_read(&fs_info->commit_root_sem);
11069                         mutex_unlock(&fs_info->chunk_mutex);
11070                         if (ret == -ENOSPC)
11071                                 ret = 0;
11072                         break;
11073                 }
11074
11075                 ret = btrfs_issue_discard(device->bdev, start, len, &bytes);
11076                 up_read(&fs_info->commit_root_sem);
11077                 mutex_unlock(&fs_info->chunk_mutex);
11078
11079                 if (ret)
11080                         break;
11081
11082                 start += len;
11083                 *trimmed += bytes;
11084
11085                 if (fatal_signal_pending(current)) {
11086                         ret = -ERESTARTSYS;
11087                         break;
11088                 }
11089
11090                 cond_resched();
11091         }
11092
11093         return ret;
11094 }
11095
11096 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
11097 {
11098         struct btrfs_fs_info *fs_info = root->fs_info;
11099         struct btrfs_block_group_cache *cache = NULL;
11100         struct btrfs_device *device;
11101         struct list_head *devices;
11102         u64 group_trimmed;
11103         u64 start;
11104         u64 end;
11105         u64 trimmed = 0;
11106         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
11107         int ret = 0;
11108
11109         /*
11110          * try to trim all FS space, our block group may start from non-zero.
11111          */
11112         if (range->len == total_bytes)
11113                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
11114         else
11115                 cache = btrfs_lookup_block_group(fs_info, range->start);
11116
11117         while (cache) {
11118                 if (cache->key.objectid >= (range->start + range->len)) {
11119                         btrfs_put_block_group(cache);
11120                         break;
11121                 }
11122
11123                 start = max(range->start, cache->key.objectid);
11124                 end = min(range->start + range->len,
11125                                 cache->key.objectid + cache->key.offset);
11126
11127                 if (end - start >= range->minlen) {
11128                         if (!block_group_cache_done(cache)) {
11129                                 ret = cache_block_group(cache, 0);
11130                                 if (ret) {
11131                                         btrfs_put_block_group(cache);
11132                                         break;
11133                                 }
11134                                 ret = wait_block_group_cache_done(cache);
11135                                 if (ret) {
11136                                         btrfs_put_block_group(cache);
11137                                         break;
11138                                 }
11139                         }
11140                         ret = btrfs_trim_block_group(cache,
11141                                                      &group_trimmed,
11142                                                      start,
11143                                                      end,
11144                                                      range->minlen);
11145
11146                         trimmed += group_trimmed;
11147                         if (ret) {
11148                                 btrfs_put_block_group(cache);
11149                                 break;
11150                         }
11151                 }
11152
11153                 cache = next_block_group(fs_info->tree_root, cache);
11154         }
11155
11156         mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
11157         devices = &root->fs_info->fs_devices->alloc_list;
11158         list_for_each_entry(device, devices, dev_alloc_list) {
11159                 ret = btrfs_trim_free_extents(device, range->minlen,
11160                                               &group_trimmed);
11161                 if (ret)
11162                         break;
11163
11164                 trimmed += group_trimmed;
11165         }
11166         mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
11167
11168         range->len = trimmed;
11169         return ret;
11170 }
11171
11172 /*
11173  * btrfs_{start,end}_write_no_snapshoting() are similar to
11174  * mnt_{want,drop}_write(), they are used to prevent some tasks from writing
11175  * data into the page cache through nocow before the subvolume is snapshoted,
11176  * but flush the data into disk after the snapshot creation, or to prevent
11177  * operations while snapshoting is ongoing and that cause the snapshot to be
11178  * inconsistent (writes followed by expanding truncates for example).
11179  */
11180 void btrfs_end_write_no_snapshoting(struct btrfs_root *root)
11181 {
11182         percpu_counter_dec(&root->subv_writers->counter);
11183         /*
11184          * Make sure counter is updated before we wake up waiters.
11185          */
11186         smp_mb();
11187         if (waitqueue_active(&root->subv_writers->wait))
11188                 wake_up(&root->subv_writers->wait);
11189 }
11190
11191 int btrfs_start_write_no_snapshoting(struct btrfs_root *root)
11192 {
11193         if (atomic_read(&root->will_be_snapshoted))
11194                 return 0;
11195
11196         percpu_counter_inc(&root->subv_writers->counter);
11197         /*
11198          * Make sure counter is updated before we check for snapshot creation.
11199          */
11200         smp_mb();
11201         if (atomic_read(&root->will_be_snapshoted)) {
11202                 btrfs_end_write_no_snapshoting(root);
11203                 return 0;
11204         }
11205         return 1;
11206 }
11207
11208 static int wait_snapshoting_atomic_t(atomic_t *a)
11209 {
11210         schedule();
11211         return 0;
11212 }
11213
11214 void btrfs_wait_for_snapshot_creation(struct btrfs_root *root)
11215 {
11216         while (true) {
11217                 int ret;
11218
11219                 ret = btrfs_start_write_no_snapshoting(root);
11220                 if (ret)
11221                         break;
11222                 wait_on_atomic_t(&root->will_be_snapshoted,
11223                                  wait_snapshoting_atomic_t,
11224                                  TASK_UNINTERRUPTIBLE);
11225         }
11226 }