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                         spin_lock(&delayed_refs->lock);
2541                         locked_ref->processing = 0;
2542                         delayed_refs->num_heads_ready++;
2543                         spin_unlock(&delayed_refs->lock);
2544                         btrfs_delayed_ref_unlock(locked_ref);
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                         spin_lock(&delayed_refs->lock);
2651                         locked_ref->processing = 0;
2652                         delayed_refs->num_heads_ready++;
2653                         spin_unlock(&delayed_refs->lock);
2654                         btrfs_delayed_ref_unlock(locked_ref);
2655                         btrfs_put_delayed_ref(ref);
2656                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d",
2657                                     ret);
2658                         return ret;
2659                 }
2660
2661                 /*
2662                  * If this node is a head, that means all the refs in this head
2663                  * have been dealt with, and we will pick the next head to deal
2664                  * with, so we must unlock the head and drop it from the cluster
2665                  * list before we release it.
2666                  */
2667                 if (btrfs_delayed_ref_is_head(ref)) {
2668                         if (locked_ref->is_data &&
2669                             locked_ref->total_ref_mod < 0) {
2670                                 spin_lock(&delayed_refs->lock);
2671                                 delayed_refs->pending_csums -= ref->num_bytes;
2672                                 spin_unlock(&delayed_refs->lock);
2673                         }
2674                         btrfs_delayed_ref_unlock(locked_ref);
2675                         locked_ref = NULL;
2676                 }
2677                 btrfs_put_delayed_ref(ref);
2678                 count++;
2679                 cond_resched();
2680         }
2681
2682         /*
2683          * We don't want to include ref heads since we can have empty ref heads
2684          * and those will drastically skew our runtime down since we just do
2685          * accounting, no actual extent tree updates.
2686          */
2687         if (actual_count > 0) {
2688                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2689                 u64 avg;
2690
2691                 /*
2692                  * We weigh the current average higher than our current runtime
2693                  * to avoid large swings in the average.
2694                  */
2695                 spin_lock(&delayed_refs->lock);
2696                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2697                 fs_info->avg_delayed_ref_runtime = avg >> 2;    /* div by 4 */
2698                 spin_unlock(&delayed_refs->lock);
2699         }
2700         return 0;
2701 }
2702
2703 #ifdef SCRAMBLE_DELAYED_REFS
2704 /*
2705  * Normally delayed refs get processed in ascending bytenr order. This
2706  * correlates in most cases to the order added. To expose dependencies on this
2707  * order, we start to process the tree in the middle instead of the beginning
2708  */
2709 static u64 find_middle(struct rb_root *root)
2710 {
2711         struct rb_node *n = root->rb_node;
2712         struct btrfs_delayed_ref_node *entry;
2713         int alt = 1;
2714         u64 middle;
2715         u64 first = 0, last = 0;
2716
2717         n = rb_first(root);
2718         if (n) {
2719                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2720                 first = entry->bytenr;
2721         }
2722         n = rb_last(root);
2723         if (n) {
2724                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2725                 last = entry->bytenr;
2726         }
2727         n = root->rb_node;
2728
2729         while (n) {
2730                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2731                 WARN_ON(!entry->in_tree);
2732
2733                 middle = entry->bytenr;
2734
2735                 if (alt)
2736                         n = n->rb_left;
2737                 else
2738                         n = n->rb_right;
2739
2740                 alt = 1 - alt;
2741         }
2742         return middle;
2743 }
2744 #endif
2745
2746 static inline u64 heads_to_leaves(struct btrfs_root *root, u64 heads)
2747 {
2748         u64 num_bytes;
2749
2750         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2751                              sizeof(struct btrfs_extent_inline_ref));
2752         if (!btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2753                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2754
2755         /*
2756          * We don't ever fill up leaves all the way so multiply by 2 just to be
2757          * closer to what we're really going to want to use.
2758          */
2759         return div_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(root));
2760 }
2761
2762 /*
2763  * Takes the number of bytes to be csumm'ed and figures out how many leaves it
2764  * would require to store the csums for that many bytes.
2765  */
2766 u64 btrfs_csum_bytes_to_leaves(struct btrfs_root *root, u64 csum_bytes)
2767 {
2768         u64 csum_size;
2769         u64 num_csums_per_leaf;
2770         u64 num_csums;
2771
2772         csum_size = BTRFS_MAX_ITEM_SIZE(root);
2773         num_csums_per_leaf = div64_u64(csum_size,
2774                         (u64)btrfs_super_csum_size(root->fs_info->super_copy));
2775         num_csums = div64_u64(csum_bytes, root->sectorsize);
2776         num_csums += num_csums_per_leaf - 1;
2777         num_csums = div64_u64(num_csums, num_csums_per_leaf);
2778         return num_csums;
2779 }
2780
2781 int btrfs_check_space_for_delayed_refs(struct btrfs_trans_handle *trans,
2782                                        struct btrfs_root *root)
2783 {
2784         struct btrfs_block_rsv *global_rsv;
2785         u64 num_heads = trans->transaction->delayed_refs.num_heads_ready;
2786         u64 csum_bytes = trans->transaction->delayed_refs.pending_csums;
2787         u64 num_dirty_bgs = trans->transaction->num_dirty_bgs;
2788         u64 num_bytes, num_dirty_bgs_bytes;
2789         int ret = 0;
2790
2791         num_bytes = btrfs_calc_trans_metadata_size(root, 1);
2792         num_heads = heads_to_leaves(root, num_heads);
2793         if (num_heads > 1)
2794                 num_bytes += (num_heads - 1) * root->nodesize;
2795         num_bytes <<= 1;
2796         num_bytes += btrfs_csum_bytes_to_leaves(root, csum_bytes) * root->nodesize;
2797         num_dirty_bgs_bytes = btrfs_calc_trans_metadata_size(root,
2798                                                              num_dirty_bgs);
2799         global_rsv = &root->fs_info->global_block_rsv;
2800
2801         /*
2802          * If we can't allocate any more chunks lets make sure we have _lots_ of
2803          * wiggle room since running delayed refs can create more delayed refs.
2804          */
2805         if (global_rsv->space_info->full) {
2806                 num_dirty_bgs_bytes <<= 1;
2807                 num_bytes <<= 1;
2808         }
2809
2810         spin_lock(&global_rsv->lock);
2811         if (global_rsv->reserved <= num_bytes + num_dirty_bgs_bytes)
2812                 ret = 1;
2813         spin_unlock(&global_rsv->lock);
2814         return ret;
2815 }
2816
2817 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans,
2818                                        struct btrfs_root *root)
2819 {
2820         struct btrfs_fs_info *fs_info = root->fs_info;
2821         u64 num_entries =
2822                 atomic_read(&trans->transaction->delayed_refs.num_entries);
2823         u64 avg_runtime;
2824         u64 val;
2825
2826         smp_mb();
2827         avg_runtime = fs_info->avg_delayed_ref_runtime;
2828         val = num_entries * avg_runtime;
2829         if (num_entries * avg_runtime >= NSEC_PER_SEC)
2830                 return 1;
2831         if (val >= NSEC_PER_SEC / 2)
2832                 return 2;
2833
2834         return btrfs_check_space_for_delayed_refs(trans, root);
2835 }
2836
2837 struct async_delayed_refs {
2838         struct btrfs_root *root;
2839         u64 transid;
2840         int count;
2841         int error;
2842         int sync;
2843         struct completion wait;
2844         struct btrfs_work work;
2845 };
2846
2847 static void delayed_ref_async_start(struct btrfs_work *work)
2848 {
2849         struct async_delayed_refs *async;
2850         struct btrfs_trans_handle *trans;
2851         int ret;
2852
2853         async = container_of(work, struct async_delayed_refs, work);
2854
2855         /* if the commit is already started, we don't need to wait here */
2856         if (btrfs_transaction_blocked(async->root->fs_info))
2857                 goto done;
2858
2859         trans = btrfs_join_transaction(async->root);
2860         if (IS_ERR(trans)) {
2861                 async->error = PTR_ERR(trans);
2862                 goto done;
2863         }
2864
2865         /*
2866          * trans->sync means that when we call end_transaction, we won't
2867          * wait on delayed refs
2868          */
2869         trans->sync = true;
2870
2871         /* Don't bother flushing if we got into a different transaction */
2872         if (trans->transid > async->transid)
2873                 goto end;
2874
2875         ret = btrfs_run_delayed_refs(trans, async->root, async->count);
2876         if (ret)
2877                 async->error = ret;
2878 end:
2879         ret = btrfs_end_transaction(trans, async->root);
2880         if (ret && !async->error)
2881                 async->error = ret;
2882 done:
2883         if (async->sync)
2884                 complete(&async->wait);
2885         else
2886                 kfree(async);
2887 }
2888
2889 int btrfs_async_run_delayed_refs(struct btrfs_root *root,
2890                                  unsigned long count, u64 transid, int wait)
2891 {
2892         struct async_delayed_refs *async;
2893         int ret;
2894
2895         async = kmalloc(sizeof(*async), GFP_NOFS);
2896         if (!async)
2897                 return -ENOMEM;
2898
2899         async->root = root->fs_info->tree_root;
2900         async->count = count;
2901         async->error = 0;
2902         async->transid = transid;
2903         if (wait)
2904                 async->sync = 1;
2905         else
2906                 async->sync = 0;
2907         init_completion(&async->wait);
2908
2909         btrfs_init_work(&async->work, btrfs_extent_refs_helper,
2910                         delayed_ref_async_start, NULL, NULL);
2911
2912         btrfs_queue_work(root->fs_info->extent_workers, &async->work);
2913
2914         if (wait) {
2915                 wait_for_completion(&async->wait);
2916                 ret = async->error;
2917                 kfree(async);
2918                 return ret;
2919         }
2920         return 0;
2921 }
2922
2923 /*
2924  * this starts processing the delayed reference count updates and
2925  * extent insertions we have queued up so far.  count can be
2926  * 0, which means to process everything in the tree at the start
2927  * of the run (but not newly added entries), or it can be some target
2928  * number you'd like to process.
2929  *
2930  * Returns 0 on success or if called with an aborted transaction
2931  * Returns <0 on error and aborts the transaction
2932  */
2933 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2934                            struct btrfs_root *root, unsigned long count)
2935 {
2936         struct rb_node *node;
2937         struct btrfs_delayed_ref_root *delayed_refs;
2938         struct btrfs_delayed_ref_head *head;
2939         int ret;
2940         int run_all = count == (unsigned long)-1;
2941         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
2942
2943         /* We'll clean this up in btrfs_cleanup_transaction */
2944         if (trans->aborted)
2945                 return 0;
2946
2947         if (test_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &root->fs_info->flags))
2948                 return 0;
2949
2950         if (root == root->fs_info->extent_root)
2951                 root = root->fs_info->tree_root;
2952
2953         delayed_refs = &trans->transaction->delayed_refs;
2954         if (count == 0)
2955                 count = atomic_read(&delayed_refs->num_entries) * 2;
2956
2957 again:
2958 #ifdef SCRAMBLE_DELAYED_REFS
2959         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2960 #endif
2961         trans->can_flush_pending_bgs = false;
2962         ret = __btrfs_run_delayed_refs(trans, root, count);
2963         if (ret < 0) {
2964                 btrfs_abort_transaction(trans, ret);
2965                 return ret;
2966         }
2967
2968         if (run_all) {
2969                 if (!list_empty(&trans->new_bgs))
2970                         btrfs_create_pending_block_groups(trans, root);
2971
2972                 spin_lock(&delayed_refs->lock);
2973                 node = rb_first(&delayed_refs->href_root);
2974                 if (!node) {
2975                         spin_unlock(&delayed_refs->lock);
2976                         goto out;
2977                 }
2978
2979                 while (node) {
2980                         head = rb_entry(node, struct btrfs_delayed_ref_head,
2981                                         href_node);
2982                         if (btrfs_delayed_ref_is_head(&head->node)) {
2983                                 struct btrfs_delayed_ref_node *ref;
2984
2985                                 ref = &head->node;
2986                                 atomic_inc(&ref->refs);
2987
2988                                 spin_unlock(&delayed_refs->lock);
2989                                 /*
2990                                  * Mutex was contended, block until it's
2991                                  * released and try again
2992                                  */
2993                                 mutex_lock(&head->mutex);
2994                                 mutex_unlock(&head->mutex);
2995
2996                                 btrfs_put_delayed_ref(ref);
2997                                 cond_resched();
2998                                 goto again;
2999                         } else {
3000                                 WARN_ON(1);
3001                         }
3002                         node = rb_next(node);
3003                 }
3004                 spin_unlock(&delayed_refs->lock);
3005                 cond_resched();
3006                 goto again;
3007         }
3008 out:
3009         assert_qgroups_uptodate(trans);
3010         trans->can_flush_pending_bgs = can_flush_pending_bgs;
3011         return 0;
3012 }
3013
3014 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
3015                                 struct btrfs_root *root,
3016                                 u64 bytenr, u64 num_bytes, u64 flags,
3017                                 int level, int is_data)
3018 {
3019         struct btrfs_delayed_extent_op *extent_op;
3020         int ret;
3021
3022         extent_op = btrfs_alloc_delayed_extent_op();
3023         if (!extent_op)
3024                 return -ENOMEM;
3025
3026         extent_op->flags_to_set = flags;
3027         extent_op->update_flags = true;
3028         extent_op->update_key = false;
3029         extent_op->is_data = is_data ? true : false;
3030         extent_op->level = level;
3031
3032         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
3033                                           num_bytes, extent_op);
3034         if (ret)
3035                 btrfs_free_delayed_extent_op(extent_op);
3036         return ret;
3037 }
3038
3039 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
3040                                       struct btrfs_root *root,
3041                                       struct btrfs_path *path,
3042                                       u64 objectid, u64 offset, u64 bytenr)
3043 {
3044         struct btrfs_delayed_ref_head *head;
3045         struct btrfs_delayed_ref_node *ref;
3046         struct btrfs_delayed_data_ref *data_ref;
3047         struct btrfs_delayed_ref_root *delayed_refs;
3048         int ret = 0;
3049
3050         delayed_refs = &trans->transaction->delayed_refs;
3051         spin_lock(&delayed_refs->lock);
3052         head = btrfs_find_delayed_ref_head(trans, bytenr);
3053         if (!head) {
3054                 spin_unlock(&delayed_refs->lock);
3055                 return 0;
3056         }
3057
3058         if (!mutex_trylock(&head->mutex)) {
3059                 atomic_inc(&head->node.refs);
3060                 spin_unlock(&delayed_refs->lock);
3061
3062                 btrfs_release_path(path);
3063
3064                 /*
3065                  * Mutex was contended, block until it's released and let
3066                  * caller try again
3067                  */
3068                 mutex_lock(&head->mutex);
3069                 mutex_unlock(&head->mutex);
3070                 btrfs_put_delayed_ref(&head->node);
3071                 return -EAGAIN;
3072         }
3073         spin_unlock(&delayed_refs->lock);
3074
3075         spin_lock(&head->lock);
3076         list_for_each_entry(ref, &head->ref_list, list) {
3077                 /* If it's a shared ref we know a cross reference exists */
3078                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
3079                         ret = 1;
3080                         break;
3081                 }
3082
3083                 data_ref = btrfs_delayed_node_to_data_ref(ref);
3084
3085                 /*
3086                  * If our ref doesn't match the one we're currently looking at
3087                  * then we have a cross reference.
3088                  */
3089                 if (data_ref->root != root->root_key.objectid ||
3090                     data_ref->objectid != objectid ||
3091                     data_ref->offset != offset) {
3092                         ret = 1;
3093                         break;
3094                 }
3095         }
3096         spin_unlock(&head->lock);
3097         mutex_unlock(&head->mutex);
3098         return ret;
3099 }
3100
3101 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
3102                                         struct btrfs_root *root,
3103                                         struct btrfs_path *path,
3104                                         u64 objectid, u64 offset, u64 bytenr)
3105 {
3106         struct btrfs_root *extent_root = root->fs_info->extent_root;
3107         struct extent_buffer *leaf;
3108         struct btrfs_extent_data_ref *ref;
3109         struct btrfs_extent_inline_ref *iref;
3110         struct btrfs_extent_item *ei;
3111         struct btrfs_key key;
3112         u32 item_size;
3113         int ret;
3114
3115         key.objectid = bytenr;
3116         key.offset = (u64)-1;
3117         key.type = BTRFS_EXTENT_ITEM_KEY;
3118
3119         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
3120         if (ret < 0)
3121                 goto out;
3122         BUG_ON(ret == 0); /* Corruption */
3123
3124         ret = -ENOENT;
3125         if (path->slots[0] == 0)
3126                 goto out;
3127
3128         path->slots[0]--;
3129         leaf = path->nodes[0];
3130         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3131
3132         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
3133                 goto out;
3134
3135         ret = 1;
3136         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
3137 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3138         if (item_size < sizeof(*ei)) {
3139                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3140                 goto out;
3141         }
3142 #endif
3143         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
3144
3145         if (item_size != sizeof(*ei) +
3146             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
3147                 goto out;
3148
3149         if (btrfs_extent_generation(leaf, ei) <=
3150             btrfs_root_last_snapshot(&root->root_item))
3151                 goto out;
3152
3153         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
3154         if (btrfs_extent_inline_ref_type(leaf, iref) !=
3155             BTRFS_EXTENT_DATA_REF_KEY)
3156                 goto out;
3157
3158         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
3159         if (btrfs_extent_refs(leaf, ei) !=
3160             btrfs_extent_data_ref_count(leaf, ref) ||
3161             btrfs_extent_data_ref_root(leaf, ref) !=
3162             root->root_key.objectid ||
3163             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
3164             btrfs_extent_data_ref_offset(leaf, ref) != offset)
3165                 goto out;
3166
3167         ret = 0;
3168 out:
3169         return ret;
3170 }
3171
3172 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
3173                           struct btrfs_root *root,
3174                           u64 objectid, u64 offset, u64 bytenr)
3175 {
3176         struct btrfs_path *path;
3177         int ret;
3178         int ret2;
3179
3180         path = btrfs_alloc_path();
3181         if (!path)
3182                 return -ENOENT;
3183
3184         do {
3185                 ret = check_committed_ref(trans, root, path, objectid,
3186                                           offset, bytenr);
3187                 if (ret && ret != -ENOENT)
3188                         goto out;
3189
3190                 ret2 = check_delayed_ref(trans, root, path, objectid,
3191                                          offset, bytenr);
3192         } while (ret2 == -EAGAIN);
3193
3194         if (ret2 && ret2 != -ENOENT) {
3195                 ret = ret2;
3196                 goto out;
3197         }
3198
3199         if (ret != -ENOENT || ret2 != -ENOENT)
3200                 ret = 0;
3201 out:
3202         btrfs_free_path(path);
3203         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
3204                 WARN_ON(ret > 0);
3205         return ret;
3206 }
3207
3208 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
3209                            struct btrfs_root *root,
3210                            struct extent_buffer *buf,
3211                            int full_backref, int inc)
3212 {
3213         u64 bytenr;
3214         u64 num_bytes;
3215         u64 parent;
3216         u64 ref_root;
3217         u32 nritems;
3218         struct btrfs_key key;
3219         struct btrfs_file_extent_item *fi;
3220         int i;
3221         int level;
3222         int ret = 0;
3223         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
3224                             u64, u64, u64, u64, u64, u64);
3225
3226
3227         if (btrfs_is_testing(root->fs_info))
3228                 return 0;
3229
3230         ref_root = btrfs_header_owner(buf);
3231         nritems = btrfs_header_nritems(buf);
3232         level = btrfs_header_level(buf);
3233
3234         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state) && level == 0)
3235                 return 0;
3236
3237         if (inc)
3238                 process_func = btrfs_inc_extent_ref;
3239         else
3240                 process_func = btrfs_free_extent;
3241
3242         if (full_backref)
3243                 parent = buf->start;
3244         else
3245                 parent = 0;
3246
3247         for (i = 0; i < nritems; i++) {
3248                 if (level == 0) {
3249                         btrfs_item_key_to_cpu(buf, &key, i);
3250                         if (key.type != BTRFS_EXTENT_DATA_KEY)
3251                                 continue;
3252                         fi = btrfs_item_ptr(buf, i,
3253                                             struct btrfs_file_extent_item);
3254                         if (btrfs_file_extent_type(buf, fi) ==
3255                             BTRFS_FILE_EXTENT_INLINE)
3256                                 continue;
3257                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3258                         if (bytenr == 0)
3259                                 continue;
3260
3261                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3262                         key.offset -= btrfs_file_extent_offset(buf, fi);
3263                         ret = process_func(trans, root, bytenr, num_bytes,
3264                                            parent, ref_root, key.objectid,
3265                                            key.offset);
3266                         if (ret)
3267                                 goto fail;
3268                 } else {
3269                         bytenr = btrfs_node_blockptr(buf, i);
3270                         num_bytes = root->nodesize;
3271                         ret = process_func(trans, root, bytenr, num_bytes,
3272                                            parent, ref_root, level - 1, 0);
3273                         if (ret)
3274                                 goto fail;
3275                 }
3276         }
3277         return 0;
3278 fail:
3279         return ret;
3280 }
3281
3282 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3283                   struct extent_buffer *buf, int full_backref)
3284 {
3285         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
3286 }
3287
3288 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3289                   struct extent_buffer *buf, int full_backref)
3290 {
3291         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
3292 }
3293
3294 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3295                                  struct btrfs_root *root,
3296                                  struct btrfs_path *path,
3297                                  struct btrfs_block_group_cache *cache)
3298 {
3299         int ret;
3300         struct btrfs_root *extent_root = root->fs_info->extent_root;
3301         unsigned long bi;
3302         struct extent_buffer *leaf;
3303
3304         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3305         if (ret) {
3306                 if (ret > 0)
3307                         ret = -ENOENT;
3308                 goto fail;
3309         }
3310
3311         leaf = path->nodes[0];
3312         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3313         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3314         btrfs_mark_buffer_dirty(leaf);
3315 fail:
3316         btrfs_release_path(path);
3317         return ret;
3318
3319 }
3320
3321 static struct btrfs_block_group_cache *
3322 next_block_group(struct btrfs_root *root,
3323                  struct btrfs_block_group_cache *cache)
3324 {
3325         struct rb_node *node;
3326
3327         spin_lock(&root->fs_info->block_group_cache_lock);
3328
3329         /* If our block group was removed, we need a full search. */
3330         if (RB_EMPTY_NODE(&cache->cache_node)) {
3331                 const u64 next_bytenr = cache->key.objectid + cache->key.offset;
3332
3333                 spin_unlock(&root->fs_info->block_group_cache_lock);
3334                 btrfs_put_block_group(cache);
3335                 cache = btrfs_lookup_first_block_group(root->fs_info,
3336                                                        next_bytenr);
3337                 return cache;
3338         }
3339         node = rb_next(&cache->cache_node);
3340         btrfs_put_block_group(cache);
3341         if (node) {
3342                 cache = rb_entry(node, struct btrfs_block_group_cache,
3343                                  cache_node);
3344                 btrfs_get_block_group(cache);
3345         } else
3346                 cache = NULL;
3347         spin_unlock(&root->fs_info->block_group_cache_lock);
3348         return cache;
3349 }
3350
3351 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3352                             struct btrfs_trans_handle *trans,
3353                             struct btrfs_path *path)
3354 {
3355         struct btrfs_root *root = block_group->fs_info->tree_root;
3356         struct inode *inode = NULL;
3357         u64 alloc_hint = 0;
3358         int dcs = BTRFS_DC_ERROR;
3359         u64 num_pages = 0;
3360         int retries = 0;
3361         int ret = 0;
3362
3363         /*
3364          * If this block group is smaller than 100 megs don't bother caching the
3365          * block group.
3366          */
3367         if (block_group->key.offset < (100 * SZ_1M)) {
3368                 spin_lock(&block_group->lock);
3369                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3370                 spin_unlock(&block_group->lock);
3371                 return 0;
3372         }
3373
3374         if (trans->aborted)
3375                 return 0;
3376 again:
3377         inode = lookup_free_space_inode(root, block_group, path);
3378         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3379                 ret = PTR_ERR(inode);
3380                 btrfs_release_path(path);
3381                 goto out;
3382         }
3383
3384         if (IS_ERR(inode)) {
3385                 BUG_ON(retries);
3386                 retries++;
3387
3388                 if (block_group->ro)
3389                         goto out_free;
3390
3391                 ret = create_free_space_inode(root, trans, block_group, path);
3392                 if (ret)
3393                         goto out_free;
3394                 goto again;
3395         }
3396
3397         /* We've already setup this transaction, go ahead and exit */
3398         if (block_group->cache_generation == trans->transid &&
3399             i_size_read(inode)) {
3400                 dcs = BTRFS_DC_SETUP;
3401                 goto out_put;
3402         }
3403
3404         /*
3405          * We want to set the generation to 0, that way if anything goes wrong
3406          * from here on out we know not to trust this cache when we load up next
3407          * time.
3408          */
3409         BTRFS_I(inode)->generation = 0;
3410         ret = btrfs_update_inode(trans, root, inode);
3411         if (ret) {
3412                 /*
3413                  * So theoretically we could recover from this, simply set the
3414                  * super cache generation to 0 so we know to invalidate the
3415                  * cache, but then we'd have to keep track of the block groups
3416                  * that fail this way so we know we _have_ to reset this cache
3417                  * before the next commit or risk reading stale cache.  So to
3418                  * limit our exposure to horrible edge cases lets just abort the
3419                  * transaction, this only happens in really bad situations
3420                  * anyway.
3421                  */
3422                 btrfs_abort_transaction(trans, ret);
3423                 goto out_put;
3424         }
3425         WARN_ON(ret);
3426
3427         if (i_size_read(inode) > 0) {
3428                 ret = btrfs_check_trunc_cache_free_space(root,
3429                                         &root->fs_info->global_block_rsv);
3430                 if (ret)
3431                         goto out_put;
3432
3433                 ret = btrfs_truncate_free_space_cache(root, trans, NULL, inode);
3434                 if (ret)
3435                         goto out_put;
3436         }
3437
3438         spin_lock(&block_group->lock);
3439         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3440             !btrfs_test_opt(root->fs_info, SPACE_CACHE)) {
3441                 /*
3442                  * don't bother trying to write stuff out _if_
3443                  * a) we're not cached,
3444                  * b) we're with nospace_cache mount option.
3445                  */
3446                 dcs = BTRFS_DC_WRITTEN;
3447                 spin_unlock(&block_group->lock);
3448                 goto out_put;
3449         }
3450         spin_unlock(&block_group->lock);
3451
3452         /*
3453          * We hit an ENOSPC when setting up the cache in this transaction, just
3454          * skip doing the setup, we've already cleared the cache so we're safe.
3455          */
3456         if (test_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags)) {
3457                 ret = -ENOSPC;
3458                 goto out_put;
3459         }
3460
3461         /*
3462          * Try to preallocate enough space based on how big the block group is.
3463          * Keep in mind this has to include any pinned space which could end up
3464          * taking up quite a bit since it's not folded into the other space
3465          * cache.
3466          */
3467         num_pages = div_u64(block_group->key.offset, SZ_256M);
3468         if (!num_pages)
3469                 num_pages = 1;
3470
3471         num_pages *= 16;
3472         num_pages *= PAGE_SIZE;
3473
3474         ret = btrfs_check_data_free_space(inode, 0, num_pages);
3475         if (ret)
3476                 goto out_put;
3477
3478         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3479                                               num_pages, num_pages,
3480                                               &alloc_hint);
3481         /*
3482          * Our cache requires contiguous chunks so that we don't modify a bunch
3483          * of metadata or split extents when writing the cache out, which means
3484          * we can enospc if we are heavily fragmented in addition to just normal
3485          * out of space conditions.  So if we hit this just skip setting up any
3486          * other block groups for this transaction, maybe we'll unpin enough
3487          * space the next time around.
3488          */
3489         if (!ret)
3490                 dcs = BTRFS_DC_SETUP;
3491         else if (ret == -ENOSPC)
3492                 set_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags);
3493
3494 out_put:
3495         iput(inode);
3496 out_free:
3497         btrfs_release_path(path);
3498 out:
3499         spin_lock(&block_group->lock);
3500         if (!ret && dcs == BTRFS_DC_SETUP)
3501                 block_group->cache_generation = trans->transid;
3502         block_group->disk_cache_state = dcs;
3503         spin_unlock(&block_group->lock);
3504
3505         return ret;
3506 }
3507
3508 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans,
3509                             struct btrfs_root *root)
3510 {
3511         struct btrfs_block_group_cache *cache, *tmp;
3512         struct btrfs_transaction *cur_trans = trans->transaction;
3513         struct btrfs_path *path;
3514
3515         if (list_empty(&cur_trans->dirty_bgs) ||
3516             !btrfs_test_opt(root->fs_info, SPACE_CACHE))
3517                 return 0;
3518
3519         path = btrfs_alloc_path();
3520         if (!path)
3521                 return -ENOMEM;
3522
3523         /* Could add new block groups, use _safe just in case */
3524         list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3525                                  dirty_list) {
3526                 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3527                         cache_save_setup(cache, trans, path);
3528         }
3529
3530         btrfs_free_path(path);
3531         return 0;
3532 }
3533
3534 /*
3535  * transaction commit does final block group cache writeback during a
3536  * critical section where nothing is allowed to change the FS.  This is
3537  * required in order for the cache to actually match the block group,
3538  * but can introduce a lot of latency into the commit.
3539  *
3540  * So, btrfs_start_dirty_block_groups is here to kick off block group
3541  * cache IO.  There's a chance we'll have to redo some of it if the
3542  * block group changes again during the commit, but it greatly reduces
3543  * the commit latency by getting rid of the easy block groups while
3544  * we're still allowing others to join the commit.
3545  */
3546 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans,
3547                                    struct btrfs_root *root)
3548 {
3549         struct btrfs_block_group_cache *cache;
3550         struct btrfs_transaction *cur_trans = trans->transaction;
3551         int ret = 0;
3552         int should_put;
3553         struct btrfs_path *path = NULL;
3554         LIST_HEAD(dirty);
3555         struct list_head *io = &cur_trans->io_bgs;
3556         int num_started = 0;
3557         int loops = 0;
3558
3559         spin_lock(&cur_trans->dirty_bgs_lock);
3560         if (list_empty(&cur_trans->dirty_bgs)) {
3561                 spin_unlock(&cur_trans->dirty_bgs_lock);
3562                 return 0;
3563         }
3564         list_splice_init(&cur_trans->dirty_bgs, &dirty);
3565         spin_unlock(&cur_trans->dirty_bgs_lock);
3566
3567 again:
3568         /*
3569          * make sure all the block groups on our dirty list actually
3570          * exist
3571          */
3572         btrfs_create_pending_block_groups(trans, root);
3573
3574         if (!path) {
3575                 path = btrfs_alloc_path();
3576                 if (!path)
3577                         return -ENOMEM;
3578         }
3579
3580         /*
3581          * cache_write_mutex is here only to save us from balance or automatic
3582          * removal of empty block groups deleting this block group while we are
3583          * writing out the cache
3584          */
3585         mutex_lock(&trans->transaction->cache_write_mutex);
3586         while (!list_empty(&dirty)) {
3587                 cache = list_first_entry(&dirty,
3588                                          struct btrfs_block_group_cache,
3589                                          dirty_list);
3590                 /*
3591                  * this can happen if something re-dirties a block
3592                  * group that is already under IO.  Just wait for it to
3593                  * finish and then do it all again
3594                  */
3595                 if (!list_empty(&cache->io_list)) {
3596                         list_del_init(&cache->io_list);
3597                         btrfs_wait_cache_io(root, trans, cache,
3598                                             &cache->io_ctl, path,
3599                                             cache->key.objectid);
3600                         btrfs_put_block_group(cache);
3601                 }
3602
3603
3604                 /*
3605                  * btrfs_wait_cache_io uses the cache->dirty_list to decide
3606                  * if it should update the cache_state.  Don't delete
3607                  * until after we wait.
3608                  *
3609                  * Since we're not running in the commit critical section
3610                  * we need the dirty_bgs_lock to protect from update_block_group
3611                  */
3612                 spin_lock(&cur_trans->dirty_bgs_lock);
3613                 list_del_init(&cache->dirty_list);
3614                 spin_unlock(&cur_trans->dirty_bgs_lock);
3615
3616                 should_put = 1;
3617
3618                 cache_save_setup(cache, trans, path);
3619
3620                 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
3621                         cache->io_ctl.inode = NULL;
3622                         ret = btrfs_write_out_cache(root, trans, cache, path);
3623                         if (ret == 0 && cache->io_ctl.inode) {
3624                                 num_started++;
3625                                 should_put = 0;
3626
3627                                 /*
3628                                  * the cache_write_mutex is protecting
3629                                  * the io_list
3630                                  */
3631                                 list_add_tail(&cache->io_list, io);
3632                         } else {
3633                                 /*
3634                                  * if we failed to write the cache, the
3635                                  * generation will be bad and life goes on
3636                                  */
3637                                 ret = 0;
3638                         }
3639                 }
3640                 if (!ret) {
3641                         ret = write_one_cache_group(trans, root, path, cache);
3642                         /*
3643                          * Our block group might still be attached to the list
3644                          * of new block groups in the transaction handle of some
3645                          * other task (struct btrfs_trans_handle->new_bgs). This
3646                          * means its block group item isn't yet in the extent
3647                          * tree. If this happens ignore the error, as we will
3648                          * try again later in the critical section of the
3649                          * transaction commit.
3650                          */
3651                         if (ret == -ENOENT) {
3652                                 ret = 0;
3653                                 spin_lock(&cur_trans->dirty_bgs_lock);
3654                                 if (list_empty(&cache->dirty_list)) {
3655                                         list_add_tail(&cache->dirty_list,
3656                                                       &cur_trans->dirty_bgs);
3657                                         btrfs_get_block_group(cache);
3658                                 }
3659                                 spin_unlock(&cur_trans->dirty_bgs_lock);
3660                         } else if (ret) {
3661                                 btrfs_abort_transaction(trans, ret);
3662                         }
3663                 }
3664
3665                 /* if its not on the io list, we need to put the block group */
3666                 if (should_put)
3667                         btrfs_put_block_group(cache);
3668
3669                 if (ret)
3670                         break;
3671
3672                 /*
3673                  * Avoid blocking other tasks for too long. It might even save
3674                  * us from writing caches for block groups that are going to be
3675                  * removed.
3676                  */
3677                 mutex_unlock(&trans->transaction->cache_write_mutex);
3678                 mutex_lock(&trans->transaction->cache_write_mutex);
3679         }
3680         mutex_unlock(&trans->transaction->cache_write_mutex);
3681
3682         /*
3683          * go through delayed refs for all the stuff we've just kicked off
3684          * and then loop back (just once)
3685          */
3686         ret = btrfs_run_delayed_refs(trans, root, 0);
3687         if (!ret && loops == 0) {
3688                 loops++;
3689                 spin_lock(&cur_trans->dirty_bgs_lock);
3690                 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3691                 /*
3692                  * dirty_bgs_lock protects us from concurrent block group
3693                  * deletes too (not just cache_write_mutex).
3694                  */
3695                 if (!list_empty(&dirty)) {
3696                         spin_unlock(&cur_trans->dirty_bgs_lock);
3697                         goto again;
3698                 }
3699                 spin_unlock(&cur_trans->dirty_bgs_lock);
3700         } else if (ret < 0) {
3701                 btrfs_cleanup_dirty_bgs(cur_trans, root);
3702         }
3703
3704         btrfs_free_path(path);
3705         return ret;
3706 }
3707
3708 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3709                                    struct btrfs_root *root)
3710 {
3711         struct btrfs_block_group_cache *cache;
3712         struct btrfs_transaction *cur_trans = trans->transaction;
3713         int ret = 0;
3714         int should_put;
3715         struct btrfs_path *path;
3716         struct list_head *io = &cur_trans->io_bgs;
3717         int num_started = 0;
3718
3719         path = btrfs_alloc_path();
3720         if (!path)
3721                 return -ENOMEM;
3722
3723         /*
3724          * Even though we are in the critical section of the transaction commit,
3725          * we can still have concurrent tasks adding elements to this
3726          * transaction's list of dirty block groups. These tasks correspond to
3727          * endio free space workers started when writeback finishes for a
3728          * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
3729          * allocate new block groups as a result of COWing nodes of the root
3730          * tree when updating the free space inode. The writeback for the space
3731          * caches is triggered by an earlier call to
3732          * btrfs_start_dirty_block_groups() and iterations of the following
3733          * loop.
3734          * Also we want to do the cache_save_setup first and then run the
3735          * delayed refs to make sure we have the best chance at doing this all
3736          * in one shot.
3737          */
3738         spin_lock(&cur_trans->dirty_bgs_lock);
3739         while (!list_empty(&cur_trans->dirty_bgs)) {
3740                 cache = list_first_entry(&cur_trans->dirty_bgs,
3741                                          struct btrfs_block_group_cache,
3742                                          dirty_list);
3743
3744                 /*
3745                  * this can happen if cache_save_setup re-dirties a block
3746                  * group that is already under IO.  Just wait for it to
3747                  * finish and then do it all again
3748                  */
3749                 if (!list_empty(&cache->io_list)) {
3750                         spin_unlock(&cur_trans->dirty_bgs_lock);
3751                         list_del_init(&cache->io_list);
3752                         btrfs_wait_cache_io(root, trans, cache,
3753                                             &cache->io_ctl, path,
3754                                             cache->key.objectid);
3755                         btrfs_put_block_group(cache);
3756                         spin_lock(&cur_trans->dirty_bgs_lock);
3757                 }
3758
3759                 /*
3760                  * don't remove from the dirty list until after we've waited
3761                  * on any pending IO
3762                  */
3763                 list_del_init(&cache->dirty_list);
3764                 spin_unlock(&cur_trans->dirty_bgs_lock);
3765                 should_put = 1;
3766
3767                 cache_save_setup(cache, trans, path);
3768
3769                 if (!ret)
3770                         ret = btrfs_run_delayed_refs(trans, root, (unsigned long) -1);
3771
3772                 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3773                         cache->io_ctl.inode = NULL;
3774                         ret = btrfs_write_out_cache(root, trans, cache, path);
3775                         if (ret == 0 && cache->io_ctl.inode) {
3776                                 num_started++;
3777                                 should_put = 0;
3778                                 list_add_tail(&cache->io_list, io);
3779                         } else {
3780                                 /*
3781                                  * if we failed to write the cache, the
3782                                  * generation will be bad and life goes on
3783                                  */
3784                                 ret = 0;
3785                         }
3786                 }
3787                 if (!ret) {
3788                         ret = write_one_cache_group(trans, root, path, cache);
3789                         /*
3790                          * One of the free space endio workers might have
3791                          * created a new block group while updating a free space
3792                          * cache's inode (at inode.c:btrfs_finish_ordered_io())
3793                          * and hasn't released its transaction handle yet, in
3794                          * which case the new block group is still attached to
3795                          * its transaction handle and its creation has not
3796                          * finished yet (no block group item in the extent tree
3797                          * yet, etc). If this is the case, wait for all free
3798                          * space endio workers to finish and retry. This is a
3799                          * a very rare case so no need for a more efficient and
3800                          * complex approach.
3801                          */
3802                         if (ret == -ENOENT) {
3803                                 wait_event(cur_trans->writer_wait,
3804                                    atomic_read(&cur_trans->num_writers) == 1);
3805                                 ret = write_one_cache_group(trans, root, path,
3806                                                             cache);
3807                         }
3808                         if (ret)
3809                                 btrfs_abort_transaction(trans, ret);
3810                 }
3811
3812                 /* if its not on the io list, we need to put the block group */
3813                 if (should_put)
3814                         btrfs_put_block_group(cache);
3815                 spin_lock(&cur_trans->dirty_bgs_lock);
3816         }
3817         spin_unlock(&cur_trans->dirty_bgs_lock);
3818
3819         while (!list_empty(io)) {
3820                 cache = list_first_entry(io, struct btrfs_block_group_cache,
3821                                          io_list);
3822                 list_del_init(&cache->io_list);
3823                 btrfs_wait_cache_io(root, trans, cache,
3824                                     &cache->io_ctl, path, cache->key.objectid);
3825                 btrfs_put_block_group(cache);
3826         }
3827
3828         btrfs_free_path(path);
3829         return ret;
3830 }
3831
3832 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3833 {
3834         struct btrfs_block_group_cache *block_group;
3835         int readonly = 0;
3836
3837         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3838         if (!block_group || block_group->ro)
3839                 readonly = 1;
3840         if (block_group)
3841                 btrfs_put_block_group(block_group);
3842         return readonly;
3843 }
3844
3845 bool btrfs_inc_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3846 {
3847         struct btrfs_block_group_cache *bg;
3848         bool ret = true;
3849
3850         bg = btrfs_lookup_block_group(fs_info, bytenr);
3851         if (!bg)
3852                 return false;
3853
3854         spin_lock(&bg->lock);
3855         if (bg->ro)
3856                 ret = false;
3857         else
3858                 atomic_inc(&bg->nocow_writers);
3859         spin_unlock(&bg->lock);
3860
3861         /* no put on block group, done by btrfs_dec_nocow_writers */
3862         if (!ret)
3863                 btrfs_put_block_group(bg);
3864
3865         return ret;
3866
3867 }
3868
3869 void btrfs_dec_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3870 {
3871         struct btrfs_block_group_cache *bg;
3872
3873         bg = btrfs_lookup_block_group(fs_info, bytenr);
3874         ASSERT(bg);
3875         if (atomic_dec_and_test(&bg->nocow_writers))
3876                 wake_up_atomic_t(&bg->nocow_writers);
3877         /*
3878          * Once for our lookup and once for the lookup done by a previous call
3879          * to btrfs_inc_nocow_writers()
3880          */
3881         btrfs_put_block_group(bg);
3882         btrfs_put_block_group(bg);
3883 }
3884
3885 static int btrfs_wait_nocow_writers_atomic_t(atomic_t *a)
3886 {
3887         schedule();
3888         return 0;
3889 }
3890
3891 void btrfs_wait_nocow_writers(struct btrfs_block_group_cache *bg)
3892 {
3893         wait_on_atomic_t(&bg->nocow_writers,
3894                          btrfs_wait_nocow_writers_atomic_t,
3895                          TASK_UNINTERRUPTIBLE);
3896 }
3897
3898 static const char *alloc_name(u64 flags)
3899 {
3900         switch (flags) {
3901         case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA:
3902                 return "mixed";
3903         case BTRFS_BLOCK_GROUP_METADATA:
3904                 return "metadata";
3905         case BTRFS_BLOCK_GROUP_DATA:
3906                 return "data";
3907         case BTRFS_BLOCK_GROUP_SYSTEM:
3908                 return "system";
3909         default:
3910                 WARN_ON(1);
3911                 return "invalid-combination";
3912         };
3913 }
3914
3915 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3916                              u64 total_bytes, u64 bytes_used,
3917                              u64 bytes_readonly,
3918                              struct btrfs_space_info **space_info)
3919 {
3920         struct btrfs_space_info *found;
3921         int i;
3922         int factor;
3923         int ret;
3924
3925         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3926                      BTRFS_BLOCK_GROUP_RAID10))
3927                 factor = 2;
3928         else
3929                 factor = 1;
3930
3931         found = __find_space_info(info, flags);
3932         if (found) {
3933                 spin_lock(&found->lock);
3934                 found->total_bytes += total_bytes;
3935                 found->disk_total += total_bytes * factor;
3936                 found->bytes_used += bytes_used;
3937                 found->disk_used += bytes_used * factor;
3938                 found->bytes_readonly += bytes_readonly;
3939                 if (total_bytes > 0)
3940                         found->full = 0;
3941                 space_info_add_new_bytes(info, found, total_bytes -
3942                                          bytes_used - bytes_readonly);
3943                 spin_unlock(&found->lock);
3944                 *space_info = found;
3945                 return 0;
3946         }
3947         found = kzalloc(sizeof(*found), GFP_NOFS);
3948         if (!found)
3949                 return -ENOMEM;
3950
3951         ret = percpu_counter_init(&found->total_bytes_pinned, 0, GFP_KERNEL);
3952         if (ret) {
3953                 kfree(found);
3954                 return ret;
3955         }
3956
3957         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3958                 INIT_LIST_HEAD(&found->block_groups[i]);
3959         init_rwsem(&found->groups_sem);
3960         spin_lock_init(&found->lock);
3961         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3962         found->total_bytes = total_bytes;
3963         found->disk_total = total_bytes * factor;
3964         found->bytes_used = bytes_used;
3965         found->disk_used = bytes_used * factor;
3966         found->bytes_pinned = 0;
3967         found->bytes_reserved = 0;
3968         found->bytes_readonly = bytes_readonly;
3969         found->bytes_may_use = 0;
3970         found->full = 0;
3971         found->max_extent_size = 0;
3972         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3973         found->chunk_alloc = 0;
3974         found->flush = 0;
3975         init_waitqueue_head(&found->wait);
3976         INIT_LIST_HEAD(&found->ro_bgs);
3977         INIT_LIST_HEAD(&found->tickets);
3978         INIT_LIST_HEAD(&found->priority_tickets);
3979
3980         ret = kobject_init_and_add(&found->kobj, &space_info_ktype,
3981                                     info->space_info_kobj, "%s",
3982                                     alloc_name(found->flags));
3983         if (ret) {
3984                 kfree(found);
3985                 return ret;
3986         }
3987
3988         *space_info = found;
3989         list_add_rcu(&found->list, &info->space_info);
3990         if (flags & BTRFS_BLOCK_GROUP_DATA)
3991                 info->data_sinfo = found;
3992
3993         return ret;
3994 }
3995
3996 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3997 {
3998         u64 extra_flags = chunk_to_extended(flags) &
3999                                 BTRFS_EXTENDED_PROFILE_MASK;
4000
4001         write_seqlock(&fs_info->profiles_lock);
4002         if (flags & BTRFS_BLOCK_GROUP_DATA)
4003                 fs_info->avail_data_alloc_bits |= extra_flags;
4004         if (flags & BTRFS_BLOCK_GROUP_METADATA)
4005                 fs_info->avail_metadata_alloc_bits |= extra_flags;
4006         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4007                 fs_info->avail_system_alloc_bits |= extra_flags;
4008         write_sequnlock(&fs_info->profiles_lock);
4009 }
4010
4011 /*
4012  * returns target flags in extended format or 0 if restripe for this
4013  * chunk_type is not in progress
4014  *
4015  * should be called with either volume_mutex or balance_lock held
4016  */
4017 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
4018 {
4019         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
4020         u64 target = 0;
4021
4022         if (!bctl)
4023                 return 0;
4024
4025         if (flags & BTRFS_BLOCK_GROUP_DATA &&
4026             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4027                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
4028         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
4029                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4030                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
4031         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
4032                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4033                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
4034         }
4035
4036         return target;
4037 }
4038
4039 /*
4040  * @flags: available profiles in extended format (see ctree.h)
4041  *
4042  * Returns reduced profile in chunk format.  If profile changing is in
4043  * progress (either running or paused) picks the target profile (if it's
4044  * already available), otherwise falls back to plain reducing.
4045  */
4046 static u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
4047 {
4048         u64 num_devices = root->fs_info->fs_devices->rw_devices;
4049         u64 target;
4050         u64 raid_type;
4051         u64 allowed = 0;
4052
4053         /*
4054          * see if restripe for this chunk_type is in progress, if so
4055          * try to reduce to the target profile
4056          */
4057         spin_lock(&root->fs_info->balance_lock);
4058         target = get_restripe_target(root->fs_info, flags);
4059         if (target) {
4060                 /* pick target profile only if it's already available */
4061                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
4062                         spin_unlock(&root->fs_info->balance_lock);
4063                         return extended_to_chunk(target);
4064                 }
4065         }
4066         spin_unlock(&root->fs_info->balance_lock);
4067
4068         /* First, mask out the RAID levels which aren't possible */
4069         for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
4070                 if (num_devices >= btrfs_raid_array[raid_type].devs_min)
4071                         allowed |= btrfs_raid_group[raid_type];
4072         }
4073         allowed &= flags;
4074
4075         if (allowed & BTRFS_BLOCK_GROUP_RAID6)
4076                 allowed = BTRFS_BLOCK_GROUP_RAID6;
4077         else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
4078                 allowed = BTRFS_BLOCK_GROUP_RAID5;
4079         else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
4080                 allowed = BTRFS_BLOCK_GROUP_RAID10;
4081         else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
4082                 allowed = BTRFS_BLOCK_GROUP_RAID1;
4083         else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
4084                 allowed = BTRFS_BLOCK_GROUP_RAID0;
4085
4086         flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
4087
4088         return extended_to_chunk(flags | allowed);
4089 }
4090
4091 static u64 get_alloc_profile(struct btrfs_root *root, u64 orig_flags)
4092 {
4093         unsigned seq;
4094         u64 flags;
4095
4096         do {
4097                 flags = orig_flags;
4098                 seq = read_seqbegin(&root->fs_info->profiles_lock);
4099
4100                 if (flags & BTRFS_BLOCK_GROUP_DATA)
4101                         flags |= root->fs_info->avail_data_alloc_bits;
4102                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4103                         flags |= root->fs_info->avail_system_alloc_bits;
4104                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
4105                         flags |= root->fs_info->avail_metadata_alloc_bits;
4106         } while (read_seqretry(&root->fs_info->profiles_lock, seq));
4107
4108         return btrfs_reduce_alloc_profile(root, flags);
4109 }
4110
4111 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
4112 {
4113         u64 flags;
4114         u64 ret;
4115
4116         if (data)
4117                 flags = BTRFS_BLOCK_GROUP_DATA;
4118         else if (root == root->fs_info->chunk_root)
4119                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
4120         else
4121                 flags = BTRFS_BLOCK_GROUP_METADATA;
4122
4123         ret = get_alloc_profile(root, flags);
4124         return ret;
4125 }
4126
4127 int btrfs_alloc_data_chunk_ondemand(struct inode *inode, u64 bytes)
4128 {
4129         struct btrfs_space_info *data_sinfo;
4130         struct btrfs_root *root = BTRFS_I(inode)->root;
4131         struct btrfs_fs_info *fs_info = root->fs_info;
4132         u64 used;
4133         int ret = 0;
4134         int need_commit = 2;
4135         int have_pinned_space;
4136
4137         /* make sure bytes are sectorsize aligned */
4138         bytes = ALIGN(bytes, root->sectorsize);
4139
4140         if (btrfs_is_free_space_inode(inode)) {
4141                 need_commit = 0;
4142                 ASSERT(current->journal_info);
4143         }
4144
4145         data_sinfo = fs_info->data_sinfo;
4146         if (!data_sinfo)
4147                 goto alloc;
4148
4149 again:
4150         /* make sure we have enough space to handle the data first */
4151         spin_lock(&data_sinfo->lock);
4152         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
4153                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
4154                 data_sinfo->bytes_may_use;
4155
4156         if (used + bytes > data_sinfo->total_bytes) {
4157                 struct btrfs_trans_handle *trans;
4158
4159                 /*
4160                  * if we don't have enough free bytes in this space then we need
4161                  * to alloc a new chunk.
4162                  */
4163                 if (!data_sinfo->full) {
4164                         u64 alloc_target;
4165
4166                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
4167                         spin_unlock(&data_sinfo->lock);
4168 alloc:
4169                         alloc_target = btrfs_get_alloc_profile(root, 1);
4170                         /*
4171                          * It is ugly that we don't call nolock join
4172                          * transaction for the free space inode case here.
4173                          * But it is safe because we only do the data space
4174                          * reservation for the free space cache in the
4175                          * transaction context, the common join transaction
4176                          * just increase the counter of the current transaction
4177                          * handler, doesn't try to acquire the trans_lock of
4178                          * the fs.
4179                          */
4180                         trans = btrfs_join_transaction(root);
4181                         if (IS_ERR(trans))
4182                                 return PTR_ERR(trans);
4183
4184                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
4185                                              alloc_target,
4186                                              CHUNK_ALLOC_NO_FORCE);
4187                         btrfs_end_transaction(trans, root);
4188                         if (ret < 0) {
4189                                 if (ret != -ENOSPC)
4190                                         return ret;
4191                                 else {
4192                                         have_pinned_space = 1;
4193                                         goto commit_trans;
4194                                 }
4195                         }
4196
4197                         if (!data_sinfo)
4198                                 data_sinfo = fs_info->data_sinfo;
4199
4200                         goto again;
4201                 }
4202
4203                 /*
4204                  * If we don't have enough pinned space to deal with this
4205                  * allocation, and no removed chunk in current transaction,
4206                  * don't bother committing the transaction.
4207                  */
4208                 have_pinned_space = percpu_counter_compare(
4209                         &data_sinfo->total_bytes_pinned,
4210                         used + bytes - data_sinfo->total_bytes);
4211                 spin_unlock(&data_sinfo->lock);
4212
4213                 /* commit the current transaction and try again */
4214 commit_trans:
4215                 if (need_commit &&
4216                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
4217                         need_commit--;
4218
4219                         if (need_commit > 0) {
4220                                 btrfs_start_delalloc_roots(fs_info, 0, -1);
4221                                 btrfs_wait_ordered_roots(fs_info, -1, 0, (u64)-1);
4222                         }
4223
4224                         trans = btrfs_join_transaction(root);
4225                         if (IS_ERR(trans))
4226                                 return PTR_ERR(trans);
4227                         if (have_pinned_space >= 0 ||
4228                             test_bit(BTRFS_TRANS_HAVE_FREE_BGS,
4229                                      &trans->transaction->flags) ||
4230                             need_commit > 0) {
4231                                 ret = btrfs_commit_transaction(trans, root);
4232                                 if (ret)
4233                                         return ret;
4234                                 /*
4235                                  * The cleaner kthread might still be doing iput
4236                                  * operations. Wait for it to finish so that
4237                                  * more space is released.
4238                                  */
4239                                 mutex_lock(&root->fs_info->cleaner_delayed_iput_mutex);
4240                                 mutex_unlock(&root->fs_info->cleaner_delayed_iput_mutex);
4241                                 goto again;
4242                         } else {
4243                                 btrfs_end_transaction(trans, root);
4244                         }
4245                 }
4246
4247                 trace_btrfs_space_reservation(root->fs_info,
4248                                               "space_info:enospc",
4249                                               data_sinfo->flags, bytes, 1);
4250                 return -ENOSPC;
4251         }
4252         data_sinfo->bytes_may_use += bytes;
4253         trace_btrfs_space_reservation(root->fs_info, "space_info",
4254                                       data_sinfo->flags, bytes, 1);
4255         spin_unlock(&data_sinfo->lock);
4256
4257         return ret;
4258 }
4259
4260 /*
4261  * New check_data_free_space() with ability for precious data reservation
4262  * Will replace old btrfs_check_data_free_space(), but for patch split,
4263  * add a new function first and then replace it.
4264  */
4265 int btrfs_check_data_free_space(struct inode *inode, u64 start, u64 len)
4266 {
4267         struct btrfs_root *root = BTRFS_I(inode)->root;
4268         int ret;
4269
4270         /* align the range */
4271         len = round_up(start + len, root->sectorsize) -
4272               round_down(start, root->sectorsize);
4273         start = round_down(start, root->sectorsize);
4274
4275         ret = btrfs_alloc_data_chunk_ondemand(inode, len);
4276         if (ret < 0)
4277                 return ret;
4278
4279         /* Use new btrfs_qgroup_reserve_data to reserve precious data space. */
4280         ret = btrfs_qgroup_reserve_data(inode, start, len);
4281         if (ret)
4282                 btrfs_free_reserved_data_space_noquota(inode, start, len);
4283         return ret;
4284 }
4285
4286 /*
4287  * Called if we need to clear a data reservation for this inode
4288  * Normally in a error case.
4289  *
4290  * This one will *NOT* use accurate qgroup reserved space API, just for case
4291  * which we can't sleep and is sure it won't affect qgroup reserved space.
4292  * Like clear_bit_hook().
4293  */
4294 void btrfs_free_reserved_data_space_noquota(struct inode *inode, u64 start,
4295                                             u64 len)
4296 {
4297         struct btrfs_root *root = BTRFS_I(inode)->root;
4298         struct btrfs_space_info *data_sinfo;
4299
4300         /* Make sure the range is aligned to sectorsize */
4301         len = round_up(start + len, root->sectorsize) -
4302               round_down(start, root->sectorsize);
4303         start = round_down(start, root->sectorsize);
4304
4305         data_sinfo = root->fs_info->data_sinfo;
4306         spin_lock(&data_sinfo->lock);
4307         if (WARN_ON(data_sinfo->bytes_may_use < len))
4308                 data_sinfo->bytes_may_use = 0;
4309         else
4310                 data_sinfo->bytes_may_use -= len;
4311         trace_btrfs_space_reservation(root->fs_info, "space_info",
4312                                       data_sinfo->flags, len, 0);
4313         spin_unlock(&data_sinfo->lock);
4314 }
4315
4316 /*
4317  * Called if we need to clear a data reservation for this inode
4318  * Normally in a error case.
4319  *
4320  * This one will handle the per-inode data rsv map for accurate reserved
4321  * space framework.
4322  */
4323 void btrfs_free_reserved_data_space(struct inode *inode, u64 start, u64 len)
4324 {
4325         btrfs_free_reserved_data_space_noquota(inode, start, len);
4326         btrfs_qgroup_free_data(inode, start, len);
4327 }
4328
4329 static void force_metadata_allocation(struct btrfs_fs_info *info)
4330 {
4331         struct list_head *head = &info->space_info;
4332         struct btrfs_space_info *found;
4333
4334         rcu_read_lock();
4335         list_for_each_entry_rcu(found, head, list) {
4336                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
4337                         found->force_alloc = CHUNK_ALLOC_FORCE;
4338         }
4339         rcu_read_unlock();
4340 }
4341
4342 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
4343 {
4344         return (global->size << 1);
4345 }
4346
4347 static int should_alloc_chunk(struct btrfs_root *root,
4348                               struct btrfs_space_info *sinfo, int force)
4349 {
4350         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4351         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
4352         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
4353         u64 thresh;
4354
4355         if (force == CHUNK_ALLOC_FORCE)
4356                 return 1;
4357
4358         /*
4359          * We need to take into account the global rsv because for all intents
4360          * and purposes it's used space.  Don't worry about locking the
4361          * global_rsv, it doesn't change except when the transaction commits.
4362          */
4363         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
4364                 num_allocated += calc_global_rsv_need_space(global_rsv);
4365
4366         /*
4367          * in limited mode, we want to have some free space up to
4368          * about 1% of the FS size.
4369          */
4370         if (force == CHUNK_ALLOC_LIMITED) {
4371                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
4372                 thresh = max_t(u64, SZ_64M, div_factor_fine(thresh, 1));
4373
4374                 if (num_bytes - num_allocated < thresh)
4375                         return 1;
4376         }
4377
4378         if (num_allocated + SZ_2M < div_factor(num_bytes, 8))
4379                 return 0;
4380         return 1;
4381 }
4382
4383 static u64 get_profile_num_devs(struct btrfs_root *root, u64 type)
4384 {
4385         u64 num_dev;
4386
4387         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
4388                     BTRFS_BLOCK_GROUP_RAID0 |
4389                     BTRFS_BLOCK_GROUP_RAID5 |
4390                     BTRFS_BLOCK_GROUP_RAID6))
4391                 num_dev = root->fs_info->fs_devices->rw_devices;
4392         else if (type & BTRFS_BLOCK_GROUP_RAID1)
4393                 num_dev = 2;
4394         else
4395                 num_dev = 1;    /* DUP or single */
4396
4397         return num_dev;
4398 }
4399
4400 /*
4401  * If @is_allocation is true, reserve space in the system space info necessary
4402  * for allocating a chunk, otherwise if it's false, reserve space necessary for
4403  * removing a chunk.
4404  */
4405 void check_system_chunk(struct btrfs_trans_handle *trans,
4406                         struct btrfs_root *root,
4407                         u64 type)
4408 {
4409         struct btrfs_space_info *info;
4410         u64 left;
4411         u64 thresh;
4412         int ret = 0;
4413         u64 num_devs;
4414
4415         /*
4416          * Needed because we can end up allocating a system chunk and for an
4417          * atomic and race free space reservation in the chunk block reserve.
4418          */
4419         ASSERT(mutex_is_locked(&root->fs_info->chunk_mutex));
4420
4421         info = __find_space_info(root->fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4422         spin_lock(&info->lock);
4423         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
4424                 info->bytes_reserved - info->bytes_readonly -
4425                 info->bytes_may_use;
4426         spin_unlock(&info->lock);
4427
4428         num_devs = get_profile_num_devs(root, type);
4429
4430         /* num_devs device items to update and 1 chunk item to add or remove */
4431         thresh = btrfs_calc_trunc_metadata_size(root, num_devs) +
4432                 btrfs_calc_trans_metadata_size(root, 1);
4433
4434         if (left < thresh && btrfs_test_opt(root->fs_info, ENOSPC_DEBUG)) {
4435                 btrfs_info(root->fs_info, "left=%llu, need=%llu, flags=%llu",
4436                         left, thresh, type);
4437                 dump_space_info(root->fs_info, info, 0, 0);
4438         }
4439
4440         if (left < thresh) {
4441                 u64 flags;
4442
4443                 flags = btrfs_get_alloc_profile(root->fs_info->chunk_root, 0);
4444                 /*
4445                  * Ignore failure to create system chunk. We might end up not
4446                  * needing it, as we might not need to COW all nodes/leafs from
4447                  * the paths we visit in the chunk tree (they were already COWed
4448                  * or created in the current transaction for example).
4449                  */
4450                 ret = btrfs_alloc_chunk(trans, root, flags);
4451         }
4452
4453         if (!ret) {
4454                 ret = btrfs_block_rsv_add(root->fs_info->chunk_root,
4455                                           &root->fs_info->chunk_block_rsv,
4456                                           thresh, BTRFS_RESERVE_NO_FLUSH);
4457                 if (!ret)
4458                         trans->chunk_bytes_reserved += thresh;
4459         }
4460 }
4461
4462 /*
4463  * If force is CHUNK_ALLOC_FORCE:
4464  *    - return 1 if it successfully allocates a chunk,
4465  *    - return errors including -ENOSPC otherwise.
4466  * If force is NOT CHUNK_ALLOC_FORCE:
4467  *    - return 0 if it doesn't need to allocate a new chunk,
4468  *    - return 1 if it successfully allocates a chunk,
4469  *    - return errors including -ENOSPC otherwise.
4470  */
4471 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
4472                           struct btrfs_root *extent_root, u64 flags, int force)
4473 {
4474         struct btrfs_space_info *space_info;
4475         struct btrfs_fs_info *fs_info = extent_root->fs_info;
4476         int wait_for_alloc = 0;
4477         int ret = 0;
4478
4479         /* Don't re-enter if we're already allocating a chunk */
4480         if (trans->allocating_chunk)
4481                 return -ENOSPC;
4482
4483         space_info = __find_space_info(extent_root->fs_info, flags);
4484         if (!space_info) {
4485                 ret = update_space_info(extent_root->fs_info, flags,
4486                                         0, 0, 0, &space_info);
4487                 BUG_ON(ret); /* -ENOMEM */
4488         }
4489         BUG_ON(!space_info); /* Logic error */
4490
4491 again:
4492         spin_lock(&space_info->lock);
4493         if (force < space_info->force_alloc)
4494                 force = space_info->force_alloc;
4495         if (space_info->full) {
4496                 if (should_alloc_chunk(extent_root, space_info, force))
4497                         ret = -ENOSPC;
4498                 else
4499                         ret = 0;
4500                 spin_unlock(&space_info->lock);
4501                 return ret;
4502         }
4503
4504         if (!should_alloc_chunk(extent_root, space_info, force)) {
4505                 spin_unlock(&space_info->lock);
4506                 return 0;
4507         } else if (space_info->chunk_alloc) {
4508                 wait_for_alloc = 1;
4509         } else {
4510                 space_info->chunk_alloc = 1;
4511         }
4512
4513         spin_unlock(&space_info->lock);
4514
4515         mutex_lock(&fs_info->chunk_mutex);
4516
4517         /*
4518          * The chunk_mutex is held throughout the entirety of a chunk
4519          * allocation, so once we've acquired the chunk_mutex we know that the
4520          * other guy is done and we need to recheck and see if we should
4521          * allocate.
4522          */
4523         if (wait_for_alloc) {
4524                 mutex_unlock(&fs_info->chunk_mutex);
4525                 wait_for_alloc = 0;
4526                 goto again;
4527         }
4528
4529         trans->allocating_chunk = true;
4530
4531         /*
4532          * If we have mixed data/metadata chunks we want to make sure we keep
4533          * allocating mixed chunks instead of individual chunks.
4534          */
4535         if (btrfs_mixed_space_info(space_info))
4536                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4537
4538         /*
4539          * if we're doing a data chunk, go ahead and make sure that
4540          * we keep a reasonable number of metadata chunks allocated in the
4541          * FS as well.
4542          */
4543         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4544                 fs_info->data_chunk_allocations++;
4545                 if (!(fs_info->data_chunk_allocations %
4546                       fs_info->metadata_ratio))
4547                         force_metadata_allocation(fs_info);
4548         }
4549
4550         /*
4551          * Check if we have enough space in SYSTEM chunk because we may need
4552          * to update devices.
4553          */
4554         check_system_chunk(trans, extent_root, flags);
4555
4556         ret = btrfs_alloc_chunk(trans, extent_root, flags);
4557         trans->allocating_chunk = false;
4558
4559         spin_lock(&space_info->lock);
4560         if (ret < 0 && ret != -ENOSPC)
4561                 goto out;
4562         if (ret)
4563                 space_info->full = 1;
4564         else
4565                 ret = 1;
4566
4567         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4568 out:
4569         space_info->chunk_alloc = 0;
4570         spin_unlock(&space_info->lock);
4571         mutex_unlock(&fs_info->chunk_mutex);
4572         /*
4573          * When we allocate a new chunk we reserve space in the chunk block
4574          * reserve to make sure we can COW nodes/leafs in the chunk tree or
4575          * add new nodes/leafs to it if we end up needing to do it when
4576          * inserting the chunk item and updating device items as part of the
4577          * second phase of chunk allocation, performed by
4578          * btrfs_finish_chunk_alloc(). So make sure we don't accumulate a
4579          * large number of new block groups to create in our transaction
4580          * handle's new_bgs list to avoid exhausting the chunk block reserve
4581          * in extreme cases - like having a single transaction create many new
4582          * block groups when starting to write out the free space caches of all
4583          * the block groups that were made dirty during the lifetime of the
4584          * transaction.
4585          */
4586         if (trans->can_flush_pending_bgs &&
4587             trans->chunk_bytes_reserved >= (u64)SZ_2M) {
4588                 btrfs_create_pending_block_groups(trans, extent_root);
4589                 btrfs_trans_release_chunk_metadata(trans);
4590         }
4591         return ret;
4592 }
4593
4594 static int can_overcommit(struct btrfs_root *root,
4595                           struct btrfs_space_info *space_info, u64 bytes,
4596                           enum btrfs_reserve_flush_enum flush)
4597 {
4598         struct btrfs_block_rsv *global_rsv;
4599         u64 profile;
4600         u64 space_size;
4601         u64 avail;
4602         u64 used;
4603
4604         /* Don't overcommit when in mixed mode. */
4605         if (space_info->flags & BTRFS_BLOCK_GROUP_DATA)
4606                 return 0;
4607
4608         BUG_ON(root->fs_info == NULL);
4609         global_rsv = &root->fs_info->global_block_rsv;
4610         profile = btrfs_get_alloc_profile(root, 0);
4611         used = space_info->bytes_used + space_info->bytes_reserved +
4612                 space_info->bytes_pinned + space_info->bytes_readonly;
4613
4614         /*
4615          * We only want to allow over committing if we have lots of actual space
4616          * free, but if we don't have enough space to handle the global reserve
4617          * space then we could end up having a real enospc problem when trying
4618          * to allocate a chunk or some other such important allocation.
4619          */
4620         spin_lock(&global_rsv->lock);
4621         space_size = calc_global_rsv_need_space(global_rsv);
4622         spin_unlock(&global_rsv->lock);
4623         if (used + space_size >= space_info->total_bytes)
4624                 return 0;
4625
4626         used += space_info->bytes_may_use;
4627
4628         spin_lock(&root->fs_info->free_chunk_lock);
4629         avail = root->fs_info->free_chunk_space;
4630         spin_unlock(&root->fs_info->free_chunk_lock);
4631
4632         /*
4633          * If we have dup, raid1 or raid10 then only half of the free
4634          * space is actually useable.  For raid56, the space info used
4635          * doesn't include the parity drive, so we don't have to
4636          * change the math
4637          */
4638         if (profile & (BTRFS_BLOCK_GROUP_DUP |
4639                        BTRFS_BLOCK_GROUP_RAID1 |
4640                        BTRFS_BLOCK_GROUP_RAID10))
4641                 avail >>= 1;
4642
4643         /*
4644          * If we aren't flushing all things, let us overcommit up to
4645          * 1/2th of the space. If we can flush, don't let us overcommit
4646          * too much, let it overcommit up to 1/8 of the space.
4647          */
4648         if (flush == BTRFS_RESERVE_FLUSH_ALL)
4649                 avail >>= 3;
4650         else
4651                 avail >>= 1;
4652
4653         if (used + bytes < space_info->total_bytes + avail)
4654                 return 1;
4655         return 0;
4656 }
4657
4658 static void btrfs_writeback_inodes_sb_nr(struct btrfs_root *root,
4659                                          unsigned long nr_pages, int nr_items)
4660 {
4661         struct super_block *sb = root->fs_info->sb;
4662
4663         if (down_read_trylock(&sb->s_umount)) {
4664                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
4665                 up_read(&sb->s_umount);
4666         } else {
4667                 /*
4668                  * We needn't worry the filesystem going from r/w to r/o though
4669                  * we don't acquire ->s_umount mutex, because the filesystem
4670                  * should guarantee the delalloc inodes list be empty after
4671                  * the filesystem is readonly(all dirty pages are written to
4672                  * the disk).
4673                  */
4674                 btrfs_start_delalloc_roots(root->fs_info, 0, nr_items);
4675                 if (!current->journal_info)
4676                         btrfs_wait_ordered_roots(root->fs_info, nr_items,
4677                                                  0, (u64)-1);
4678         }
4679 }
4680
4681 static inline int calc_reclaim_items_nr(struct btrfs_root *root, u64 to_reclaim)
4682 {
4683         u64 bytes;
4684         int nr;
4685
4686         bytes = btrfs_calc_trans_metadata_size(root, 1);
4687         nr = (int)div64_u64(to_reclaim, bytes);
4688         if (!nr)
4689                 nr = 1;
4690         return nr;
4691 }
4692
4693 #define EXTENT_SIZE_PER_ITEM    SZ_256K
4694
4695 /*
4696  * shrink metadata reservation for delalloc
4697  */
4698 static void shrink_delalloc(struct btrfs_root *root, u64 to_reclaim, u64 orig,
4699                             bool wait_ordered)
4700 {
4701         struct btrfs_block_rsv *block_rsv;
4702         struct btrfs_space_info *space_info;
4703         struct btrfs_trans_handle *trans;
4704         u64 delalloc_bytes;
4705         u64 max_reclaim;
4706         long time_left;
4707         unsigned long nr_pages;
4708         int loops;
4709         int items;
4710         enum btrfs_reserve_flush_enum flush;
4711
4712         /* Calc the number of the pages we need flush for space reservation */
4713         items = calc_reclaim_items_nr(root, to_reclaim);
4714         to_reclaim = (u64)items * EXTENT_SIZE_PER_ITEM;
4715
4716         trans = (struct btrfs_trans_handle *)current->journal_info;
4717         block_rsv = &root->fs_info->delalloc_block_rsv;
4718         space_info = block_rsv->space_info;
4719
4720         delalloc_bytes = percpu_counter_sum_positive(
4721                                                 &root->fs_info->delalloc_bytes);
4722         if (delalloc_bytes == 0) {
4723                 if (trans)
4724                         return;
4725                 if (wait_ordered)
4726                         btrfs_wait_ordered_roots(root->fs_info, items,
4727                                                  0, (u64)-1);
4728                 return;
4729         }
4730
4731         loops = 0;
4732         while (delalloc_bytes && loops < 3) {
4733                 max_reclaim = min(delalloc_bytes, to_reclaim);
4734                 nr_pages = max_reclaim >> PAGE_SHIFT;
4735                 btrfs_writeback_inodes_sb_nr(root, nr_pages, items);
4736                 /*
4737                  * We need to wait for the async pages to actually start before
4738                  * we do anything.
4739                  */
4740                 max_reclaim = atomic_read(&root->fs_info->async_delalloc_pages);
4741                 if (!max_reclaim)
4742                         goto skip_async;
4743
4744                 if (max_reclaim <= nr_pages)
4745                         max_reclaim = 0;
4746                 else
4747                         max_reclaim -= nr_pages;
4748
4749                 wait_event(root->fs_info->async_submit_wait,
4750                            atomic_read(&root->fs_info->async_delalloc_pages) <=
4751                            (int)max_reclaim);
4752 skip_async:
4753                 if (!trans)
4754                         flush = BTRFS_RESERVE_FLUSH_ALL;
4755                 else
4756                         flush = BTRFS_RESERVE_NO_FLUSH;
4757                 spin_lock(&space_info->lock);
4758                 if (can_overcommit(root, space_info, orig, flush)) {
4759                         spin_unlock(&space_info->lock);
4760                         break;
4761                 }
4762                 if (list_empty(&space_info->tickets) &&
4763                     list_empty(&space_info->priority_tickets)) {
4764                         spin_unlock(&space_info->lock);
4765                         break;
4766                 }
4767                 spin_unlock(&space_info->lock);
4768
4769                 loops++;
4770                 if (wait_ordered && !trans) {
4771                         btrfs_wait_ordered_roots(root->fs_info, items,
4772                                                  0, (u64)-1);
4773                 } else {
4774                         time_left = schedule_timeout_killable(1);
4775                         if (time_left)
4776                                 break;
4777                 }
4778                 delalloc_bytes = percpu_counter_sum_positive(
4779                                                 &root->fs_info->delalloc_bytes);
4780         }
4781 }
4782
4783 /**
4784  * maybe_commit_transaction - possibly commit the transaction if its ok to
4785  * @root - the root we're allocating for
4786  * @bytes - the number of bytes we want to reserve
4787  * @force - force the commit
4788  *
4789  * This will check to make sure that committing the transaction will actually
4790  * get us somewhere and then commit the transaction if it does.  Otherwise it
4791  * will return -ENOSPC.
4792  */
4793 static int may_commit_transaction(struct btrfs_root *root,
4794                                   struct btrfs_space_info *space_info,
4795                                   u64 bytes, int force)
4796 {
4797         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
4798         struct btrfs_trans_handle *trans;
4799
4800         trans = (struct btrfs_trans_handle *)current->journal_info;
4801         if (trans)
4802                 return -EAGAIN;
4803
4804         if (force)
4805                 goto commit;
4806
4807         /* See if there is enough pinned space to make this reservation */
4808         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4809                                    bytes) >= 0)
4810                 goto commit;
4811
4812         /*
4813          * See if there is some space in the delayed insertion reservation for
4814          * this reservation.
4815          */
4816         if (space_info != delayed_rsv->space_info)
4817                 return -ENOSPC;
4818
4819         spin_lock(&delayed_rsv->lock);
4820         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4821                                    bytes - delayed_rsv->size) >= 0) {
4822                 spin_unlock(&delayed_rsv->lock);
4823                 return -ENOSPC;
4824         }
4825         spin_unlock(&delayed_rsv->lock);
4826
4827 commit:
4828         trans = btrfs_join_transaction(root);
4829         if (IS_ERR(trans))
4830                 return -ENOSPC;
4831
4832         return btrfs_commit_transaction(trans, root);
4833 }
4834
4835 struct reserve_ticket {
4836         u64 bytes;
4837         int error;
4838         struct list_head list;
4839         wait_queue_head_t wait;
4840 };
4841
4842 static int flush_space(struct btrfs_root *root,
4843                        struct btrfs_space_info *space_info, u64 num_bytes,
4844                        u64 orig_bytes, int state)
4845 {
4846         struct btrfs_trans_handle *trans;
4847         int nr;
4848         int ret = 0;
4849
4850         switch (state) {
4851         case FLUSH_DELAYED_ITEMS_NR:
4852         case FLUSH_DELAYED_ITEMS:
4853                 if (state == FLUSH_DELAYED_ITEMS_NR)
4854                         nr = calc_reclaim_items_nr(root, num_bytes) * 2;
4855                 else
4856                         nr = -1;
4857
4858                 trans = btrfs_join_transaction(root);
4859                 if (IS_ERR(trans)) {
4860                         ret = PTR_ERR(trans);
4861                         break;
4862                 }
4863                 ret = btrfs_run_delayed_items_nr(trans, root, nr);
4864                 btrfs_end_transaction(trans, root);
4865                 break;
4866         case FLUSH_DELALLOC:
4867         case FLUSH_DELALLOC_WAIT:
4868                 shrink_delalloc(root, num_bytes * 2, orig_bytes,
4869                                 state == FLUSH_DELALLOC_WAIT);
4870                 break;
4871         case ALLOC_CHUNK:
4872                 trans = btrfs_join_transaction(root);
4873                 if (IS_ERR(trans)) {
4874                         ret = PTR_ERR(trans);
4875                         break;
4876                 }
4877                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
4878                                      btrfs_get_alloc_profile(root, 0),
4879                                      CHUNK_ALLOC_NO_FORCE);
4880                 btrfs_end_transaction(trans, root);
4881                 if (ret > 0 || ret == -ENOSPC)
4882                         ret = 0;
4883                 break;
4884         case COMMIT_TRANS:
4885                 ret = may_commit_transaction(root, space_info, orig_bytes, 0);
4886                 break;
4887         default:
4888                 ret = -ENOSPC;
4889                 break;
4890         }
4891
4892         trace_btrfs_flush_space(root->fs_info, space_info->flags, num_bytes,
4893                                 orig_bytes, state, ret);
4894         return ret;
4895 }
4896
4897 static inline u64
4898 btrfs_calc_reclaim_metadata_size(struct btrfs_root *root,
4899                                  struct btrfs_space_info *space_info)
4900 {
4901         struct reserve_ticket *ticket;
4902         u64 used;
4903         u64 expected;
4904         u64 to_reclaim = 0;
4905
4906         list_for_each_entry(ticket, &space_info->tickets, list)
4907                 to_reclaim += ticket->bytes;
4908         list_for_each_entry(ticket, &space_info->priority_tickets, list)
4909                 to_reclaim += ticket->bytes;
4910         if (to_reclaim)
4911                 return to_reclaim;
4912
4913         to_reclaim = min_t(u64, num_online_cpus() * SZ_1M, SZ_16M);
4914         if (can_overcommit(root, space_info, to_reclaim,
4915                            BTRFS_RESERVE_FLUSH_ALL))
4916                 return 0;
4917
4918         used = space_info->bytes_used + space_info->bytes_reserved +
4919                space_info->bytes_pinned + space_info->bytes_readonly +
4920                space_info->bytes_may_use;
4921         if (can_overcommit(root, space_info, SZ_1M, BTRFS_RESERVE_FLUSH_ALL))
4922                 expected = div_factor_fine(space_info->total_bytes, 95);
4923         else
4924                 expected = div_factor_fine(space_info->total_bytes, 90);
4925
4926         if (used > expected)
4927                 to_reclaim = used - expected;
4928         else
4929                 to_reclaim = 0;
4930         to_reclaim = min(to_reclaim, space_info->bytes_may_use +
4931                                      space_info->bytes_reserved);
4932         return to_reclaim;
4933 }
4934
4935 static inline int need_do_async_reclaim(struct btrfs_space_info *space_info,
4936                                         struct btrfs_root *root, u64 used)
4937 {
4938         u64 thresh = div_factor_fine(space_info->total_bytes, 98);
4939
4940         /* If we're just plain full then async reclaim just slows us down. */
4941         if ((space_info->bytes_used + space_info->bytes_reserved) >= thresh)
4942                 return 0;
4943
4944         if (!btrfs_calc_reclaim_metadata_size(root, space_info))
4945                 return 0;
4946
4947         return (used >= thresh && !btrfs_fs_closing(root->fs_info) &&
4948                 !test_bit(BTRFS_FS_STATE_REMOUNTING,
4949                           &root->fs_info->fs_state));
4950 }
4951
4952 static void wake_all_tickets(struct list_head *head)
4953 {
4954         struct reserve_ticket *ticket;
4955
4956         while (!list_empty(head)) {
4957                 ticket = list_first_entry(head, struct reserve_ticket, list);
4958                 list_del_init(&ticket->list);
4959                 ticket->error = -ENOSPC;
4960                 wake_up(&ticket->wait);
4961         }
4962 }
4963
4964 /*
4965  * This is for normal flushers, we can wait all goddamned day if we want to.  We
4966  * will loop and continuously try to flush as long as we are making progress.
4967  * We count progress as clearing off tickets each time we have to loop.
4968  */
4969 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
4970 {
4971         struct btrfs_fs_info *fs_info;
4972         struct btrfs_space_info *space_info;
4973         u64 to_reclaim;
4974         int flush_state;
4975         int commit_cycles = 0;
4976         u64 last_tickets_id;
4977
4978         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
4979         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4980
4981         spin_lock(&space_info->lock);
4982         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info->fs_root,
4983                                                       space_info);
4984         if (!to_reclaim) {
4985                 space_info->flush = 0;
4986                 spin_unlock(&space_info->lock);
4987                 return;
4988         }
4989         last_tickets_id = space_info->tickets_id;
4990         spin_unlock(&space_info->lock);
4991
4992         flush_state = FLUSH_DELAYED_ITEMS_NR;
4993         do {
4994                 struct reserve_ticket *ticket;
4995                 int ret;
4996
4997                 ret = flush_space(fs_info->fs_root, space_info, to_reclaim,
4998                             to_reclaim, flush_state);
4999                 spin_lock(&space_info->lock);
5000                 if (list_empty(&space_info->tickets)) {
5001                         space_info->flush = 0;
5002                         spin_unlock(&space_info->lock);
5003                         return;
5004                 }
5005                 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info->fs_root,
5006                                                               space_info);
5007                 ticket = list_first_entry(&space_info->tickets,
5008                                           struct reserve_ticket, list);
5009                 if (last_tickets_id == space_info->tickets_id) {
5010                         flush_state++;
5011                 } else {
5012                         last_tickets_id = space_info->tickets_id;
5013                         flush_state = FLUSH_DELAYED_ITEMS_NR;
5014                         if (commit_cycles)
5015                                 commit_cycles--;
5016                 }
5017
5018                 if (flush_state > COMMIT_TRANS) {
5019                         commit_cycles++;
5020                         if (commit_cycles > 2) {
5021                                 wake_all_tickets(&space_info->tickets);
5022                                 space_info->flush = 0;
5023                         } else {
5024                                 flush_state = FLUSH_DELAYED_ITEMS_NR;
5025                         }
5026                 }
5027                 spin_unlock(&space_info->lock);
5028         } while (flush_state <= COMMIT_TRANS);
5029 }
5030
5031 void btrfs_init_async_reclaim_work(struct work_struct *work)
5032 {
5033         INIT_WORK(work, btrfs_async_reclaim_metadata_space);
5034 }
5035
5036 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info,
5037                                             struct btrfs_space_info *space_info,
5038                                             struct reserve_ticket *ticket)
5039 {
5040         u64 to_reclaim;
5041         int flush_state = FLUSH_DELAYED_ITEMS_NR;
5042
5043         spin_lock(&space_info->lock);
5044         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info->fs_root,
5045                                                       space_info);
5046         if (!to_reclaim) {
5047                 spin_unlock(&space_info->lock);
5048                 return;
5049         }
5050         spin_unlock(&space_info->lock);
5051
5052         do {
5053                 flush_space(fs_info->fs_root, space_info, to_reclaim,
5054                             to_reclaim, flush_state);
5055                 flush_state++;
5056                 spin_lock(&space_info->lock);
5057                 if (ticket->bytes == 0) {
5058                         spin_unlock(&space_info->lock);
5059                         return;
5060                 }
5061                 spin_unlock(&space_info->lock);
5062
5063                 /*
5064                  * Priority flushers can't wait on delalloc without
5065                  * deadlocking.
5066                  */
5067                 if (flush_state == FLUSH_DELALLOC ||
5068                     flush_state == FLUSH_DELALLOC_WAIT)
5069                         flush_state = ALLOC_CHUNK;
5070         } while (flush_state < COMMIT_TRANS);
5071 }
5072
5073 static int wait_reserve_ticket(struct btrfs_fs_info *fs_info,
5074                                struct btrfs_space_info *space_info,
5075                                struct reserve_ticket *ticket, u64 orig_bytes)
5076
5077 {
5078         DEFINE_WAIT(wait);
5079         int ret = 0;
5080
5081         spin_lock(&space_info->lock);
5082         while (ticket->bytes > 0 && ticket->error == 0) {
5083                 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE);
5084                 if (ret) {
5085                         ret = -EINTR;
5086                         break;
5087                 }
5088                 spin_unlock(&space_info->lock);
5089
5090                 schedule();
5091
5092                 finish_wait(&ticket->wait, &wait);
5093                 spin_lock(&space_info->lock);
5094         }
5095         if (!ret)
5096                 ret = ticket->error;
5097         if (!list_empty(&ticket->list))
5098                 list_del_init(&ticket->list);
5099         if (ticket->bytes && ticket->bytes < orig_bytes) {
5100                 u64 num_bytes = orig_bytes - ticket->bytes;
5101                 space_info->bytes_may_use -= num_bytes;
5102                 trace_btrfs_space_reservation(fs_info, "space_info",
5103                                               space_info->flags, num_bytes, 0);
5104         }
5105         spin_unlock(&space_info->lock);
5106
5107         return ret;
5108 }
5109
5110 /**
5111  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5112  * @root - the root we're allocating for
5113  * @space_info - the space info we want to allocate from
5114  * @orig_bytes - the number of bytes we want
5115  * @flush - whether or not we can flush to make our reservation
5116  *
5117  * This will reserve orig_bytes number of bytes from the space info associated
5118  * with the block_rsv.  If there is not enough space it will make an attempt to
5119  * flush out space to make room.  It will do this by flushing delalloc if
5120  * possible or committing the transaction.  If flush is 0 then no attempts to
5121  * regain reservations will be made and this will fail if there is not enough
5122  * space already.
5123  */
5124 static int __reserve_metadata_bytes(struct btrfs_root *root,
5125                                     struct btrfs_space_info *space_info,
5126                                     u64 orig_bytes,
5127                                     enum btrfs_reserve_flush_enum flush)
5128 {
5129         struct reserve_ticket ticket;
5130         u64 used;
5131         int ret = 0;
5132
5133         ASSERT(orig_bytes);
5134         ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL);
5135
5136         spin_lock(&space_info->lock);
5137         ret = -ENOSPC;
5138         used = space_info->bytes_used + space_info->bytes_reserved +
5139                 space_info->bytes_pinned + space_info->bytes_readonly +
5140                 space_info->bytes_may_use;
5141
5142         /*
5143          * If we have enough space then hooray, make our reservation and carry
5144          * on.  If not see if we can overcommit, and if we can, hooray carry on.
5145          * If not things get more complicated.
5146          */
5147         if (used + orig_bytes <= space_info->total_bytes) {
5148                 space_info->bytes_may_use += orig_bytes;
5149                 trace_btrfs_space_reservation(root->fs_info, "space_info",
5150                                               space_info->flags, orig_bytes,
5151                                               1);
5152                 ret = 0;
5153         } else if (can_overcommit(root, space_info, orig_bytes, flush)) {
5154                 space_info->bytes_may_use += orig_bytes;
5155                 trace_btrfs_space_reservation(root->fs_info, "space_info",
5156                                               space_info->flags, orig_bytes,
5157                                               1);
5158                 ret = 0;
5159         }
5160
5161         /*
5162          * If we couldn't make a reservation then setup our reservation ticket
5163          * and kick the async worker if it's not already running.
5164          *
5165          * If we are a priority flusher then we just need to add our ticket to
5166          * the list and we will do our own flushing further down.
5167          */
5168         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
5169                 ticket.bytes = orig_bytes;
5170                 ticket.error = 0;
5171                 init_waitqueue_head(&ticket.wait);
5172                 if (flush == BTRFS_RESERVE_FLUSH_ALL) {
5173                         list_add_tail(&ticket.list, &space_info->tickets);
5174                         if (!space_info->flush) {
5175                                 space_info->flush = 1;
5176                                 trace_btrfs_trigger_flush(root->fs_info,
5177                                                           space_info->flags,
5178                                                           orig_bytes, flush,
5179                                                           "enospc");
5180                                 queue_work(system_unbound_wq,
5181                                            &root->fs_info->async_reclaim_work);
5182                         }
5183                 } else {
5184                         list_add_tail(&ticket.list,
5185                                       &space_info->priority_tickets);
5186                 }
5187         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
5188                 used += orig_bytes;
5189                 /*
5190                  * We will do the space reservation dance during log replay,
5191                  * which means we won't have fs_info->fs_root set, so don't do
5192                  * the async reclaim as we will panic.
5193                  */
5194                 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &root->fs_info->flags) &&
5195                     need_do_async_reclaim(space_info, root, used) &&
5196                     !work_busy(&root->fs_info->async_reclaim_work)) {
5197                         trace_btrfs_trigger_flush(root->fs_info,
5198                                                   space_info->flags,
5199                                                   orig_bytes, flush,
5200                                                   "preempt");
5201                         queue_work(system_unbound_wq,
5202                                    &root->fs_info->async_reclaim_work);
5203                 }
5204         }
5205         spin_unlock(&space_info->lock);
5206         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
5207                 return ret;
5208
5209         if (flush == BTRFS_RESERVE_FLUSH_ALL)
5210                 return wait_reserve_ticket(root->fs_info, space_info, &ticket,
5211                                            orig_bytes);
5212
5213         ret = 0;
5214         priority_reclaim_metadata_space(root->fs_info, space_info, &ticket);
5215         spin_lock(&space_info->lock);
5216         if (ticket.bytes) {
5217                 if (ticket.bytes < orig_bytes) {
5218                         u64 num_bytes = orig_bytes - ticket.bytes;
5219                         space_info->bytes_may_use -= num_bytes;
5220                         trace_btrfs_space_reservation(root->fs_info,
5221                                         "space_info", space_info->flags,
5222                                         num_bytes, 0);
5223
5224                 }
5225                 list_del_init(&ticket.list);
5226                 ret = -ENOSPC;
5227         }
5228         spin_unlock(&space_info->lock);
5229         ASSERT(list_empty(&ticket.list));
5230         return ret;
5231 }
5232
5233 /**
5234  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5235  * @root - the root we're allocating for
5236  * @block_rsv - the block_rsv we're allocating for
5237  * @orig_bytes - the number of bytes we want
5238  * @flush - whether or not we can flush to make our reservation
5239  *
5240  * This will reserve orgi_bytes number of bytes from the space info associated
5241  * with the block_rsv.  If there is not enough space it will make an attempt to
5242  * flush out space to make room.  It will do this by flushing delalloc if
5243  * possible or committing the transaction.  If flush is 0 then no attempts to
5244  * regain reservations will be made and this will fail if there is not enough
5245  * space already.
5246  */
5247 static int reserve_metadata_bytes(struct btrfs_root *root,
5248                                   struct btrfs_block_rsv *block_rsv,
5249                                   u64 orig_bytes,
5250                                   enum btrfs_reserve_flush_enum flush)
5251 {
5252         int ret;
5253
5254         ret = __reserve_metadata_bytes(root, block_rsv->space_info, orig_bytes,
5255                                        flush);
5256         if (ret == -ENOSPC &&
5257             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
5258                 struct btrfs_block_rsv *global_rsv =
5259                         &root->fs_info->global_block_rsv;
5260
5261                 if (block_rsv != global_rsv &&
5262                     !block_rsv_use_bytes(global_rsv, orig_bytes))
5263                         ret = 0;
5264         }
5265         if (ret == -ENOSPC)
5266                 trace_btrfs_space_reservation(root->fs_info,
5267                                               "space_info:enospc",
5268                                               block_rsv->space_info->flags,
5269                                               orig_bytes, 1);
5270         return ret;
5271 }
5272
5273 static struct btrfs_block_rsv *get_block_rsv(
5274                                         const struct btrfs_trans_handle *trans,
5275                                         const struct btrfs_root *root)
5276 {
5277         struct btrfs_block_rsv *block_rsv = NULL;
5278
5279         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
5280             (root == root->fs_info->csum_root && trans->adding_csums) ||
5281              (root == root->fs_info->uuid_root))
5282                 block_rsv = trans->block_rsv;
5283
5284         if (!block_rsv)
5285                 block_rsv = root->block_rsv;
5286
5287         if (!block_rsv)
5288                 block_rsv = &root->fs_info->empty_block_rsv;
5289
5290         return block_rsv;
5291 }
5292
5293 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
5294                                u64 num_bytes)
5295 {
5296         int ret = -ENOSPC;
5297         spin_lock(&block_rsv->lock);
5298         if (block_rsv->reserved >= num_bytes) {
5299                 block_rsv->reserved -= num_bytes;
5300                 if (block_rsv->reserved < block_rsv->size)
5301                         block_rsv->full = 0;
5302                 ret = 0;
5303         }
5304         spin_unlock(&block_rsv->lock);
5305         return ret;
5306 }
5307
5308 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
5309                                 u64 num_bytes, int update_size)
5310 {
5311         spin_lock(&block_rsv->lock);
5312         block_rsv->reserved += num_bytes;
5313         if (update_size)
5314                 block_rsv->size += num_bytes;
5315         else if (block_rsv->reserved >= block_rsv->size)
5316                 block_rsv->full = 1;
5317         spin_unlock(&block_rsv->lock);
5318 }
5319
5320 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
5321                              struct btrfs_block_rsv *dest, u64 num_bytes,
5322                              int min_factor)
5323 {
5324         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5325         u64 min_bytes;
5326
5327         if (global_rsv->space_info != dest->space_info)
5328                 return -ENOSPC;
5329
5330         spin_lock(&global_rsv->lock);
5331         min_bytes = div_factor(global_rsv->size, min_factor);
5332         if (global_rsv->reserved < min_bytes + num_bytes) {
5333                 spin_unlock(&global_rsv->lock);
5334                 return -ENOSPC;
5335         }
5336         global_rsv->reserved -= num_bytes;
5337         if (global_rsv->reserved < global_rsv->size)
5338                 global_rsv->full = 0;
5339         spin_unlock(&global_rsv->lock);
5340
5341         block_rsv_add_bytes(dest, num_bytes, 1);
5342         return 0;
5343 }
5344
5345 /*
5346  * This is for space we already have accounted in space_info->bytes_may_use, so
5347  * basically when we're returning space from block_rsv's.
5348  */
5349 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
5350                                      struct btrfs_space_info *space_info,
5351                                      u64 num_bytes)
5352 {
5353         struct reserve_ticket *ticket;
5354         struct list_head *head;
5355         u64 used;
5356         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH;
5357         bool check_overcommit = false;
5358
5359         spin_lock(&space_info->lock);
5360         head = &space_info->priority_tickets;
5361
5362         /*
5363          * If we are over our limit then we need to check and see if we can
5364          * overcommit, and if we can't then we just need to free up our space
5365          * and not satisfy any requests.
5366          */
5367         used = space_info->bytes_used + space_info->bytes_reserved +
5368                 space_info->bytes_pinned + space_info->bytes_readonly +
5369                 space_info->bytes_may_use;
5370         if (used - num_bytes >= space_info->total_bytes)
5371                 check_overcommit = true;
5372 again:
5373         while (!list_empty(head) && num_bytes) {
5374                 ticket = list_first_entry(head, struct reserve_ticket,
5375                                           list);
5376                 /*
5377                  * We use 0 bytes because this space is already reserved, so
5378                  * adding the ticket space would be a double count.
5379                  */
5380                 if (check_overcommit &&
5381                     !can_overcommit(fs_info->extent_root, space_info, 0,
5382                                     flush))
5383                         break;
5384                 if (num_bytes >= ticket->bytes) {
5385                         list_del_init(&ticket->list);
5386                         num_bytes -= ticket->bytes;
5387                         ticket->bytes = 0;
5388                         space_info->tickets_id++;
5389                         wake_up(&ticket->wait);
5390                 } else {
5391                         ticket->bytes -= num_bytes;
5392                         num_bytes = 0;
5393                 }
5394         }
5395
5396         if (num_bytes && head == &space_info->priority_tickets) {
5397                 head = &space_info->tickets;
5398                 flush = BTRFS_RESERVE_FLUSH_ALL;
5399                 goto again;
5400         }
5401         space_info->bytes_may_use -= num_bytes;
5402         trace_btrfs_space_reservation(fs_info, "space_info",
5403                                       space_info->flags, num_bytes, 0);
5404         spin_unlock(&space_info->lock);
5405 }
5406
5407 /*
5408  * This is for newly allocated space that isn't accounted in
5409  * space_info->bytes_may_use yet.  So if we allocate a chunk or unpin an extent
5410  * we use this helper.
5411  */
5412 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
5413                                      struct btrfs_space_info *space_info,
5414                                      u64 num_bytes)
5415 {
5416         struct reserve_ticket *ticket;
5417         struct list_head *head = &space_info->priority_tickets;
5418
5419 again:
5420         while (!list_empty(head) && num_bytes) {
5421                 ticket = list_first_entry(head, struct reserve_ticket,
5422                                           list);
5423                 if (num_bytes >= ticket->bytes) {
5424                         trace_btrfs_space_reservation(fs_info, "space_info",
5425                                                       space_info->flags,
5426                                                       ticket->bytes, 1);
5427                         list_del_init(&ticket->list);
5428                         num_bytes -= ticket->bytes;
5429                         space_info->bytes_may_use += ticket->bytes;
5430                         ticket->bytes = 0;
5431                         space_info->tickets_id++;
5432                         wake_up(&ticket->wait);
5433                 } else {
5434                         trace_btrfs_space_reservation(fs_info, "space_info",
5435                                                       space_info->flags,
5436                                                       num_bytes, 1);
5437                         space_info->bytes_may_use += num_bytes;
5438                         ticket->bytes -= num_bytes;
5439                         num_bytes = 0;
5440                 }
5441         }
5442
5443         if (num_bytes && head == &space_info->priority_tickets) {
5444                 head = &space_info->tickets;
5445                 goto again;
5446         }
5447 }
5448
5449 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
5450                                     struct btrfs_block_rsv *block_rsv,
5451                                     struct btrfs_block_rsv *dest, u64 num_bytes)
5452 {
5453         struct btrfs_space_info *space_info = block_rsv->space_info;
5454
5455         spin_lock(&block_rsv->lock);
5456         if (num_bytes == (u64)-1)
5457                 num_bytes = block_rsv->size;
5458         block_rsv->size -= num_bytes;
5459         if (block_rsv->reserved >= block_rsv->size) {
5460                 num_bytes = block_rsv->reserved - block_rsv->size;
5461                 block_rsv->reserved = block_rsv->size;
5462                 block_rsv->full = 1;
5463         } else {
5464                 num_bytes = 0;
5465         }
5466         spin_unlock(&block_rsv->lock);
5467
5468         if (num_bytes > 0) {
5469                 if (dest) {
5470                         spin_lock(&dest->lock);
5471                         if (!dest->full) {
5472                                 u64 bytes_to_add;
5473
5474                                 bytes_to_add = dest->size - dest->reserved;
5475                                 bytes_to_add = min(num_bytes, bytes_to_add);
5476                                 dest->reserved += bytes_to_add;
5477                                 if (dest->reserved >= dest->size)
5478                                         dest->full = 1;
5479                                 num_bytes -= bytes_to_add;
5480                         }
5481                         spin_unlock(&dest->lock);
5482                 }
5483                 if (num_bytes)
5484                         space_info_add_old_bytes(fs_info, space_info,
5485                                                  num_bytes);
5486         }
5487 }
5488
5489 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src,
5490                             struct btrfs_block_rsv *dst, u64 num_bytes,
5491                             int update_size)
5492 {
5493         int ret;
5494
5495         ret = block_rsv_use_bytes(src, num_bytes);
5496         if (ret)
5497                 return ret;
5498
5499         block_rsv_add_bytes(dst, num_bytes, update_size);
5500         return 0;
5501 }
5502
5503 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
5504 {
5505         memset(rsv, 0, sizeof(*rsv));
5506         spin_lock_init(&rsv->lock);
5507         rsv->type = type;
5508 }
5509
5510 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root,
5511                                               unsigned short type)
5512 {
5513         struct btrfs_block_rsv *block_rsv;
5514         struct btrfs_fs_info *fs_info = root->fs_info;
5515
5516         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
5517         if (!block_rsv)
5518                 return NULL;
5519
5520         btrfs_init_block_rsv(block_rsv, type);
5521         block_rsv->space_info = __find_space_info(fs_info,
5522                                                   BTRFS_BLOCK_GROUP_METADATA);
5523         return block_rsv;
5524 }
5525
5526 void btrfs_free_block_rsv(struct btrfs_root *root,
5527                           struct btrfs_block_rsv *rsv)
5528 {
5529         if (!rsv)
5530                 return;
5531         btrfs_block_rsv_release(root, rsv, (u64)-1);
5532         kfree(rsv);
5533 }
5534
5535 void __btrfs_free_block_rsv(struct btrfs_block_rsv *rsv)
5536 {
5537         kfree(rsv);
5538 }
5539
5540 int btrfs_block_rsv_add(struct btrfs_root *root,
5541                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
5542                         enum btrfs_reserve_flush_enum flush)
5543 {
5544         int ret;
5545
5546         if (num_bytes == 0)
5547                 return 0;
5548
5549         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5550         if (!ret) {
5551                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
5552                 return 0;
5553         }
5554
5555         return ret;
5556 }
5557
5558 int btrfs_block_rsv_check(struct btrfs_root *root,
5559                           struct btrfs_block_rsv *block_rsv, int min_factor)
5560 {
5561         u64 num_bytes = 0;
5562         int ret = -ENOSPC;
5563
5564         if (!block_rsv)
5565                 return 0;
5566
5567         spin_lock(&block_rsv->lock);
5568         num_bytes = div_factor(block_rsv->size, min_factor);
5569         if (block_rsv->reserved >= num_bytes)
5570                 ret = 0;
5571         spin_unlock(&block_rsv->lock);
5572
5573         return ret;
5574 }
5575
5576 int btrfs_block_rsv_refill(struct btrfs_root *root,
5577                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
5578                            enum btrfs_reserve_flush_enum flush)
5579 {
5580         u64 num_bytes = 0;
5581         int ret = -ENOSPC;
5582
5583         if (!block_rsv)
5584                 return 0;
5585
5586         spin_lock(&block_rsv->lock);
5587         num_bytes = min_reserved;
5588         if (block_rsv->reserved >= num_bytes)
5589                 ret = 0;
5590         else
5591                 num_bytes -= block_rsv->reserved;
5592         spin_unlock(&block_rsv->lock);
5593
5594         if (!ret)
5595                 return 0;
5596
5597         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5598         if (!ret) {
5599                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
5600                 return 0;
5601         }
5602
5603         return ret;
5604 }
5605
5606 void btrfs_block_rsv_release(struct btrfs_root *root,
5607                              struct btrfs_block_rsv *block_rsv,
5608                              u64 num_bytes)
5609 {
5610         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
5611         if (global_rsv == block_rsv ||
5612             block_rsv->space_info != global_rsv->space_info)
5613                 global_rsv = NULL;
5614         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
5615                                 num_bytes);
5616 }
5617
5618 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
5619 {
5620         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
5621         struct btrfs_space_info *sinfo = block_rsv->space_info;
5622         u64 num_bytes;
5623
5624         /*
5625          * The global block rsv is based on the size of the extent tree, the
5626          * checksum tree and the root tree.  If the fs is empty we want to set
5627          * it to a minimal amount for safety.
5628          */
5629         num_bytes = btrfs_root_used(&fs_info->extent_root->root_item) +
5630                 btrfs_root_used(&fs_info->csum_root->root_item) +
5631                 btrfs_root_used(&fs_info->tree_root->root_item);
5632         num_bytes = max_t(u64, num_bytes, SZ_16M);
5633
5634         spin_lock(&sinfo->lock);
5635         spin_lock(&block_rsv->lock);
5636
5637         block_rsv->size = min_t(u64, num_bytes, SZ_512M);
5638
5639         if (block_rsv->reserved < block_rsv->size) {
5640                 num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
5641                         sinfo->bytes_reserved + sinfo->bytes_readonly +
5642                         sinfo->bytes_may_use;
5643                 if (sinfo->total_bytes > num_bytes) {
5644                         num_bytes = sinfo->total_bytes - num_bytes;
5645                         num_bytes = min(num_bytes,
5646                                         block_rsv->size - block_rsv->reserved);
5647                         block_rsv->reserved += num_bytes;
5648                         sinfo->bytes_may_use += num_bytes;
5649                         trace_btrfs_space_reservation(fs_info, "space_info",
5650                                                       sinfo->flags, num_bytes,
5651                                                       1);
5652                 }
5653         } else if (block_rsv->reserved > block_rsv->size) {
5654                 num_bytes = block_rsv->reserved - block_rsv->size;
5655                 sinfo->bytes_may_use -= num_bytes;
5656                 trace_btrfs_space_reservation(fs_info, "space_info",
5657                                       sinfo->flags, num_bytes, 0);
5658                 block_rsv->reserved = block_rsv->size;
5659         }
5660
5661         if (block_rsv->reserved == block_rsv->size)
5662                 block_rsv->full = 1;
5663         else
5664                 block_rsv->full = 0;
5665
5666         spin_unlock(&block_rsv->lock);
5667         spin_unlock(&sinfo->lock);
5668 }
5669
5670 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
5671 {
5672         struct btrfs_space_info *space_info;
5673
5674         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
5675         fs_info->chunk_block_rsv.space_info = space_info;
5676
5677         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5678         fs_info->global_block_rsv.space_info = space_info;
5679         fs_info->delalloc_block_rsv.space_info = space_info;
5680         fs_info->trans_block_rsv.space_info = space_info;
5681         fs_info->empty_block_rsv.space_info = space_info;
5682         fs_info->delayed_block_rsv.space_info = space_info;
5683
5684         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
5685         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
5686         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
5687         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
5688         if (fs_info->quota_root)
5689                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
5690         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
5691
5692         update_global_block_rsv(fs_info);
5693 }
5694
5695 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
5696 {
5697         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
5698                                 (u64)-1);
5699         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
5700         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
5701         WARN_ON(fs_info->trans_block_rsv.size > 0);
5702         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
5703         WARN_ON(fs_info->chunk_block_rsv.size > 0);
5704         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
5705         WARN_ON(fs_info->delayed_block_rsv.size > 0);
5706         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
5707 }
5708
5709 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
5710                                   struct btrfs_root *root)
5711 {
5712         if (!trans->block_rsv)
5713                 return;
5714
5715         if (!trans->bytes_reserved)
5716                 return;
5717
5718         trace_btrfs_space_reservation(root->fs_info, "transaction",
5719                                       trans->transid, trans->bytes_reserved, 0);
5720         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
5721         trans->bytes_reserved = 0;
5722 }
5723
5724 /*
5725  * To be called after all the new block groups attached to the transaction
5726  * handle have been created (btrfs_create_pending_block_groups()).
5727  */
5728 void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans)
5729 {
5730         struct btrfs_fs_info *fs_info = trans->fs_info;
5731
5732         if (!trans->chunk_bytes_reserved)
5733                 return;
5734
5735         WARN_ON_ONCE(!list_empty(&trans->new_bgs));
5736
5737         block_rsv_release_bytes(fs_info, &fs_info->chunk_block_rsv, NULL,
5738                                 trans->chunk_bytes_reserved);
5739         trans->chunk_bytes_reserved = 0;
5740 }
5741
5742 /* Can only return 0 or -ENOSPC */
5743 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
5744                                   struct inode *inode)
5745 {
5746         struct btrfs_root *root = BTRFS_I(inode)->root;
5747         /*
5748          * We always use trans->block_rsv here as we will have reserved space
5749          * for our orphan when starting the transaction, using get_block_rsv()
5750          * here will sometimes make us choose the wrong block rsv as we could be
5751          * doing a reloc inode for a non refcounted root.
5752          */
5753         struct btrfs_block_rsv *src_rsv = trans->block_rsv;
5754         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
5755
5756         /*
5757          * We need to hold space in order to delete our orphan item once we've
5758          * added it, so this takes the reservation so we can release it later
5759          * when we are truly done with the orphan item.
5760          */
5761         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
5762         trace_btrfs_space_reservation(root->fs_info, "orphan",
5763                                       btrfs_ino(inode), num_bytes, 1);
5764         return btrfs_block_rsv_migrate(src_rsv, dst_rsv, num_bytes, 1);
5765 }
5766
5767 void btrfs_orphan_release_metadata(struct inode *inode)
5768 {
5769         struct btrfs_root *root = BTRFS_I(inode)->root;
5770         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
5771         trace_btrfs_space_reservation(root->fs_info, "orphan",
5772                                       btrfs_ino(inode), num_bytes, 0);
5773         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
5774 }
5775
5776 /*
5777  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
5778  * root: the root of the parent directory
5779  * rsv: block reservation
5780  * items: the number of items that we need do reservation
5781  * qgroup_reserved: used to return the reserved size in qgroup
5782  *
5783  * This function is used to reserve the space for snapshot/subvolume
5784  * creation and deletion. Those operations are different with the
5785  * common file/directory operations, they change two fs/file trees
5786  * and root tree, the number of items that the qgroup reserves is
5787  * different with the free space reservation. So we can not use
5788  * the space reservation mechanism in start_transaction().
5789  */
5790 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
5791                                      struct btrfs_block_rsv *rsv,
5792                                      int items,
5793                                      u64 *qgroup_reserved,
5794                                      bool use_global_rsv)
5795 {
5796         u64 num_bytes;
5797         int ret;
5798         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
5799
5800         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &root->fs_info->flags)) {
5801                 /* One for parent inode, two for dir entries */
5802                 num_bytes = 3 * root->nodesize;
5803                 ret = btrfs_qgroup_reserve_meta(root, num_bytes);
5804                 if (ret)
5805                         return ret;
5806         } else {
5807                 num_bytes = 0;
5808         }
5809
5810         *qgroup_reserved = num_bytes;
5811
5812         num_bytes = btrfs_calc_trans_metadata_size(root, items);
5813         rsv->space_info = __find_space_info(root->fs_info,
5814                                             BTRFS_BLOCK_GROUP_METADATA);
5815         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
5816                                   BTRFS_RESERVE_FLUSH_ALL);
5817
5818         if (ret == -ENOSPC && use_global_rsv)
5819                 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes, 1);
5820
5821         if (ret && *qgroup_reserved)
5822                 btrfs_qgroup_free_meta(root, *qgroup_reserved);
5823
5824         return ret;
5825 }
5826
5827 void btrfs_subvolume_release_metadata(struct btrfs_root *root,
5828                                       struct btrfs_block_rsv *rsv,
5829                                       u64 qgroup_reserved)
5830 {
5831         btrfs_block_rsv_release(root, rsv, (u64)-1);
5832 }
5833
5834 /**
5835  * drop_outstanding_extent - drop an outstanding extent
5836  * @inode: the inode we're dropping the extent for
5837  * @num_bytes: the number of bytes we're releasing.
5838  *
5839  * This is called when we are freeing up an outstanding extent, either called
5840  * after an error or after an extent is written.  This will return the number of
5841  * reserved extents that need to be freed.  This must be called with
5842  * BTRFS_I(inode)->lock held.
5843  */
5844 static unsigned drop_outstanding_extent(struct inode *inode, u64 num_bytes)
5845 {
5846         unsigned drop_inode_space = 0;
5847         unsigned dropped_extents = 0;
5848         unsigned num_extents = 0;
5849
5850         num_extents = (unsigned)div64_u64(num_bytes +
5851                                           BTRFS_MAX_EXTENT_SIZE - 1,
5852                                           BTRFS_MAX_EXTENT_SIZE);
5853         ASSERT(num_extents);
5854         ASSERT(BTRFS_I(inode)->outstanding_extents >= num_extents);
5855         BTRFS_I(inode)->outstanding_extents -= num_extents;
5856
5857         if (BTRFS_I(inode)->outstanding_extents == 0 &&
5858             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5859                                &BTRFS_I(inode)->runtime_flags))
5860                 drop_inode_space = 1;
5861
5862         /*
5863          * If we have more or the same amount of outstanding extents than we have
5864          * reserved then we need to leave the reserved extents count alone.
5865          */
5866         if (BTRFS_I(inode)->outstanding_extents >=
5867             BTRFS_I(inode)->reserved_extents)
5868                 return drop_inode_space;
5869
5870         dropped_extents = BTRFS_I(inode)->reserved_extents -
5871                 BTRFS_I(inode)->outstanding_extents;
5872         BTRFS_I(inode)->reserved_extents -= dropped_extents;
5873         return dropped_extents + drop_inode_space;
5874 }
5875
5876 /**
5877  * calc_csum_metadata_size - return the amount of metadata space that must be
5878  *      reserved/freed for the given bytes.
5879  * @inode: the inode we're manipulating
5880  * @num_bytes: the number of bytes in question
5881  * @reserve: 1 if we are reserving space, 0 if we are freeing space
5882  *
5883  * This adjusts the number of csum_bytes in the inode and then returns the
5884  * correct amount of metadata that must either be reserved or freed.  We
5885  * calculate how many checksums we can fit into one leaf and then divide the
5886  * number of bytes that will need to be checksumed by this value to figure out
5887  * how many checksums will be required.  If we are adding bytes then the number
5888  * may go up and we will return the number of additional bytes that must be
5889  * reserved.  If it is going down we will return the number of bytes that must
5890  * be freed.
5891  *
5892  * This must be called with BTRFS_I(inode)->lock held.
5893  */
5894 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
5895                                    int reserve)
5896 {
5897         struct btrfs_root *root = BTRFS_I(inode)->root;
5898         u64 old_csums, num_csums;
5899
5900         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
5901             BTRFS_I(inode)->csum_bytes == 0)
5902                 return 0;
5903
5904         old_csums = btrfs_csum_bytes_to_leaves(root, BTRFS_I(inode)->csum_bytes);
5905         if (reserve)
5906                 BTRFS_I(inode)->csum_bytes += num_bytes;
5907         else
5908                 BTRFS_I(inode)->csum_bytes -= num_bytes;
5909         num_csums = btrfs_csum_bytes_to_leaves(root, BTRFS_I(inode)->csum_bytes);
5910
5911         /* No change, no need to reserve more */
5912         if (old_csums == num_csums)
5913                 return 0;
5914
5915         if (reserve)
5916                 return btrfs_calc_trans_metadata_size(root,
5917                                                       num_csums - old_csums);
5918
5919         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
5920 }
5921
5922 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
5923 {
5924         struct btrfs_root *root = BTRFS_I(inode)->root;
5925         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
5926         u64 to_reserve = 0;
5927         u64 csum_bytes;
5928         unsigned nr_extents = 0;
5929         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
5930         int ret = 0;
5931         bool delalloc_lock = true;
5932         u64 to_free = 0;
5933         unsigned dropped;
5934         bool release_extra = false;
5935
5936         /* If we are a free space inode we need to not flush since we will be in
5937          * the middle of a transaction commit.  We also don't need the delalloc
5938          * mutex since we won't race with anybody.  We need this mostly to make
5939          * lockdep shut its filthy mouth.
5940          *
5941          * If we have a transaction open (can happen if we call truncate_block
5942          * from truncate), then we need FLUSH_LIMIT so we don't deadlock.
5943          */
5944         if (btrfs_is_free_space_inode(inode)) {
5945                 flush = BTRFS_RESERVE_NO_FLUSH;
5946                 delalloc_lock = false;
5947         } else if (current->journal_info) {
5948                 flush = BTRFS_RESERVE_FLUSH_LIMIT;
5949         }
5950
5951         if (flush != BTRFS_RESERVE_NO_FLUSH &&
5952             btrfs_transaction_in_commit(root->fs_info))
5953                 schedule_timeout(1);
5954
5955         if (delalloc_lock)
5956                 mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
5957
5958         num_bytes = ALIGN(num_bytes, root->sectorsize);
5959
5960         spin_lock(&BTRFS_I(inode)->lock);
5961         nr_extents = (unsigned)div64_u64(num_bytes +
5962                                          BTRFS_MAX_EXTENT_SIZE - 1,
5963                                          BTRFS_MAX_EXTENT_SIZE);
5964         BTRFS_I(inode)->outstanding_extents += nr_extents;
5965
5966         nr_extents = 0;
5967         if (BTRFS_I(inode)->outstanding_extents >
5968             BTRFS_I(inode)->reserved_extents)
5969                 nr_extents += BTRFS_I(inode)->outstanding_extents -
5970                         BTRFS_I(inode)->reserved_extents;
5971
5972         /* We always want to reserve a slot for updating the inode. */
5973         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents + 1);
5974         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
5975         csum_bytes = BTRFS_I(inode)->csum_bytes;
5976         spin_unlock(&BTRFS_I(inode)->lock);
5977
5978         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &root->fs_info->flags)) {
5979                 ret = btrfs_qgroup_reserve_meta(root,
5980                                 nr_extents * root->nodesize);
5981                 if (ret)
5982                         goto out_fail;
5983         }
5984
5985         ret = btrfs_block_rsv_add(root, block_rsv, to_reserve, flush);
5986         if (unlikely(ret)) {
5987                 btrfs_qgroup_free_meta(root, nr_extents * root->nodesize);
5988                 goto out_fail;
5989         }
5990
5991         spin_lock(&BTRFS_I(inode)->lock);
5992         if (test_and_set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5993                              &BTRFS_I(inode)->runtime_flags)) {
5994                 to_reserve -= btrfs_calc_trans_metadata_size(root, 1);
5995                 release_extra = true;
5996         }
5997         BTRFS_I(inode)->reserved_extents += nr_extents;
5998         spin_unlock(&BTRFS_I(inode)->lock);
5999
6000         if (delalloc_lock)
6001                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
6002
6003         if (to_reserve)
6004                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
6005                                               btrfs_ino(inode), to_reserve, 1);
6006         if (release_extra)
6007                 btrfs_block_rsv_release(root, block_rsv,
6008                                         btrfs_calc_trans_metadata_size(root,
6009                                                                        1));
6010         return 0;
6011
6012 out_fail:
6013         spin_lock(&BTRFS_I(inode)->lock);
6014         dropped = drop_outstanding_extent(inode, num_bytes);
6015         /*
6016          * If the inodes csum_bytes is the same as the original
6017          * csum_bytes then we know we haven't raced with any free()ers
6018          * so we can just reduce our inodes csum bytes and carry on.
6019          */
6020         if (BTRFS_I(inode)->csum_bytes == csum_bytes) {
6021                 calc_csum_metadata_size(inode, num_bytes, 0);
6022         } else {
6023                 u64 orig_csum_bytes = BTRFS_I(inode)->csum_bytes;
6024                 u64 bytes;
6025
6026                 /*
6027                  * This is tricky, but first we need to figure out how much we
6028                  * freed from any free-ers that occurred during this
6029                  * reservation, so we reset ->csum_bytes to the csum_bytes
6030                  * before we dropped our lock, and then call the free for the
6031                  * number of bytes that were freed while we were trying our
6032                  * reservation.
6033                  */
6034                 bytes = csum_bytes - BTRFS_I(inode)->csum_bytes;
6035                 BTRFS_I(inode)->csum_bytes = csum_bytes;
6036                 to_free = calc_csum_metadata_size(inode, bytes, 0);
6037
6038
6039                 /*
6040                  * Now we need to see how much we would have freed had we not
6041                  * been making this reservation and our ->csum_bytes were not
6042                  * artificially inflated.
6043                  */
6044                 BTRFS_I(inode)->csum_bytes = csum_bytes - num_bytes;
6045                 bytes = csum_bytes - orig_csum_bytes;
6046                 bytes = calc_csum_metadata_size(inode, bytes, 0);
6047
6048                 /*
6049                  * Now reset ->csum_bytes to what it should be.  If bytes is
6050                  * more than to_free then we would have freed more space had we
6051                  * not had an artificially high ->csum_bytes, so we need to free
6052                  * the remainder.  If bytes is the same or less then we don't
6053                  * need to do anything, the other free-ers did the correct
6054                  * thing.
6055                  */
6056                 BTRFS_I(inode)->csum_bytes = orig_csum_bytes - num_bytes;
6057                 if (bytes > to_free)
6058                         to_free = bytes - to_free;
6059                 else
6060                         to_free = 0;
6061         }
6062         spin_unlock(&BTRFS_I(inode)->lock);
6063         if (dropped)
6064                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
6065
6066         if (to_free) {
6067                 btrfs_block_rsv_release(root, block_rsv, to_free);
6068                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
6069                                               btrfs_ino(inode), to_free, 0);
6070         }
6071         if (delalloc_lock)
6072                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
6073         return ret;
6074 }
6075
6076 /**
6077  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
6078  * @inode: the inode to release the reservation for
6079  * @num_bytes: the number of bytes we're releasing
6080  *
6081  * This will release the metadata reservation for an inode.  This can be called
6082  * once we complete IO for a given set of bytes to release their metadata
6083  * reservations.
6084  */
6085 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
6086 {
6087         struct btrfs_root *root = BTRFS_I(inode)->root;
6088         u64 to_free = 0;
6089         unsigned dropped;
6090
6091         num_bytes = ALIGN(num_bytes, root->sectorsize);
6092         spin_lock(&BTRFS_I(inode)->lock);
6093         dropped = drop_outstanding_extent(inode, num_bytes);
6094
6095         if (num_bytes)
6096                 to_free = calc_csum_metadata_size(inode, num_bytes, 0);
6097         spin_unlock(&BTRFS_I(inode)->lock);
6098         if (dropped > 0)
6099                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
6100
6101         if (btrfs_is_testing(root->fs_info))
6102                 return;
6103
6104         trace_btrfs_space_reservation(root->fs_info, "delalloc",
6105                                       btrfs_ino(inode), to_free, 0);
6106
6107         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
6108                                 to_free);
6109 }
6110
6111 /**
6112  * btrfs_delalloc_reserve_space - reserve data and metadata space for
6113  * delalloc
6114  * @inode: inode we're writing to
6115  * @start: start range we are writing to
6116  * @len: how long the range we are writing to
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                 goto out_unlock;
8894
8895         if (unlikely(wc->refs[level - 1] == 0)) {
8896                 btrfs_err(root->fs_info, "Missing references.");
8897                 ret = -EIO;
8898                 goto out_unlock;
8899         }
8900         *lookup_info = 0;
8901
8902         if (wc->stage == DROP_REFERENCE) {
8903                 if (wc->refs[level - 1] > 1) {
8904                         need_account = true;
8905                         if (level == 1 &&
8906                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8907                                 goto skip;
8908
8909                         if (!wc->update_ref ||
8910                             generation <= root->root_key.offset)
8911                                 goto skip;
8912
8913                         btrfs_node_key_to_cpu(path->nodes[level], &key,
8914                                               path->slots[level]);
8915                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
8916                         if (ret < 0)
8917                                 goto skip;
8918
8919                         wc->stage = UPDATE_BACKREF;
8920                         wc->shared_level = level - 1;
8921                 }
8922         } else {
8923                 if (level == 1 &&
8924                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8925                         goto skip;
8926         }
8927
8928         if (!btrfs_buffer_uptodate(next, generation, 0)) {
8929                 btrfs_tree_unlock(next);
8930                 free_extent_buffer(next);
8931                 next = NULL;
8932                 *lookup_info = 1;
8933         }
8934
8935         if (!next) {
8936                 if (reada && level == 1)
8937                         reada_walk_down(trans, root, wc, path);
8938                 next = read_tree_block(root, bytenr, generation);
8939                 if (IS_ERR(next)) {
8940                         return PTR_ERR(next);
8941                 } else if (!extent_buffer_uptodate(next)) {
8942                         free_extent_buffer(next);
8943                         return -EIO;
8944                 }
8945                 btrfs_tree_lock(next);
8946                 btrfs_set_lock_blocking(next);
8947         }
8948
8949         level--;
8950         ASSERT(level == btrfs_header_level(next));
8951         if (level != btrfs_header_level(next)) {
8952                 btrfs_err(root->fs_info, "mismatched level");
8953                 ret = -EIO;
8954                 goto out_unlock;
8955         }
8956         path->nodes[level] = next;
8957         path->slots[level] = 0;
8958         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8959         wc->level = level;
8960         if (wc->level == 1)
8961                 wc->reada_slot = 0;
8962         return 0;
8963 skip:
8964         wc->refs[level - 1] = 0;
8965         wc->flags[level - 1] = 0;
8966         if (wc->stage == DROP_REFERENCE) {
8967                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
8968                         parent = path->nodes[level]->start;
8969                 } else {
8970                         ASSERT(root->root_key.objectid ==
8971                                btrfs_header_owner(path->nodes[level]));
8972                         if (root->root_key.objectid !=
8973                             btrfs_header_owner(path->nodes[level])) {
8974                                 btrfs_err(root->fs_info,
8975                                                 "mismatched block owner");
8976                                 ret = -EIO;
8977                                 goto out_unlock;
8978                         }
8979                         parent = 0;
8980                 }
8981
8982                 if (need_account) {
8983                         ret = account_shared_subtree(trans, root, next,
8984                                                      generation, level - 1);
8985                         if (ret) {
8986                                 btrfs_err_rl(root->fs_info,
8987                                              "Error %d accounting shared subtree. Quota is out of sync, rescan required.",
8988                                              ret);
8989                         }
8990                 }
8991                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
8992                                 root->root_key.objectid, level - 1, 0);
8993                 if (ret)
8994                         goto out_unlock;
8995         }
8996
8997         *lookup_info = 1;
8998         ret = 1;
8999
9000 out_unlock:
9001         btrfs_tree_unlock(next);
9002         free_extent_buffer(next);
9003
9004         return ret;
9005 }
9006
9007 /*
9008  * helper to process tree block while walking up the tree.
9009  *
9010  * when wc->stage == DROP_REFERENCE, this function drops
9011  * reference count on the block.
9012  *
9013  * when wc->stage == UPDATE_BACKREF, this function changes
9014  * wc->stage back to DROP_REFERENCE if we changed wc->stage
9015  * to UPDATE_BACKREF previously while processing the block.
9016  *
9017  * NOTE: return value 1 means we should stop walking up.
9018  */
9019 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
9020                                  struct btrfs_root *root,
9021                                  struct btrfs_path *path,
9022                                  struct walk_control *wc)
9023 {
9024         int ret;
9025         int level = wc->level;
9026         struct extent_buffer *eb = path->nodes[level];
9027         u64 parent = 0;
9028
9029         if (wc->stage == UPDATE_BACKREF) {
9030                 BUG_ON(wc->shared_level < level);
9031                 if (level < wc->shared_level)
9032                         goto out;
9033
9034                 ret = find_next_key(path, level + 1, &wc->update_progress);
9035                 if (ret > 0)
9036                         wc->update_ref = 0;
9037
9038                 wc->stage = DROP_REFERENCE;
9039                 wc->shared_level = -1;
9040                 path->slots[level] = 0;
9041
9042                 /*
9043                  * check reference count again if the block isn't locked.
9044                  * we should start walking down the tree again if reference
9045                  * count is one.
9046                  */
9047                 if (!path->locks[level]) {
9048                         BUG_ON(level == 0);
9049                         btrfs_tree_lock(eb);
9050                         btrfs_set_lock_blocking(eb);
9051                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9052
9053                         ret = btrfs_lookup_extent_info(trans, root,
9054                                                        eb->start, level, 1,
9055                                                        &wc->refs[level],
9056                                                        &wc->flags[level]);
9057                         if (ret < 0) {
9058                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
9059                                 path->locks[level] = 0;
9060                                 return ret;
9061                         }
9062                         BUG_ON(wc->refs[level] == 0);
9063                         if (wc->refs[level] == 1) {
9064                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
9065                                 path->locks[level] = 0;
9066                                 return 1;
9067                         }
9068                 }
9069         }
9070
9071         /* wc->stage == DROP_REFERENCE */
9072         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
9073
9074         if (wc->refs[level] == 1) {
9075                 if (level == 0) {
9076                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
9077                                 ret = btrfs_dec_ref(trans, root, eb, 1);
9078                         else
9079                                 ret = btrfs_dec_ref(trans, root, eb, 0);
9080                         BUG_ON(ret); /* -ENOMEM */
9081                         ret = account_leaf_items(trans, root, eb);
9082                         if (ret) {
9083                                 btrfs_err_rl(root->fs_info,
9084                                              "error %d accounting leaf items. Quota is out of sync, rescan required.",
9085                                              ret);
9086                         }
9087                 }
9088                 /* make block locked assertion in clean_tree_block happy */
9089                 if (!path->locks[level] &&
9090                     btrfs_header_generation(eb) == trans->transid) {
9091                         btrfs_tree_lock(eb);
9092                         btrfs_set_lock_blocking(eb);
9093                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9094                 }
9095                 clean_tree_block(trans, root->fs_info, eb);
9096         }
9097
9098         if (eb == root->node) {
9099                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
9100                         parent = eb->start;
9101                 else
9102                         BUG_ON(root->root_key.objectid !=
9103                                btrfs_header_owner(eb));
9104         } else {
9105                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
9106                         parent = path->nodes[level + 1]->start;
9107                 else
9108                         BUG_ON(root->root_key.objectid !=
9109                                btrfs_header_owner(path->nodes[level + 1]));
9110         }
9111
9112         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
9113 out:
9114         wc->refs[level] = 0;
9115         wc->flags[level] = 0;
9116         return 0;
9117 }
9118
9119 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
9120                                    struct btrfs_root *root,
9121                                    struct btrfs_path *path,
9122                                    struct walk_control *wc)
9123 {
9124         int level = wc->level;
9125         int lookup_info = 1;
9126         int ret;
9127
9128         while (level >= 0) {
9129                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
9130                 if (ret > 0)
9131                         break;
9132
9133                 if (level == 0)
9134                         break;
9135
9136                 if (path->slots[level] >=
9137                     btrfs_header_nritems(path->nodes[level]))
9138                         break;
9139
9140                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
9141                 if (ret > 0) {
9142                         path->slots[level]++;
9143                         continue;
9144                 } else if (ret < 0)
9145                         return ret;
9146                 level = wc->level;
9147         }
9148         return 0;
9149 }
9150
9151 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
9152                                  struct btrfs_root *root,
9153                                  struct btrfs_path *path,
9154                                  struct walk_control *wc, int max_level)
9155 {
9156         int level = wc->level;
9157         int ret;
9158
9159         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
9160         while (level < max_level && path->nodes[level]) {
9161                 wc->level = level;
9162                 if (path->slots[level] + 1 <
9163                     btrfs_header_nritems(path->nodes[level])) {
9164                         path->slots[level]++;
9165                         return 0;
9166                 } else {
9167                         ret = walk_up_proc(trans, root, path, wc);
9168                         if (ret > 0)
9169                                 return 0;
9170
9171                         if (path->locks[level]) {
9172                                 btrfs_tree_unlock_rw(path->nodes[level],
9173                                                      path->locks[level]);
9174                                 path->locks[level] = 0;
9175                         }
9176                         free_extent_buffer(path->nodes[level]);
9177                         path->nodes[level] = NULL;
9178                         level++;
9179                 }
9180         }
9181         return 1;
9182 }
9183
9184 /*
9185  * drop a subvolume tree.
9186  *
9187  * this function traverses the tree freeing any blocks that only
9188  * referenced by the tree.
9189  *
9190  * when a shared tree block is found. this function decreases its
9191  * reference count by one. if update_ref is true, this function
9192  * also make sure backrefs for the shared block and all lower level
9193  * blocks are properly updated.
9194  *
9195  * If called with for_reloc == 0, may exit early with -EAGAIN
9196  */
9197 int btrfs_drop_snapshot(struct btrfs_root *root,
9198                          struct btrfs_block_rsv *block_rsv, int update_ref,
9199                          int for_reloc)
9200 {
9201         struct btrfs_fs_info *fs_info = root->fs_info;
9202         struct btrfs_path *path;
9203         struct btrfs_trans_handle *trans;
9204         struct btrfs_root *tree_root = fs_info->tree_root;
9205         struct btrfs_root_item *root_item = &root->root_item;
9206         struct walk_control *wc;
9207         struct btrfs_key key;
9208         int err = 0;
9209         int ret;
9210         int level;
9211         bool root_dropped = false;
9212
9213         btrfs_debug(fs_info, "Drop subvolume %llu", root->objectid);
9214
9215         path = btrfs_alloc_path();
9216         if (!path) {
9217                 err = -ENOMEM;
9218                 goto out;
9219         }
9220
9221         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9222         if (!wc) {
9223                 btrfs_free_path(path);
9224                 err = -ENOMEM;
9225                 goto out;
9226         }
9227
9228         trans = btrfs_start_transaction(tree_root, 0);
9229         if (IS_ERR(trans)) {
9230                 err = PTR_ERR(trans);
9231                 goto out_free;
9232         }
9233
9234         if (block_rsv)
9235                 trans->block_rsv = block_rsv;
9236
9237         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
9238                 level = btrfs_header_level(root->node);
9239                 path->nodes[level] = btrfs_lock_root_node(root);
9240                 btrfs_set_lock_blocking(path->nodes[level]);
9241                 path->slots[level] = 0;
9242                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9243                 memset(&wc->update_progress, 0,
9244                        sizeof(wc->update_progress));
9245         } else {
9246                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
9247                 memcpy(&wc->update_progress, &key,
9248                        sizeof(wc->update_progress));
9249
9250                 level = root_item->drop_level;
9251                 BUG_ON(level == 0);
9252                 path->lowest_level = level;
9253                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
9254                 path->lowest_level = 0;
9255                 if (ret < 0) {
9256                         err = ret;
9257                         goto out_end_trans;
9258                 }
9259                 WARN_ON(ret > 0);
9260
9261                 /*
9262                  * unlock our path, this is safe because only this
9263                  * function is allowed to delete this snapshot
9264                  */
9265                 btrfs_unlock_up_safe(path, 0);
9266
9267                 level = btrfs_header_level(root->node);
9268                 while (1) {
9269                         btrfs_tree_lock(path->nodes[level]);
9270                         btrfs_set_lock_blocking(path->nodes[level]);
9271                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9272
9273                         ret = btrfs_lookup_extent_info(trans, root,
9274                                                 path->nodes[level]->start,
9275                                                 level, 1, &wc->refs[level],
9276                                                 &wc->flags[level]);
9277                         if (ret < 0) {
9278                                 err = ret;
9279                                 goto out_end_trans;
9280                         }
9281                         BUG_ON(wc->refs[level] == 0);
9282
9283                         if (level == root_item->drop_level)
9284                                 break;
9285
9286                         btrfs_tree_unlock(path->nodes[level]);
9287                         path->locks[level] = 0;
9288                         WARN_ON(wc->refs[level] != 1);
9289                         level--;
9290                 }
9291         }
9292
9293         wc->level = level;
9294         wc->shared_level = -1;
9295         wc->stage = DROP_REFERENCE;
9296         wc->update_ref = update_ref;
9297         wc->keep_locks = 0;
9298         wc->for_reloc = for_reloc;
9299         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
9300
9301         while (1) {
9302
9303                 ret = walk_down_tree(trans, root, path, wc);
9304                 if (ret < 0) {
9305                         err = ret;
9306                         break;
9307                 }
9308
9309                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
9310                 if (ret < 0) {
9311                         err = ret;
9312                         break;
9313                 }
9314
9315                 if (ret > 0) {
9316                         BUG_ON(wc->stage != DROP_REFERENCE);
9317                         break;
9318                 }
9319
9320                 if (wc->stage == DROP_REFERENCE) {
9321                         level = wc->level;
9322                         btrfs_node_key(path->nodes[level],
9323                                        &root_item->drop_progress,
9324                                        path->slots[level]);
9325                         root_item->drop_level = level;
9326                 }
9327
9328                 BUG_ON(wc->level == 0);
9329                 if (btrfs_should_end_transaction(trans, tree_root) ||
9330                     (!for_reloc && btrfs_need_cleaner_sleep(root))) {
9331                         ret = btrfs_update_root(trans, tree_root,
9332                                                 &root->root_key,
9333                                                 root_item);
9334                         if (ret) {
9335                                 btrfs_abort_transaction(trans, ret);
9336                                 err = ret;
9337                                 goto out_end_trans;
9338                         }
9339
9340                         btrfs_end_transaction_throttle(trans, tree_root);
9341                         if (!for_reloc && btrfs_need_cleaner_sleep(root)) {
9342                                 btrfs_debug(fs_info,
9343                                             "drop snapshot early exit");
9344                                 err = -EAGAIN;
9345                                 goto out_free;
9346                         }
9347
9348                         trans = btrfs_start_transaction(tree_root, 0);
9349                         if (IS_ERR(trans)) {
9350                                 err = PTR_ERR(trans);
9351                                 goto out_free;
9352                         }
9353                         if (block_rsv)
9354                                 trans->block_rsv = block_rsv;
9355                 }
9356         }
9357         btrfs_release_path(path);
9358         if (err)
9359                 goto out_end_trans;
9360
9361         ret = btrfs_del_root(trans, tree_root, &root->root_key);
9362         if (ret) {
9363                 btrfs_abort_transaction(trans, ret);
9364                 goto out_end_trans;
9365         }
9366
9367         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
9368                 ret = btrfs_find_root(tree_root, &root->root_key, path,
9369                                       NULL, NULL);
9370                 if (ret < 0) {
9371                         btrfs_abort_transaction(trans, ret);
9372                         err = ret;
9373                         goto out_end_trans;
9374                 } else if (ret > 0) {
9375                         /* if we fail to delete the orphan item this time
9376                          * around, it'll get picked up the next time.
9377                          *
9378                          * The most common failure here is just -ENOENT.
9379                          */
9380                         btrfs_del_orphan_item(trans, tree_root,
9381                                               root->root_key.objectid);
9382                 }
9383         }
9384
9385         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
9386                 btrfs_add_dropped_root(trans, root);
9387         } else {
9388                 free_extent_buffer(root->node);
9389                 free_extent_buffer(root->commit_root);
9390                 btrfs_put_fs_root(root);
9391         }
9392         root_dropped = true;
9393 out_end_trans:
9394         btrfs_end_transaction_throttle(trans, tree_root);
9395 out_free:
9396         kfree(wc);
9397         btrfs_free_path(path);
9398 out:
9399         /*
9400          * So if we need to stop dropping the snapshot for whatever reason we
9401          * need to make sure to add it back to the dead root list so that we
9402          * keep trying to do the work later.  This also cleans up roots if we
9403          * don't have it in the radix (like when we recover after a power fail
9404          * or unmount) so we don't leak memory.
9405          */
9406         if (!for_reloc && root_dropped == false)
9407                 btrfs_add_dead_root(root);
9408         if (err && err != -EAGAIN)
9409                 btrfs_handle_fs_error(fs_info, err, NULL);
9410         return err;
9411 }
9412
9413 /*
9414  * drop subtree rooted at tree block 'node'.
9415  *
9416  * NOTE: this function will unlock and release tree block 'node'
9417  * only used by relocation code
9418  */
9419 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
9420                         struct btrfs_root *root,
9421                         struct extent_buffer *node,
9422                         struct extent_buffer *parent)
9423 {
9424         struct btrfs_path *path;
9425         struct walk_control *wc;
9426         int level;
9427         int parent_level;
9428         int ret = 0;
9429         int wret;
9430
9431         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
9432
9433         path = btrfs_alloc_path();
9434         if (!path)
9435                 return -ENOMEM;
9436
9437         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9438         if (!wc) {
9439                 btrfs_free_path(path);
9440                 return -ENOMEM;
9441         }
9442
9443         btrfs_assert_tree_locked(parent);
9444         parent_level = btrfs_header_level(parent);
9445         extent_buffer_get(parent);
9446         path->nodes[parent_level] = parent;
9447         path->slots[parent_level] = btrfs_header_nritems(parent);
9448
9449         btrfs_assert_tree_locked(node);
9450         level = btrfs_header_level(node);
9451         path->nodes[level] = node;
9452         path->slots[level] = 0;
9453         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9454
9455         wc->refs[parent_level] = 1;
9456         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
9457         wc->level = level;
9458         wc->shared_level = -1;
9459         wc->stage = DROP_REFERENCE;
9460         wc->update_ref = 0;
9461         wc->keep_locks = 1;
9462         wc->for_reloc = 1;
9463         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
9464
9465         while (1) {
9466                 wret = walk_down_tree(trans, root, path, wc);
9467                 if (wret < 0) {
9468                         ret = wret;
9469                         break;
9470                 }
9471
9472                 wret = walk_up_tree(trans, root, path, wc, parent_level);
9473                 if (wret < 0)
9474                         ret = wret;
9475                 if (wret != 0)
9476                         break;
9477         }
9478
9479         kfree(wc);
9480         btrfs_free_path(path);
9481         return ret;
9482 }
9483
9484 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
9485 {
9486         u64 num_devices;
9487         u64 stripped;
9488
9489         /*
9490          * if restripe for this chunk_type is on pick target profile and
9491          * return, otherwise do the usual balance
9492          */
9493         stripped = get_restripe_target(root->fs_info, flags);
9494         if (stripped)
9495                 return extended_to_chunk(stripped);
9496
9497         num_devices = root->fs_info->fs_devices->rw_devices;
9498
9499         stripped = BTRFS_BLOCK_GROUP_RAID0 |
9500                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
9501                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
9502
9503         if (num_devices == 1) {
9504                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9505                 stripped = flags & ~stripped;
9506
9507                 /* turn raid0 into single device chunks */
9508                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
9509                         return stripped;
9510
9511                 /* turn mirroring into duplication */
9512                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
9513                              BTRFS_BLOCK_GROUP_RAID10))
9514                         return stripped | BTRFS_BLOCK_GROUP_DUP;
9515         } else {
9516                 /* they already had raid on here, just return */
9517                 if (flags & stripped)
9518                         return flags;
9519
9520                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9521                 stripped = flags & ~stripped;
9522
9523                 /* switch duplicated blocks with raid1 */
9524                 if (flags & BTRFS_BLOCK_GROUP_DUP)
9525                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
9526
9527                 /* this is drive concat, leave it alone */
9528         }
9529
9530         return flags;
9531 }
9532
9533 static int inc_block_group_ro(struct btrfs_block_group_cache *cache, int force)
9534 {
9535         struct btrfs_space_info *sinfo = cache->space_info;
9536         u64 num_bytes;
9537         u64 min_allocable_bytes;
9538         int ret = -ENOSPC;
9539
9540         /*
9541          * We need some metadata space and system metadata space for
9542          * allocating chunks in some corner cases until we force to set
9543          * it to be readonly.
9544          */
9545         if ((sinfo->flags &
9546              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
9547             !force)
9548                 min_allocable_bytes = SZ_1M;
9549         else
9550                 min_allocable_bytes = 0;
9551
9552         spin_lock(&sinfo->lock);
9553         spin_lock(&cache->lock);
9554
9555         if (cache->ro) {
9556                 cache->ro++;
9557                 ret = 0;
9558                 goto out;
9559         }
9560
9561         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
9562                     cache->bytes_super - btrfs_block_group_used(&cache->item);
9563
9564         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
9565             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
9566             min_allocable_bytes <= sinfo->total_bytes) {
9567                 sinfo->bytes_readonly += num_bytes;
9568                 cache->ro++;
9569                 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
9570                 ret = 0;
9571         }
9572 out:
9573         spin_unlock(&cache->lock);
9574         spin_unlock(&sinfo->lock);
9575         return ret;
9576 }
9577
9578 int btrfs_inc_block_group_ro(struct btrfs_root *root,
9579                              struct btrfs_block_group_cache *cache)
9580
9581 {
9582         struct btrfs_trans_handle *trans;
9583         u64 alloc_flags;
9584         int ret;
9585
9586 again:
9587         trans = btrfs_join_transaction(root);
9588         if (IS_ERR(trans))
9589                 return PTR_ERR(trans);
9590
9591         /*
9592          * we're not allowed to set block groups readonly after the dirty
9593          * block groups cache has started writing.  If it already started,
9594          * back off and let this transaction commit
9595          */
9596         mutex_lock(&root->fs_info->ro_block_group_mutex);
9597         if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
9598                 u64 transid = trans->transid;
9599
9600                 mutex_unlock(&root->fs_info->ro_block_group_mutex);
9601                 btrfs_end_transaction(trans, root);
9602
9603                 ret = btrfs_wait_for_commit(root, transid);
9604                 if (ret)
9605                         return ret;
9606                 goto again;
9607         }
9608
9609         /*
9610          * if we are changing raid levels, try to allocate a corresponding
9611          * block group with the new raid level.
9612          */
9613         alloc_flags = update_block_group_flags(root, cache->flags);
9614         if (alloc_flags != cache->flags) {
9615                 ret = do_chunk_alloc(trans, root, alloc_flags,
9616                                      CHUNK_ALLOC_FORCE);
9617                 /*
9618                  * ENOSPC is allowed here, we may have enough space
9619                  * already allocated at the new raid level to
9620                  * carry on
9621                  */
9622                 if (ret == -ENOSPC)
9623                         ret = 0;
9624                 if (ret < 0)
9625                         goto out;
9626         }
9627
9628         ret = inc_block_group_ro(cache, 0);
9629         if (!ret)
9630                 goto out;
9631         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
9632         ret = do_chunk_alloc(trans, root, alloc_flags,
9633                              CHUNK_ALLOC_FORCE);
9634         if (ret < 0)
9635                 goto out;
9636         ret = inc_block_group_ro(cache, 0);
9637 out:
9638         if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
9639                 alloc_flags = update_block_group_flags(root, cache->flags);
9640                 lock_chunks(root->fs_info->chunk_root);
9641                 check_system_chunk(trans, root, alloc_flags);
9642                 unlock_chunks(root->fs_info->chunk_root);
9643         }
9644         mutex_unlock(&root->fs_info->ro_block_group_mutex);
9645
9646         btrfs_end_transaction(trans, root);
9647         return ret;
9648 }
9649
9650 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
9651                             struct btrfs_root *root, u64 type)
9652 {
9653         u64 alloc_flags = get_alloc_profile(root, type);
9654         return do_chunk_alloc(trans, root, alloc_flags,
9655                               CHUNK_ALLOC_FORCE);
9656 }
9657
9658 /*
9659  * helper to account the unused space of all the readonly block group in the
9660  * space_info. takes mirrors into account.
9661  */
9662 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
9663 {
9664         struct btrfs_block_group_cache *block_group;
9665         u64 free_bytes = 0;
9666         int factor;
9667
9668         /* It's df, we don't care if it's racy */
9669         if (list_empty(&sinfo->ro_bgs))
9670                 return 0;
9671
9672         spin_lock(&sinfo->lock);
9673         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
9674                 spin_lock(&block_group->lock);
9675
9676                 if (!block_group->ro) {
9677                         spin_unlock(&block_group->lock);
9678                         continue;
9679                 }
9680
9681                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
9682                                           BTRFS_BLOCK_GROUP_RAID10 |
9683                                           BTRFS_BLOCK_GROUP_DUP))
9684                         factor = 2;
9685                 else
9686                         factor = 1;
9687
9688                 free_bytes += (block_group->key.offset -
9689                                btrfs_block_group_used(&block_group->item)) *
9690                                factor;
9691
9692                 spin_unlock(&block_group->lock);
9693         }
9694         spin_unlock(&sinfo->lock);
9695
9696         return free_bytes;
9697 }
9698
9699 void btrfs_dec_block_group_ro(struct btrfs_root *root,
9700                               struct btrfs_block_group_cache *cache)
9701 {
9702         struct btrfs_space_info *sinfo = cache->space_info;
9703         u64 num_bytes;
9704
9705         BUG_ON(!cache->ro);
9706
9707         spin_lock(&sinfo->lock);
9708         spin_lock(&cache->lock);
9709         if (!--cache->ro) {
9710                 num_bytes = cache->key.offset - cache->reserved -
9711                             cache->pinned - cache->bytes_super -
9712                             btrfs_block_group_used(&cache->item);
9713                 sinfo->bytes_readonly -= num_bytes;
9714                 list_del_init(&cache->ro_list);
9715         }
9716         spin_unlock(&cache->lock);
9717         spin_unlock(&sinfo->lock);
9718 }
9719
9720 /*
9721  * checks to see if its even possible to relocate this block group.
9722  *
9723  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
9724  * ok to go ahead and try.
9725  */
9726 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
9727 {
9728         struct btrfs_block_group_cache *block_group;
9729         struct btrfs_space_info *space_info;
9730         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
9731         struct btrfs_device *device;
9732         struct btrfs_trans_handle *trans;
9733         u64 min_free;
9734         u64 dev_min = 1;
9735         u64 dev_nr = 0;
9736         u64 target;
9737         int debug;
9738         int index;
9739         int full = 0;
9740         int ret = 0;
9741
9742         debug = btrfs_test_opt(root->fs_info, ENOSPC_DEBUG);
9743
9744         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
9745
9746         /* odd, couldn't find the block group, leave it alone */
9747         if (!block_group) {
9748                 if (debug)
9749                         btrfs_warn(root->fs_info,
9750                                    "can't find block group for bytenr %llu",
9751                                    bytenr);
9752                 return -1;
9753         }
9754
9755         min_free = btrfs_block_group_used(&block_group->item);
9756
9757         /* no bytes used, we're good */
9758         if (!min_free)
9759                 goto out;
9760
9761         space_info = block_group->space_info;
9762         spin_lock(&space_info->lock);
9763
9764         full = space_info->full;
9765
9766         /*
9767          * if this is the last block group we have in this space, we can't
9768          * relocate it unless we're able to allocate a new chunk below.
9769          *
9770          * Otherwise, we need to make sure we have room in the space to handle
9771          * all of the extents from this block group.  If we can, we're good
9772          */
9773         if ((space_info->total_bytes != block_group->key.offset) &&
9774             (space_info->bytes_used + space_info->bytes_reserved +
9775              space_info->bytes_pinned + space_info->bytes_readonly +
9776              min_free < space_info->total_bytes)) {
9777                 spin_unlock(&space_info->lock);
9778                 goto out;
9779         }
9780         spin_unlock(&space_info->lock);
9781
9782         /*
9783          * ok we don't have enough space, but maybe we have free space on our
9784          * devices to allocate new chunks for relocation, so loop through our
9785          * alloc devices and guess if we have enough space.  if this block
9786          * group is going to be restriped, run checks against the target
9787          * profile instead of the current one.
9788          */
9789         ret = -1;
9790
9791         /*
9792          * index:
9793          *      0: raid10
9794          *      1: raid1
9795          *      2: dup
9796          *      3: raid0
9797          *      4: single
9798          */
9799         target = get_restripe_target(root->fs_info, block_group->flags);
9800         if (target) {
9801                 index = __get_raid_index(extended_to_chunk(target));
9802         } else {
9803                 /*
9804                  * this is just a balance, so if we were marked as full
9805                  * we know there is no space for a new chunk
9806                  */
9807                 if (full) {
9808                         if (debug)
9809                                 btrfs_warn(root->fs_info,
9810                                         "no space to alloc new chunk for block group %llu",
9811                                         block_group->key.objectid);
9812                         goto out;
9813                 }
9814
9815                 index = get_block_group_index(block_group);
9816         }
9817
9818         if (index == BTRFS_RAID_RAID10) {
9819                 dev_min = 4;
9820                 /* Divide by 2 */
9821                 min_free >>= 1;
9822         } else if (index == BTRFS_RAID_RAID1) {
9823                 dev_min = 2;
9824         } else if (index == BTRFS_RAID_DUP) {
9825                 /* Multiply by 2 */
9826                 min_free <<= 1;
9827         } else if (index == BTRFS_RAID_RAID0) {
9828                 dev_min = fs_devices->rw_devices;
9829                 min_free = div64_u64(min_free, dev_min);
9830         }
9831
9832         /* We need to do this so that we can look at pending chunks */
9833         trans = btrfs_join_transaction(root);
9834         if (IS_ERR(trans)) {
9835                 ret = PTR_ERR(trans);
9836                 goto out;
9837         }
9838
9839         mutex_lock(&root->fs_info->chunk_mutex);
9840         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
9841                 u64 dev_offset;
9842
9843                 /*
9844                  * check to make sure we can actually find a chunk with enough
9845                  * space to fit our block group in.
9846                  */
9847                 if (device->total_bytes > device->bytes_used + min_free &&
9848                     !device->is_tgtdev_for_dev_replace) {
9849                         ret = find_free_dev_extent(trans, device, min_free,
9850                                                    &dev_offset, NULL);
9851                         if (!ret)
9852                                 dev_nr++;
9853
9854                         if (dev_nr >= dev_min)
9855                                 break;
9856
9857                         ret = -1;
9858                 }
9859         }
9860         if (debug && ret == -1)
9861                 btrfs_warn(root->fs_info,
9862                         "no space to allocate a new chunk for block group %llu",
9863                         block_group->key.objectid);
9864         mutex_unlock(&root->fs_info->chunk_mutex);
9865         btrfs_end_transaction(trans, root);
9866 out:
9867         btrfs_put_block_group(block_group);
9868         return ret;
9869 }
9870
9871 static int find_first_block_group(struct btrfs_root *root,
9872                 struct btrfs_path *path, struct btrfs_key *key)
9873 {
9874         int ret = 0;
9875         struct btrfs_key found_key;
9876         struct extent_buffer *leaf;
9877         int slot;
9878
9879         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
9880         if (ret < 0)
9881                 goto out;
9882
9883         while (1) {
9884                 slot = path->slots[0];
9885                 leaf = path->nodes[0];
9886                 if (slot >= btrfs_header_nritems(leaf)) {
9887                         ret = btrfs_next_leaf(root, path);
9888                         if (ret == 0)
9889                                 continue;
9890                         if (ret < 0)
9891                                 goto out;
9892                         break;
9893                 }
9894                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
9895
9896                 if (found_key.objectid >= key->objectid &&
9897                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
9898                         struct extent_map_tree *em_tree;
9899                         struct extent_map *em;
9900
9901                         em_tree = &root->fs_info->mapping_tree.map_tree;
9902                         read_lock(&em_tree->lock);
9903                         em = lookup_extent_mapping(em_tree, found_key.objectid,
9904                                                    found_key.offset);
9905                         read_unlock(&em_tree->lock);
9906                         if (!em) {
9907                                 btrfs_err(root->fs_info,
9908                         "logical %llu len %llu found bg but no related chunk",
9909                                           found_key.objectid, found_key.offset);
9910                                 ret = -ENOENT;
9911                         } else {
9912                                 ret = 0;
9913                         }
9914                         free_extent_map(em);
9915                         goto out;
9916                 }
9917                 path->slots[0]++;
9918         }
9919 out:
9920         return ret;
9921 }
9922
9923 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
9924 {
9925         struct btrfs_block_group_cache *block_group;
9926         u64 last = 0;
9927
9928         while (1) {
9929                 struct inode *inode;
9930
9931                 block_group = btrfs_lookup_first_block_group(info, last);
9932                 while (block_group) {
9933                         spin_lock(&block_group->lock);
9934                         if (block_group->iref)
9935                                 break;
9936                         spin_unlock(&block_group->lock);
9937                         block_group = next_block_group(info->tree_root,
9938                                                        block_group);
9939                 }
9940                 if (!block_group) {
9941                         if (last == 0)
9942                                 break;
9943                         last = 0;
9944                         continue;
9945                 }
9946
9947                 inode = block_group->inode;
9948                 block_group->iref = 0;
9949                 block_group->inode = NULL;
9950                 spin_unlock(&block_group->lock);
9951                 ASSERT(block_group->io_ctl.inode == NULL);
9952                 iput(inode);
9953                 last = block_group->key.objectid + block_group->key.offset;
9954                 btrfs_put_block_group(block_group);
9955         }
9956 }
9957
9958 int btrfs_free_block_groups(struct btrfs_fs_info *info)
9959 {
9960         struct btrfs_block_group_cache *block_group;
9961         struct btrfs_space_info *space_info;
9962         struct btrfs_caching_control *caching_ctl;
9963         struct rb_node *n;
9964
9965         down_write(&info->commit_root_sem);
9966         while (!list_empty(&info->caching_block_groups)) {
9967                 caching_ctl = list_entry(info->caching_block_groups.next,
9968                                          struct btrfs_caching_control, list);
9969                 list_del(&caching_ctl->list);
9970                 put_caching_control(caching_ctl);
9971         }
9972         up_write(&info->commit_root_sem);
9973
9974         spin_lock(&info->unused_bgs_lock);
9975         while (!list_empty(&info->unused_bgs)) {
9976                 block_group = list_first_entry(&info->unused_bgs,
9977                                                struct btrfs_block_group_cache,
9978                                                bg_list);
9979                 list_del_init(&block_group->bg_list);
9980                 btrfs_put_block_group(block_group);
9981         }
9982         spin_unlock(&info->unused_bgs_lock);
9983
9984         spin_lock(&info->block_group_cache_lock);
9985         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
9986                 block_group = rb_entry(n, struct btrfs_block_group_cache,
9987                                        cache_node);
9988                 rb_erase(&block_group->cache_node,
9989                          &info->block_group_cache_tree);
9990                 RB_CLEAR_NODE(&block_group->cache_node);
9991                 spin_unlock(&info->block_group_cache_lock);
9992
9993                 down_write(&block_group->space_info->groups_sem);
9994                 list_del(&block_group->list);
9995                 up_write(&block_group->space_info->groups_sem);
9996
9997                 if (block_group->cached == BTRFS_CACHE_STARTED)
9998                         wait_block_group_cache_done(block_group);
9999
10000                 /*
10001                  * We haven't cached this block group, which means we could
10002                  * possibly have excluded extents on this block group.
10003                  */
10004                 if (block_group->cached == BTRFS_CACHE_NO ||
10005                     block_group->cached == BTRFS_CACHE_ERROR)
10006                         free_excluded_extents(info->extent_root, block_group);
10007
10008                 btrfs_remove_free_space_cache(block_group);
10009                 ASSERT(list_empty(&block_group->dirty_list));
10010                 ASSERT(list_empty(&block_group->io_list));
10011                 ASSERT(list_empty(&block_group->bg_list));
10012                 ASSERT(atomic_read(&block_group->count) == 1);
10013                 btrfs_put_block_group(block_group);
10014
10015                 spin_lock(&info->block_group_cache_lock);
10016         }
10017         spin_unlock(&info->block_group_cache_lock);
10018
10019         /* now that all the block groups are freed, go through and
10020          * free all the space_info structs.  This is only called during
10021          * the final stages of unmount, and so we know nobody is
10022          * using them.  We call synchronize_rcu() once before we start,
10023          * just to be on the safe side.
10024          */
10025         synchronize_rcu();
10026
10027         release_global_block_rsv(info);
10028
10029         while (!list_empty(&info->space_info)) {
10030                 int i;
10031
10032                 space_info = list_entry(info->space_info.next,
10033                                         struct btrfs_space_info,
10034                                         list);
10035
10036                 /*
10037                  * Do not hide this behind enospc_debug, this is actually
10038                  * important and indicates a real bug if this happens.
10039                  */
10040                 if (WARN_ON(space_info->bytes_pinned > 0 ||
10041                             space_info->bytes_reserved > 0 ||
10042                             space_info->bytes_may_use > 0))
10043                         dump_space_info(info, space_info, 0, 0);
10044                 list_del(&space_info->list);
10045                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
10046                         struct kobject *kobj;
10047                         kobj = space_info->block_group_kobjs[i];
10048                         space_info->block_group_kobjs[i] = NULL;
10049                         if (kobj) {
10050                                 kobject_del(kobj);
10051                                 kobject_put(kobj);
10052                         }
10053                 }
10054                 kobject_del(&space_info->kobj);
10055                 kobject_put(&space_info->kobj);
10056         }
10057         return 0;
10058 }
10059
10060 static void __link_block_group(struct btrfs_space_info *space_info,
10061                                struct btrfs_block_group_cache *cache)
10062 {
10063         int index = get_block_group_index(cache);
10064         bool first = false;
10065
10066         down_write(&space_info->groups_sem);
10067         if (list_empty(&space_info->block_groups[index]))
10068                 first = true;
10069         list_add_tail(&cache->list, &space_info->block_groups[index]);
10070         up_write(&space_info->groups_sem);
10071
10072         if (first) {
10073                 struct raid_kobject *rkobj;
10074                 int ret;
10075
10076                 rkobj = kzalloc(sizeof(*rkobj), GFP_NOFS);
10077                 if (!rkobj)
10078                         goto out_err;
10079                 rkobj->raid_type = index;
10080                 kobject_init(&rkobj->kobj, &btrfs_raid_ktype);
10081                 ret = kobject_add(&rkobj->kobj, &space_info->kobj,
10082                                   "%s", get_raid_name(index));
10083                 if (ret) {
10084                         kobject_put(&rkobj->kobj);
10085                         goto out_err;
10086                 }
10087                 space_info->block_group_kobjs[index] = &rkobj->kobj;
10088         }
10089
10090         return;
10091 out_err:
10092         btrfs_warn(cache->fs_info,
10093                    "failed to add kobject for block cache, ignoring");
10094 }
10095
10096 static struct btrfs_block_group_cache *
10097 btrfs_create_block_group_cache(struct btrfs_root *root, u64 start, u64 size)
10098 {
10099         struct btrfs_block_group_cache *cache;
10100
10101         cache = kzalloc(sizeof(*cache), GFP_NOFS);
10102         if (!cache)
10103                 return NULL;
10104
10105         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
10106                                         GFP_NOFS);
10107         if (!cache->free_space_ctl) {
10108                 kfree(cache);
10109                 return NULL;
10110         }
10111
10112         cache->key.objectid = start;
10113         cache->key.offset = size;
10114         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
10115
10116         cache->sectorsize = root->sectorsize;
10117         cache->fs_info = root->fs_info;
10118         cache->full_stripe_len = btrfs_full_stripe_len(root,
10119                                                &root->fs_info->mapping_tree,
10120                                                start);
10121         set_free_space_tree_thresholds(cache);
10122
10123         atomic_set(&cache->count, 1);
10124         spin_lock_init(&cache->lock);
10125         init_rwsem(&cache->data_rwsem);
10126         INIT_LIST_HEAD(&cache->list);
10127         INIT_LIST_HEAD(&cache->cluster_list);
10128         INIT_LIST_HEAD(&cache->bg_list);
10129         INIT_LIST_HEAD(&cache->ro_list);
10130         INIT_LIST_HEAD(&cache->dirty_list);
10131         INIT_LIST_HEAD(&cache->io_list);
10132         btrfs_init_free_space_ctl(cache);
10133         atomic_set(&cache->trimming, 0);
10134         mutex_init(&cache->free_space_lock);
10135
10136         return cache;
10137 }
10138
10139 int btrfs_read_block_groups(struct btrfs_root *root)
10140 {
10141         struct btrfs_path *path;
10142         int ret;
10143         struct btrfs_block_group_cache *cache;
10144         struct btrfs_fs_info *info = root->fs_info;
10145         struct btrfs_space_info *space_info;
10146         struct btrfs_key key;
10147         struct btrfs_key found_key;
10148         struct extent_buffer *leaf;
10149         int need_clear = 0;
10150         u64 cache_gen;
10151         u64 feature;
10152         int mixed;
10153
10154         feature = btrfs_super_incompat_flags(info->super_copy);
10155         mixed = !!(feature & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS);
10156
10157         root = info->extent_root;
10158         key.objectid = 0;
10159         key.offset = 0;
10160         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
10161         path = btrfs_alloc_path();
10162         if (!path)
10163                 return -ENOMEM;
10164         path->reada = READA_FORWARD;
10165
10166         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
10167         if (btrfs_test_opt(root->fs_info, SPACE_CACHE) &&
10168             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
10169                 need_clear = 1;
10170         if (btrfs_test_opt(root->fs_info, CLEAR_CACHE))
10171                 need_clear = 1;
10172
10173         while (1) {
10174                 ret = find_first_block_group(root, path, &key);
10175                 if (ret > 0)
10176                         break;
10177                 if (ret != 0)
10178                         goto error;
10179
10180                 leaf = path->nodes[0];
10181                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
10182
10183                 cache = btrfs_create_block_group_cache(root, found_key.objectid,
10184                                                        found_key.offset);
10185                 if (!cache) {
10186                         ret = -ENOMEM;
10187                         goto error;
10188                 }
10189
10190                 if (need_clear) {
10191                         /*
10192                          * When we mount with old space cache, we need to
10193                          * set BTRFS_DC_CLEAR and set dirty flag.
10194                          *
10195                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
10196                          *    truncate the old free space cache inode and
10197                          *    setup a new one.
10198                          * b) Setting 'dirty flag' makes sure that we flush
10199                          *    the new space cache info onto disk.
10200                          */
10201                         if (btrfs_test_opt(root->fs_info, SPACE_CACHE))
10202                                 cache->disk_cache_state = BTRFS_DC_CLEAR;
10203                 }
10204
10205                 read_extent_buffer(leaf, &cache->item,
10206                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
10207                                    sizeof(cache->item));
10208                 cache->flags = btrfs_block_group_flags(&cache->item);
10209                 if (!mixed &&
10210                     ((cache->flags & BTRFS_BLOCK_GROUP_METADATA) &&
10211                     (cache->flags & BTRFS_BLOCK_GROUP_DATA))) {
10212                         btrfs_err(info,
10213 "bg %llu is a mixed block group but filesystem hasn't enabled mixed block groups",
10214                                   cache->key.objectid);
10215                         ret = -EINVAL;
10216                         goto error;
10217                 }
10218
10219                 key.objectid = found_key.objectid + found_key.offset;
10220                 btrfs_release_path(path);
10221
10222                 /*
10223                  * We need to exclude the super stripes now so that the space
10224                  * info has super bytes accounted for, otherwise we'll think
10225                  * we have more space than we actually do.
10226                  */
10227                 ret = exclude_super_stripes(root, cache);
10228                 if (ret) {
10229                         /*
10230                          * We may have excluded something, so call this just in
10231                          * case.
10232                          */
10233                         free_excluded_extents(root, cache);
10234                         btrfs_put_block_group(cache);
10235                         goto error;
10236                 }
10237
10238                 /*
10239                  * check for two cases, either we are full, and therefore
10240                  * don't need to bother with the caching work since we won't
10241                  * find any space, or we are empty, and we can just add all
10242                  * the space in and be done with it.  This saves us _alot_ of
10243                  * time, particularly in the full case.
10244                  */
10245                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
10246                         cache->last_byte_to_unpin = (u64)-1;
10247                         cache->cached = BTRFS_CACHE_FINISHED;
10248                         free_excluded_extents(root, cache);
10249                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10250                         cache->last_byte_to_unpin = (u64)-1;
10251                         cache->cached = BTRFS_CACHE_FINISHED;
10252                         add_new_free_space(cache, root->fs_info,
10253                                            found_key.objectid,
10254                                            found_key.objectid +
10255                                            found_key.offset);
10256                         free_excluded_extents(root, cache);
10257                 }
10258
10259                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
10260                 if (ret) {
10261                         btrfs_remove_free_space_cache(cache);
10262                         btrfs_put_block_group(cache);
10263                         goto error;
10264                 }
10265
10266                 trace_btrfs_add_block_group(root->fs_info, cache, 0);
10267                 ret = update_space_info(info, cache->flags, found_key.offset,
10268                                         btrfs_block_group_used(&cache->item),
10269                                         cache->bytes_super, &space_info);
10270                 if (ret) {
10271                         btrfs_remove_free_space_cache(cache);
10272                         spin_lock(&info->block_group_cache_lock);
10273                         rb_erase(&cache->cache_node,
10274                                  &info->block_group_cache_tree);
10275                         RB_CLEAR_NODE(&cache->cache_node);
10276                         spin_unlock(&info->block_group_cache_lock);
10277                         btrfs_put_block_group(cache);
10278                         goto error;
10279                 }
10280
10281                 cache->space_info = space_info;
10282
10283                 __link_block_group(space_info, cache);
10284
10285                 set_avail_alloc_bits(root->fs_info, cache->flags);
10286                 if (btrfs_chunk_readonly(root, cache->key.objectid)) {
10287                         inc_block_group_ro(cache, 1);
10288                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10289                         spin_lock(&info->unused_bgs_lock);
10290                         /* Should always be true but just in case. */
10291                         if (list_empty(&cache->bg_list)) {
10292                                 btrfs_get_block_group(cache);
10293                                 list_add_tail(&cache->bg_list,
10294                                               &info->unused_bgs);
10295                         }
10296                         spin_unlock(&info->unused_bgs_lock);
10297                 }
10298         }
10299
10300         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
10301                 if (!(get_alloc_profile(root, space_info->flags) &
10302                       (BTRFS_BLOCK_GROUP_RAID10 |
10303                        BTRFS_BLOCK_GROUP_RAID1 |
10304                        BTRFS_BLOCK_GROUP_RAID5 |
10305                        BTRFS_BLOCK_GROUP_RAID6 |
10306                        BTRFS_BLOCK_GROUP_DUP)))
10307                         continue;
10308                 /*
10309                  * avoid allocating from un-mirrored block group if there are
10310                  * mirrored block groups.
10311                  */
10312                 list_for_each_entry(cache,
10313                                 &space_info->block_groups[BTRFS_RAID_RAID0],
10314                                 list)
10315                         inc_block_group_ro(cache, 1);
10316                 list_for_each_entry(cache,
10317                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
10318                                 list)
10319                         inc_block_group_ro(cache, 1);
10320         }
10321
10322         init_global_block_rsv(info);
10323         ret = 0;
10324 error:
10325         btrfs_free_path(path);
10326         return ret;
10327 }
10328
10329 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans,
10330                                        struct btrfs_root *root)
10331 {
10332         struct btrfs_block_group_cache *block_group, *tmp;
10333         struct btrfs_root *extent_root = root->fs_info->extent_root;
10334         struct btrfs_block_group_item item;
10335         struct btrfs_key key;
10336         int ret = 0;
10337         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
10338
10339         trans->can_flush_pending_bgs = false;
10340         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) {
10341                 if (ret)
10342                         goto next;
10343
10344                 spin_lock(&block_group->lock);
10345                 memcpy(&item, &block_group->item, sizeof(item));
10346                 memcpy(&key, &block_group->key, sizeof(key));
10347                 spin_unlock(&block_group->lock);
10348
10349                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
10350                                         sizeof(item));
10351                 if (ret)
10352                         btrfs_abort_transaction(trans, ret);
10353                 ret = btrfs_finish_chunk_alloc(trans, extent_root,
10354                                                key.objectid, key.offset);
10355                 if (ret)
10356                         btrfs_abort_transaction(trans, ret);
10357                 add_block_group_free_space(trans, root->fs_info, block_group);
10358                 /* already aborted the transaction if it failed. */
10359 next:
10360                 list_del_init(&block_group->bg_list);
10361         }
10362         trans->can_flush_pending_bgs = can_flush_pending_bgs;
10363 }
10364
10365 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
10366                            struct btrfs_root *root, u64 bytes_used,
10367                            u64 type, u64 chunk_objectid, u64 chunk_offset,
10368                            u64 size)
10369 {
10370         int ret;
10371         struct btrfs_root *extent_root;
10372         struct btrfs_block_group_cache *cache;
10373         extent_root = root->fs_info->extent_root;
10374
10375         btrfs_set_log_full_commit(root->fs_info, trans);
10376
10377         cache = btrfs_create_block_group_cache(root, chunk_offset, size);
10378         if (!cache)
10379                 return -ENOMEM;
10380
10381         btrfs_set_block_group_used(&cache->item, bytes_used);
10382         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
10383         btrfs_set_block_group_flags(&cache->item, type);
10384
10385         cache->flags = type;
10386         cache->last_byte_to_unpin = (u64)-1;
10387         cache->cached = BTRFS_CACHE_FINISHED;
10388         cache->needs_free_space = 1;
10389         ret = exclude_super_stripes(root, cache);
10390         if (ret) {
10391                 /*
10392                  * We may have excluded something, so call this just in
10393                  * case.
10394                  */
10395                 free_excluded_extents(root, cache);
10396                 btrfs_put_block_group(cache);
10397                 return ret;
10398         }
10399
10400         add_new_free_space(cache, root->fs_info, chunk_offset,
10401                            chunk_offset + size);
10402
10403         free_excluded_extents(root, cache);
10404
10405 #ifdef CONFIG_BTRFS_DEBUG
10406         if (btrfs_should_fragment_free_space(root, cache)) {
10407                 u64 new_bytes_used = size - bytes_used;
10408
10409                 bytes_used += new_bytes_used >> 1;
10410                 fragment_free_space(root, cache);
10411         }
10412 #endif
10413         /*
10414          * Call to ensure the corresponding space_info object is created and
10415          * assigned to our block group, but don't update its counters just yet.
10416          * We want our bg to be added to the rbtree with its ->space_info set.
10417          */
10418         ret = update_space_info(root->fs_info, cache->flags, 0, 0, 0,
10419                                 &cache->space_info);
10420         if (ret) {
10421                 btrfs_remove_free_space_cache(cache);
10422                 btrfs_put_block_group(cache);
10423                 return ret;
10424         }
10425
10426         ret = btrfs_add_block_group_cache(root->fs_info, cache);
10427         if (ret) {
10428                 btrfs_remove_free_space_cache(cache);
10429                 btrfs_put_block_group(cache);
10430                 return ret;
10431         }
10432
10433         /*
10434          * Now that our block group has its ->space_info set and is inserted in
10435          * the rbtree, update the space info's counters.
10436          */
10437         trace_btrfs_add_block_group(root->fs_info, cache, 1);
10438         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
10439                                 cache->bytes_super, &cache->space_info);
10440         if (ret) {
10441                 btrfs_remove_free_space_cache(cache);
10442                 spin_lock(&root->fs_info->block_group_cache_lock);
10443                 rb_erase(&cache->cache_node,
10444                          &root->fs_info->block_group_cache_tree);
10445                 RB_CLEAR_NODE(&cache->cache_node);
10446                 spin_unlock(&root->fs_info->block_group_cache_lock);
10447                 btrfs_put_block_group(cache);
10448                 return ret;
10449         }
10450         update_global_block_rsv(root->fs_info);
10451
10452         __link_block_group(cache->space_info, cache);
10453
10454         list_add_tail(&cache->bg_list, &trans->new_bgs);
10455
10456         set_avail_alloc_bits(extent_root->fs_info, type);
10457         return 0;
10458 }
10459
10460 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
10461 {
10462         u64 extra_flags = chunk_to_extended(flags) &
10463                                 BTRFS_EXTENDED_PROFILE_MASK;
10464
10465         write_seqlock(&fs_info->profiles_lock);
10466         if (flags & BTRFS_BLOCK_GROUP_DATA)
10467                 fs_info->avail_data_alloc_bits &= ~extra_flags;
10468         if (flags & BTRFS_BLOCK_GROUP_METADATA)
10469                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
10470         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
10471                 fs_info->avail_system_alloc_bits &= ~extra_flags;
10472         write_sequnlock(&fs_info->profiles_lock);
10473 }
10474
10475 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
10476                              struct btrfs_root *root, u64 group_start,
10477                              struct extent_map *em)
10478 {
10479         struct btrfs_path *path;
10480         struct btrfs_block_group_cache *block_group;
10481         struct btrfs_free_cluster *cluster;
10482         struct btrfs_root *tree_root = root->fs_info->tree_root;
10483         struct btrfs_key key;
10484         struct inode *inode;
10485         struct kobject *kobj = NULL;
10486         int ret;
10487         int index;
10488         int factor;
10489         struct btrfs_caching_control *caching_ctl = NULL;
10490         bool remove_em;
10491
10492         root = root->fs_info->extent_root;
10493
10494         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
10495         BUG_ON(!block_group);
10496         BUG_ON(!block_group->ro);
10497
10498         /*
10499          * Free the reserved super bytes from this block group before
10500          * remove it.
10501          */
10502         free_excluded_extents(root, block_group);
10503
10504         memcpy(&key, &block_group->key, sizeof(key));
10505         index = get_block_group_index(block_group);
10506         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
10507                                   BTRFS_BLOCK_GROUP_RAID1 |
10508                                   BTRFS_BLOCK_GROUP_RAID10))
10509                 factor = 2;
10510         else
10511                 factor = 1;
10512
10513         /* make sure this block group isn't part of an allocation cluster */
10514         cluster = &root->fs_info->data_alloc_cluster;
10515         spin_lock(&cluster->refill_lock);
10516         btrfs_return_cluster_to_free_space(block_group, cluster);
10517         spin_unlock(&cluster->refill_lock);
10518
10519         /*
10520          * make sure this block group isn't part of a metadata
10521          * allocation cluster
10522          */
10523         cluster = &root->fs_info->meta_alloc_cluster;
10524         spin_lock(&cluster->refill_lock);
10525         btrfs_return_cluster_to_free_space(block_group, cluster);
10526         spin_unlock(&cluster->refill_lock);
10527
10528         path = btrfs_alloc_path();
10529         if (!path) {
10530                 ret = -ENOMEM;
10531                 goto out;
10532         }
10533
10534         /*
10535          * get the inode first so any iput calls done for the io_list
10536          * aren't the final iput (no unlinks allowed now)
10537          */
10538         inode = lookup_free_space_inode(tree_root, block_group, path);
10539
10540         mutex_lock(&trans->transaction->cache_write_mutex);
10541         /*
10542          * make sure our free spache cache IO is done before remove the
10543          * free space inode
10544          */
10545         spin_lock(&trans->transaction->dirty_bgs_lock);
10546         if (!list_empty(&block_group->io_list)) {
10547                 list_del_init(&block_group->io_list);
10548
10549                 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
10550
10551                 spin_unlock(&trans->transaction->dirty_bgs_lock);
10552                 btrfs_wait_cache_io(root, trans, block_group,
10553                                     &block_group->io_ctl, path,
10554                                     block_group->key.objectid);
10555                 btrfs_put_block_group(block_group);
10556                 spin_lock(&trans->transaction->dirty_bgs_lock);
10557         }
10558
10559         if (!list_empty(&block_group->dirty_list)) {
10560                 list_del_init(&block_group->dirty_list);
10561                 btrfs_put_block_group(block_group);
10562         }
10563         spin_unlock(&trans->transaction->dirty_bgs_lock);
10564         mutex_unlock(&trans->transaction->cache_write_mutex);
10565
10566         if (!IS_ERR(inode)) {
10567                 ret = btrfs_orphan_add(trans, inode);
10568                 if (ret) {
10569                         btrfs_add_delayed_iput(inode);
10570                         goto out;
10571                 }
10572                 clear_nlink(inode);
10573                 /* One for the block groups ref */
10574                 spin_lock(&block_group->lock);
10575                 if (block_group->iref) {
10576                         block_group->iref = 0;
10577                         block_group->inode = NULL;
10578                         spin_unlock(&block_group->lock);
10579                         iput(inode);
10580                 } else {
10581                         spin_unlock(&block_group->lock);
10582                 }
10583                 /* One for our lookup ref */
10584                 btrfs_add_delayed_iput(inode);
10585         }
10586
10587         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
10588         key.offset = block_group->key.objectid;
10589         key.type = 0;
10590
10591         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
10592         if (ret < 0)
10593                 goto out;
10594         if (ret > 0)
10595                 btrfs_release_path(path);
10596         if (ret == 0) {
10597                 ret = btrfs_del_item(trans, tree_root, path);
10598                 if (ret)
10599                         goto out;
10600                 btrfs_release_path(path);
10601         }
10602
10603         spin_lock(&root->fs_info->block_group_cache_lock);
10604         rb_erase(&block_group->cache_node,
10605                  &root->fs_info->block_group_cache_tree);
10606         RB_CLEAR_NODE(&block_group->cache_node);
10607
10608         if (root->fs_info->first_logical_byte == block_group->key.objectid)
10609                 root->fs_info->first_logical_byte = (u64)-1;
10610         spin_unlock(&root->fs_info->block_group_cache_lock);
10611
10612         down_write(&block_group->space_info->groups_sem);
10613         /*
10614          * we must use list_del_init so people can check to see if they
10615          * are still on the list after taking the semaphore
10616          */
10617         list_del_init(&block_group->list);
10618         if (list_empty(&block_group->space_info->block_groups[index])) {
10619                 kobj = block_group->space_info->block_group_kobjs[index];
10620                 block_group->space_info->block_group_kobjs[index] = NULL;
10621                 clear_avail_alloc_bits(root->fs_info, block_group->flags);
10622         }
10623         up_write(&block_group->space_info->groups_sem);
10624         if (kobj) {
10625                 kobject_del(kobj);
10626                 kobject_put(kobj);
10627         }
10628
10629         if (block_group->has_caching_ctl)
10630                 caching_ctl = get_caching_control(block_group);
10631         if (block_group->cached == BTRFS_CACHE_STARTED)
10632                 wait_block_group_cache_done(block_group);
10633         if (block_group->has_caching_ctl) {
10634                 down_write(&root->fs_info->commit_root_sem);
10635                 if (!caching_ctl) {
10636                         struct btrfs_caching_control *ctl;
10637
10638                         list_for_each_entry(ctl,
10639                                     &root->fs_info->caching_block_groups, list)
10640                                 if (ctl->block_group == block_group) {
10641                                         caching_ctl = ctl;
10642                                         atomic_inc(&caching_ctl->count);
10643                                         break;
10644                                 }
10645                 }
10646                 if (caching_ctl)
10647                         list_del_init(&caching_ctl->list);
10648                 up_write(&root->fs_info->commit_root_sem);
10649                 if (caching_ctl) {
10650                         /* Once for the caching bgs list and once for us. */
10651                         put_caching_control(caching_ctl);
10652                         put_caching_control(caching_ctl);
10653                 }
10654         }
10655
10656         spin_lock(&trans->transaction->dirty_bgs_lock);
10657         if (!list_empty(&block_group->dirty_list)) {
10658                 WARN_ON(1);
10659         }
10660         if (!list_empty(&block_group->io_list)) {
10661                 WARN_ON(1);
10662         }
10663         spin_unlock(&trans->transaction->dirty_bgs_lock);
10664         btrfs_remove_free_space_cache(block_group);
10665
10666         spin_lock(&block_group->space_info->lock);
10667         list_del_init(&block_group->ro_list);
10668
10669         if (btrfs_test_opt(root->fs_info, ENOSPC_DEBUG)) {
10670                 WARN_ON(block_group->space_info->total_bytes
10671                         < block_group->key.offset);
10672                 WARN_ON(block_group->space_info->bytes_readonly
10673                         < block_group->key.offset);
10674                 WARN_ON(block_group->space_info->disk_total
10675                         < block_group->key.offset * factor);
10676         }
10677         block_group->space_info->total_bytes -= block_group->key.offset;
10678         block_group->space_info->bytes_readonly -= block_group->key.offset;
10679         block_group->space_info->disk_total -= block_group->key.offset * factor;
10680
10681         spin_unlock(&block_group->space_info->lock);
10682
10683         memcpy(&key, &block_group->key, sizeof(key));
10684
10685         lock_chunks(root);
10686         if (!list_empty(&em->list)) {
10687                 /* We're in the transaction->pending_chunks list. */
10688                 free_extent_map(em);
10689         }
10690         spin_lock(&block_group->lock);
10691         block_group->removed = 1;
10692         /*
10693          * At this point trimming can't start on this block group, because we
10694          * removed the block group from the tree fs_info->block_group_cache_tree
10695          * so no one can't find it anymore and even if someone already got this
10696          * block group before we removed it from the rbtree, they have already
10697          * incremented block_group->trimming - if they didn't, they won't find
10698          * any free space entries because we already removed them all when we
10699          * called btrfs_remove_free_space_cache().
10700          *
10701          * And we must not remove the extent map from the fs_info->mapping_tree
10702          * to prevent the same logical address range and physical device space
10703          * ranges from being reused for a new block group. This is because our
10704          * fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
10705          * completely transactionless, so while it is trimming a range the
10706          * currently running transaction might finish and a new one start,
10707          * allowing for new block groups to be created that can reuse the same
10708          * physical device locations unless we take this special care.
10709          *
10710          * There may also be an implicit trim operation if the file system
10711          * is mounted with -odiscard. The same protections must remain
10712          * in place until the extents have been discarded completely when
10713          * the transaction commit has completed.
10714          */
10715         remove_em = (atomic_read(&block_group->trimming) == 0);
10716         /*
10717          * Make sure a trimmer task always sees the em in the pinned_chunks list
10718          * if it sees block_group->removed == 1 (needs to lock block_group->lock
10719          * before checking block_group->removed).
10720          */
10721         if (!remove_em) {
10722                 /*
10723                  * Our em might be in trans->transaction->pending_chunks which
10724                  * is protected by fs_info->chunk_mutex ([lock|unlock]_chunks),
10725                  * and so is the fs_info->pinned_chunks list.
10726                  *
10727                  * So at this point we must be holding the chunk_mutex to avoid
10728                  * any races with chunk allocation (more specifically at
10729                  * volumes.c:contains_pending_extent()), to ensure it always
10730                  * sees the em, either in the pending_chunks list or in the
10731                  * pinned_chunks list.
10732                  */
10733                 list_move_tail(&em->list, &root->fs_info->pinned_chunks);
10734         }
10735         spin_unlock(&block_group->lock);
10736
10737         if (remove_em) {
10738                 struct extent_map_tree *em_tree;
10739
10740                 em_tree = &root->fs_info->mapping_tree.map_tree;
10741                 write_lock(&em_tree->lock);
10742                 /*
10743                  * The em might be in the pending_chunks list, so make sure the
10744                  * chunk mutex is locked, since remove_extent_mapping() will
10745                  * delete us from that list.
10746                  */
10747                 remove_extent_mapping(em_tree, em);
10748                 write_unlock(&em_tree->lock);
10749                 /* once for the tree */
10750                 free_extent_map(em);
10751         }
10752
10753         unlock_chunks(root);
10754
10755         ret = remove_block_group_free_space(trans, root->fs_info, block_group);
10756         if (ret)
10757                 goto out;
10758
10759         btrfs_put_block_group(block_group);
10760         btrfs_put_block_group(block_group);
10761
10762         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
10763         if (ret > 0)
10764                 ret = -EIO;
10765         if (ret < 0)
10766                 goto out;
10767
10768         ret = btrfs_del_item(trans, root, path);
10769 out:
10770         btrfs_free_path(path);
10771         return ret;
10772 }
10773
10774 struct btrfs_trans_handle *
10775 btrfs_start_trans_remove_block_group(struct btrfs_fs_info *fs_info,
10776                                      const u64 chunk_offset)
10777 {
10778         struct extent_map_tree *em_tree = &fs_info->mapping_tree.map_tree;
10779         struct extent_map *em;
10780         struct map_lookup *map;
10781         unsigned int num_items;
10782
10783         read_lock(&em_tree->lock);
10784         em = lookup_extent_mapping(em_tree, chunk_offset, 1);
10785         read_unlock(&em_tree->lock);
10786         ASSERT(em && em->start == chunk_offset);
10787
10788         /*
10789          * We need to reserve 3 + N units from the metadata space info in order
10790          * to remove a block group (done at btrfs_remove_chunk() and at
10791          * btrfs_remove_block_group()), which are used for:
10792          *
10793          * 1 unit for adding the free space inode's orphan (located in the tree
10794          * of tree roots).
10795          * 1 unit for deleting the block group item (located in the extent
10796          * tree).
10797          * 1 unit for deleting the free space item (located in tree of tree
10798          * roots).
10799          * N units for deleting N device extent items corresponding to each
10800          * stripe (located in the device tree).
10801          *
10802          * In order to remove a block group we also need to reserve units in the
10803          * system space info in order to update the chunk tree (update one or
10804          * more device items and remove one chunk item), but this is done at
10805          * btrfs_remove_chunk() through a call to check_system_chunk().
10806          */
10807         map = em->map_lookup;
10808         num_items = 3 + map->num_stripes;
10809         free_extent_map(em);
10810
10811         return btrfs_start_transaction_fallback_global_rsv(fs_info->extent_root,
10812                                                            num_items, 1);
10813 }
10814
10815 /*
10816  * Process the unused_bgs list and remove any that don't have any allocated
10817  * space inside of them.
10818  */
10819 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
10820 {
10821         struct btrfs_block_group_cache *block_group;
10822         struct btrfs_space_info *space_info;
10823         struct btrfs_root *root = fs_info->extent_root;
10824         struct btrfs_trans_handle *trans;
10825         int ret = 0;
10826
10827         if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
10828                 return;
10829
10830         spin_lock(&fs_info->unused_bgs_lock);
10831         while (!list_empty(&fs_info->unused_bgs)) {
10832                 u64 start, end;
10833                 int trimming;
10834
10835                 block_group = list_first_entry(&fs_info->unused_bgs,
10836                                                struct btrfs_block_group_cache,
10837                                                bg_list);
10838                 list_del_init(&block_group->bg_list);
10839
10840                 space_info = block_group->space_info;
10841
10842                 if (ret || btrfs_mixed_space_info(space_info)) {
10843                         btrfs_put_block_group(block_group);
10844                         continue;
10845                 }
10846                 spin_unlock(&fs_info->unused_bgs_lock);
10847
10848                 mutex_lock(&fs_info->delete_unused_bgs_mutex);
10849
10850                 /* Don't want to race with allocators so take the groups_sem */
10851                 down_write(&space_info->groups_sem);
10852                 spin_lock(&block_group->lock);
10853                 if (block_group->reserved ||
10854                     btrfs_block_group_used(&block_group->item) ||
10855                     block_group->ro ||
10856                     list_is_singular(&block_group->list)) {
10857                         /*
10858                          * We want to bail if we made new allocations or have
10859                          * outstanding allocations in this block group.  We do
10860                          * the ro check in case balance is currently acting on
10861                          * this block group.
10862                          */
10863                         spin_unlock(&block_group->lock);
10864                         up_write(&space_info->groups_sem);
10865                         goto next;
10866                 }
10867                 spin_unlock(&block_group->lock);
10868
10869                 /* We don't want to force the issue, only flip if it's ok. */
10870                 ret = inc_block_group_ro(block_group, 0);
10871                 up_write(&space_info->groups_sem);
10872                 if (ret < 0) {
10873                         ret = 0;
10874                         goto next;
10875                 }
10876
10877                 /*
10878                  * Want to do this before we do anything else so we can recover
10879                  * properly if we fail to join the transaction.
10880                  */
10881                 trans = btrfs_start_trans_remove_block_group(fs_info,
10882                                                      block_group->key.objectid);
10883                 if (IS_ERR(trans)) {
10884                         btrfs_dec_block_group_ro(root, block_group);
10885                         ret = PTR_ERR(trans);
10886                         goto next;
10887                 }
10888
10889                 /*
10890                  * We could have pending pinned extents for this block group,
10891                  * just delete them, we don't care about them anymore.
10892                  */
10893                 start = block_group->key.objectid;
10894                 end = start + block_group->key.offset - 1;
10895                 /*
10896                  * Hold the unused_bg_unpin_mutex lock to avoid racing with
10897                  * btrfs_finish_extent_commit(). If we are at transaction N,
10898                  * another task might be running finish_extent_commit() for the
10899                  * previous transaction N - 1, and have seen a range belonging
10900                  * to the block group in freed_extents[] before we were able to
10901                  * clear the whole block group range from freed_extents[]. This
10902                  * means that task can lookup for the block group after we
10903                  * unpinned it from freed_extents[] and removed it, leading to
10904                  * a BUG_ON() at btrfs_unpin_extent_range().
10905                  */
10906                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
10907                 ret = clear_extent_bits(&fs_info->freed_extents[0], start, end,
10908                                   EXTENT_DIRTY);
10909                 if (ret) {
10910                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10911                         btrfs_dec_block_group_ro(root, block_group);
10912                         goto end_trans;
10913                 }
10914                 ret = clear_extent_bits(&fs_info->freed_extents[1], start, end,
10915                                   EXTENT_DIRTY);
10916                 if (ret) {
10917                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10918                         btrfs_dec_block_group_ro(root, block_group);
10919                         goto end_trans;
10920                 }
10921                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10922
10923                 /* Reset pinned so btrfs_put_block_group doesn't complain */
10924                 spin_lock(&space_info->lock);
10925                 spin_lock(&block_group->lock);
10926
10927                 space_info->bytes_pinned -= block_group->pinned;
10928                 space_info->bytes_readonly += block_group->pinned;
10929                 percpu_counter_add(&space_info->total_bytes_pinned,
10930                                    -block_group->pinned);
10931                 block_group->pinned = 0;
10932
10933                 spin_unlock(&block_group->lock);
10934                 spin_unlock(&space_info->lock);
10935
10936                 /* DISCARD can flip during remount */
10937                 trimming = btrfs_test_opt(root->fs_info, DISCARD);
10938
10939                 /* Implicit trim during transaction commit. */
10940                 if (trimming)
10941                         btrfs_get_block_group_trimming(block_group);
10942
10943                 /*
10944                  * Btrfs_remove_chunk will abort the transaction if things go
10945                  * horribly wrong.
10946                  */
10947                 ret = btrfs_remove_chunk(trans, root,
10948                                          block_group->key.objectid);
10949
10950                 if (ret) {
10951                         if (trimming)
10952                                 btrfs_put_block_group_trimming(block_group);
10953                         goto end_trans;
10954                 }
10955
10956                 /*
10957                  * If we're not mounted with -odiscard, we can just forget
10958                  * about this block group. Otherwise we'll need to wait
10959                  * until transaction commit to do the actual discard.
10960                  */
10961                 if (trimming) {
10962                         spin_lock(&fs_info->unused_bgs_lock);
10963                         /*
10964                          * A concurrent scrub might have added us to the list
10965                          * fs_info->unused_bgs, so use a list_move operation
10966                          * to add the block group to the deleted_bgs list.
10967                          */
10968                         list_move(&block_group->bg_list,
10969                                   &trans->transaction->deleted_bgs);
10970                         spin_unlock(&fs_info->unused_bgs_lock);
10971                         btrfs_get_block_group(block_group);
10972                 }
10973 end_trans:
10974                 btrfs_end_transaction(trans, root);
10975 next:
10976                 mutex_unlock(&fs_info->delete_unused_bgs_mutex);
10977                 btrfs_put_block_group(block_group);
10978                 spin_lock(&fs_info->unused_bgs_lock);
10979         }
10980         spin_unlock(&fs_info->unused_bgs_lock);
10981 }
10982
10983 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
10984 {
10985         struct btrfs_space_info *space_info;
10986         struct btrfs_super_block *disk_super;
10987         u64 features;
10988         u64 flags;
10989         int mixed = 0;
10990         int ret;
10991
10992         disk_super = fs_info->super_copy;
10993         if (!btrfs_super_root(disk_super))
10994                 return -EINVAL;
10995
10996         features = btrfs_super_incompat_flags(disk_super);
10997         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
10998                 mixed = 1;
10999
11000         flags = BTRFS_BLOCK_GROUP_SYSTEM;
11001         ret = update_space_info(fs_info, flags, 0, 0, 0, &space_info);
11002         if (ret)
11003                 goto out;
11004
11005         if (mixed) {
11006                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
11007                 ret = update_space_info(fs_info, flags, 0, 0, 0, &space_info);
11008         } else {
11009                 flags = BTRFS_BLOCK_GROUP_METADATA;
11010                 ret = update_space_info(fs_info, flags, 0, 0, 0, &space_info);
11011                 if (ret)
11012                         goto out;
11013
11014                 flags = BTRFS_BLOCK_GROUP_DATA;
11015                 ret = update_space_info(fs_info, flags, 0, 0, 0, &space_info);
11016         }
11017 out:
11018         return ret;
11019 }
11020
11021 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
11022 {
11023         return unpin_extent_range(root, start, end, false);
11024 }
11025
11026 /*
11027  * It used to be that old block groups would be left around forever.
11028  * Iterating over them would be enough to trim unused space.  Since we
11029  * now automatically remove them, we also need to iterate over unallocated
11030  * space.
11031  *
11032  * We don't want a transaction for this since the discard may take a
11033  * substantial amount of time.  We don't require that a transaction be
11034  * running, but we do need to take a running transaction into account
11035  * to ensure that we're not discarding chunks that were released in
11036  * the current transaction.
11037  *
11038  * Holding the chunks lock will prevent other threads from allocating
11039  * or releasing chunks, but it won't prevent a running transaction
11040  * from committing and releasing the memory that the pending chunks
11041  * list head uses.  For that, we need to take a reference to the
11042  * transaction.
11043  */
11044 static int btrfs_trim_free_extents(struct btrfs_device *device,
11045                                    u64 minlen, u64 *trimmed)
11046 {
11047         u64 start = 0, len = 0;
11048         int ret;
11049
11050         *trimmed = 0;
11051
11052         /* Not writeable = nothing to do. */
11053         if (!device->writeable)
11054                 return 0;
11055
11056         /* No free space = nothing to do. */
11057         if (device->total_bytes <= device->bytes_used)
11058                 return 0;
11059
11060         ret = 0;
11061
11062         while (1) {
11063                 struct btrfs_fs_info *fs_info = device->dev_root->fs_info;
11064                 struct btrfs_transaction *trans;
11065                 u64 bytes;
11066
11067                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
11068                 if (ret)
11069                         return ret;
11070
11071                 down_read(&fs_info->commit_root_sem);
11072
11073                 spin_lock(&fs_info->trans_lock);
11074                 trans = fs_info->running_transaction;
11075                 if (trans)
11076                         atomic_inc(&trans->use_count);
11077                 spin_unlock(&fs_info->trans_lock);
11078
11079                 ret = find_free_dev_extent_start(trans, device, minlen, start,
11080                                                  &start, &len);
11081                 if (trans)
11082                         btrfs_put_transaction(trans);
11083
11084                 if (ret) {
11085                         up_read(&fs_info->commit_root_sem);
11086                         mutex_unlock(&fs_info->chunk_mutex);
11087                         if (ret == -ENOSPC)
11088                                 ret = 0;
11089                         break;
11090                 }
11091
11092                 ret = btrfs_issue_discard(device->bdev, start, len, &bytes);
11093                 up_read(&fs_info->commit_root_sem);
11094                 mutex_unlock(&fs_info->chunk_mutex);
11095
11096                 if (ret)
11097                         break;
11098
11099                 start += len;
11100                 *trimmed += bytes;
11101
11102                 if (fatal_signal_pending(current)) {
11103                         ret = -ERESTARTSYS;
11104                         break;
11105                 }
11106
11107                 cond_resched();
11108         }
11109
11110         return ret;
11111 }
11112
11113 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
11114 {
11115         struct btrfs_fs_info *fs_info = root->fs_info;
11116         struct btrfs_block_group_cache *cache = NULL;
11117         struct btrfs_device *device;
11118         struct list_head *devices;
11119         u64 group_trimmed;
11120         u64 start;
11121         u64 end;
11122         u64 trimmed = 0;
11123         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
11124         int ret = 0;
11125
11126         /*
11127          * try to trim all FS space, our block group may start from non-zero.
11128          */
11129         if (range->len == total_bytes)
11130                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
11131         else
11132                 cache = btrfs_lookup_block_group(fs_info, range->start);
11133
11134         while (cache) {
11135                 if (cache->key.objectid >= (range->start + range->len)) {
11136                         btrfs_put_block_group(cache);
11137                         break;
11138                 }
11139
11140                 start = max(range->start, cache->key.objectid);
11141                 end = min(range->start + range->len,
11142                                 cache->key.objectid + cache->key.offset);
11143
11144                 if (end - start >= range->minlen) {
11145                         if (!block_group_cache_done(cache)) {
11146                                 ret = cache_block_group(cache, 0);
11147                                 if (ret) {
11148                                         btrfs_put_block_group(cache);
11149                                         break;
11150                                 }
11151                                 ret = wait_block_group_cache_done(cache);
11152                                 if (ret) {
11153                                         btrfs_put_block_group(cache);
11154                                         break;
11155                                 }
11156                         }
11157                         ret = btrfs_trim_block_group(cache,
11158                                                      &group_trimmed,
11159                                                      start,
11160                                                      end,
11161                                                      range->minlen);
11162
11163                         trimmed += group_trimmed;
11164                         if (ret) {
11165                                 btrfs_put_block_group(cache);
11166                                 break;
11167                         }
11168                 }
11169
11170                 cache = next_block_group(fs_info->tree_root, cache);
11171         }
11172
11173         mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
11174         devices = &root->fs_info->fs_devices->alloc_list;
11175         list_for_each_entry(device, devices, dev_alloc_list) {
11176                 ret = btrfs_trim_free_extents(device, range->minlen,
11177                                               &group_trimmed);
11178                 if (ret)
11179                         break;
11180
11181                 trimmed += group_trimmed;
11182         }
11183         mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
11184
11185         range->len = trimmed;
11186         return ret;
11187 }
11188
11189 /*
11190  * btrfs_{start,end}_write_no_snapshoting() are similar to
11191  * mnt_{want,drop}_write(), they are used to prevent some tasks from writing
11192  * data into the page cache through nocow before the subvolume is snapshoted,
11193  * but flush the data into disk after the snapshot creation, or to prevent
11194  * operations while snapshoting is ongoing and that cause the snapshot to be
11195  * inconsistent (writes followed by expanding truncates for example).
11196  */
11197 void btrfs_end_write_no_snapshoting(struct btrfs_root *root)
11198 {
11199         percpu_counter_dec(&root->subv_writers->counter);
11200         /*
11201          * Make sure counter is updated before we wake up waiters.
11202          */
11203         smp_mb();
11204         if (waitqueue_active(&root->subv_writers->wait))
11205                 wake_up(&root->subv_writers->wait);
11206 }
11207
11208 int btrfs_start_write_no_snapshoting(struct btrfs_root *root)
11209 {
11210         if (atomic_read(&root->will_be_snapshoted))
11211                 return 0;
11212
11213         percpu_counter_inc(&root->subv_writers->counter);
11214         /*
11215          * Make sure counter is updated before we check for snapshot creation.
11216          */
11217         smp_mb();
11218         if (atomic_read(&root->will_be_snapshoted)) {
11219                 btrfs_end_write_no_snapshoting(root);
11220                 return 0;
11221         }
11222         return 1;
11223 }
11224
11225 static int wait_snapshoting_atomic_t(atomic_t *a)
11226 {
11227         schedule();
11228         return 0;
11229 }
11230
11231 void btrfs_wait_for_snapshot_creation(struct btrfs_root *root)
11232 {
11233         while (true) {
11234                 int ret;
11235
11236                 ret = btrfs_start_write_no_snapshoting(root);
11237                 if (ret)
11238                         break;
11239                 wait_on_atomic_t(&root->will_be_snapshoted,
11240                                  wait_snapshoting_atomic_t,
11241                                  TASK_UNINTERRUPTIBLE);
11242         }
11243 }