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_fs_info *fs_info, u64 bytenr,
  65                               u64 num_bytes, int alloc);
  66 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  67                                struct btrfs_fs_info *fs_info,
  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_fs_info *fs_info,
  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_fs_info *fs_info,
  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_fs_info *fs_info, 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 static int __reserve_metadata_bytes(struct btrfs_root *root,
 100                                     struct btrfs_space_info *space_info,
 101                                     u64 orig_bytes,
 102                                     enum btrfs_reserve_flush_enum flush);
 103 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
 104                                      struct btrfs_space_info *space_info,
 105                                      u64 num_bytes);
 106 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
 107                                      struct btrfs_space_info *space_info,
 108                                      u64 num_bytes);
 109
 110 static noinline int
 111 block_group_cache_done(struct btrfs_block_group_cache *cache)
 112 {
 113         smp_mb();
 114         return cache->cached == BTRFS_CACHE_FINISHED ||
 115                 cache->cached == BTRFS_CACHE_ERROR;
 116 }
 117
 118 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
 119 {
 120         return (cache->flags & bits) == bits;
 121 }
 122
 123 void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 124 {
 125         atomic_inc(&cache->count);
 126 }
 127
 128 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 129 {
 130         if (atomic_dec_and_test(&cache->count)) {
 131                 WARN_ON(cache->pinned > 0);
 132                 WARN_ON(cache->reserved > 0);
 133                 kfree(cache->free_space_ctl);
 134                 kfree(cache);
 135         }
 136 }
 137
 138 /*
 139  * this adds the block group to the fs_info rb tree for the block group
 140  * cache
 141  */
 142 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 143                                 struct btrfs_block_group_cache *block_group)
 144 {
 145         struct rb_node **p;
 146         struct rb_node *parent = NULL;
 147         struct btrfs_block_group_cache *cache;
 148
 149         spin_lock(&info->block_group_cache_lock);
 150         p = &info->block_group_cache_tree.rb_node;
 151
 152         while (*p) {
 153                 parent = *p;
 154                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 155                                  cache_node);
 156                 if (block_group->key.objectid < cache->key.objectid) {
 157                         p = &(*p)->rb_left;
 158                 } else if (block_group->key.objectid > cache->key.objectid) {
 159                         p = &(*p)->rb_right;
 160                 } else {
 161                         spin_unlock(&info->block_group_cache_lock);
 162                         return -EEXIST;
 163                 }
 164         }
 165
 166         rb_link_node(&block_group->cache_node, parent, p);
 167         rb_insert_color(&block_group->cache_node,
 168                         &info->block_group_cache_tree);
 169
 170         if (info->first_logical_byte > block_group->key.objectid)
 171                 info->first_logical_byte = block_group->key.objectid;
 172
 173         spin_unlock(&info->block_group_cache_lock);
 174
 175         return 0;
 176 }
 177
 178 /*
 179  * This will return the block group at or after bytenr if contains is 0, else
 180  * it will return the block group that contains the bytenr
 181  */
 182 static struct btrfs_block_group_cache *
 183 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 184                               int contains)
 185 {
 186         struct btrfs_block_group_cache *cache, *ret = NULL;
 187         struct rb_node *n;
 188         u64 end, start;
 189
 190         spin_lock(&info->block_group_cache_lock);
 191         n = info->block_group_cache_tree.rb_node;
 192
 193         while (n) {
 194                 cache = rb_entry(n, struct btrfs_block_group_cache,
 195                                  cache_node);
 196                 end = cache->key.objectid + cache->key.offset - 1;
 197                 start = cache->key.objectid;
 198
 199                 if (bytenr < start) {
 200                         if (!contains && (!ret || start < ret->key.objectid))
 201                                 ret = cache;
 202                         n = n->rb_left;
 203                 } else if (bytenr > start) {
 204                         if (contains && bytenr <= end) {
 205                                 ret = cache;
 206                                 break;
 207                         }
 208                         n = n->rb_right;
 209                 } else {
 210                         ret = cache;
 211                         break;
 212                 }
 213         }
 214         if (ret) {
 215                 btrfs_get_block_group(ret);
 216                 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
 217                         info->first_logical_byte = ret->key.objectid;
 218         }
 219         spin_unlock(&info->block_group_cache_lock);
 220
 221         return ret;
 222 }
 223
 224 static int add_excluded_extent(struct btrfs_fs_info *fs_info,
 225                                u64 start, u64 num_bytes)
 226 {
 227         u64 end = start + num_bytes - 1;
 228         set_extent_bits(&fs_info->freed_extents[0],
 229                         start, end, EXTENT_UPTODATE);
 230         set_extent_bits(&fs_info->freed_extents[1],
 231                         start, end, EXTENT_UPTODATE);
 232         return 0;
 233 }
 234
 235 static void free_excluded_extents(struct btrfs_fs_info *fs_info,
 236                                   struct btrfs_block_group_cache *cache)
 237 {
 238         u64 start, end;
 239
 240         start = cache->key.objectid;
 241         end = start + cache->key.offset - 1;
 242
 243         clear_extent_bits(&fs_info->freed_extents[0],
 244                           start, end, EXTENT_UPTODATE);
 245         clear_extent_bits(&fs_info->freed_extents[1],
 246                           start, end, EXTENT_UPTODATE);
 247 }
 248
 249 static int exclude_super_stripes(struct btrfs_fs_info *fs_info,
 250                                  struct btrfs_block_group_cache *cache)
 251 {
 252         u64 bytenr;
 253         u64 *logical;
 254         int stripe_len;
 255         int i, nr, ret;
 256
 257         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 258                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 259                 cache->bytes_super += stripe_len;
 260                 ret = add_excluded_extent(fs_info, cache->key.objectid,
 261                                           stripe_len);
 262                 if (ret)
 263                         return ret;
 264         }
 265
 266         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 267                 bytenr = btrfs_sb_offset(i);
 268                 ret = btrfs_rmap_block(fs_info, cache->key.objectid,
 269                                        bytenr, 0, &logical, &nr, &stripe_len);
 270                 if (ret)
 271                         return ret;
 272
 273                 while (nr--) {
 274                         u64 start, len;
 275
 276                         if (logical[nr] > cache->key.objectid +
 277                             cache->key.offset)
 278                                 continue;
 279
 280                         if (logical[nr] + stripe_len <= cache->key.objectid)
 281                                 continue;
 282
 283                         start = logical[nr];
 284                         if (start < cache->key.objectid) {
 285                                 start = cache->key.objectid;
 286                                 len = (logical[nr] + stripe_len) - start;
 287                         } else {
 288                                 len = min_t(u64, stripe_len,
 289                                             cache->key.objectid +
 290                                             cache->key.offset - start);
 291                         }
 292
 293                         cache->bytes_super += len;
 294                         ret = add_excluded_extent(fs_info, start, len);
 295                         if (ret) {
 296                                 kfree(logical);
 297                                 return ret;
 298                         }
 299                 }
 300
 301                 kfree(logical);
 302         }
 303         return 0;
 304 }
 305
 306 static struct btrfs_caching_control *
 307 get_caching_control(struct btrfs_block_group_cache *cache)
 308 {
 309         struct btrfs_caching_control *ctl;
 310
 311         spin_lock(&cache->lock);
 312         if (!cache->caching_ctl) {
 313                 spin_unlock(&cache->lock);
 314                 return NULL;
 315         }
 316
 317         ctl = cache->caching_ctl;
 318         atomic_inc(&ctl->count);
 319         spin_unlock(&cache->lock);
 320         return ctl;
 321 }
 322
 323 static void put_caching_control(struct btrfs_caching_control *ctl)
 324 {
 325         if (atomic_dec_and_test(&ctl->count))
 326                 kfree(ctl);
 327 }
 328
 329 #ifdef CONFIG_BTRFS_DEBUG
 330 static void fragment_free_space(struct btrfs_block_group_cache *block_group)
 331 {
 332         struct btrfs_fs_info *fs_info = block_group->fs_info;
 333         u64 start = block_group->key.objectid;
 334         u64 len = block_group->key.offset;
 335         u64 chunk = block_group->flags & BTRFS_BLOCK_GROUP_METADATA ?
 336                 fs_info->nodesize : fs_info->sectorsize;
 337         u64 step = chunk << 1;
 338
 339         while (len > chunk) {
 340                 btrfs_remove_free_space(block_group, start, chunk);
 341                 start += step;
 342                 if (len < step)
 343                         len = 0;
 344                 else
 345                         len -= step;
 346         }
 347 }
 348 #endif
 349
 350 /*
 351  * this is only called by cache_block_group, since we could have freed extents
 352  * we need to check the pinned_extents for any extents that can't be used yet
 353  * since their free space will be released as soon as the transaction commits.
 354  */
 355 u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 356                        struct btrfs_fs_info *info, u64 start, u64 end)
 357 {
 358         u64 extent_start, extent_end, size, total_added = 0;
 359         int ret;
 360
 361         while (start < end) {
 362                 ret = find_first_extent_bit(info->pinned_extents, start,
 363                                             &extent_start, &extent_end,
 364                                             EXTENT_DIRTY | EXTENT_UPTODATE,
 365                                             NULL);
 366                 if (ret)
 367                         break;
 368
 369                 if (extent_start <= start) {
 370                         start = extent_end + 1;
 371                 } else if (extent_start > start && extent_start < end) {
 372                         size = extent_start - start;
 373                         total_added += size;
 374                         ret = btrfs_add_free_space(block_group, start,
 375                                                    size);
 376                         BUG_ON(ret); /* -ENOMEM or logic error */
 377                         start = extent_end + 1;
 378                 } else {
 379                         break;
 380                 }
 381         }
 382
 383         if (start < end) {
 384                 size = end - start;
 385                 total_added += size;
 386                 ret = btrfs_add_free_space(block_group, start, size);
 387                 BUG_ON(ret); /* -ENOMEM or logic error */
 388         }
 389
 390         return total_added;
 391 }
 392
 393 static int load_extent_tree_free(struct btrfs_caching_control *caching_ctl)
 394 {
 395         struct btrfs_block_group_cache *block_group = caching_ctl->block_group;
 396         struct btrfs_fs_info *fs_info = block_group->fs_info;
 397         struct btrfs_root *extent_root = fs_info->extent_root;
 398         struct btrfs_path *path;
 399         struct extent_buffer *leaf;
 400         struct btrfs_key key;
 401         u64 total_found = 0;
 402         u64 last = 0;
 403         u32 nritems;
 404         int ret;
 405         bool wakeup = true;
 406
 407         path = btrfs_alloc_path();
 408         if (!path)
 409                 return -ENOMEM;
 410
 411         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 412
 413 #ifdef CONFIG_BTRFS_DEBUG
 414         /*
 415          * If we're fragmenting we don't want to make anybody think we can
 416          * allocate from this block group until we've had a chance to fragment
 417          * the free space.
 418          */
 419         if (btrfs_should_fragment_free_space(block_group))
 420                 wakeup = false;
 421 #endif
 422         /*
 423          * We don't want to deadlock with somebody trying to allocate a new
 424          * extent for the extent root while also trying to search the extent
 425          * root to add free space.  So we skip locking and search the commit
 426          * root, since its read-only
 427          */
 428         path->skip_locking = 1;
 429         path->search_commit_root = 1;
 430         path->reada = READA_FORWARD;
 431
 432         key.objectid = last;
 433         key.offset = 0;
 434         key.type = BTRFS_EXTENT_ITEM_KEY;
 435
 436 next:
 437         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 438         if (ret < 0)
 439                 goto out;
 440
 441         leaf = path->nodes[0];
 442         nritems = btrfs_header_nritems(leaf);
 443
 444         while (1) {
 445                 if (btrfs_fs_closing(fs_info) > 1) {
 446                         last = (u64)-1;
 447                         break;
 448                 }
 449
 450                 if (path->slots[0] < nritems) {
 451                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 452                 } else {
 453                         ret = find_next_key(path, 0, &key);
 454                         if (ret)
 455                                 break;
 456
 457                         if (need_resched() ||
 458                             rwsem_is_contended(&fs_info->commit_root_sem)) {
 459                                 if (wakeup)
 460                                         caching_ctl->progress = last;
 461                                 btrfs_release_path(path);
 462                                 up_read(&fs_info->commit_root_sem);
 463                                 mutex_unlock(&caching_ctl->mutex);
 464                                 cond_resched();
 465                                 mutex_lock(&caching_ctl->mutex);
 466                                 down_read(&fs_info->commit_root_sem);
 467                                 goto next;
 468                         }
 469
 470                         ret = btrfs_next_leaf(extent_root, path);
 471                         if (ret < 0)
 472                                 goto out;
 473                         if (ret)
 474                                 break;
 475                         leaf = path->nodes[0];
 476                         nritems = btrfs_header_nritems(leaf);
 477                         continue;
 478                 }
 479
 480                 if (key.objectid < last) {
 481                         key.objectid = last;
 482                         key.offset = 0;
 483                         key.type = BTRFS_EXTENT_ITEM_KEY;
 484
 485                         if (wakeup)
 486                                 caching_ctl->progress = last;
 487                         btrfs_release_path(path);
 488                         goto next;
 489                 }
 490
 491                 if (key.objectid < block_group->key.objectid) {
 492                         path->slots[0]++;
 493                         continue;
 494                 }
 495
 496                 if (key.objectid >= block_group->key.objectid +
 497                     block_group->key.offset)
 498                         break;
 499
 500                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 501                     key.type == BTRFS_METADATA_ITEM_KEY) {
 502                         total_found += add_new_free_space(block_group,
 503                                                           fs_info, last,
 504                                                           key.objectid);
 505                         if (key.type == BTRFS_METADATA_ITEM_KEY)
 506                                 last = key.objectid +
 507                                         fs_info->nodesize;
 508                         else
 509                                 last = key.objectid + key.offset;
 510
 511                         if (total_found > CACHING_CTL_WAKE_UP) {
 512                                 total_found = 0;
 513                                 if (wakeup)
 514                                         wake_up(&caching_ctl->wait);
 515                         }
 516                 }
 517                 path->slots[0]++;
 518         }
 519         ret = 0;
 520
 521         total_found += add_new_free_space(block_group, fs_info, last,
 522                                           block_group->key.objectid +
 523                                           block_group->key.offset);
 524         caching_ctl->progress = (u64)-1;
 525
 526 out:
 527         btrfs_free_path(path);
 528         return ret;
 529 }
 530
 531 static noinline void caching_thread(struct btrfs_work *work)
 532 {
 533         struct btrfs_block_group_cache *block_group;
 534         struct btrfs_fs_info *fs_info;
 535         struct btrfs_caching_control *caching_ctl;
 536         struct btrfs_root *extent_root;
 537         int ret;
 538
 539         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 540         block_group = caching_ctl->block_group;
 541         fs_info = block_group->fs_info;
 542         extent_root = fs_info->extent_root;
 543
 544         mutex_lock(&caching_ctl->mutex);
 545         down_read(&fs_info->commit_root_sem);
 546
 547         if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
 548                 ret = load_free_space_tree(caching_ctl);
 549         else
 550                 ret = load_extent_tree_free(caching_ctl);
 551
 552         spin_lock(&block_group->lock);
 553         block_group->caching_ctl = NULL;
 554         block_group->cached = ret ? BTRFS_CACHE_ERROR : BTRFS_CACHE_FINISHED;
 555         spin_unlock(&block_group->lock);
 556
 557 #ifdef CONFIG_BTRFS_DEBUG
 558         if (btrfs_should_fragment_free_space(block_group)) {
 559                 u64 bytes_used;
 560
 561                 spin_lock(&block_group->space_info->lock);
 562                 spin_lock(&block_group->lock);
 563                 bytes_used = block_group->key.offset -
 564                         btrfs_block_group_used(&block_group->item);
 565                 block_group->space_info->bytes_used += bytes_used >> 1;
 566                 spin_unlock(&block_group->lock);
 567                 spin_unlock(&block_group->space_info->lock);
 568                 fragment_free_space(block_group);
 569         }
 570 #endif
 571
 572         caching_ctl->progress = (u64)-1;
 573
 574         up_read(&fs_info->commit_root_sem);
 575         free_excluded_extents(fs_info, block_group);
 576         mutex_unlock(&caching_ctl->mutex);
 577
 578         wake_up(&caching_ctl->wait);
 579
 580         put_caching_control(caching_ctl);
 581         btrfs_put_block_group(block_group);
 582 }
 583
 584 static int cache_block_group(struct btrfs_block_group_cache *cache,
 585                              int load_cache_only)
 586 {
 587         DEFINE_WAIT(wait);
 588         struct btrfs_fs_info *fs_info = cache->fs_info;
 589         struct btrfs_caching_control *caching_ctl;
 590         int ret = 0;
 591
 592         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 593         if (!caching_ctl)
 594                 return -ENOMEM;
 595
 596         INIT_LIST_HEAD(&caching_ctl->list);
 597         mutex_init(&caching_ctl->mutex);
 598         init_waitqueue_head(&caching_ctl->wait);
 599         caching_ctl->block_group = cache;
 600         caching_ctl->progress = cache->key.objectid;
 601         atomic_set(&caching_ctl->count, 1);
 602         btrfs_init_work(&caching_ctl->work, btrfs_cache_helper,
 603                         caching_thread, NULL, NULL);
 604
 605         spin_lock(&cache->lock);
 606         /*
 607          * This should be a rare occasion, but this could happen I think in the
 608          * case where one thread starts to load the space cache info, and then
 609          * some other thread starts a transaction commit which tries to do an
 610          * allocation while the other thread is still loading the space cache
 611          * info.  The previous loop should have kept us from choosing this block
 612          * group, but if we've moved to the state where we will wait on caching
 613          * block groups we need to first check if we're doing a fast load here,
 614          * so we can wait for it to finish, otherwise we could end up allocating
 615          * from a block group who's cache gets evicted for one reason or
 616          * another.
 617          */
 618         while (cache->cached == BTRFS_CACHE_FAST) {
 619                 struct btrfs_caching_control *ctl;
 620
 621                 ctl = cache->caching_ctl;
 622                 atomic_inc(&ctl->count);
 623                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 624                 spin_unlock(&cache->lock);
 625
 626                 schedule();
 627
 628                 finish_wait(&ctl->wait, &wait);
 629                 put_caching_control(ctl);
 630                 spin_lock(&cache->lock);
 631         }
 632
 633         if (cache->cached != BTRFS_CACHE_NO) {
 634                 spin_unlock(&cache->lock);
 635                 kfree(caching_ctl);
 636                 return 0;
 637         }
 638         WARN_ON(cache->caching_ctl);
 639         cache->caching_ctl = caching_ctl;
 640         cache->cached = BTRFS_CACHE_FAST;
 641         spin_unlock(&cache->lock);
 642
 643         if (fs_info->mount_opt & BTRFS_MOUNT_SPACE_CACHE) {
 644                 mutex_lock(&caching_ctl->mutex);
 645                 ret = load_free_space_cache(fs_info, cache);
 646
 647                 spin_lock(&cache->lock);
 648                 if (ret == 1) {
 649                         cache->caching_ctl = NULL;
 650                         cache->cached = BTRFS_CACHE_FINISHED;
 651                         cache->last_byte_to_unpin = (u64)-1;
 652                         caching_ctl->progress = (u64)-1;
 653                 } else {
 654                         if (load_cache_only) {
 655                                 cache->caching_ctl = NULL;
 656                                 cache->cached = BTRFS_CACHE_NO;
 657                         } else {
 658                                 cache->cached = BTRFS_CACHE_STARTED;
 659                                 cache->has_caching_ctl = 1;
 660                         }
 661                 }
 662                 spin_unlock(&cache->lock);
 663 #ifdef CONFIG_BTRFS_DEBUG
 664                 if (ret == 1 &&
 665                     btrfs_should_fragment_free_space(cache)) {
 666                         u64 bytes_used;
 667
 668                         spin_lock(&cache->space_info->lock);
 669                         spin_lock(&cache->lock);
 670                         bytes_used = cache->key.offset -
 671                                 btrfs_block_group_used(&cache->item);
 672                         cache->space_info->bytes_used += bytes_used >> 1;
 673                         spin_unlock(&cache->lock);
 674                         spin_unlock(&cache->space_info->lock);
 675                         fragment_free_space(cache);
 676                 }
 677 #endif
 678                 mutex_unlock(&caching_ctl->mutex);
 679
 680                 wake_up(&caching_ctl->wait);
 681                 if (ret == 1) {
 682                         put_caching_control(caching_ctl);
 683                         free_excluded_extents(fs_info, cache);
 684                         return 0;
 685                 }
 686         } else {
 687                 /*
 688                  * We're either using the free space tree or no caching at all.
 689                  * Set cached to the appropriate value and wakeup any waiters.
 690                  */
 691                 spin_lock(&cache->lock);
 692                 if (load_cache_only) {
 693                         cache->caching_ctl = NULL;
 694                         cache->cached = BTRFS_CACHE_NO;
 695                 } else {
 696                         cache->cached = BTRFS_CACHE_STARTED;
 697                         cache->has_caching_ctl = 1;
 698                 }
 699                 spin_unlock(&cache->lock);
 700                 wake_up(&caching_ctl->wait);
 701         }
 702
 703         if (load_cache_only) {
 704                 put_caching_control(caching_ctl);
 705                 return 0;
 706         }
 707
 708         down_write(&fs_info->commit_root_sem);
 709         atomic_inc(&caching_ctl->count);
 710         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 711         up_write(&fs_info->commit_root_sem);
 712
 713         btrfs_get_block_group(cache);
 714
 715         btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
 716
 717         return ret;
 718 }
 719
 720 /*
 721  * return the block group that starts at or after bytenr
 722  */
 723 static struct btrfs_block_group_cache *
 724 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 725 {
 726         return block_group_cache_tree_search(info, bytenr, 0);
 727 }
 728
 729 /*
 730  * return the block group that contains the given bytenr
 731  */
 732 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 733                                                  struct btrfs_fs_info *info,
 734                                                  u64 bytenr)
 735 {
 736         return block_group_cache_tree_search(info, bytenr, 1);
 737 }
 738
 739 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 740                                                   u64 flags)
 741 {
 742         struct list_head *head = &info->space_info;
 743         struct btrfs_space_info *found;
 744
 745         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 746
 747         rcu_read_lock();
 748         list_for_each_entry_rcu(found, head, list) {
 749                 if (found->flags & flags) {
 750                         rcu_read_unlock();
 751                         return found;
 752                 }
 753         }
 754         rcu_read_unlock();
 755         return NULL;
 756 }
 757
 758 /*
 759  * after adding space to the filesystem, we need to clear the full flags
 760  * on all the space infos.
 761  */
 762 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 763 {
 764         struct list_head *head = &info->space_info;
 765         struct btrfs_space_info *found;
 766
 767         rcu_read_lock();
 768         list_for_each_entry_rcu(found, head, list)
 769                 found->full = 0;
 770         rcu_read_unlock();
 771 }
 772
 773 /* simple helper to search for an existing data extent at a given offset */
 774 int btrfs_lookup_data_extent(struct btrfs_fs_info *fs_info, u64 start, u64 len)
 775 {
 776         int ret;
 777         struct btrfs_key key;
 778         struct btrfs_path *path;
 779
 780         path = btrfs_alloc_path();
 781         if (!path)
 782                 return -ENOMEM;
 783
 784         key.objectid = start;
 785         key.offset = len;
 786         key.type = BTRFS_EXTENT_ITEM_KEY;
 787         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
 788         btrfs_free_path(path);
 789         return ret;
 790 }
 791
 792 /*
 793  * helper function to lookup reference count and flags of a tree block.
 794  *
 795  * the head node for delayed ref is used to store the sum of all the
 796  * reference count modifications queued up in the rbtree. the head
 797  * node may also store the extent flags to set. This way you can check
 798  * to see what the reference count and extent flags would be if all of
 799  * the delayed refs are not processed.
 800  */
 801 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 802                              struct btrfs_fs_info *fs_info, u64 bytenr,
 803                              u64 offset, int metadata, u64 *refs, u64 *flags)
 804 {
 805         struct btrfs_delayed_ref_head *head;
 806         struct btrfs_delayed_ref_root *delayed_refs;
 807         struct btrfs_path *path;
 808         struct btrfs_extent_item *ei;
 809         struct extent_buffer *leaf;
 810         struct btrfs_key key;
 811         u32 item_size;
 812         u64 num_refs;
 813         u64 extent_flags;
 814         int ret;
 815
 816         /*
 817          * If we don't have skinny metadata, don't bother doing anything
 818          * different
 819          */
 820         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA)) {
 821                 offset = fs_info->nodesize;
 822                 metadata = 0;
 823         }
 824
 825         path = btrfs_alloc_path();
 826         if (!path)
 827                 return -ENOMEM;
 828
 829         if (!trans) {
 830                 path->skip_locking = 1;
 831                 path->search_commit_root = 1;
 832         }
 833
 834 search_again:
 835         key.objectid = bytenr;
 836         key.offset = offset;
 837         if (metadata)
 838                 key.type = BTRFS_METADATA_ITEM_KEY;
 839         else
 840                 key.type = BTRFS_EXTENT_ITEM_KEY;
 841
 842         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 0);
 843         if (ret < 0)
 844                 goto out_free;
 845
 846         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 847                 if (path->slots[0]) {
 848                         path->slots[0]--;
 849                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 850                                               path->slots[0]);
 851                         if (key.objectid == bytenr &&
 852                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 853                             key.offset == fs_info->nodesize)
 854                                 ret = 0;
 855                 }
 856         }
 857
 858         if (ret == 0) {
 859                 leaf = path->nodes[0];
 860                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 861                 if (item_size >= sizeof(*ei)) {
 862                         ei = btrfs_item_ptr(leaf, path->slots[0],
 863                                             struct btrfs_extent_item);
 864                         num_refs = btrfs_extent_refs(leaf, ei);
 865                         extent_flags = btrfs_extent_flags(leaf, ei);
 866                 } else {
 867 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 868                         struct btrfs_extent_item_v0 *ei0;
 869                         BUG_ON(item_size != sizeof(*ei0));
 870                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 871                                              struct btrfs_extent_item_v0);
 872                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
 873                         /* FIXME: this isn't correct for data */
 874                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 875 #else
 876                         BUG();
 877 #endif
 878                 }
 879                 BUG_ON(num_refs == 0);
 880         } else {
 881                 num_refs = 0;
 882                 extent_flags = 0;
 883                 ret = 0;
 884         }
 885
 886         if (!trans)
 887                 goto out;
 888
 889         delayed_refs = &trans->transaction->delayed_refs;
 890         spin_lock(&delayed_refs->lock);
 891         head = btrfs_find_delayed_ref_head(trans, bytenr);
 892         if (head) {
 893                 if (!mutex_trylock(&head->mutex)) {
 894                         atomic_inc(&head->node.refs);
 895                         spin_unlock(&delayed_refs->lock);
 896
 897                         btrfs_release_path(path);
 898
 899                         /*
 900                          * Mutex was contended, block until it's released and try
 901                          * again
 902                          */
 903                         mutex_lock(&head->mutex);
 904                         mutex_unlock(&head->mutex);
 905                         btrfs_put_delayed_ref(&head->node);
 906                         goto search_again;
 907                 }
 908                 spin_lock(&head->lock);
 909                 if (head->extent_op && head->extent_op->update_flags)
 910                         extent_flags |= head->extent_op->flags_to_set;
 911                 else
 912                         BUG_ON(num_refs == 0);
 913
 914                 num_refs += head->node.ref_mod;
 915                 spin_unlock(&head->lock);
 916                 mutex_unlock(&head->mutex);
 917         }
 918         spin_unlock(&delayed_refs->lock);
 919 out:
 920         WARN_ON(num_refs == 0);
 921         if (refs)
 922                 *refs = num_refs;
 923         if (flags)
 924                 *flags = extent_flags;
 925 out_free:
 926         btrfs_free_path(path);
 927         return ret;
 928 }
 929
 930 /*
 931  * Back reference rules.  Back refs have three main goals:
 932  *
 933  * 1) differentiate between all holders of references to an extent so that
 934  *    when a reference is dropped we can make sure it was a valid reference
 935  *    before freeing the extent.
 936  *
 937  * 2) Provide enough information to quickly find the holders of an extent
 938  *    if we notice a given block is corrupted or bad.
 939  *
 940  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 941  *    maintenance.  This is actually the same as #2, but with a slightly
 942  *    different use case.
 943  *
 944  * There are two kinds of back refs. The implicit back refs is optimized
 945  * for pointers in non-shared tree blocks. For a given pointer in a block,
 946  * back refs of this kind provide information about the block's owner tree
 947  * and the pointer's key. These information allow us to find the block by
 948  * b-tree searching. The full back refs is for pointers in tree blocks not
 949  * referenced by their owner trees. The location of tree block is recorded
 950  * in the back refs. Actually the full back refs is generic, and can be
 951  * used in all cases the implicit back refs is used. The major shortcoming
 952  * of the full back refs is its overhead. Every time a tree block gets
 953  * COWed, we have to update back refs entry for all pointers in it.
 954  *
 955  * For a newly allocated tree block, we use implicit back refs for
 956  * pointers in it. This means most tree related operations only involve
 957  * implicit back refs. For a tree block created in old transaction, the
 958  * only way to drop a reference to it is COW it. So we can detect the
 959  * event that tree block loses its owner tree's reference and do the
 960  * back refs conversion.
 961  *
 962  * When a tree block is COWed through a tree, there are four cases:
 963  *
 964  * The reference count of the block is one and the tree is the block's
 965  * owner tree. Nothing to do in this case.
 966  *
 967  * The reference count of the block is one and the tree is not the
 968  * block's owner tree. In this case, full back refs is used for pointers
 969  * in the block. Remove these full back refs, add implicit back refs for
 970  * every pointers in the new block.
 971  *
 972  * The reference count of the block is greater than one and the tree is
 973  * the block's owner tree. In this case, implicit back refs is used for
 974  * pointers in the block. Add full back refs for every pointers in the
 975  * block, increase lower level extents' reference counts. The original
 976  * implicit back refs are entailed to the new block.
 977  *
 978  * The reference count of the block is greater than one and the tree is
 979  * not the block's owner tree. Add implicit back refs for every pointer in
 980  * the new block, increase lower level extents' reference count.
 981  *
 982  * Back Reference Key composing:
 983  *
 984  * The key objectid corresponds to the first byte in the extent,
 985  * The key type is used to differentiate between types of back refs.
 986  * There are different meanings of the key offset for different types
 987  * of back refs.
 988  *
 989  * File extents can be referenced by:
 990  *
 991  * - multiple snapshots, subvolumes, or different generations in one subvol
 992  * - different files inside a single subvolume
 993  * - different offsets inside a file (bookend extents in file.c)
 994  *
 995  * The extent ref structure for the implicit back refs has fields for:
 996  *
 997  * - Objectid of the subvolume root
 998  * - objectid of the file holding the reference
 999  * - original offset in the file
1000  * - how many bookend extents
1001  *
1002  * The key offset for the implicit back refs is hash of the first
1003  * three fields.
1004  *
1005  * The extent ref structure for the full back refs has field for:
1006  *
1007  * - number of pointers in the tree leaf
1008  *
1009  * The key offset for the implicit back refs is the first byte of
1010  * the tree leaf
1011  *
1012  * When a file extent is allocated, The implicit back refs is used.
1013  * the fields are filled in:
1014  *
1015  *     (root_key.objectid, inode objectid, offset in file, 1)
1016  *
1017  * When a file extent is removed file truncation, we find the
1018  * corresponding implicit back refs and check the following fields:
1019  *
1020  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
1021  *
1022  * Btree extents can be referenced by:
1023  *
1024  * - Different subvolumes
1025  *
1026  * Both the implicit back refs and the full back refs for tree blocks
1027  * only consist of key. The key offset for the implicit back refs is
1028  * objectid of block's owner tree. The key offset for the full back refs
1029  * is the first byte of parent block.
1030  *
1031  * When implicit back refs is used, information about the lowest key and
1032  * level of the tree block are required. These information are stored in
1033  * tree block info structure.
1034  */
1035
1036 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1037 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
1038                                   struct btrfs_root *root,
1039                                   struct btrfs_path *path,
1040                                   u64 owner, u32 extra_size)
1041 {
1042         struct btrfs_extent_item *item;
1043         struct btrfs_extent_item_v0 *ei0;
1044         struct btrfs_extent_ref_v0 *ref0;
1045         struct btrfs_tree_block_info *bi;
1046         struct extent_buffer *leaf;
1047         struct btrfs_key key;
1048         struct btrfs_key found_key;
1049         u32 new_size = sizeof(*item);
1050         u64 refs;
1051         int ret;
1052
1053         leaf = path->nodes[0];
1054         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
1055
1056         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1057         ei0 = btrfs_item_ptr(leaf, path->slots[0],
1058                              struct btrfs_extent_item_v0);
1059         refs = btrfs_extent_refs_v0(leaf, ei0);
1060
1061         if (owner == (u64)-1) {
1062                 while (1) {
1063                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1064                                 ret = btrfs_next_leaf(root, path);
1065                                 if (ret < 0)
1066                                         return ret;
1067                                 BUG_ON(ret > 0); /* Corruption */
1068                                 leaf = path->nodes[0];
1069                         }
1070                         btrfs_item_key_to_cpu(leaf, &found_key,
1071                                               path->slots[0]);
1072                         BUG_ON(key.objectid != found_key.objectid);
1073                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
1074                                 path->slots[0]++;
1075                                 continue;
1076                         }
1077                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1078                                               struct btrfs_extent_ref_v0);
1079                         owner = btrfs_ref_objectid_v0(leaf, ref0);
1080                         break;
1081                 }
1082         }
1083         btrfs_release_path(path);
1084
1085         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1086                 new_size += sizeof(*bi);
1087
1088         new_size -= sizeof(*ei0);
1089         ret = btrfs_search_slot(trans, root, &key, path,
1090                                 new_size + extra_size, 1);
1091         if (ret < 0)
1092                 return ret;
1093         BUG_ON(ret); /* Corruption */
1094
1095         btrfs_extend_item(root->fs_info, path, new_size);
1096
1097         leaf = path->nodes[0];
1098         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1099         btrfs_set_extent_refs(leaf, item, refs);
1100         /* FIXME: get real generation */
1101         btrfs_set_extent_generation(leaf, item, 0);
1102         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1103                 btrfs_set_extent_flags(leaf, item,
1104                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1105                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1106                 bi = (struct btrfs_tree_block_info *)(item + 1);
1107                 /* FIXME: get first key of the block */
1108                 memzero_extent_buffer(leaf, (unsigned long)bi, sizeof(*bi));
1109                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1110         } else {
1111                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1112         }
1113         btrfs_mark_buffer_dirty(leaf);
1114         return 0;
1115 }
1116 #endif
1117
1118 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1119 {
1120         u32 high_crc = ~(u32)0;
1121         u32 low_crc = ~(u32)0;
1122         __le64 lenum;
1123
1124         lenum = cpu_to_le64(root_objectid);
1125         high_crc = btrfs_crc32c(high_crc, &lenum, sizeof(lenum));
1126         lenum = cpu_to_le64(owner);
1127         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
1128         lenum = cpu_to_le64(offset);
1129         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
1130
1131         return ((u64)high_crc << 31) ^ (u64)low_crc;
1132 }
1133
1134 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1135                                      struct btrfs_extent_data_ref *ref)
1136 {
1137         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1138                                     btrfs_extent_data_ref_objectid(leaf, ref),
1139                                     btrfs_extent_data_ref_offset(leaf, ref));
1140 }
1141
1142 static int match_extent_data_ref(struct extent_buffer *leaf,
1143                                  struct btrfs_extent_data_ref *ref,
1144                                  u64 root_objectid, u64 owner, u64 offset)
1145 {
1146         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1147             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1148             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1149                 return 0;
1150         return 1;
1151 }
1152
1153 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1154                                            struct btrfs_root *root,
1155                                            struct btrfs_path *path,
1156                                            u64 bytenr, u64 parent,
1157                                            u64 root_objectid,
1158                                            u64 owner, u64 offset)
1159 {
1160         struct btrfs_key key;
1161         struct btrfs_extent_data_ref *ref;
1162         struct extent_buffer *leaf;
1163         u32 nritems;
1164         int ret;
1165         int recow;
1166         int err = -ENOENT;
1167
1168         key.objectid = bytenr;
1169         if (parent) {
1170                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1171                 key.offset = parent;
1172         } else {
1173                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1174                 key.offset = hash_extent_data_ref(root_objectid,
1175                                                   owner, offset);
1176         }
1177 again:
1178         recow = 0;
1179         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1180         if (ret < 0) {
1181                 err = ret;
1182                 goto fail;
1183         }
1184
1185         if (parent) {
1186                 if (!ret)
1187                         return 0;
1188 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1189                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1190                 btrfs_release_path(path);
1191                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1192                 if (ret < 0) {
1193                         err = ret;
1194                         goto fail;
1195                 }
1196                 if (!ret)
1197                         return 0;
1198 #endif
1199                 goto fail;
1200         }
1201
1202         leaf = path->nodes[0];
1203         nritems = btrfs_header_nritems(leaf);
1204         while (1) {
1205                 if (path->slots[0] >= nritems) {
1206                         ret = btrfs_next_leaf(root, path);
1207                         if (ret < 0)
1208                                 err = ret;
1209                         if (ret)
1210                                 goto fail;
1211
1212                         leaf = path->nodes[0];
1213                         nritems = btrfs_header_nritems(leaf);
1214                         recow = 1;
1215                 }
1216
1217                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1218                 if (key.objectid != bytenr ||
1219                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1220                         goto fail;
1221
1222                 ref = btrfs_item_ptr(leaf, path->slots[0],
1223                                      struct btrfs_extent_data_ref);
1224
1225                 if (match_extent_data_ref(leaf, ref, root_objectid,
1226                                           owner, offset)) {
1227                         if (recow) {
1228                                 btrfs_release_path(path);
1229                                 goto again;
1230                         }
1231                         err = 0;
1232                         break;
1233                 }
1234                 path->slots[0]++;
1235         }
1236 fail:
1237         return err;
1238 }
1239
1240 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1241                                            struct btrfs_root *root,
1242                                            struct btrfs_path *path,
1243                                            u64 bytenr, u64 parent,
1244                                            u64 root_objectid, u64 owner,
1245                                            u64 offset, int refs_to_add)
1246 {
1247         struct btrfs_key key;
1248         struct extent_buffer *leaf;
1249         u32 size;
1250         u32 num_refs;
1251         int ret;
1252
1253         key.objectid = bytenr;
1254         if (parent) {
1255                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1256                 key.offset = parent;
1257                 size = sizeof(struct btrfs_shared_data_ref);
1258         } else {
1259                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1260                 key.offset = hash_extent_data_ref(root_objectid,
1261                                                   owner, offset);
1262                 size = sizeof(struct btrfs_extent_data_ref);
1263         }
1264
1265         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1266         if (ret && ret != -EEXIST)
1267                 goto fail;
1268
1269         leaf = path->nodes[0];
1270         if (parent) {
1271                 struct btrfs_shared_data_ref *ref;
1272                 ref = btrfs_item_ptr(leaf, path->slots[0],
1273                                      struct btrfs_shared_data_ref);
1274                 if (ret == 0) {
1275                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1276                 } else {
1277                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1278                         num_refs += refs_to_add;
1279                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1280                 }
1281         } else {
1282                 struct btrfs_extent_data_ref *ref;
1283                 while (ret == -EEXIST) {
1284                         ref = btrfs_item_ptr(leaf, path->slots[0],
1285                                              struct btrfs_extent_data_ref);
1286                         if (match_extent_data_ref(leaf, ref, root_objectid,
1287                                                   owner, offset))
1288                                 break;
1289                         btrfs_release_path(path);
1290                         key.offset++;
1291                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1292                                                       size);
1293                         if (ret && ret != -EEXIST)
1294                                 goto fail;
1295
1296                         leaf = path->nodes[0];
1297                 }
1298                 ref = btrfs_item_ptr(leaf, path->slots[0],
1299                                      struct btrfs_extent_data_ref);
1300                 if (ret == 0) {
1301                         btrfs_set_extent_data_ref_root(leaf, ref,
1302                                                        root_objectid);
1303                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1304                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1305                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1306                 } else {
1307                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1308                         num_refs += refs_to_add;
1309                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1310                 }
1311         }
1312         btrfs_mark_buffer_dirty(leaf);
1313         ret = 0;
1314 fail:
1315         btrfs_release_path(path);
1316         return ret;
1317 }
1318
1319 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1320                                            struct btrfs_root *root,
1321                                            struct btrfs_path *path,
1322                                            int refs_to_drop, int *last_ref)
1323 {
1324         struct btrfs_key key;
1325         struct btrfs_extent_data_ref *ref1 = NULL;
1326         struct btrfs_shared_data_ref *ref2 = NULL;
1327         struct extent_buffer *leaf;
1328         u32 num_refs = 0;
1329         int ret = 0;
1330
1331         leaf = path->nodes[0];
1332         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1333
1334         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1335                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1336                                       struct btrfs_extent_data_ref);
1337                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1338         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1339                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1340                                       struct btrfs_shared_data_ref);
1341                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1342 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1343         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1344                 struct btrfs_extent_ref_v0 *ref0;
1345                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1346                                       struct btrfs_extent_ref_v0);
1347                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1348 #endif
1349         } else {
1350                 BUG();
1351         }
1352
1353         BUG_ON(num_refs < refs_to_drop);
1354         num_refs -= refs_to_drop;
1355
1356         if (num_refs == 0) {
1357                 ret = btrfs_del_item(trans, root, path);
1358                 *last_ref = 1;
1359         } else {
1360                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1361                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1362                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1363                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1364 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1365                 else {
1366                         struct btrfs_extent_ref_v0 *ref0;
1367                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1368                                         struct btrfs_extent_ref_v0);
1369                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1370                 }
1371 #endif
1372                 btrfs_mark_buffer_dirty(leaf);
1373         }
1374         return ret;
1375 }
1376
1377 static noinline u32 extent_data_ref_count(struct btrfs_path *path,
1378                                           struct btrfs_extent_inline_ref *iref)
1379 {
1380         struct btrfs_key key;
1381         struct extent_buffer *leaf;
1382         struct btrfs_extent_data_ref *ref1;
1383         struct btrfs_shared_data_ref *ref2;
1384         u32 num_refs = 0;
1385
1386         leaf = path->nodes[0];
1387         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1388         if (iref) {
1389                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1390                     BTRFS_EXTENT_DATA_REF_KEY) {
1391                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1392                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1393                 } else {
1394                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1395                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1396                 }
1397         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1398                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1399                                       struct btrfs_extent_data_ref);
1400                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1401         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1402                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1403                                       struct btrfs_shared_data_ref);
1404                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1405 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1406         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1407                 struct btrfs_extent_ref_v0 *ref0;
1408                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1409                                       struct btrfs_extent_ref_v0);
1410                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1411 #endif
1412         } else {
1413                 WARN_ON(1);
1414         }
1415         return num_refs;
1416 }
1417
1418 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1419                                           struct btrfs_root *root,
1420                                           struct btrfs_path *path,
1421                                           u64 bytenr, u64 parent,
1422                                           u64 root_objectid)
1423 {
1424         struct btrfs_key key;
1425         int ret;
1426
1427         key.objectid = bytenr;
1428         if (parent) {
1429                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1430                 key.offset = parent;
1431         } else {
1432                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1433                 key.offset = root_objectid;
1434         }
1435
1436         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1437         if (ret > 0)
1438                 ret = -ENOENT;
1439 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1440         if (ret == -ENOENT && parent) {
1441                 btrfs_release_path(path);
1442                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1443                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1444                 if (ret > 0)
1445                         ret = -ENOENT;
1446         }
1447 #endif
1448         return ret;
1449 }
1450
1451 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1452                                           struct btrfs_root *root,
1453                                           struct btrfs_path *path,
1454                                           u64 bytenr, u64 parent,
1455                                           u64 root_objectid)
1456 {
1457         struct btrfs_key key;
1458         int ret;
1459
1460         key.objectid = bytenr;
1461         if (parent) {
1462                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1463                 key.offset = parent;
1464         } else {
1465                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1466                 key.offset = root_objectid;
1467         }
1468
1469         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1470         btrfs_release_path(path);
1471         return ret;
1472 }
1473
1474 static inline int extent_ref_type(u64 parent, u64 owner)
1475 {
1476         int type;
1477         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1478                 if (parent > 0)
1479                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1480                 else
1481                         type = BTRFS_TREE_BLOCK_REF_KEY;
1482         } else {
1483                 if (parent > 0)
1484                         type = BTRFS_SHARED_DATA_REF_KEY;
1485                 else
1486                         type = BTRFS_EXTENT_DATA_REF_KEY;
1487         }
1488         return type;
1489 }
1490
1491 static int find_next_key(struct btrfs_path *path, int level,
1492                          struct btrfs_key *key)
1493
1494 {
1495         for (; level < BTRFS_MAX_LEVEL; level++) {
1496                 if (!path->nodes[level])
1497                         break;
1498                 if (path->slots[level] + 1 >=
1499                     btrfs_header_nritems(path->nodes[level]))
1500                         continue;
1501                 if (level == 0)
1502                         btrfs_item_key_to_cpu(path->nodes[level], key,
1503                                               path->slots[level] + 1);
1504                 else
1505                         btrfs_node_key_to_cpu(path->nodes[level], key,
1506                                               path->slots[level] + 1);
1507                 return 0;
1508         }
1509         return 1;
1510 }
1511
1512 /*
1513  * look for inline back ref. if back ref is found, *ref_ret is set
1514  * to the address of inline back ref, and 0 is returned.
1515  *
1516  * if back ref isn't found, *ref_ret is set to the address where it
1517  * should be inserted, and -ENOENT is returned.
1518  *
1519  * if insert is true and there are too many inline back refs, the path
1520  * points to the extent item, and -EAGAIN is returned.
1521  *
1522  * NOTE: inline back refs are ordered in the same way that back ref
1523  *       items in the tree are ordered.
1524  */
1525 static noinline_for_stack
1526 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1527                                  struct btrfs_root *root,
1528                                  struct btrfs_path *path,
1529                                  struct btrfs_extent_inline_ref **ref_ret,
1530                                  u64 bytenr, u64 num_bytes,
1531                                  u64 parent, u64 root_objectid,
1532                                  u64 owner, u64 offset, int insert)
1533 {
1534         struct btrfs_fs_info *fs_info = root->fs_info;
1535         struct btrfs_key key;
1536         struct extent_buffer *leaf;
1537         struct btrfs_extent_item *ei;
1538         struct btrfs_extent_inline_ref *iref;
1539         u64 flags;
1540         u64 item_size;
1541         unsigned long ptr;
1542         unsigned long end;
1543         int extra_size;
1544         int type;
1545         int want;
1546         int ret;
1547         int err = 0;
1548         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
1549
1550         key.objectid = bytenr;
1551         key.type = BTRFS_EXTENT_ITEM_KEY;
1552         key.offset = num_bytes;
1553
1554         want = extent_ref_type(parent, owner);
1555         if (insert) {
1556                 extra_size = btrfs_extent_inline_ref_size(want);
1557                 path->keep_locks = 1;
1558         } else
1559                 extra_size = -1;
1560
1561         /*
1562          * Owner is our parent level, so we can just add one to get the level
1563          * for the block we are interested in.
1564          */
1565         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
1566                 key.type = BTRFS_METADATA_ITEM_KEY;
1567                 key.offset = owner;
1568         }
1569
1570 again:
1571         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1572         if (ret < 0) {
1573                 err = ret;
1574                 goto out;
1575         }
1576
1577         /*
1578          * We may be a newly converted file system which still has the old fat
1579          * extent entries for metadata, so try and see if we have one of those.
1580          */
1581         if (ret > 0 && skinny_metadata) {
1582                 skinny_metadata = false;
1583                 if (path->slots[0]) {
1584                         path->slots[0]--;
1585                         btrfs_item_key_to_cpu(path->nodes[0], &key,
1586                                               path->slots[0]);
1587                         if (key.objectid == bytenr &&
1588                             key.type == BTRFS_EXTENT_ITEM_KEY &&
1589                             key.offset == num_bytes)
1590                                 ret = 0;
1591                 }
1592                 if (ret) {
1593                         key.objectid = bytenr;
1594                         key.type = BTRFS_EXTENT_ITEM_KEY;
1595                         key.offset = num_bytes;
1596                         btrfs_release_path(path);
1597                         goto again;
1598                 }
1599         }
1600
1601         if (ret && !insert) {
1602                 err = -ENOENT;
1603                 goto out;
1604         } else if (WARN_ON(ret)) {
1605                 err = -EIO;
1606                 goto out;
1607         }
1608
1609         leaf = path->nodes[0];
1610         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1611 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1612         if (item_size < sizeof(*ei)) {
1613                 if (!insert) {
1614                         err = -ENOENT;
1615                         goto out;
1616                 }
1617                 ret = convert_extent_item_v0(trans, root, path, owner,
1618                                              extra_size);
1619                 if (ret < 0) {
1620                         err = ret;
1621                         goto out;
1622                 }
1623                 leaf = path->nodes[0];
1624                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1625         }
1626 #endif
1627         BUG_ON(item_size < sizeof(*ei));
1628
1629         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1630         flags = btrfs_extent_flags(leaf, ei);
1631
1632         ptr = (unsigned long)(ei + 1);
1633         end = (unsigned long)ei + item_size;
1634
1635         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
1636                 ptr += sizeof(struct btrfs_tree_block_info);
1637                 BUG_ON(ptr > end);
1638         }
1639
1640         err = -ENOENT;
1641         while (1) {
1642                 if (ptr >= end) {
1643                         WARN_ON(ptr > end);
1644                         break;
1645                 }
1646                 iref = (struct btrfs_extent_inline_ref *)ptr;
1647                 type = btrfs_extent_inline_ref_type(leaf, iref);
1648                 if (want < type)
1649                         break;
1650                 if (want > type) {
1651                         ptr += btrfs_extent_inline_ref_size(type);
1652                         continue;
1653                 }
1654
1655                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1656                         struct btrfs_extent_data_ref *dref;
1657                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1658                         if (match_extent_data_ref(leaf, dref, root_objectid,
1659                                                   owner, offset)) {
1660                                 err = 0;
1661                                 break;
1662                         }
1663                         if (hash_extent_data_ref_item(leaf, dref) <
1664                             hash_extent_data_ref(root_objectid, owner, offset))
1665                                 break;
1666                 } else {
1667                         u64 ref_offset;
1668                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1669                         if (parent > 0) {
1670                                 if (parent == ref_offset) {
1671                                         err = 0;
1672                                         break;
1673                                 }
1674                                 if (ref_offset < parent)
1675                                         break;
1676                         } else {
1677                                 if (root_objectid == ref_offset) {
1678                                         err = 0;
1679                                         break;
1680                                 }
1681                                 if (ref_offset < root_objectid)
1682                                         break;
1683                         }
1684                 }
1685                 ptr += btrfs_extent_inline_ref_size(type);
1686         }
1687         if (err == -ENOENT && insert) {
1688                 if (item_size + extra_size >=
1689                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1690                         err = -EAGAIN;
1691                         goto out;
1692                 }
1693                 /*
1694                  * To add new inline back ref, we have to make sure
1695                  * there is no corresponding back ref item.
1696                  * For simplicity, we just do not add new inline back
1697                  * ref if there is any kind of item for this block
1698                  */
1699                 if (find_next_key(path, 0, &key) == 0 &&
1700                     key.objectid == bytenr &&
1701                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1702                         err = -EAGAIN;
1703                         goto out;
1704                 }
1705         }
1706         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1707 out:
1708         if (insert) {
1709                 path->keep_locks = 0;
1710                 btrfs_unlock_up_safe(path, 1);
1711         }
1712         return err;
1713 }
1714
1715 /*
1716  * helper to add new inline back ref
1717  */
1718 static noinline_for_stack
1719 void setup_inline_extent_backref(struct btrfs_root *root,
1720                                  struct btrfs_path *path,
1721                                  struct btrfs_extent_inline_ref *iref,
1722                                  u64 parent, u64 root_objectid,
1723                                  u64 owner, u64 offset, int refs_to_add,
1724                                  struct btrfs_delayed_extent_op *extent_op)
1725 {
1726         struct extent_buffer *leaf;
1727         struct btrfs_extent_item *ei;
1728         unsigned long ptr;
1729         unsigned long end;
1730         unsigned long item_offset;
1731         u64 refs;
1732         int size;
1733         int type;
1734
1735         leaf = path->nodes[0];
1736         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1737         item_offset = (unsigned long)iref - (unsigned long)ei;
1738
1739         type = extent_ref_type(parent, owner);
1740         size = btrfs_extent_inline_ref_size(type);
1741
1742         btrfs_extend_item(root->fs_info, path, size);
1743
1744         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1745         refs = btrfs_extent_refs(leaf, ei);
1746         refs += refs_to_add;
1747         btrfs_set_extent_refs(leaf, ei, refs);
1748         if (extent_op)
1749                 __run_delayed_extent_op(extent_op, leaf, ei);
1750
1751         ptr = (unsigned long)ei + item_offset;
1752         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1753         if (ptr < end - size)
1754                 memmove_extent_buffer(leaf, ptr + size, ptr,
1755                                       end - size - ptr);
1756
1757         iref = (struct btrfs_extent_inline_ref *)ptr;
1758         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1759         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1760                 struct btrfs_extent_data_ref *dref;
1761                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1762                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1763                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1764                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1765                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1766         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1767                 struct btrfs_shared_data_ref *sref;
1768                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1769                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1770                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1771         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1772                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1773         } else {
1774                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1775         }
1776         btrfs_mark_buffer_dirty(leaf);
1777 }
1778
1779 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1780                                  struct btrfs_root *root,
1781                                  struct btrfs_path *path,
1782                                  struct btrfs_extent_inline_ref **ref_ret,
1783                                  u64 bytenr, u64 num_bytes, u64 parent,
1784                                  u64 root_objectid, u64 owner, u64 offset)
1785 {
1786         int ret;
1787
1788         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1789                                            bytenr, num_bytes, parent,
1790                                            root_objectid, owner, offset, 0);
1791         if (ret != -ENOENT)
1792                 return ret;
1793
1794         btrfs_release_path(path);
1795         *ref_ret = NULL;
1796
1797         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1798                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1799                                             root_objectid);
1800         } else {
1801                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1802                                              root_objectid, owner, offset);
1803         }
1804         return ret;
1805 }
1806
1807 /*
1808  * helper to update/remove inline back ref
1809  */
1810 static noinline_for_stack
1811 void update_inline_extent_backref(struct btrfs_root *root,
1812                                   struct btrfs_path *path,
1813                                   struct btrfs_extent_inline_ref *iref,
1814                                   int refs_to_mod,
1815                                   struct btrfs_delayed_extent_op *extent_op,
1816                                   int *last_ref)
1817 {
1818         struct extent_buffer *leaf;
1819         struct btrfs_extent_item *ei;
1820         struct btrfs_extent_data_ref *dref = NULL;
1821         struct btrfs_shared_data_ref *sref = NULL;
1822         unsigned long ptr;
1823         unsigned long end;
1824         u32 item_size;
1825         int size;
1826         int type;
1827         u64 refs;
1828
1829         leaf = path->nodes[0];
1830         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1831         refs = btrfs_extent_refs(leaf, ei);
1832         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1833         refs += refs_to_mod;
1834         btrfs_set_extent_refs(leaf, ei, refs);
1835         if (extent_op)
1836                 __run_delayed_extent_op(extent_op, leaf, ei);
1837
1838         type = btrfs_extent_inline_ref_type(leaf, iref);
1839
1840         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1841                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1842                 refs = btrfs_extent_data_ref_count(leaf, dref);
1843         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1844                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1845                 refs = btrfs_shared_data_ref_count(leaf, sref);
1846         } else {
1847                 refs = 1;
1848                 BUG_ON(refs_to_mod != -1);
1849         }
1850
1851         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1852         refs += refs_to_mod;
1853
1854         if (refs > 0) {
1855                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1856                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1857                 else
1858                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1859         } else {
1860                 *last_ref = 1;
1861                 size =  btrfs_extent_inline_ref_size(type);
1862                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1863                 ptr = (unsigned long)iref;
1864                 end = (unsigned long)ei + item_size;
1865                 if (ptr + size < end)
1866                         memmove_extent_buffer(leaf, ptr, ptr + size,
1867                                               end - ptr - size);
1868                 item_size -= size;
1869                 btrfs_truncate_item(root->fs_info, path, item_size, 1);
1870         }
1871         btrfs_mark_buffer_dirty(leaf);
1872 }
1873
1874 static noinline_for_stack
1875 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1876                                  struct btrfs_root *root,
1877                                  struct btrfs_path *path,
1878                                  u64 bytenr, u64 num_bytes, u64 parent,
1879                                  u64 root_objectid, u64 owner,
1880                                  u64 offset, int refs_to_add,
1881                                  struct btrfs_delayed_extent_op *extent_op)
1882 {
1883         struct btrfs_extent_inline_ref *iref;
1884         int ret;
1885
1886         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1887                                            bytenr, num_bytes, parent,
1888                                            root_objectid, owner, offset, 1);
1889         if (ret == 0) {
1890                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1891                 update_inline_extent_backref(root, path, iref,
1892                                              refs_to_add, extent_op, NULL);
1893         } else if (ret == -ENOENT) {
1894                 setup_inline_extent_backref(root, path, iref, parent,
1895                                             root_objectid, owner, offset,
1896                                             refs_to_add, extent_op);
1897                 ret = 0;
1898         }
1899         return ret;
1900 }
1901
1902 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1903                                  struct btrfs_root *root,
1904                                  struct btrfs_path *path,
1905                                  u64 bytenr, u64 parent, u64 root_objectid,
1906                                  u64 owner, u64 offset, int refs_to_add)
1907 {
1908         int ret;
1909         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1910                 BUG_ON(refs_to_add != 1);
1911                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1912                                             parent, root_objectid);
1913         } else {
1914                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1915                                              parent, root_objectid,
1916                                              owner, offset, refs_to_add);
1917         }
1918         return ret;
1919 }
1920
1921 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1922                                  struct btrfs_root *root,
1923                                  struct btrfs_path *path,
1924                                  struct btrfs_extent_inline_ref *iref,
1925                                  int refs_to_drop, int is_data, int *last_ref)
1926 {
1927         int ret = 0;
1928
1929         BUG_ON(!is_data && refs_to_drop != 1);
1930         if (iref) {
1931                 update_inline_extent_backref(root, path, iref,
1932                                              -refs_to_drop, NULL, last_ref);
1933         } else if (is_data) {
1934                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop,
1935                                              last_ref);
1936         } else {
1937                 *last_ref = 1;
1938                 ret = btrfs_del_item(trans, root, path);
1939         }
1940         return ret;
1941 }
1942
1943 #define in_range(b, first, len)        ((b) >= (first) && (b) < (first) + (len))
1944 static int btrfs_issue_discard(struct block_device *bdev, u64 start, u64 len,
1945                                u64 *discarded_bytes)
1946 {
1947         int j, ret = 0;
1948         u64 bytes_left, end;
1949         u64 aligned_start = ALIGN(start, 1 << 9);
1950
1951         if (WARN_ON(start != aligned_start)) {
1952                 len -= aligned_start - start;
1953                 len = round_down(len, 1 << 9);
1954                 start = aligned_start;
1955         }
1956
1957         *discarded_bytes = 0;
1958
1959         if (!len)
1960                 return 0;
1961
1962         end = start + len;
1963         bytes_left = len;
1964
1965         /* Skip any superblocks on this device. */
1966         for (j = 0; j < BTRFS_SUPER_MIRROR_MAX; j++) {
1967                 u64 sb_start = btrfs_sb_offset(j);
1968                 u64 sb_end = sb_start + BTRFS_SUPER_INFO_SIZE;
1969                 u64 size = sb_start - start;
1970
1971                 if (!in_range(sb_start, start, bytes_left) &&
1972                     !in_range(sb_end, start, bytes_left) &&
1973                     !in_range(start, sb_start, BTRFS_SUPER_INFO_SIZE))
1974                         continue;
1975
1976                 /*
1977                  * Superblock spans beginning of range.  Adjust start and
1978                  * try again.
1979                  */
1980                 if (sb_start <= start) {
1981                         start += sb_end - start;
1982                         if (start > end) {
1983                                 bytes_left = 0;
1984                                 break;
1985                         }
1986                         bytes_left = end - start;
1987                         continue;
1988                 }
1989
1990                 if (size) {
1991                         ret = blkdev_issue_discard(bdev, start >> 9, size >> 9,
1992                                                    GFP_NOFS, 0);
1993                         if (!ret)
1994                                 *discarded_bytes += size;
1995                         else if (ret != -EOPNOTSUPP)
1996                                 return ret;
1997                 }
1998
1999                 start = sb_end;
2000                 if (start > end) {
2001                         bytes_left = 0;
2002                         break;
2003                 }
2004                 bytes_left = end - start;
2005         }
2006
2007         if (bytes_left) {
2008                 ret = blkdev_issue_discard(bdev, start >> 9, bytes_left >> 9,
2009                                            GFP_NOFS, 0);
2010                 if (!ret)
2011                         *discarded_bytes += bytes_left;
2012         }
2013         return ret;
2014 }
2015
2016 int btrfs_discard_extent(struct btrfs_fs_info *fs_info, u64 bytenr,
2017                          u64 num_bytes, u64 *actual_bytes)
2018 {
2019         int ret;
2020         u64 discarded_bytes = 0;
2021         struct btrfs_bio *bbio = NULL;
2022
2023
2024         /*
2025          * Avoid races with device replace and make sure our bbio has devices
2026          * associated to its stripes that don't go away while we are discarding.
2027          */
2028         btrfs_bio_counter_inc_blocked(fs_info);
2029         /* Tell the block device(s) that the sectors can be discarded */
2030         ret = btrfs_map_block(fs_info, BTRFS_MAP_DISCARD, bytenr, &num_bytes,
2031                               &bbio, 0);
2032         /* Error condition is -ENOMEM */
2033         if (!ret) {
2034                 struct btrfs_bio_stripe *stripe = bbio->stripes;
2035                 int i;
2036
2037
2038                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
2039                         u64 bytes;
2040                         if (!stripe->dev->can_discard)
2041                                 continue;
2042
2043                         ret = btrfs_issue_discard(stripe->dev->bdev,
2044                                                   stripe->physical,
2045                                                   stripe->length,
2046                                                   &bytes);
2047                         if (!ret)
2048                                 discarded_bytes += bytes;
2049                         else if (ret != -EOPNOTSUPP)
2050                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
2051
2052                         /*
2053                          * Just in case we get back EOPNOTSUPP for some reason,
2054                          * just ignore the return value so we don't screw up
2055                          * people calling discard_extent.
2056                          */
2057                         ret = 0;
2058                 }
2059                 btrfs_put_bbio(bbio);
2060         }
2061         btrfs_bio_counter_dec(fs_info);
2062
2063         if (actual_bytes)
2064                 *actual_bytes = discarded_bytes;
2065
2066
2067         if (ret == -EOPNOTSUPP)
2068                 ret = 0;
2069         return ret;
2070 }
2071
2072 /* Can return -ENOMEM */
2073 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2074                          struct btrfs_fs_info *fs_info,
2075                          u64 bytenr, u64 num_bytes, u64 parent,
2076                          u64 root_objectid, u64 owner, u64 offset)
2077 {
2078         int ret;
2079
2080         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
2081                root_objectid == BTRFS_TREE_LOG_OBJECTID);
2082
2083         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
2084                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
2085                                         num_bytes,
2086                                         parent, root_objectid, (int)owner,
2087                                         BTRFS_ADD_DELAYED_REF, NULL);
2088         } else {
2089                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
2090                                         num_bytes, parent, root_objectid,
2091                                         owner, offset, 0,
2092                                         BTRFS_ADD_DELAYED_REF, NULL);
2093         }
2094         return ret;
2095 }
2096
2097 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2098                                   struct btrfs_fs_info *fs_info,
2099                                   struct btrfs_delayed_ref_node *node,
2100                                   u64 parent, u64 root_objectid,
2101                                   u64 owner, u64 offset, int refs_to_add,
2102                                   struct btrfs_delayed_extent_op *extent_op)
2103 {
2104         struct btrfs_path *path;
2105         struct extent_buffer *leaf;
2106         struct btrfs_extent_item *item;
2107         struct btrfs_key key;
2108         u64 bytenr = node->bytenr;
2109         u64 num_bytes = node->num_bytes;
2110         u64 refs;
2111         int ret;
2112
2113         path = btrfs_alloc_path();
2114         if (!path)
2115                 return -ENOMEM;
2116
2117         path->reada = READA_FORWARD;
2118         path->leave_spinning = 1;
2119         /* this will setup the path even if it fails to insert the back ref */
2120         ret = insert_inline_extent_backref(trans, fs_info->extent_root, path,
2121                                            bytenr, num_bytes, parent,
2122                                            root_objectid, owner, offset,
2123                                            refs_to_add, extent_op);
2124         if ((ret < 0 && ret != -EAGAIN) || !ret)
2125                 goto out;
2126
2127         /*
2128          * Ok we had -EAGAIN which means we didn't have space to insert and
2129          * inline extent ref, so just update the reference count and add a
2130          * normal backref.
2131          */
2132         leaf = path->nodes[0];
2133         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2134         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2135         refs = btrfs_extent_refs(leaf, item);
2136         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
2137         if (extent_op)
2138                 __run_delayed_extent_op(extent_op, leaf, item);
2139
2140         btrfs_mark_buffer_dirty(leaf);
2141         btrfs_release_path(path);
2142
2143         path->reada = READA_FORWARD;
2144         path->leave_spinning = 1;
2145         /* now insert the actual backref */
2146         ret = insert_extent_backref(trans, fs_info->extent_root,
2147                                     path, bytenr, parent, root_objectid,
2148                                     owner, offset, refs_to_add);
2149         if (ret)
2150                 btrfs_abort_transaction(trans, ret);
2151 out:
2152         btrfs_free_path(path);
2153         return ret;
2154 }
2155
2156 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
2157                                 struct btrfs_fs_info *fs_info,
2158                                 struct btrfs_delayed_ref_node *node,
2159                                 struct btrfs_delayed_extent_op *extent_op,
2160                                 int insert_reserved)
2161 {
2162         int ret = 0;
2163         struct btrfs_delayed_data_ref *ref;
2164         struct btrfs_key ins;
2165         u64 parent = 0;
2166         u64 ref_root = 0;
2167         u64 flags = 0;
2168
2169         ins.objectid = node->bytenr;
2170         ins.offset = node->num_bytes;
2171         ins.type = BTRFS_EXTENT_ITEM_KEY;
2172
2173         ref = btrfs_delayed_node_to_data_ref(node);
2174         trace_run_delayed_data_ref(fs_info, node, ref, node->action);
2175
2176         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2177                 parent = ref->parent;
2178         ref_root = ref->root;
2179
2180         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2181                 if (extent_op)
2182                         flags |= extent_op->flags_to_set;
2183                 ret = alloc_reserved_file_extent(trans, fs_info,
2184                                                  parent, ref_root, flags,
2185                                                  ref->objectid, ref->offset,
2186                                                  &ins, node->ref_mod);
2187         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2188                 ret = __btrfs_inc_extent_ref(trans, fs_info, node, parent,
2189                                              ref_root, ref->objectid,
2190                                              ref->offset, node->ref_mod,
2191                                              extent_op);
2192         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2193                 ret = __btrfs_free_extent(trans, fs_info, node, parent,
2194                                           ref_root, ref->objectid,
2195                                           ref->offset, node->ref_mod,
2196                                           extent_op);
2197         } else {
2198                 BUG();
2199         }
2200         return ret;
2201 }
2202
2203 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2204                                     struct extent_buffer *leaf,
2205                                     struct btrfs_extent_item *ei)
2206 {
2207         u64 flags = btrfs_extent_flags(leaf, ei);
2208         if (extent_op->update_flags) {
2209                 flags |= extent_op->flags_to_set;
2210                 btrfs_set_extent_flags(leaf, ei, flags);
2211         }
2212
2213         if (extent_op->update_key) {
2214                 struct btrfs_tree_block_info *bi;
2215                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2216                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2217                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2218         }
2219 }
2220
2221 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2222                                  struct btrfs_fs_info *fs_info,
2223                                  struct btrfs_delayed_ref_node *node,
2224                                  struct btrfs_delayed_extent_op *extent_op)
2225 {
2226         struct btrfs_key key;
2227         struct btrfs_path *path;
2228         struct btrfs_extent_item *ei;
2229         struct extent_buffer *leaf;
2230         u32 item_size;
2231         int ret;
2232         int err = 0;
2233         int metadata = !extent_op->is_data;
2234
2235         if (trans->aborted)
2236                 return 0;
2237
2238         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2239                 metadata = 0;
2240
2241         path = btrfs_alloc_path();
2242         if (!path)
2243                 return -ENOMEM;
2244
2245         key.objectid = node->bytenr;
2246
2247         if (metadata) {
2248                 key.type = BTRFS_METADATA_ITEM_KEY;
2249                 key.offset = extent_op->level;
2250         } else {
2251                 key.type = BTRFS_EXTENT_ITEM_KEY;
2252                 key.offset = node->num_bytes;
2253         }
2254
2255 again:
2256         path->reada = READA_FORWARD;
2257         path->leave_spinning = 1;
2258         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 1);
2259         if (ret < 0) {
2260                 err = ret;
2261                 goto out;
2262         }
2263         if (ret > 0) {
2264                 if (metadata) {
2265                         if (path->slots[0] > 0) {
2266                                 path->slots[0]--;
2267                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
2268                                                       path->slots[0]);
2269                                 if (key.objectid == node->bytenr &&
2270                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
2271                                     key.offset == node->num_bytes)
2272                                         ret = 0;
2273                         }
2274                         if (ret > 0) {
2275                                 btrfs_release_path(path);
2276                                 metadata = 0;
2277
2278                                 key.objectid = node->bytenr;
2279                                 key.offset = node->num_bytes;
2280                                 key.type = BTRFS_EXTENT_ITEM_KEY;
2281                                 goto again;
2282                         }
2283                 } else {
2284                         err = -EIO;
2285                         goto out;
2286                 }
2287         }
2288
2289         leaf = path->nodes[0];
2290         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2291 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2292         if (item_size < sizeof(*ei)) {
2293                 ret = convert_extent_item_v0(trans, fs_info->extent_root,
2294                                              path, (u64)-1, 0);
2295                 if (ret < 0) {
2296                         err = ret;
2297                         goto out;
2298                 }
2299                 leaf = path->nodes[0];
2300                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2301         }
2302 #endif
2303         BUG_ON(item_size < sizeof(*ei));
2304         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2305         __run_delayed_extent_op(extent_op, leaf, ei);
2306
2307         btrfs_mark_buffer_dirty(leaf);
2308 out:
2309         btrfs_free_path(path);
2310         return err;
2311 }
2312
2313 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2314                                 struct btrfs_fs_info *fs_info,
2315                                 struct btrfs_delayed_ref_node *node,
2316                                 struct btrfs_delayed_extent_op *extent_op,
2317                                 int insert_reserved)
2318 {
2319         int ret = 0;
2320         struct btrfs_delayed_tree_ref *ref;
2321         struct btrfs_key ins;
2322         u64 parent = 0;
2323         u64 ref_root = 0;
2324         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
2325
2326         ref = btrfs_delayed_node_to_tree_ref(node);
2327         trace_run_delayed_tree_ref(fs_info, node, ref, node->action);
2328
2329         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2330                 parent = ref->parent;
2331         ref_root = ref->root;
2332
2333         ins.objectid = node->bytenr;
2334         if (skinny_metadata) {
2335                 ins.offset = ref->level;
2336                 ins.type = BTRFS_METADATA_ITEM_KEY;
2337         } else {
2338                 ins.offset = node->num_bytes;
2339                 ins.type = BTRFS_EXTENT_ITEM_KEY;
2340         }
2341
2342         if (node->ref_mod != 1) {
2343                 btrfs_err(fs_info,
2344         "btree block(%llu) has %d references rather than 1: action %d ref_root %llu parent %llu",
2345                           node->bytenr, node->ref_mod, node->action, ref_root,
2346                           parent);
2347                 return -EIO;
2348         }
2349         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2350                 BUG_ON(!extent_op || !extent_op->update_flags);
2351                 ret = alloc_reserved_tree_block(trans, fs_info,
2352                                                 parent, ref_root,
2353                                                 extent_op->flags_to_set,
2354                                                 &extent_op->key,
2355                                                 ref->level, &ins);
2356         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2357                 ret = __btrfs_inc_extent_ref(trans, fs_info, node,
2358                                              parent, ref_root,
2359                                              ref->level, 0, 1,
2360                                              extent_op);
2361         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2362                 ret = __btrfs_free_extent(trans, fs_info, node,
2363                                           parent, ref_root,
2364                                           ref->level, 0, 1, extent_op);
2365         } else {
2366                 BUG();
2367         }
2368         return ret;
2369 }
2370
2371 /* helper function to actually process a single delayed ref entry */
2372 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2373                                struct btrfs_fs_info *fs_info,
2374                                struct btrfs_delayed_ref_node *node,
2375                                struct btrfs_delayed_extent_op *extent_op,
2376                                int insert_reserved)
2377 {
2378         int ret = 0;
2379
2380         if (trans->aborted) {
2381                 if (insert_reserved)
2382                         btrfs_pin_extent(fs_info, node->bytenr,
2383                                          node->num_bytes, 1);
2384                 return 0;
2385         }
2386
2387         if (btrfs_delayed_ref_is_head(node)) {
2388                 struct btrfs_delayed_ref_head *head;
2389                 /*
2390                  * we've hit the end of the chain and we were supposed
2391                  * to insert this extent into the tree.  But, it got
2392                  * deleted before we ever needed to insert it, so all
2393                  * we have to do is clean up the accounting
2394                  */
2395                 BUG_ON(extent_op);
2396                 head = btrfs_delayed_node_to_head(node);
2397                 trace_run_delayed_ref_head(fs_info, node, head, node->action);
2398
2399                 if (insert_reserved) {
2400                         btrfs_pin_extent(fs_info, node->bytenr,
2401                                          node->num_bytes, 1);
2402                         if (head->is_data) {
2403                                 ret = btrfs_del_csums(trans, fs_info,
2404                                                       node->bytenr,
2405                                                       node->num_bytes);
2406                         }
2407                 }
2408
2409                 /* Also free its reserved qgroup space */
2410                 btrfs_qgroup_free_delayed_ref(fs_info, head->qgroup_ref_root,
2411                                               head->qgroup_reserved);
2412                 return ret;
2413         }
2414
2415         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2416             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2417                 ret = run_delayed_tree_ref(trans, fs_info, node, extent_op,
2418                                            insert_reserved);
2419         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2420                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2421                 ret = run_delayed_data_ref(trans, fs_info, node, extent_op,
2422                                            insert_reserved);
2423         else
2424                 BUG();
2425         return ret;
2426 }
2427
2428 static inline struct btrfs_delayed_ref_node *
2429 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2430 {
2431         struct btrfs_delayed_ref_node *ref;
2432
2433         if (list_empty(&head->ref_list))
2434                 return NULL;
2435
2436         /*
2437          * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
2438          * This is to prevent a ref count from going down to zero, which deletes
2439          * the extent item from the extent tree, when there still are references
2440          * to add, which would fail because they would not find the extent item.
2441          */
2442         if (!list_empty(&head->ref_add_list))
2443                 return list_first_entry(&head->ref_add_list,
2444                                 struct btrfs_delayed_ref_node, add_list);
2445
2446         ref = list_first_entry(&head->ref_list, struct btrfs_delayed_ref_node,
2447                                list);
2448         ASSERT(list_empty(&ref->add_list));
2449         return ref;
2450 }
2451
2452 /*
2453  * Returns 0 on success or if called with an already aborted transaction.
2454  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2455  */
2456 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2457                                              struct btrfs_fs_info *fs_info,
2458                                              unsigned long nr)
2459 {
2460         struct btrfs_delayed_ref_root *delayed_refs;
2461         struct btrfs_delayed_ref_node *ref;
2462         struct btrfs_delayed_ref_head *locked_ref = NULL;
2463         struct btrfs_delayed_extent_op *extent_op;
2464         ktime_t start = ktime_get();
2465         int ret;
2466         unsigned long count = 0;
2467         unsigned long actual_count = 0;
2468         int must_insert_reserved = 0;
2469
2470         delayed_refs = &trans->transaction->delayed_refs;
2471         while (1) {
2472                 if (!locked_ref) {
2473                         if (count >= nr)
2474                                 break;
2475
2476                         spin_lock(&delayed_refs->lock);
2477                         locked_ref = btrfs_select_ref_head(trans);
2478                         if (!locked_ref) {
2479                                 spin_unlock(&delayed_refs->lock);
2480                                 break;
2481                         }
2482
2483                         /* grab the lock that says we are going to process
2484                          * all the refs for this head */
2485                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2486                         spin_unlock(&delayed_refs->lock);
2487                         /*
2488                          * we may have dropped the spin lock to get the head
2489                          * mutex lock, and that might have given someone else
2490                          * time to free the head.  If that's true, it has been
2491                          * removed from our list and we can move on.
2492                          */
2493                         if (ret == -EAGAIN) {
2494                                 locked_ref = NULL;
2495                                 count++;
2496                                 continue;
2497                         }
2498                 }
2499
2500                 /*
2501                  * We need to try and merge add/drops of the same ref since we
2502                  * can run into issues with relocate dropping the implicit ref
2503                  * and then it being added back again before the drop can
2504                  * finish.  If we merged anything we need to re-loop so we can
2505                  * get a good ref.
2506                  * Or we can get node references of the same type that weren't
2507                  * merged when created due to bumps in the tree mod seq, and
2508                  * we need to merge them to prevent adding an inline extent
2509                  * backref before dropping it (triggering a BUG_ON at
2510                  * insert_inline_extent_backref()).
2511                  */
2512                 spin_lock(&locked_ref->lock);
2513                 btrfs_merge_delayed_refs(trans, fs_info, delayed_refs,
2514                                          locked_ref);
2515
2516                 /*
2517                  * locked_ref is the head node, so we have to go one
2518                  * node back for any delayed ref updates
2519                  */
2520                 ref = select_delayed_ref(locked_ref);
2521
2522                 if (ref && ref->seq &&
2523                     btrfs_check_delayed_seq(fs_info, delayed_refs, ref->seq)) {
2524                         spin_unlock(&locked_ref->lock);
2525                         btrfs_delayed_ref_unlock(locked_ref);
2526                         spin_lock(&delayed_refs->lock);
2527                         locked_ref->processing = 0;
2528                         delayed_refs->num_heads_ready++;
2529                         spin_unlock(&delayed_refs->lock);
2530                         locked_ref = NULL;
2531                         cond_resched();
2532                         count++;
2533                         continue;
2534                 }
2535
2536                 /*
2537                  * record the must insert reserved flag before we
2538                  * drop the spin lock.
2539                  */
2540                 must_insert_reserved = locked_ref->must_insert_reserved;
2541                 locked_ref->must_insert_reserved = 0;
2542
2543                 extent_op = locked_ref->extent_op;
2544                 locked_ref->extent_op = NULL;
2545
2546                 if (!ref) {
2547
2548
2549                         /* All delayed refs have been processed, Go ahead
2550                          * and send the head node to run_one_delayed_ref,
2551                          * so that any accounting fixes can happen
2552                          */
2553                         ref = &locked_ref->node;
2554
2555                         if (extent_op && must_insert_reserved) {
2556                                 btrfs_free_delayed_extent_op(extent_op);
2557                                 extent_op = NULL;
2558                         }
2559
2560                         if (extent_op) {
2561                                 spin_unlock(&locked_ref->lock);
2562                                 ret = run_delayed_extent_op(trans, fs_info,
2563                                                             ref, extent_op);
2564                                 btrfs_free_delayed_extent_op(extent_op);
2565
2566                                 if (ret) {
2567                                         /*
2568                                          * Need to reset must_insert_reserved if
2569                                          * there was an error so the abort stuff
2570                                          * can cleanup the reserved space
2571                                          * properly.
2572                                          */
2573                                         if (must_insert_reserved)
2574                                                 locked_ref->must_insert_reserved = 1;
2575                                         locked_ref->processing = 0;
2576                                         btrfs_debug(fs_info,
2577                                                     "run_delayed_extent_op returned %d",
2578                                                     ret);
2579                                         btrfs_delayed_ref_unlock(locked_ref);
2580                                         return ret;
2581                                 }
2582                                 continue;
2583                         }
2584
2585                         /*
2586                          * Need to drop our head ref lock and re-acquire the
2587                          * delayed ref lock and then re-check to make sure
2588                          * nobody got added.
2589                          */
2590                         spin_unlock(&locked_ref->lock);
2591                         spin_lock(&delayed_refs->lock);
2592                         spin_lock(&locked_ref->lock);
2593                         if (!list_empty(&locked_ref->ref_list) ||
2594                             locked_ref->extent_op) {
2595                                 spin_unlock(&locked_ref->lock);
2596                                 spin_unlock(&delayed_refs->lock);
2597                                 continue;
2598                         }
2599                         ref->in_tree = 0;
2600                         delayed_refs->num_heads--;
2601                         rb_erase(&locked_ref->href_node,
2602                                  &delayed_refs->href_root);
2603                         spin_unlock(&delayed_refs->lock);
2604                 } else {
2605                         actual_count++;
2606                         ref->in_tree = 0;
2607                         list_del(&ref->list);
2608                         if (!list_empty(&ref->add_list))
2609                                 list_del(&ref->add_list);
2610                 }
2611                 atomic_dec(&delayed_refs->num_entries);
2612
2613                 if (!btrfs_delayed_ref_is_head(ref)) {
2614                         /*
2615                          * when we play the delayed ref, also correct the
2616                          * ref_mod on head
2617                          */
2618                         switch (ref->action) {
2619                         case BTRFS_ADD_DELAYED_REF:
2620                         case BTRFS_ADD_DELAYED_EXTENT:
2621                                 locked_ref->node.ref_mod -= ref->ref_mod;
2622                                 break;
2623                         case BTRFS_DROP_DELAYED_REF:
2624                                 locked_ref->node.ref_mod += ref->ref_mod;
2625                                 break;
2626                         default:
2627                                 WARN_ON(1);
2628                         }
2629                 }
2630                 spin_unlock(&locked_ref->lock);
2631
2632                 ret = run_one_delayed_ref(trans, fs_info, ref, extent_op,
2633                                           must_insert_reserved);
2634
2635                 btrfs_free_delayed_extent_op(extent_op);
2636                 if (ret) {
2637                         spin_lock(&delayed_refs->lock);
2638                         locked_ref->processing = 0;
2639                         delayed_refs->num_heads_ready++;
2640                         spin_unlock(&delayed_refs->lock);
2641                         btrfs_delayed_ref_unlock(locked_ref);
2642                         btrfs_put_delayed_ref(ref);
2643                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d",
2644                                     ret);
2645                         return ret;
2646                 }
2647
2648                 /*
2649                  * If this node is a head, that means all the refs in this head
2650                  * have been dealt with, and we will pick the next head to deal
2651                  * with, so we must unlock the head and drop it from the cluster
2652                  * list before we release it.
2653                  */
2654                 if (btrfs_delayed_ref_is_head(ref)) {
2655                         if (locked_ref->is_data &&
2656                             locked_ref->total_ref_mod < 0) {
2657                                 spin_lock(&delayed_refs->lock);
2658                                 delayed_refs->pending_csums -= ref->num_bytes;
2659                                 spin_unlock(&delayed_refs->lock);
2660                         }
2661                         btrfs_delayed_ref_unlock(locked_ref);
2662                         locked_ref = NULL;
2663                 }
2664                 btrfs_put_delayed_ref(ref);
2665                 count++;
2666                 cond_resched();
2667         }
2668
2669         /*
2670          * We don't want to include ref heads since we can have empty ref heads
2671          * and those will drastically skew our runtime down since we just do
2672          * accounting, no actual extent tree updates.
2673          */
2674         if (actual_count > 0) {
2675                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2676                 u64 avg;
2677
2678                 /*
2679                  * We weigh the current average higher than our current runtime
2680                  * to avoid large swings in the average.
2681                  */
2682                 spin_lock(&delayed_refs->lock);
2683                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2684                 fs_info->avg_delayed_ref_runtime = avg >> 2;    /* div by 4 */
2685                 spin_unlock(&delayed_refs->lock);
2686         }
2687         return 0;
2688 }
2689
2690 #ifdef SCRAMBLE_DELAYED_REFS
2691 /*
2692  * Normally delayed refs get processed in ascending bytenr order. This
2693  * correlates in most cases to the order added. To expose dependencies on this
2694  * order, we start to process the tree in the middle instead of the beginning
2695  */
2696 static u64 find_middle(struct rb_root *root)
2697 {
2698         struct rb_node *n = root->rb_node;
2699         struct btrfs_delayed_ref_node *entry;
2700         int alt = 1;
2701         u64 middle;
2702         u64 first = 0, last = 0;
2703
2704         n = rb_first(root);
2705         if (n) {
2706                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2707                 first = entry->bytenr;
2708         }
2709         n = rb_last(root);
2710         if (n) {
2711                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2712                 last = entry->bytenr;
2713         }
2714         n = root->rb_node;
2715
2716         while (n) {
2717                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2718                 WARN_ON(!entry->in_tree);
2719
2720                 middle = entry->bytenr;
2721
2722                 if (alt)
2723                         n = n->rb_left;
2724                 else
2725                         n = n->rb_right;
2726
2727                 alt = 1 - alt;
2728         }
2729         return middle;
2730 }
2731 #endif
2732
2733 static inline u64 heads_to_leaves(struct btrfs_fs_info *fs_info, u64 heads)
2734 {
2735         u64 num_bytes;
2736
2737         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2738                              sizeof(struct btrfs_extent_inline_ref));
2739         if (!btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2740                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2741
2742         /*
2743          * We don't ever fill up leaves all the way so multiply by 2 just to be
2744          * closer to what we're really going to want to use.
2745          */
2746         return div_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(fs_info));
2747 }
2748
2749 /*
2750  * Takes the number of bytes to be csumm'ed and figures out how many leaves it
2751  * would require to store the csums for that many bytes.
2752  */
2753 u64 btrfs_csum_bytes_to_leaves(struct btrfs_fs_info *fs_info, u64 csum_bytes)
2754 {
2755         u64 csum_size;
2756         u64 num_csums_per_leaf;
2757         u64 num_csums;
2758
2759         csum_size = BTRFS_MAX_ITEM_SIZE(fs_info);
2760         num_csums_per_leaf = div64_u64(csum_size,
2761                         (u64)btrfs_super_csum_size(fs_info->super_copy));
2762         num_csums = div64_u64(csum_bytes, fs_info->sectorsize);
2763         num_csums += num_csums_per_leaf - 1;
2764         num_csums = div64_u64(num_csums, num_csums_per_leaf);
2765         return num_csums;
2766 }
2767
2768 int btrfs_check_space_for_delayed_refs(struct btrfs_trans_handle *trans,
2769                                        struct btrfs_fs_info *fs_info)
2770 {
2771         struct btrfs_block_rsv *global_rsv;
2772         u64 num_heads = trans->transaction->delayed_refs.num_heads_ready;
2773         u64 csum_bytes = trans->transaction->delayed_refs.pending_csums;
2774         u64 num_dirty_bgs = trans->transaction->num_dirty_bgs;
2775         u64 num_bytes, num_dirty_bgs_bytes;
2776         int ret = 0;
2777
2778         num_bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
2779         num_heads = heads_to_leaves(fs_info, num_heads);
2780         if (num_heads > 1)
2781                 num_bytes += (num_heads - 1) * fs_info->nodesize;
2782         num_bytes <<= 1;
2783         num_bytes += btrfs_csum_bytes_to_leaves(fs_info, csum_bytes) *
2784                                                         fs_info->nodesize;
2785         num_dirty_bgs_bytes = btrfs_calc_trans_metadata_size(fs_info,
2786                                                              num_dirty_bgs);
2787         global_rsv = &fs_info->global_block_rsv;
2788
2789         /*
2790          * If we can't allocate any more chunks lets make sure we have _lots_ of
2791          * wiggle room since running delayed refs can create more delayed refs.
2792          */
2793         if (global_rsv->space_info->full) {
2794                 num_dirty_bgs_bytes <<= 1;
2795                 num_bytes <<= 1;
2796         }
2797
2798         spin_lock(&global_rsv->lock);
2799         if (global_rsv->reserved <= num_bytes + num_dirty_bgs_bytes)
2800                 ret = 1;
2801         spin_unlock(&global_rsv->lock);
2802         return ret;
2803 }
2804
2805 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans,
2806                                        struct btrfs_fs_info *fs_info)
2807 {
2808         u64 num_entries =
2809                 atomic_read(&trans->transaction->delayed_refs.num_entries);
2810         u64 avg_runtime;
2811         u64 val;
2812
2813         smp_mb();
2814         avg_runtime = fs_info->avg_delayed_ref_runtime;
2815         val = num_entries * avg_runtime;
2816         if (val >= NSEC_PER_SEC)
2817                 return 1;
2818         if (val >= NSEC_PER_SEC / 2)
2819                 return 2;
2820
2821         return btrfs_check_space_for_delayed_refs(trans, fs_info);
2822 }
2823
2824 struct async_delayed_refs {
2825         struct btrfs_root *root;
2826         u64 transid;
2827         int count;
2828         int error;
2829         int sync;
2830         struct completion wait;
2831         struct btrfs_work work;
2832 };
2833
2834 static inline struct async_delayed_refs *
2835 to_async_delayed_refs(struct btrfs_work *work)
2836 {
2837         return container_of(work, struct async_delayed_refs, work);
2838 }
2839
2840 static void delayed_ref_async_start(struct btrfs_work *work)
2841 {
2842         struct async_delayed_refs *async = to_async_delayed_refs(work);
2843         struct btrfs_trans_handle *trans;
2844         struct btrfs_fs_info *fs_info = async->root->fs_info;
2845         int ret;
2846
2847         /* if the commit is already started, we don't need to wait here */
2848         if (btrfs_transaction_blocked(fs_info))
2849                 goto done;
2850
2851         trans = btrfs_join_transaction(async->root);
2852         if (IS_ERR(trans)) {
2853                 async->error = PTR_ERR(trans);
2854                 goto done;
2855         }
2856
2857         /*
2858          * trans->sync means that when we call end_transaction, we won't
2859          * wait on delayed refs
2860          */
2861         trans->sync = true;
2862
2863         /* Don't bother flushing if we got into a different transaction */
2864         if (trans->transid > async->transid)
2865                 goto end;
2866
2867         ret = btrfs_run_delayed_refs(trans, fs_info, async->count);
2868         if (ret)
2869                 async->error = ret;
2870 end:
2871         ret = btrfs_end_transaction(trans);
2872         if (ret && !async->error)
2873                 async->error = ret;
2874 done:
2875         if (async->sync)
2876                 complete(&async->wait);
2877         else
2878                 kfree(async);
2879 }
2880
2881 int btrfs_async_run_delayed_refs(struct btrfs_fs_info *fs_info,
2882                                  unsigned long count, u64 transid, int wait)
2883 {
2884         struct async_delayed_refs *async;
2885         int ret;
2886
2887         async = kmalloc(sizeof(*async), GFP_NOFS);
2888         if (!async)
2889                 return -ENOMEM;
2890
2891         async->root = fs_info->tree_root;
2892         async->count = count;
2893         async->error = 0;
2894         async->transid = transid;
2895         if (wait)
2896                 async->sync = 1;
2897         else
2898                 async->sync = 0;
2899         init_completion(&async->wait);
2900
2901         btrfs_init_work(&async->work, btrfs_extent_refs_helper,
2902                         delayed_ref_async_start, NULL, NULL);
2903
2904         btrfs_queue_work(fs_info->extent_workers, &async->work);
2905
2906         if (wait) {
2907                 wait_for_completion(&async->wait);
2908                 ret = async->error;
2909                 kfree(async);
2910                 return ret;
2911         }
2912         return 0;
2913 }
2914
2915 /*
2916  * this starts processing the delayed reference count updates and
2917  * extent insertions we have queued up so far.  count can be
2918  * 0, which means to process everything in the tree at the start
2919  * of the run (but not newly added entries), or it can be some target
2920  * number you'd like to process.
2921  *
2922  * Returns 0 on success or if called with an aborted transaction
2923  * Returns <0 on error and aborts the transaction
2924  */
2925 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2926                            struct btrfs_fs_info *fs_info, unsigned long count)
2927 {
2928         struct rb_node *node;
2929         struct btrfs_delayed_ref_root *delayed_refs;
2930         struct btrfs_delayed_ref_head *head;
2931         int ret;
2932         int run_all = count == (unsigned long)-1;
2933         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
2934
2935         /* We'll clean this up in btrfs_cleanup_transaction */
2936         if (trans->aborted)
2937                 return 0;
2938
2939         if (test_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags))
2940                 return 0;
2941
2942         delayed_refs = &trans->transaction->delayed_refs;
2943         if (count == 0)
2944                 count = atomic_read(&delayed_refs->num_entries) * 2;
2945
2946 again:
2947 #ifdef SCRAMBLE_DELAYED_REFS
2948         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2949 #endif
2950         trans->can_flush_pending_bgs = false;
2951         ret = __btrfs_run_delayed_refs(trans, fs_info, count);
2952         if (ret < 0) {
2953                 btrfs_abort_transaction(trans, ret);
2954                 return ret;
2955         }
2956
2957         if (run_all) {
2958                 if (!list_empty(&trans->new_bgs))
2959                         btrfs_create_pending_block_groups(trans, fs_info);
2960
2961                 spin_lock(&delayed_refs->lock);
2962                 node = rb_first(&delayed_refs->href_root);
2963                 if (!node) {
2964                         spin_unlock(&delayed_refs->lock);
2965                         goto out;
2966                 }
2967
2968                 while (node) {
2969                         head = rb_entry(node, struct btrfs_delayed_ref_head,
2970                                         href_node);
2971                         if (btrfs_delayed_ref_is_head(&head->node)) {
2972                                 struct btrfs_delayed_ref_node *ref;
2973
2974                                 ref = &head->node;
2975                                 atomic_inc(&ref->refs);
2976
2977                                 spin_unlock(&delayed_refs->lock);
2978                                 /*
2979                                  * Mutex was contended, block until it's
2980                                  * released and try again
2981                                  */
2982                                 mutex_lock(&head->mutex);
2983                                 mutex_unlock(&head->mutex);
2984
2985                                 btrfs_put_delayed_ref(ref);
2986                                 cond_resched();
2987                                 goto again;
2988                         } else {
2989                                 WARN_ON(1);
2990                         }
2991                         node = rb_next(node);
2992                 }
2993                 spin_unlock(&delayed_refs->lock);
2994                 cond_resched();
2995                 goto again;
2996         }
2997 out:
2998         assert_qgroups_uptodate(trans);
2999         trans->can_flush_pending_bgs = can_flush_pending_bgs;
3000         return 0;
3001 }
3002
3003 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
3004                                 struct btrfs_fs_info *fs_info,
3005                                 u64 bytenr, u64 num_bytes, u64 flags,
3006                                 int level, int is_data)
3007 {
3008         struct btrfs_delayed_extent_op *extent_op;
3009         int ret;
3010
3011         extent_op = btrfs_alloc_delayed_extent_op();
3012         if (!extent_op)
3013                 return -ENOMEM;
3014
3015         extent_op->flags_to_set = flags;
3016         extent_op->update_flags = true;
3017         extent_op->update_key = false;
3018         extent_op->is_data = is_data ? true : false;
3019         extent_op->level = level;
3020
3021         ret = btrfs_add_delayed_extent_op(fs_info, trans, bytenr,
3022                                           num_bytes, extent_op);
3023         if (ret)
3024                 btrfs_free_delayed_extent_op(extent_op);
3025         return ret;
3026 }
3027
3028 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
3029                                       struct btrfs_root *root,
3030                                       struct btrfs_path *path,
3031                                       u64 objectid, u64 offset, u64 bytenr)
3032 {
3033         struct btrfs_delayed_ref_head *head;
3034         struct btrfs_delayed_ref_node *ref;
3035         struct btrfs_delayed_data_ref *data_ref;
3036         struct btrfs_delayed_ref_root *delayed_refs;
3037         int ret = 0;
3038
3039         delayed_refs = &trans->transaction->delayed_refs;
3040         spin_lock(&delayed_refs->lock);
3041         head = btrfs_find_delayed_ref_head(trans, bytenr);
3042         if (!head) {
3043                 spin_unlock(&delayed_refs->lock);
3044                 return 0;
3045         }
3046
3047         if (!mutex_trylock(&head->mutex)) {
3048                 atomic_inc(&head->node.refs);
3049                 spin_unlock(&delayed_refs->lock);
3050
3051                 btrfs_release_path(path);
3052
3053                 /*
3054                  * Mutex was contended, block until it's released and let
3055                  * caller try again
3056                  */
3057                 mutex_lock(&head->mutex);
3058                 mutex_unlock(&head->mutex);
3059                 btrfs_put_delayed_ref(&head->node);
3060                 return -EAGAIN;
3061         }
3062         spin_unlock(&delayed_refs->lock);
3063
3064         spin_lock(&head->lock);
3065         list_for_each_entry(ref, &head->ref_list, list) {
3066                 /* If it's a shared ref we know a cross reference exists */
3067                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
3068                         ret = 1;
3069                         break;
3070                 }
3071
3072                 data_ref = btrfs_delayed_node_to_data_ref(ref);
3073
3074                 /*
3075                  * If our ref doesn't match the one we're currently looking at
3076                  * then we have a cross reference.
3077                  */
3078                 if (data_ref->root != root->root_key.objectid ||
3079                     data_ref->objectid != objectid ||
3080                     data_ref->offset != offset) {
3081                         ret = 1;
3082                         break;
3083                 }
3084         }
3085         spin_unlock(&head->lock);
3086         mutex_unlock(&head->mutex);
3087         return ret;
3088 }
3089
3090 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
3091                                         struct btrfs_root *root,
3092                                         struct btrfs_path *path,
3093                                         u64 objectid, u64 offset, u64 bytenr)
3094 {
3095         struct btrfs_fs_info *fs_info = root->fs_info;
3096         struct btrfs_root *extent_root = fs_info->extent_root;
3097         struct extent_buffer *leaf;
3098         struct btrfs_extent_data_ref *ref;
3099         struct btrfs_extent_inline_ref *iref;
3100         struct btrfs_extent_item *ei;
3101         struct btrfs_key key;
3102         u32 item_size;
3103         int ret;
3104
3105         key.objectid = bytenr;
3106         key.offset = (u64)-1;
3107         key.type = BTRFS_EXTENT_ITEM_KEY;
3108
3109         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
3110         if (ret < 0)
3111                 goto out;
3112         BUG_ON(ret == 0); /* Corruption */
3113
3114         ret = -ENOENT;
3115         if (path->slots[0] == 0)
3116                 goto out;
3117
3118         path->slots[0]--;
3119         leaf = path->nodes[0];
3120         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3121
3122         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
3123                 goto out;
3124
3125         ret = 1;
3126         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
3127 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3128         if (item_size < sizeof(*ei)) {
3129                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3130                 goto out;
3131         }
3132 #endif
3133         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
3134
3135         if (item_size != sizeof(*ei) +
3136             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
3137                 goto out;
3138
3139         if (btrfs_extent_generation(leaf, ei) <=
3140             btrfs_root_last_snapshot(&root->root_item))
3141                 goto out;
3142
3143         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
3144         if (btrfs_extent_inline_ref_type(leaf, iref) !=
3145             BTRFS_EXTENT_DATA_REF_KEY)
3146                 goto out;
3147
3148         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
3149         if (btrfs_extent_refs(leaf, ei) !=
3150             btrfs_extent_data_ref_count(leaf, ref) ||
3151             btrfs_extent_data_ref_root(leaf, ref) !=
3152             root->root_key.objectid ||
3153             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
3154             btrfs_extent_data_ref_offset(leaf, ref) != offset)
3155                 goto out;
3156
3157         ret = 0;
3158 out:
3159         return ret;
3160 }
3161
3162 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
3163                           struct btrfs_root *root,
3164                           u64 objectid, u64 offset, u64 bytenr)
3165 {
3166         struct btrfs_path *path;
3167         int ret;
3168         int ret2;
3169
3170         path = btrfs_alloc_path();
3171         if (!path)
3172                 return -ENOENT;
3173
3174         do {
3175                 ret = check_committed_ref(trans, root, path, objectid,
3176                                           offset, bytenr);
3177                 if (ret && ret != -ENOENT)
3178                         goto out;
3179
3180                 ret2 = check_delayed_ref(trans, root, path, objectid,
3181                                          offset, bytenr);
3182         } while (ret2 == -EAGAIN);
3183
3184         if (ret2 && ret2 != -ENOENT) {
3185                 ret = ret2;
3186                 goto out;
3187         }
3188
3189         if (ret != -ENOENT || ret2 != -ENOENT)
3190                 ret = 0;
3191 out:
3192         btrfs_free_path(path);
3193         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
3194                 WARN_ON(ret > 0);
3195         return ret;
3196 }
3197
3198 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
3199                            struct btrfs_root *root,
3200                            struct extent_buffer *buf,
3201                            int full_backref, int inc)
3202 {
3203         struct btrfs_fs_info *fs_info = root->fs_info;
3204         u64 bytenr;
3205         u64 num_bytes;
3206         u64 parent;
3207         u64 ref_root;
3208         u32 nritems;
3209         struct btrfs_key key;
3210         struct btrfs_file_extent_item *fi;
3211         int i;
3212         int level;
3213         int ret = 0;
3214         int (*process_func)(struct btrfs_trans_handle *,
3215                             struct btrfs_fs_info *,
3216                             u64, u64, u64, u64, u64, u64);
3217
3218
3219         if (btrfs_is_testing(fs_info))
3220                 return 0;
3221
3222         ref_root = btrfs_header_owner(buf);
3223         nritems = btrfs_header_nritems(buf);
3224         level = btrfs_header_level(buf);
3225
3226         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state) && level == 0)
3227                 return 0;
3228
3229         if (inc)
3230                 process_func = btrfs_inc_extent_ref;
3231         else
3232                 process_func = btrfs_free_extent;
3233
3234         if (full_backref)
3235                 parent = buf->start;
3236         else
3237                 parent = 0;
3238
3239         for (i = 0; i < nritems; i++) {
3240                 if (level == 0) {
3241                         btrfs_item_key_to_cpu(buf, &key, i);
3242                         if (key.type != BTRFS_EXTENT_DATA_KEY)
3243                                 continue;
3244                         fi = btrfs_item_ptr(buf, i,
3245                                             struct btrfs_file_extent_item);
3246                         if (btrfs_file_extent_type(buf, fi) ==
3247                             BTRFS_FILE_EXTENT_INLINE)
3248                                 continue;
3249                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3250                         if (bytenr == 0)
3251                                 continue;
3252
3253                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3254                         key.offset -= btrfs_file_extent_offset(buf, fi);
3255                         ret = process_func(trans, fs_info, bytenr, num_bytes,
3256                                            parent, ref_root, key.objectid,
3257                                            key.offset);
3258                         if (ret)
3259                                 goto fail;
3260                 } else {
3261                         bytenr = btrfs_node_blockptr(buf, i);
3262                         num_bytes = fs_info->nodesize;
3263                         ret = process_func(trans, fs_info, bytenr, num_bytes,
3264                                            parent, ref_root, level - 1, 0);
3265                         if (ret)
3266                                 goto fail;
3267                 }
3268         }
3269         return 0;
3270 fail:
3271         return ret;
3272 }
3273
3274 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3275                   struct extent_buffer *buf, int full_backref)
3276 {
3277         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
3278 }
3279
3280 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3281                   struct extent_buffer *buf, int full_backref)
3282 {
3283         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
3284 }
3285
3286 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3287                                  struct btrfs_fs_info *fs_info,
3288                                  struct btrfs_path *path,
3289                                  struct btrfs_block_group_cache *cache)
3290 {
3291         int ret;
3292         struct btrfs_root *extent_root = fs_info->extent_root;
3293         unsigned long bi;
3294         struct extent_buffer *leaf;
3295
3296         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3297         if (ret) {
3298                 if (ret > 0)
3299                         ret = -ENOENT;
3300                 goto fail;
3301         }
3302
3303         leaf = path->nodes[0];
3304         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3305         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3306         btrfs_mark_buffer_dirty(leaf);
3307 fail:
3308         btrfs_release_path(path);
3309         return ret;
3310
3311 }
3312
3313 static struct btrfs_block_group_cache *
3314 next_block_group(struct btrfs_fs_info *fs_info,
3315                  struct btrfs_block_group_cache *cache)
3316 {
3317         struct rb_node *node;
3318
3319         spin_lock(&fs_info->block_group_cache_lock);
3320
3321         /* If our block group was removed, we need a full search. */
3322         if (RB_EMPTY_NODE(&cache->cache_node)) {
3323                 const u64 next_bytenr = cache->key.objectid + cache->key.offset;
3324
3325                 spin_unlock(&fs_info->block_group_cache_lock);
3326                 btrfs_put_block_group(cache);
3327                 cache = btrfs_lookup_first_block_group(fs_info, next_bytenr); return cache;
3328         }
3329         node = rb_next(&cache->cache_node);
3330         btrfs_put_block_group(cache);
3331         if (node) {
3332                 cache = rb_entry(node, struct btrfs_block_group_cache,
3333                                  cache_node);
3334                 btrfs_get_block_group(cache);
3335         } else
3336                 cache = NULL;
3337         spin_unlock(&fs_info->block_group_cache_lock);
3338         return cache;
3339 }
3340
3341 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3342                             struct btrfs_trans_handle *trans,
3343                             struct btrfs_path *path)
3344 {
3345         struct btrfs_fs_info *fs_info = block_group->fs_info;
3346         struct btrfs_root *root = fs_info->tree_root;
3347         struct inode *inode = NULL;
3348         u64 alloc_hint = 0;
3349         int dcs = BTRFS_DC_ERROR;
3350         u64 num_pages = 0;
3351         int retries = 0;
3352         int ret = 0;
3353
3354         /*
3355          * If this block group is smaller than 100 megs don't bother caching the
3356          * block group.
3357          */
3358         if (block_group->key.offset < (100 * SZ_1M)) {
3359                 spin_lock(&block_group->lock);
3360                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3361                 spin_unlock(&block_group->lock);
3362                 return 0;
3363         }
3364
3365         if (trans->aborted)
3366                 return 0;
3367 again:
3368         inode = lookup_free_space_inode(root, block_group, path);
3369         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3370                 ret = PTR_ERR(inode);
3371                 btrfs_release_path(path);
3372                 goto out;
3373         }
3374
3375         if (IS_ERR(inode)) {
3376                 BUG_ON(retries);
3377                 retries++;
3378
3379                 if (block_group->ro)
3380                         goto out_free;
3381
3382                 ret = create_free_space_inode(root, trans, block_group, path);
3383                 if (ret)
3384                         goto out_free;
3385                 goto again;
3386         }
3387
3388         /* We've already setup this transaction, go ahead and exit */
3389         if (block_group->cache_generation == trans->transid &&
3390             i_size_read(inode)) {
3391                 dcs = BTRFS_DC_SETUP;
3392                 goto out_put;
3393         }
3394
3395         /*
3396          * We want to set the generation to 0, that way if anything goes wrong
3397          * from here on out we know not to trust this cache when we load up next
3398          * time.
3399          */
3400         BTRFS_I(inode)->generation = 0;
3401         ret = btrfs_update_inode(trans, root, inode);
3402         if (ret) {
3403                 /*
3404                  * So theoretically we could recover from this, simply set the
3405                  * super cache generation to 0 so we know to invalidate the
3406                  * cache, but then we'd have to keep track of the block groups
3407                  * that fail this way so we know we _have_ to reset this cache
3408                  * before the next commit or risk reading stale cache.  So to
3409                  * limit our exposure to horrible edge cases lets just abort the
3410                  * transaction, this only happens in really bad situations
3411                  * anyway.
3412                  */
3413                 btrfs_abort_transaction(trans, ret);
3414                 goto out_put;
3415         }
3416         WARN_ON(ret);
3417
3418         if (i_size_read(inode) > 0) {
3419                 ret = btrfs_check_trunc_cache_free_space(fs_info,
3420                                         &fs_info->global_block_rsv);
3421                 if (ret)
3422                         goto out_put;
3423
3424                 ret = btrfs_truncate_free_space_cache(root, trans, NULL, inode);
3425                 if (ret)
3426                         goto out_put;
3427         }
3428
3429         spin_lock(&block_group->lock);
3430         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3431             !btrfs_test_opt(fs_info, SPACE_CACHE)) {
3432                 /*
3433                  * don't bother trying to write stuff out _if_
3434                  * a) we're not cached,
3435                  * b) we're with nospace_cache mount option.
3436                  */
3437                 dcs = BTRFS_DC_WRITTEN;
3438                 spin_unlock(&block_group->lock);
3439                 goto out_put;
3440         }
3441         spin_unlock(&block_group->lock);
3442
3443         /*
3444          * We hit an ENOSPC when setting up the cache in this transaction, just
3445          * skip doing the setup, we've already cleared the cache so we're safe.
3446          */
3447         if (test_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags)) {
3448                 ret = -ENOSPC;
3449                 goto out_put;
3450         }
3451
3452         /*
3453          * Try to preallocate enough space based on how big the block group is.
3454          * Keep in mind this has to include any pinned space which could end up
3455          * taking up quite a bit since it's not folded into the other space
3456          * cache.
3457          */
3458         num_pages = div_u64(block_group->key.offset, SZ_256M);
3459         if (!num_pages)
3460                 num_pages = 1;
3461
3462         num_pages *= 16;
3463         num_pages *= PAGE_SIZE;
3464
3465         ret = btrfs_check_data_free_space(inode, 0, num_pages);
3466         if (ret)
3467                 goto out_put;
3468
3469         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3470                                               num_pages, num_pages,
3471                                               &alloc_hint);
3472         /*
3473          * Our cache requires contiguous chunks so that we don't modify a bunch
3474          * of metadata or split extents when writing the cache out, which means
3475          * we can enospc if we are heavily fragmented in addition to just normal
3476          * out of space conditions.  So if we hit this just skip setting up any
3477          * other block groups for this transaction, maybe we'll unpin enough
3478          * space the next time around.
3479          */
3480         if (!ret)
3481                 dcs = BTRFS_DC_SETUP;
3482         else if (ret == -ENOSPC)
3483                 set_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags);
3484
3485 out_put:
3486         iput(inode);
3487 out_free:
3488         btrfs_release_path(path);
3489 out:
3490         spin_lock(&block_group->lock);
3491         if (!ret && dcs == BTRFS_DC_SETUP)
3492                 block_group->cache_generation = trans->transid;
3493         block_group->disk_cache_state = dcs;
3494         spin_unlock(&block_group->lock);
3495
3496         return ret;
3497 }
3498
3499 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans,
3500                             struct btrfs_fs_info *fs_info)
3501 {
3502         struct btrfs_block_group_cache *cache, *tmp;
3503         struct btrfs_transaction *cur_trans = trans->transaction;
3504         struct btrfs_path *path;
3505
3506         if (list_empty(&cur_trans->dirty_bgs) ||
3507             !btrfs_test_opt(fs_info, SPACE_CACHE))
3508                 return 0;
3509
3510         path = btrfs_alloc_path();
3511         if (!path)
3512                 return -ENOMEM;
3513
3514         /* Could add new block groups, use _safe just in case */
3515         list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3516                                  dirty_list) {
3517                 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3518                         cache_save_setup(cache, trans, path);
3519         }
3520
3521         btrfs_free_path(path);
3522         return 0;
3523 }
3524
3525 /*
3526  * transaction commit does final block group cache writeback during a
3527  * critical section where nothing is allowed to change the FS.  This is
3528  * required in order for the cache to actually match the block group,
3529  * but can introduce a lot of latency into the commit.
3530  *
3531  * So, btrfs_start_dirty_block_groups is here to kick off block group
3532  * cache IO.  There's a chance we'll have to redo some of it if the
3533  * block group changes again during the commit, but it greatly reduces
3534  * the commit latency by getting rid of the easy block groups while
3535  * we're still allowing others to join the commit.
3536  */
3537 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans,
3538                                    struct btrfs_fs_info *fs_info)
3539 {
3540         struct btrfs_block_group_cache *cache;
3541         struct btrfs_transaction *cur_trans = trans->transaction;
3542         int ret = 0;
3543         int should_put;
3544         struct btrfs_path *path = NULL;
3545         LIST_HEAD(dirty);
3546         struct list_head *io = &cur_trans->io_bgs;
3547         int num_started = 0;
3548         int loops = 0;
3549
3550         spin_lock(&cur_trans->dirty_bgs_lock);
3551         if (list_empty(&cur_trans->dirty_bgs)) {
3552                 spin_unlock(&cur_trans->dirty_bgs_lock);
3553                 return 0;
3554         }
3555         list_splice_init(&cur_trans->dirty_bgs, &dirty);
3556         spin_unlock(&cur_trans->dirty_bgs_lock);
3557
3558 again:
3559         /*
3560          * make sure all the block groups on our dirty list actually
3561          * exist
3562          */
3563         btrfs_create_pending_block_groups(trans, fs_info);
3564
3565         if (!path) {
3566                 path = btrfs_alloc_path();
3567                 if (!path)
3568                         return -ENOMEM;
3569         }
3570
3571         /*
3572          * cache_write_mutex is here only to save us from balance or automatic
3573          * removal of empty block groups deleting this block group while we are
3574          * writing out the cache
3575          */
3576         mutex_lock(&trans->transaction->cache_write_mutex);
3577         while (!list_empty(&dirty)) {
3578                 cache = list_first_entry(&dirty,
3579                                          struct btrfs_block_group_cache,
3580                                          dirty_list);
3581                 /*
3582                  * this can happen if something re-dirties a block
3583                  * group that is already under IO.  Just wait for it to
3584                  * finish and then do it all again
3585                  */
3586                 if (!list_empty(&cache->io_list)) {
3587                         list_del_init(&cache->io_list);
3588                         btrfs_wait_cache_io(trans, cache, path);
3589                         btrfs_put_block_group(cache);
3590                 }
3591
3592
3593                 /*
3594                  * btrfs_wait_cache_io uses the cache->dirty_list to decide
3595                  * if it should update the cache_state.  Don't delete
3596                  * until after we wait.
3597                  *
3598                  * Since we're not running in the commit critical section
3599                  * we need the dirty_bgs_lock to protect from update_block_group
3600                  */
3601                 spin_lock(&cur_trans->dirty_bgs_lock);
3602                 list_del_init(&cache->dirty_list);
3603                 spin_unlock(&cur_trans->dirty_bgs_lock);
3604
3605                 should_put = 1;
3606
3607                 cache_save_setup(cache, trans, path);
3608
3609                 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
3610                         cache->io_ctl.inode = NULL;
3611                         ret = btrfs_write_out_cache(fs_info, trans,
3612                                                     cache, path);
3613                         if (ret == 0 && cache->io_ctl.inode) {
3614                                 num_started++;
3615                                 should_put = 0;
3616
3617                                 /*
3618                                  * the cache_write_mutex is protecting
3619                                  * the io_list
3620                                  */
3621                                 list_add_tail(&cache->io_list, io);
3622                         } else {
3623                                 /*
3624                                  * if we failed to write the cache, the
3625                                  * generation will be bad and life goes on
3626                                  */
3627                                 ret = 0;
3628                         }
3629                 }
3630                 if (!ret) {
3631                         ret = write_one_cache_group(trans, fs_info,
3632                                                     path, cache);
3633                         /*
3634                          * Our block group might still be attached to the list
3635                          * of new block groups in the transaction handle of some
3636                          * other task (struct btrfs_trans_handle->new_bgs). This
3637                          * means its block group item isn't yet in the extent
3638                          * tree. If this happens ignore the error, as we will
3639                          * try again later in the critical section of the
3640                          * transaction commit.
3641                          */
3642                         if (ret == -ENOENT) {
3643                                 ret = 0;
3644                                 spin_lock(&cur_trans->dirty_bgs_lock);
3645                                 if (list_empty(&cache->dirty_list)) {
3646                                         list_add_tail(&cache->dirty_list,
3647                                                       &cur_trans->dirty_bgs);
3648                                         btrfs_get_block_group(cache);
3649                                 }
3650                                 spin_unlock(&cur_trans->dirty_bgs_lock);
3651                         } else if (ret) {
3652                                 btrfs_abort_transaction(trans, ret);
3653                         }
3654                 }
3655
3656                 /* if its not on the io list, we need to put the block group */
3657                 if (should_put)
3658                         btrfs_put_block_group(cache);
3659
3660                 if (ret)
3661                         break;
3662
3663                 /*
3664                  * Avoid blocking other tasks for too long. It might even save
3665                  * us from writing caches for block groups that are going to be
3666                  * removed.
3667                  */
3668                 mutex_unlock(&trans->transaction->cache_write_mutex);
3669                 mutex_lock(&trans->transaction->cache_write_mutex);
3670         }
3671         mutex_unlock(&trans->transaction->cache_write_mutex);
3672
3673         /*
3674          * go through delayed refs for all the stuff we've just kicked off
3675          * and then loop back (just once)
3676          */
3677         ret = btrfs_run_delayed_refs(trans, fs_info, 0);
3678         if (!ret && loops == 0) {
3679                 loops++;
3680                 spin_lock(&cur_trans->dirty_bgs_lock);
3681                 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3682                 /*
3683                  * dirty_bgs_lock protects us from concurrent block group
3684                  * deletes too (not just cache_write_mutex).
3685                  */
3686                 if (!list_empty(&dirty)) {
3687                         spin_unlock(&cur_trans->dirty_bgs_lock);
3688                         goto again;
3689                 }
3690                 spin_unlock(&cur_trans->dirty_bgs_lock);
3691         } else if (ret < 0) {
3692                 btrfs_cleanup_dirty_bgs(cur_trans, fs_info);
3693         }
3694
3695         btrfs_free_path(path);
3696         return ret;
3697 }
3698
3699 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3700                                    struct btrfs_fs_info *fs_info)
3701 {
3702         struct btrfs_block_group_cache *cache;
3703         struct btrfs_transaction *cur_trans = trans->transaction;
3704         int ret = 0;
3705         int should_put;
3706         struct btrfs_path *path;
3707         struct list_head *io = &cur_trans->io_bgs;
3708         int num_started = 0;
3709
3710         path = btrfs_alloc_path();
3711         if (!path)
3712                 return -ENOMEM;
3713
3714         /*
3715          * Even though we are in the critical section of the transaction commit,
3716          * we can still have concurrent tasks adding elements to this
3717          * transaction's list of dirty block groups. These tasks correspond to
3718          * endio free space workers started when writeback finishes for a
3719          * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
3720          * allocate new block groups as a result of COWing nodes of the root
3721          * tree when updating the free space inode. The writeback for the space
3722          * caches is triggered by an earlier call to
3723          * btrfs_start_dirty_block_groups() and iterations of the following
3724          * loop.
3725          * Also we want to do the cache_save_setup first and then run the
3726          * delayed refs to make sure we have the best chance at doing this all
3727          * in one shot.
3728          */
3729         spin_lock(&cur_trans->dirty_bgs_lock);
3730         while (!list_empty(&cur_trans->dirty_bgs)) {
3731                 cache = list_first_entry(&cur_trans->dirty_bgs,
3732                                          struct btrfs_block_group_cache,
3733                                          dirty_list);
3734
3735                 /*
3736                  * this can happen if cache_save_setup re-dirties a block
3737                  * group that is already under IO.  Just wait for it to
3738                  * finish and then do it all again
3739                  */
3740                 if (!list_empty(&cache->io_list)) {
3741                         spin_unlock(&cur_trans->dirty_bgs_lock);
3742                         list_del_init(&cache->io_list);
3743                         btrfs_wait_cache_io(trans, cache, path);
3744                         btrfs_put_block_group(cache);
3745                         spin_lock(&cur_trans->dirty_bgs_lock);
3746                 }
3747
3748                 /*
3749                  * don't remove from the dirty list until after we've waited
3750                  * on any pending IO
3751                  */
3752                 list_del_init(&cache->dirty_list);
3753                 spin_unlock(&cur_trans->dirty_bgs_lock);
3754                 should_put = 1;
3755
3756                 cache_save_setup(cache, trans, path);
3757
3758                 if (!ret)
3759                         ret = btrfs_run_delayed_refs(trans, fs_info,
3760                                                      (unsigned long) -1);
3761
3762                 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3763                         cache->io_ctl.inode = NULL;
3764                         ret = btrfs_write_out_cache(fs_info, trans,
3765                                                     cache, path);
3766                         if (ret == 0 && cache->io_ctl.inode) {
3767                                 num_started++;
3768                                 should_put = 0;
3769                                 list_add_tail(&cache->io_list, io);
3770                         } else {
3771                                 /*
3772                                  * if we failed to write the cache, the
3773                                  * generation will be bad and life goes on
3774                                  */
3775                                 ret = 0;
3776                         }
3777                 }
3778                 if (!ret) {
3779                         ret = write_one_cache_group(trans, fs_info,
3780                                                     path, cache);
3781                         /*
3782                          * One of the free space endio workers might have
3783                          * created a new block group while updating a free space
3784                          * cache's inode (at inode.c:btrfs_finish_ordered_io())
3785                          * and hasn't released its transaction handle yet, in
3786                          * which case the new block group is still attached to
3787                          * its transaction handle and its creation has not
3788                          * finished yet (no block group item in the extent tree
3789                          * yet, etc). If this is the case, wait for all free
3790                          * space endio workers to finish and retry. This is a
3791                          * a very rare case so no need for a more efficient and
3792                          * complex approach.
3793                          */
3794                         if (ret == -ENOENT) {
3795                                 wait_event(cur_trans->writer_wait,
3796                                    atomic_read(&cur_trans->num_writers) == 1);
3797                                 ret = write_one_cache_group(trans, fs_info,
3798                                                             path, cache);
3799                         }
3800                         if (ret)
3801                                 btrfs_abort_transaction(trans, ret);
3802                 }
3803
3804                 /* if its not on the io list, we need to put the block group */
3805                 if (should_put)
3806                         btrfs_put_block_group(cache);
3807                 spin_lock(&cur_trans->dirty_bgs_lock);
3808         }
3809         spin_unlock(&cur_trans->dirty_bgs_lock);
3810
3811         while (!list_empty(io)) {
3812                 cache = list_first_entry(io, struct btrfs_block_group_cache,
3813                                          io_list);
3814                 list_del_init(&cache->io_list);
3815                 btrfs_wait_cache_io(trans, cache, path);
3816                 btrfs_put_block_group(cache);
3817         }
3818
3819         btrfs_free_path(path);
3820         return ret;
3821 }
3822
3823 int btrfs_extent_readonly(struct btrfs_fs_info *fs_info, u64 bytenr)
3824 {
3825         struct btrfs_block_group_cache *block_group;
3826         int readonly = 0;
3827
3828         block_group = btrfs_lookup_block_group(fs_info, bytenr);
3829         if (!block_group || block_group->ro)
3830                 readonly = 1;
3831         if (block_group)
3832                 btrfs_put_block_group(block_group);
3833         return readonly;
3834 }
3835
3836 bool btrfs_inc_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3837 {
3838         struct btrfs_block_group_cache *bg;
3839         bool ret = true;
3840
3841         bg = btrfs_lookup_block_group(fs_info, bytenr);
3842         if (!bg)
3843                 return false;
3844
3845         spin_lock(&bg->lock);
3846         if (bg->ro)
3847                 ret = false;
3848         else
3849                 atomic_inc(&bg->nocow_writers);
3850         spin_unlock(&bg->lock);
3851
3852         /* no put on block group, done by btrfs_dec_nocow_writers */
3853         if (!ret)
3854                 btrfs_put_block_group(bg);
3855
3856         return ret;
3857
3858 }
3859
3860 void btrfs_dec_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3861 {
3862         struct btrfs_block_group_cache *bg;
3863
3864         bg = btrfs_lookup_block_group(fs_info, bytenr);
3865         ASSERT(bg);
3866         if (atomic_dec_and_test(&bg->nocow_writers))
3867                 wake_up_atomic_t(&bg->nocow_writers);
3868         /*
3869          * Once for our lookup and once for the lookup done by a previous call
3870          * to btrfs_inc_nocow_writers()
3871          */
3872         btrfs_put_block_group(bg);
3873         btrfs_put_block_group(bg);
3874 }
3875
3876 static int btrfs_wait_nocow_writers_atomic_t(atomic_t *a)
3877 {
3878         schedule();
3879         return 0;
3880 }
3881
3882 void btrfs_wait_nocow_writers(struct btrfs_block_group_cache *bg)
3883 {
3884         wait_on_atomic_t(&bg->nocow_writers,
3885                          btrfs_wait_nocow_writers_atomic_t,
3886                          TASK_UNINTERRUPTIBLE);
3887 }
3888
3889 static const char *alloc_name(u64 flags)
3890 {
3891         switch (flags) {
3892         case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA:
3893                 return "mixed";
3894         case BTRFS_BLOCK_GROUP_METADATA:
3895                 return "metadata";
3896         case BTRFS_BLOCK_GROUP_DATA:
3897                 return "data";
3898         case BTRFS_BLOCK_GROUP_SYSTEM:
3899                 return "system";
3900         default:
3901                 WARN_ON(1);
3902                 return "invalid-combination";
3903         };
3904 }
3905
3906 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3907                              u64 total_bytes, u64 bytes_used,
3908                              u64 bytes_readonly,
3909                              struct btrfs_space_info **space_info)
3910 {
3911         struct btrfs_space_info *found;
3912         int i;
3913         int factor;
3914         int ret;
3915
3916         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3917                      BTRFS_BLOCK_GROUP_RAID10))
3918                 factor = 2;
3919         else
3920                 factor = 1;
3921
3922         found = __find_space_info(info, flags);
3923         if (found) {
3924                 spin_lock(&found->lock);
3925                 found->total_bytes += total_bytes;
3926                 found->disk_total += total_bytes * factor;
3927                 found->bytes_used += bytes_used;
3928                 found->disk_used += bytes_used * factor;
3929                 found->bytes_readonly += bytes_readonly;
3930                 if (total_bytes > 0)
3931                         found->full = 0;
3932                 space_info_add_new_bytes(info, found, total_bytes -
3933                                          bytes_used - bytes_readonly);
3934                 spin_unlock(&found->lock);
3935                 *space_info = found;
3936                 return 0;
3937         }
3938         found = kzalloc(sizeof(*found), GFP_NOFS);
3939         if (!found)
3940                 return -ENOMEM;
3941
3942         ret = percpu_counter_init(&found->total_bytes_pinned, 0, GFP_KERNEL);
3943         if (ret) {
3944                 kfree(found);
3945                 return ret;
3946         }
3947
3948         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3949                 INIT_LIST_HEAD(&found->block_groups[i]);
3950         init_rwsem(&found->groups_sem);
3951         spin_lock_init(&found->lock);
3952         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3953         found->total_bytes = total_bytes;
3954         found->disk_total = total_bytes * factor;
3955         found->bytes_used = bytes_used;
3956         found->disk_used = bytes_used * factor;
3957         found->bytes_pinned = 0;
3958         found->bytes_reserved = 0;
3959         found->bytes_readonly = bytes_readonly;
3960         found->bytes_may_use = 0;
3961         found->full = 0;
3962         found->max_extent_size = 0;
3963         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3964         found->chunk_alloc = 0;
3965         found->flush = 0;
3966         init_waitqueue_head(&found->wait);
3967         INIT_LIST_HEAD(&found->ro_bgs);
3968         INIT_LIST_HEAD(&found->tickets);
3969         INIT_LIST_HEAD(&found->priority_tickets);
3970
3971         ret = kobject_init_and_add(&found->kobj, &space_info_ktype,
3972                                     info->space_info_kobj, "%s",
3973                                     alloc_name(found->flags));
3974         if (ret) {
3975                 kfree(found);
3976                 return ret;
3977         }
3978
3979         *space_info = found;
3980         list_add_rcu(&found->list, &info->space_info);
3981         if (flags & BTRFS_BLOCK_GROUP_DATA)
3982                 info->data_sinfo = found;
3983
3984         return ret;
3985 }
3986
3987 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3988 {
3989         u64 extra_flags = chunk_to_extended(flags) &
3990                                 BTRFS_EXTENDED_PROFILE_MASK;
3991
3992         write_seqlock(&fs_info->profiles_lock);
3993         if (flags & BTRFS_BLOCK_GROUP_DATA)
3994                 fs_info->avail_data_alloc_bits |= extra_flags;
3995         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3996                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3997         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3998                 fs_info->avail_system_alloc_bits |= extra_flags;
3999         write_sequnlock(&fs_info->profiles_lock);
4000 }
4001
4002 /*
4003  * returns target flags in extended format or 0 if restripe for this
4004  * chunk_type is not in progress
4005  *
4006  * should be called with either volume_mutex or balance_lock held
4007  */
4008 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
4009 {
4010         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
4011         u64 target = 0;
4012
4013         if (!bctl)
4014                 return 0;
4015
4016         if (flags & BTRFS_BLOCK_GROUP_DATA &&
4017             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4018                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
4019         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
4020                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4021                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
4022         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
4023                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4024                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
4025         }
4026
4027         return target;
4028 }
4029
4030 /*
4031  * @flags: available profiles in extended format (see ctree.h)
4032  *
4033  * Returns reduced profile in chunk format.  If profile changing is in
4034  * progress (either running or paused) picks the target profile (if it's
4035  * already available), otherwise falls back to plain reducing.
4036  */
4037 static u64 btrfs_reduce_alloc_profile(struct btrfs_fs_info *fs_info, u64 flags)
4038 {
4039         u64 num_devices = fs_info->fs_devices->rw_devices;
4040         u64 target;
4041         u64 raid_type;
4042         u64 allowed = 0;
4043
4044         /*
4045          * see if restripe for this chunk_type is in progress, if so
4046          * try to reduce to the target profile
4047          */
4048         spin_lock(&fs_info->balance_lock);
4049         target = get_restripe_target(fs_info, flags);
4050         if (target) {
4051                 /* pick target profile only if it's already available */
4052                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
4053                         spin_unlock(&fs_info->balance_lock);
4054                         return extended_to_chunk(target);
4055                 }
4056         }
4057         spin_unlock(&fs_info->balance_lock);
4058
4059         /* First, mask out the RAID levels which aren't possible */
4060         for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
4061                 if (num_devices >= btrfs_raid_array[raid_type].devs_min)
4062                         allowed |= btrfs_raid_group[raid_type];
4063         }
4064         allowed &= flags;
4065
4066         if (allowed & BTRFS_BLOCK_GROUP_RAID6)
4067                 allowed = BTRFS_BLOCK_GROUP_RAID6;
4068         else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
4069                 allowed = BTRFS_BLOCK_GROUP_RAID5;
4070         else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
4071                 allowed = BTRFS_BLOCK_GROUP_RAID10;
4072         else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
4073                 allowed = BTRFS_BLOCK_GROUP_RAID1;
4074         else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
4075                 allowed = BTRFS_BLOCK_GROUP_RAID0;
4076
4077         flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
4078
4079         return extended_to_chunk(flags | allowed);
4080 }
4081
4082 static u64 get_alloc_profile(struct btrfs_fs_info *fs_info, u64 orig_flags)
4083 {
4084         unsigned seq;
4085         u64 flags;
4086
4087         do {
4088                 flags = orig_flags;
4089                 seq = read_seqbegin(&fs_info->profiles_lock);
4090
4091                 if (flags & BTRFS_BLOCK_GROUP_DATA)
4092                         flags |= fs_info->avail_data_alloc_bits;
4093                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4094                         flags |= fs_info->avail_system_alloc_bits;
4095                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
4096                         flags |= fs_info->avail_metadata_alloc_bits;
4097         } while (read_seqretry(&fs_info->profiles_lock, seq));
4098
4099         return btrfs_reduce_alloc_profile(fs_info, flags);
4100 }
4101
4102 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
4103 {
4104         struct btrfs_fs_info *fs_info = root->fs_info;
4105         u64 flags;
4106         u64 ret;
4107
4108         if (data)
4109                 flags = BTRFS_BLOCK_GROUP_DATA;
4110         else if (root == fs_info->chunk_root)
4111                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
4112         else
4113                 flags = BTRFS_BLOCK_GROUP_METADATA;
4114
4115         ret = get_alloc_profile(fs_info, flags);
4116         return ret;
4117 }
4118
4119 int btrfs_alloc_data_chunk_ondemand(struct inode *inode, u64 bytes)
4120 {
4121         struct btrfs_space_info *data_sinfo;
4122         struct btrfs_root *root = BTRFS_I(inode)->root;
4123         struct btrfs_fs_info *fs_info = root->fs_info;
4124         u64 used;
4125         int ret = 0;
4126         int need_commit = 2;
4127         int have_pinned_space;
4128
4129         /* make sure bytes are sectorsize aligned */
4130         bytes = ALIGN(bytes, fs_info->sectorsize);
4131
4132         if (btrfs_is_free_space_inode(inode)) {
4133                 need_commit = 0;
4134                 ASSERT(current->journal_info);
4135         }
4136
4137         data_sinfo = fs_info->data_sinfo;
4138         if (!data_sinfo)
4139                 goto alloc;
4140
4141 again:
4142         /* make sure we have enough space to handle the data first */
4143         spin_lock(&data_sinfo->lock);
4144         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
4145                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
4146                 data_sinfo->bytes_may_use;
4147
4148         if (used + bytes > data_sinfo->total_bytes) {
4149                 struct btrfs_trans_handle *trans;
4150
4151                 /*
4152                  * if we don't have enough free bytes in this space then we need
4153                  * to alloc a new chunk.
4154                  */
4155                 if (!data_sinfo->full) {
4156                         u64 alloc_target;
4157
4158                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
4159                         spin_unlock(&data_sinfo->lock);
4160 alloc:
4161                         alloc_target = btrfs_get_alloc_profile(root, 1);
4162                         /*
4163                          * It is ugly that we don't call nolock join
4164                          * transaction for the free space inode case here.
4165                          * But it is safe because we only do the data space
4166                          * reservation for the free space cache in the
4167                          * transaction context, the common join transaction
4168                          * just increase the counter of the current transaction
4169                          * handler, doesn't try to acquire the trans_lock of
4170                          * the fs.
4171                          */
4172                         trans = btrfs_join_transaction(root);
4173                         if (IS_ERR(trans))
4174                                 return PTR_ERR(trans);
4175
4176                         ret = do_chunk_alloc(trans, fs_info, alloc_target,
4177                                              CHUNK_ALLOC_NO_FORCE);
4178                         btrfs_end_transaction(trans);
4179                         if (ret < 0) {
4180                                 if (ret != -ENOSPC)
4181                                         return ret;
4182                                 else {
4183                                         have_pinned_space = 1;
4184                                         goto commit_trans;
4185                                 }
4186                         }
4187
4188                         if (!data_sinfo)
4189                                 data_sinfo = fs_info->data_sinfo;
4190
4191                         goto again;
4192                 }
4193
4194                 /*
4195                  * If we don't have enough pinned space to deal with this
4196                  * allocation, and no removed chunk in current transaction,
4197                  * don't bother committing the transaction.
4198                  */
4199                 have_pinned_space = percpu_counter_compare(
4200                         &data_sinfo->total_bytes_pinned,
4201                         used + bytes - data_sinfo->total_bytes);
4202                 spin_unlock(&data_sinfo->lock);
4203
4204                 /* commit the current transaction and try again */
4205 commit_trans:
4206                 if (need_commit &&
4207                     !atomic_read(&fs_info->open_ioctl_trans)) {
4208                         need_commit--;
4209
4210                         if (need_commit > 0) {
4211                                 btrfs_start_delalloc_roots(fs_info, 0, -1);
4212                                 btrfs_wait_ordered_roots(fs_info, -1, 0,
4213                                                          (u64)-1);
4214                         }
4215
4216                         trans = btrfs_join_transaction(root);
4217                         if (IS_ERR(trans))
4218                                 return PTR_ERR(trans);
4219                         if (have_pinned_space >= 0 ||
4220                             test_bit(BTRFS_TRANS_HAVE_FREE_BGS,
4221                                      &trans->transaction->flags) ||
4222                             need_commit > 0) {
4223                                 ret = btrfs_commit_transaction(trans);
4224                                 if (ret)
4225                                         return ret;
4226                                 /*
4227                                  * The cleaner kthread might still be doing iput
4228                                  * operations. Wait for it to finish so that
4229                                  * more space is released.
4230                                  */
4231                                 mutex_lock(&fs_info->cleaner_delayed_iput_mutex);
4232                                 mutex_unlock(&fs_info->cleaner_delayed_iput_mutex);
4233                                 goto again;
4234                         } else {
4235                                 btrfs_end_transaction(trans);
4236                         }
4237                 }
4238
4239                 trace_btrfs_space_reservation(fs_info,
4240                                               "space_info:enospc",
4241                                               data_sinfo->flags, bytes, 1);
4242                 return -ENOSPC;
4243         }
4244         data_sinfo->bytes_may_use += bytes;
4245         trace_btrfs_space_reservation(fs_info, "space_info",
4246                                       data_sinfo->flags, bytes, 1);
4247         spin_unlock(&data_sinfo->lock);
4248
4249         return ret;
4250 }
4251
4252 /*
4253  * New check_data_free_space() with ability for precious data reservation
4254  * Will replace old btrfs_check_data_free_space(), but for patch split,
4255  * add a new function first and then replace it.
4256  */
4257 int btrfs_check_data_free_space(struct inode *inode, u64 start, u64 len)
4258 {
4259         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4260         int ret;
4261
4262         /* align the range */
4263         len = round_up(start + len, fs_info->sectorsize) -
4264               round_down(start, fs_info->sectorsize);
4265         start = round_down(start, fs_info->sectorsize);
4266
4267         ret = btrfs_alloc_data_chunk_ondemand(inode, len);
4268         if (ret < 0)
4269                 return ret;
4270
4271         /* Use new btrfs_qgroup_reserve_data to reserve precious data space. */
4272         ret = btrfs_qgroup_reserve_data(inode, start, len);
4273         if (ret)
4274                 btrfs_free_reserved_data_space_noquota(inode, start, len);
4275         return ret;
4276 }
4277
4278 /*
4279  * Called if we need to clear a data reservation for this inode
4280  * Normally in a error case.
4281  *
4282  * This one will *NOT* use accurate qgroup reserved space API, just for case
4283  * which we can't sleep and is sure it won't affect qgroup reserved space.
4284  * Like clear_bit_hook().
4285  */
4286 void btrfs_free_reserved_data_space_noquota(struct inode *inode, u64 start,
4287                                             u64 len)
4288 {
4289         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4290         struct btrfs_space_info *data_sinfo;
4291
4292         /* Make sure the range is aligned to sectorsize */
4293         len = round_up(start + len, fs_info->sectorsize) -
4294               round_down(start, fs_info->sectorsize);
4295         start = round_down(start, fs_info->sectorsize);
4296
4297         data_sinfo = fs_info->data_sinfo;
4298         spin_lock(&data_sinfo->lock);
4299         if (WARN_ON(data_sinfo->bytes_may_use < len))
4300                 data_sinfo->bytes_may_use = 0;
4301         else
4302                 data_sinfo->bytes_may_use -= len;
4303         trace_btrfs_space_reservation(fs_info, "space_info",
4304                                       data_sinfo->flags, len, 0);
4305         spin_unlock(&data_sinfo->lock);
4306 }
4307
4308 /*
4309  * Called if we need to clear a data reservation for this inode
4310  * Normally in a error case.
4311  *
4312  * This one will handle the per-inode data rsv map for accurate reserved
4313  * space framework.
4314  */
4315 void btrfs_free_reserved_data_space(struct inode *inode, u64 start, u64 len)
4316 {
4317         struct btrfs_root *root = BTRFS_I(inode)->root;
4318
4319         /* Make sure the range is aligned to sectorsize */
4320         len = round_up(start + len, root->fs_info->sectorsize) -
4321               round_down(start, root->fs_info->sectorsize);
4322         start = round_down(start, root->fs_info->sectorsize);
4323
4324         btrfs_free_reserved_data_space_noquota(inode, start, len);
4325         btrfs_qgroup_free_data(inode, start, len);
4326 }
4327
4328 static void force_metadata_allocation(struct btrfs_fs_info *info)
4329 {
4330         struct list_head *head = &info->space_info;
4331         struct btrfs_space_info *found;
4332
4333         rcu_read_lock();
4334         list_for_each_entry_rcu(found, head, list) {
4335                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
4336                         found->force_alloc = CHUNK_ALLOC_FORCE;
4337         }
4338         rcu_read_unlock();
4339 }
4340
4341 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
4342 {
4343         return (global->size << 1);
4344 }
4345
4346 static int should_alloc_chunk(struct btrfs_fs_info *fs_info,
4347                               struct btrfs_space_info *sinfo, int force)
4348 {
4349         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4350         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
4351         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
4352         u64 thresh;
4353
4354         if (force == CHUNK_ALLOC_FORCE)
4355                 return 1;
4356
4357         /*
4358          * We need to take into account the global rsv because for all intents
4359          * and purposes it's used space.  Don't worry about locking the
4360          * global_rsv, it doesn't change except when the transaction commits.
4361          */
4362         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
4363                 num_allocated += calc_global_rsv_need_space(global_rsv);
4364
4365         /*
4366          * in limited mode, we want to have some free space up to
4367          * about 1% of the FS size.
4368          */
4369         if (force == CHUNK_ALLOC_LIMITED) {
4370                 thresh = btrfs_super_total_bytes(fs_info->super_copy);
4371                 thresh = max_t(u64, SZ_64M, div_factor_fine(thresh, 1));
4372
4373                 if (num_bytes - num_allocated < thresh)
4374                         return 1;
4375         }
4376
4377         if (num_allocated + SZ_2M < div_factor(num_bytes, 8))
4378                 return 0;
4379         return 1;
4380 }
4381
4382 static u64 get_profile_num_devs(struct btrfs_fs_info *fs_info, u64 type)
4383 {
4384         u64 num_dev;
4385
4386         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
4387                     BTRFS_BLOCK_GROUP_RAID0 |
4388                     BTRFS_BLOCK_GROUP_RAID5 |
4389                     BTRFS_BLOCK_GROUP_RAID6))
4390                 num_dev = fs_info->fs_devices->rw_devices;
4391         else if (type & BTRFS_BLOCK_GROUP_RAID1)
4392                 num_dev = 2;
4393         else
4394                 num_dev = 1;    /* DUP or single */
4395
4396         return num_dev;
4397 }
4398
4399 /*
4400  * If @is_allocation is true, reserve space in the system space info necessary
4401  * for allocating a chunk, otherwise if it's false, reserve space necessary for
4402  * removing a chunk.
4403  */
4404 void check_system_chunk(struct btrfs_trans_handle *trans,
4405                         struct btrfs_fs_info *fs_info, u64 type)
4406 {
4407         struct btrfs_space_info *info;
4408         u64 left;
4409         u64 thresh;
4410         int ret = 0;
4411         u64 num_devs;
4412
4413         /*
4414          * Needed because we can end up allocating a system chunk and for an
4415          * atomic and race free space reservation in the chunk block reserve.
4416          */
4417         ASSERT(mutex_is_locked(&fs_info->chunk_mutex));
4418
4419         info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4420         spin_lock(&info->lock);
4421         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
4422                 info->bytes_reserved - info->bytes_readonly -
4423                 info->bytes_may_use;
4424         spin_unlock(&info->lock);
4425
4426         num_devs = get_profile_num_devs(fs_info, type);
4427
4428         /* num_devs device items to update and 1 chunk item to add or remove */
4429         thresh = btrfs_calc_trunc_metadata_size(fs_info, num_devs) +
4430                 btrfs_calc_trans_metadata_size(fs_info, 1);
4431
4432         if (left < thresh && btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
4433                 btrfs_info(fs_info, "left=%llu, need=%llu, flags=%llu",
4434                            left, thresh, type);
4435                 dump_space_info(fs_info, info, 0, 0);
4436         }
4437
4438         if (left < thresh) {
4439                 u64 flags;
4440
4441                 flags = btrfs_get_alloc_profile(fs_info->chunk_root, 0);
4442                 /*
4443                  * Ignore failure to create system chunk. We might end up not
4444                  * needing it, as we might not need to COW all nodes/leafs from
4445                  * the paths we visit in the chunk tree (they were already COWed
4446                  * or created in the current transaction for example).
4447                  */
4448                 ret = btrfs_alloc_chunk(trans, fs_info, flags);
4449         }
4450
4451         if (!ret) {
4452                 ret = btrfs_block_rsv_add(fs_info->chunk_root,
4453                                           &fs_info->chunk_block_rsv,
4454                                           thresh, BTRFS_RESERVE_NO_FLUSH);
4455                 if (!ret)
4456                         trans->chunk_bytes_reserved += thresh;
4457         }
4458 }
4459
4460 /*
4461  * If force is CHUNK_ALLOC_FORCE:
4462  *    - return 1 if it successfully allocates a chunk,
4463  *    - return errors including -ENOSPC otherwise.
4464  * If force is NOT CHUNK_ALLOC_FORCE:
4465  *    - return 0 if it doesn't need to allocate a new chunk,
4466  *    - return 1 if it successfully allocates a chunk,
4467  *    - return errors including -ENOSPC otherwise.
4468  */
4469 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
4470                           struct btrfs_fs_info *fs_info, u64 flags, int force)
4471 {
4472         struct btrfs_space_info *space_info;
4473         int wait_for_alloc = 0;
4474         int ret = 0;
4475
4476         /* Don't re-enter if we're already allocating a chunk */
4477         if (trans->allocating_chunk)
4478                 return -ENOSPC;
4479
4480         space_info = __find_space_info(fs_info, flags);
4481         if (!space_info) {
4482                 ret = update_space_info(fs_info, flags, 0, 0, 0, &space_info);
4483                 BUG_ON(ret); /* -ENOMEM */
4484         }
4485         BUG_ON(!space_info); /* Logic error */
4486
4487 again:
4488         spin_lock(&space_info->lock);
4489         if (force < space_info->force_alloc)
4490                 force = space_info->force_alloc;
4491         if (space_info->full) {
4492                 if (should_alloc_chunk(fs_info, space_info, force))
4493                         ret = -ENOSPC;
4494                 else
4495                         ret = 0;
4496                 spin_unlock(&space_info->lock);
4497                 return ret;
4498         }
4499
4500         if (!should_alloc_chunk(fs_info, space_info, force)) {
4501                 spin_unlock(&space_info->lock);
4502                 return 0;
4503         } else if (space_info->chunk_alloc) {
4504                 wait_for_alloc = 1;
4505         } else {
4506                 space_info->chunk_alloc = 1;
4507         }
4508
4509         spin_unlock(&space_info->lock);
4510
4511         mutex_lock(&fs_info->chunk_mutex);
4512
4513         /*
4514          * The chunk_mutex is held throughout the entirety of a chunk
4515          * allocation, so once we've acquired the chunk_mutex we know that the
4516          * other guy is done and we need to recheck and see if we should
4517          * allocate.
4518          */
4519         if (wait_for_alloc) {
4520                 mutex_unlock(&fs_info->chunk_mutex);
4521                 wait_for_alloc = 0;
4522                 goto again;
4523         }
4524
4525         trans->allocating_chunk = true;
4526
4527         /*
4528          * If we have mixed data/metadata chunks we want to make sure we keep
4529          * allocating mixed chunks instead of individual chunks.
4530          */
4531         if (btrfs_mixed_space_info(space_info))
4532                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4533
4534         /*
4535          * if we're doing a data chunk, go ahead and make sure that
4536          * we keep a reasonable number of metadata chunks allocated in the
4537          * FS as well.
4538          */
4539         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4540                 fs_info->data_chunk_allocations++;
4541                 if (!(fs_info->data_chunk_allocations %
4542                       fs_info->metadata_ratio))
4543                         force_metadata_allocation(fs_info);
4544         }
4545
4546         /*
4547          * Check if we have enough space in SYSTEM chunk because we may need
4548          * to update devices.
4549          */
4550         check_system_chunk(trans, fs_info, flags);
4551
4552         ret = btrfs_alloc_chunk(trans, fs_info, flags);
4553         trans->allocating_chunk = false;
4554
4555         spin_lock(&space_info->lock);
4556         if (ret < 0 && ret != -ENOSPC)
4557                 goto out;
4558         if (ret)
4559                 space_info->full = 1;
4560         else
4561                 ret = 1;
4562
4563         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4564 out:
4565         space_info->chunk_alloc = 0;
4566         spin_unlock(&space_info->lock);
4567         mutex_unlock(&fs_info->chunk_mutex);
4568         /*
4569          * When we allocate a new chunk we reserve space in the chunk block
4570          * reserve to make sure we can COW nodes/leafs in the chunk tree or
4571          * add new nodes/leafs to it if we end up needing to do it when
4572          * inserting the chunk item and updating device items as part of the
4573          * second phase of chunk allocation, performed by
4574          * btrfs_finish_chunk_alloc(). So make sure we don't accumulate a
4575          * large number of new block groups to create in our transaction
4576          * handle's new_bgs list to avoid exhausting the chunk block reserve
4577          * in extreme cases - like having a single transaction create many new
4578          * block groups when starting to write out the free space caches of all
4579          * the block groups that were made dirty during the lifetime of the
4580          * transaction.
4581          */
4582         if (trans->can_flush_pending_bgs &&
4583             trans->chunk_bytes_reserved >= (u64)SZ_2M) {
4584                 btrfs_create_pending_block_groups(trans, fs_info);
4585                 btrfs_trans_release_chunk_metadata(trans);
4586         }
4587         return ret;
4588 }
4589
4590 static int can_overcommit(struct btrfs_root *root,
4591                           struct btrfs_space_info *space_info, u64 bytes,
4592                           enum btrfs_reserve_flush_enum flush)
4593 {
4594         struct btrfs_fs_info *fs_info = root->fs_info;
4595         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4596         u64 profile;
4597         u64 space_size;
4598         u64 avail;
4599         u64 used;
4600
4601         /* Don't overcommit when in mixed mode. */
4602         if (space_info->flags & BTRFS_BLOCK_GROUP_DATA)
4603                 return 0;
4604
4605         profile = btrfs_get_alloc_profile(root, 0);
4606         used = space_info->bytes_used + space_info->bytes_reserved +
4607                 space_info->bytes_pinned + space_info->bytes_readonly;
4608
4609         /*
4610          * We only want to allow over committing if we have lots of actual space
4611          * free, but if we don't have enough space to handle the global reserve
4612          * space then we could end up having a real enospc problem when trying
4613          * to allocate a chunk or some other such important allocation.
4614          */
4615         spin_lock(&global_rsv->lock);
4616         space_size = calc_global_rsv_need_space(global_rsv);
4617         spin_unlock(&global_rsv->lock);
4618         if (used + space_size >= space_info->total_bytes)
4619                 return 0;
4620
4621         used += space_info->bytes_may_use;
4622
4623         spin_lock(&fs_info->free_chunk_lock);
4624         avail = fs_info->free_chunk_space;
4625         spin_unlock(&fs_info->free_chunk_lock);
4626
4627         /*
4628          * If we have dup, raid1 or raid10 then only half of the free
4629          * space is actually useable.  For raid56, the space info used
4630          * doesn't include the parity drive, so we don't have to
4631          * change the math
4632          */
4633         if (profile & (BTRFS_BLOCK_GROUP_DUP |
4634                        BTRFS_BLOCK_GROUP_RAID1 |
4635                        BTRFS_BLOCK_GROUP_RAID10))
4636                 avail >>= 1;
4637
4638         /*
4639          * If we aren't flushing all things, let us overcommit up to
4640          * 1/2th of the space. If we can flush, don't let us overcommit
4641          * too much, let it overcommit up to 1/8 of the space.
4642          */
4643         if (flush == BTRFS_RESERVE_FLUSH_ALL)
4644                 avail >>= 3;
4645         else
4646                 avail >>= 1;
4647
4648         if (used + bytes < space_info->total_bytes + avail)
4649                 return 1;
4650         return 0;
4651 }
4652
4653 static void btrfs_writeback_inodes_sb_nr(struct btrfs_fs_info *fs_info,
4654                                          unsigned long nr_pages, int nr_items)
4655 {
4656         struct super_block *sb = fs_info->sb;
4657
4658         if (down_read_trylock(&sb->s_umount)) {
4659                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
4660                 up_read(&sb->s_umount);
4661         } else {
4662                 /*
4663                  * We needn't worry the filesystem going from r/w to r/o though
4664                  * we don't acquire ->s_umount mutex, because the filesystem
4665                  * should guarantee the delalloc inodes list be empty after
4666                  * the filesystem is readonly(all dirty pages are written to
4667                  * the disk).
4668                  */
4669                 btrfs_start_delalloc_roots(fs_info, 0, nr_items);
4670                 if (!current->journal_info)
4671                         btrfs_wait_ordered_roots(fs_info, nr_items, 0, (u64)-1);
4672         }
4673 }
4674
4675 static inline int calc_reclaim_items_nr(struct btrfs_fs_info *fs_info,
4676                                         u64 to_reclaim)
4677 {
4678         u64 bytes;
4679         int nr;
4680
4681         bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
4682         nr = (int)div64_u64(to_reclaim, bytes);
4683         if (!nr)
4684                 nr = 1;
4685         return nr;
4686 }
4687
4688 #define EXTENT_SIZE_PER_ITEM    SZ_256K
4689
4690 /*
4691  * shrink metadata reservation for delalloc
4692  */
4693 static void shrink_delalloc(struct btrfs_root *root, u64 to_reclaim, u64 orig,
4694                             bool wait_ordered)
4695 {
4696         struct btrfs_fs_info *fs_info = root->fs_info;
4697         struct btrfs_block_rsv *block_rsv;
4698         struct btrfs_space_info *space_info;
4699         struct btrfs_trans_handle *trans;
4700         u64 delalloc_bytes;
4701         u64 max_reclaim;
4702         long time_left;
4703         unsigned long nr_pages;
4704         int loops;
4705         int items;
4706         enum btrfs_reserve_flush_enum flush;
4707
4708         /* Calc the number of the pages we need flush for space reservation */
4709         items = calc_reclaim_items_nr(fs_info, to_reclaim);
4710         to_reclaim = (u64)items * EXTENT_SIZE_PER_ITEM;
4711
4712         trans = (struct btrfs_trans_handle *)current->journal_info;
4713         block_rsv = &fs_info->delalloc_block_rsv;
4714         space_info = block_rsv->space_info;
4715
4716         delalloc_bytes = percpu_counter_sum_positive(
4717                                                 &fs_info->delalloc_bytes);
4718         if (delalloc_bytes == 0) {
4719                 if (trans)
4720                         return;
4721                 if (wait_ordered)
4722                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
4723                 return;
4724         }
4725
4726         loops = 0;
4727         while (delalloc_bytes && loops < 3) {
4728                 max_reclaim = min(delalloc_bytes, to_reclaim);
4729                 nr_pages = max_reclaim >> PAGE_SHIFT;
4730                 btrfs_writeback_inodes_sb_nr(fs_info, nr_pages, items);
4731                 /*
4732                  * We need to wait for the async pages to actually start before
4733                  * we do anything.
4734                  */
4735                 max_reclaim = atomic_read(&fs_info->async_delalloc_pages);
4736                 if (!max_reclaim)
4737                         goto skip_async;
4738
4739                 if (max_reclaim <= nr_pages)
4740                         max_reclaim = 0;
4741                 else
4742                         max_reclaim -= nr_pages;
4743
4744                 wait_event(fs_info->async_submit_wait,
4745                            atomic_read(&fs_info->async_delalloc_pages) <=
4746                            (int)max_reclaim);
4747 skip_async:
4748                 if (!trans)
4749                         flush = BTRFS_RESERVE_FLUSH_ALL;
4750                 else
4751                         flush = BTRFS_RESERVE_NO_FLUSH;
4752                 spin_lock(&space_info->lock);
4753                 if (can_overcommit(root, space_info, orig, flush)) {
4754                         spin_unlock(&space_info->lock);
4755                         break;
4756                 }
4757                 if (list_empty(&space_info->tickets) &&
4758                     list_empty(&space_info->priority_tickets)) {
4759                         spin_unlock(&space_info->lock);
4760                         break;
4761                 }
4762                 spin_unlock(&space_info->lock);
4763
4764                 loops++;
4765                 if (wait_ordered && !trans) {
4766                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
4767                 } else {
4768                         time_left = schedule_timeout_killable(1);
4769                         if (time_left)
4770                                 break;
4771                 }
4772                 delalloc_bytes = percpu_counter_sum_positive(
4773                                                 &fs_info->delalloc_bytes);
4774         }
4775 }
4776
4777 /**
4778  * maybe_commit_transaction - possibly commit the transaction if its ok to
4779  * @root - the root we're allocating for
4780  * @bytes - the number of bytes we want to reserve
4781  * @force - force the commit
4782  *
4783  * This will check to make sure that committing the transaction will actually
4784  * get us somewhere and then commit the transaction if it does.  Otherwise it
4785  * will return -ENOSPC.
4786  */
4787 static int may_commit_transaction(struct btrfs_root *root,
4788                                   struct btrfs_space_info *space_info,
4789                                   u64 bytes, int force)
4790 {
4791         struct btrfs_fs_info *fs_info = root->fs_info;
4792         struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_block_rsv;
4793         struct btrfs_trans_handle *trans;
4794
4795         trans = (struct btrfs_trans_handle *)current->journal_info;
4796         if (trans)
4797                 return -EAGAIN;
4798
4799         if (force)
4800                 goto commit;
4801
4802         /* See if there is enough pinned space to make this reservation */
4803         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4804                                    bytes) >= 0)
4805                 goto commit;
4806
4807         /*
4808          * See if there is some space in the delayed insertion reservation for
4809          * this reservation.
4810          */
4811         if (space_info != delayed_rsv->space_info)
4812                 return -ENOSPC;
4813
4814         spin_lock(&delayed_rsv->lock);
4815         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4816                                    bytes - delayed_rsv->size) >= 0) {
4817                 spin_unlock(&delayed_rsv->lock);
4818                 return -ENOSPC;
4819         }
4820         spin_unlock(&delayed_rsv->lock);
4821
4822 commit:
4823         trans = btrfs_join_transaction(root);
4824         if (IS_ERR(trans))
4825                 return -ENOSPC;
4826
4827         return btrfs_commit_transaction(trans);
4828 }
4829
4830 struct reserve_ticket {
4831         u64 bytes;
4832         int error;
4833         struct list_head list;
4834         wait_queue_head_t wait;
4835 };
4836
4837 static int flush_space(struct btrfs_root *root,
4838                        struct btrfs_space_info *space_info, u64 num_bytes,
4839                        u64 orig_bytes, int state)
4840 {
4841         struct btrfs_fs_info *fs_info = root->fs_info;
4842         struct btrfs_trans_handle *trans;
4843         int nr;
4844         int ret = 0;
4845
4846         switch (state) {
4847         case FLUSH_DELAYED_ITEMS_NR:
4848         case FLUSH_DELAYED_ITEMS:
4849                 if (state == FLUSH_DELAYED_ITEMS_NR)
4850                         nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2;
4851                 else
4852                         nr = -1;
4853
4854                 trans = btrfs_join_transaction(root);
4855                 if (IS_ERR(trans)) {
4856                         ret = PTR_ERR(trans);
4857                         break;
4858                 }
4859                 ret = btrfs_run_delayed_items_nr(trans, fs_info, nr);
4860                 btrfs_end_transaction(trans);
4861                 break;
4862         case FLUSH_DELALLOC:
4863         case FLUSH_DELALLOC_WAIT:
4864                 shrink_delalloc(root, num_bytes * 2, orig_bytes,
4865                                 state == FLUSH_DELALLOC_WAIT);
4866                 break;
4867         case ALLOC_CHUNK:
4868                 trans = btrfs_join_transaction(root);
4869                 if (IS_ERR(trans)) {
4870                         ret = PTR_ERR(trans);
4871                         break;
4872                 }
4873                 ret = do_chunk_alloc(trans, fs_info,
4874                                      btrfs_get_alloc_profile(root, 0),
4875                                      CHUNK_ALLOC_NO_FORCE);
4876                 btrfs_end_transaction(trans);
4877                 if (ret > 0 || ret == -ENOSPC)
4878                         ret = 0;
4879                 break;
4880         case COMMIT_TRANS:
4881                 ret = may_commit_transaction(root, space_info, orig_bytes, 0);
4882                 break;
4883         default:
4884                 ret = -ENOSPC;
4885                 break;
4886         }
4887
4888         trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes,
4889                                 orig_bytes, state, ret);
4890         return ret;
4891 }
4892
4893 static inline u64
4894 btrfs_calc_reclaim_metadata_size(struct btrfs_root *root,
4895                                  struct btrfs_space_info *space_info)
4896 {
4897         struct reserve_ticket *ticket;
4898         u64 used;
4899         u64 expected;
4900         u64 to_reclaim = 0;
4901
4902         list_for_each_entry(ticket, &space_info->tickets, list)
4903                 to_reclaim += ticket->bytes;
4904         list_for_each_entry(ticket, &space_info->priority_tickets, list)
4905                 to_reclaim += ticket->bytes;
4906         if (to_reclaim)
4907                 return to_reclaim;
4908
4909         to_reclaim = min_t(u64, num_online_cpus() * SZ_1M, SZ_16M);
4910         if (can_overcommit(root, space_info, to_reclaim,
4911                            BTRFS_RESERVE_FLUSH_ALL))
4912                 return 0;
4913
4914         used = space_info->bytes_used + space_info->bytes_reserved +
4915                space_info->bytes_pinned + space_info->bytes_readonly +
4916                space_info->bytes_may_use;
4917         if (can_overcommit(root, space_info, SZ_1M, BTRFS_RESERVE_FLUSH_ALL))
4918                 expected = div_factor_fine(space_info->total_bytes, 95);
4919         else
4920                 expected = div_factor_fine(space_info->total_bytes, 90);
4921
4922         if (used > expected)
4923                 to_reclaim = used - expected;
4924         else
4925                 to_reclaim = 0;
4926         to_reclaim = min(to_reclaim, space_info->bytes_may_use +
4927                                      space_info->bytes_reserved);
4928         return to_reclaim;
4929 }
4930
4931 static inline int need_do_async_reclaim(struct btrfs_space_info *space_info,
4932                                         struct btrfs_root *root, u64 used)
4933 {
4934         struct btrfs_fs_info *fs_info = root->fs_info;
4935         u64 thresh = div_factor_fine(space_info->total_bytes, 98);
4936
4937         /* If we're just plain full then async reclaim just slows us down. */
4938         if ((space_info->bytes_used + space_info->bytes_reserved) >= thresh)
4939                 return 0;
4940
4941         if (!btrfs_calc_reclaim_metadata_size(root, space_info))
4942                 return 0;
4943
4944         return (used >= thresh && !btrfs_fs_closing(fs_info) &&
4945                 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
4946 }
4947
4948 static void wake_all_tickets(struct list_head *head)
4949 {
4950         struct reserve_ticket *ticket;
4951
4952         while (!list_empty(head)) {
4953                 ticket = list_first_entry(head, struct reserve_ticket, list);
4954                 list_del_init(&ticket->list);
4955                 ticket->error = -ENOSPC;
4956                 wake_up(&ticket->wait);
4957         }
4958 }
4959
4960 /*
4961  * This is for normal flushers, we can wait all goddamned day if we want to.  We
4962  * will loop and continuously try to flush as long as we are making progress.
4963  * We count progress as clearing off tickets each time we have to loop.
4964  */
4965 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
4966 {
4967         struct btrfs_fs_info *fs_info;
4968         struct btrfs_space_info *space_info;
4969         u64 to_reclaim;
4970         int flush_state;
4971         int commit_cycles = 0;
4972         u64 last_tickets_id;
4973
4974         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
4975         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4976
4977         spin_lock(&space_info->lock);
4978         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info->fs_root,
4979                                                       space_info);
4980         if (!to_reclaim) {
4981                 space_info->flush = 0;
4982                 spin_unlock(&space_info->lock);
4983                 return;
4984         }
4985         last_tickets_id = space_info->tickets_id;
4986         spin_unlock(&space_info->lock);
4987
4988         flush_state = FLUSH_DELAYED_ITEMS_NR;
4989         do {
4990                 struct reserve_ticket *ticket;
4991                 int ret;
4992
4993                 ret = flush_space(fs_info->fs_root, space_info, to_reclaim,
4994                             to_reclaim, flush_state);
4995                 spin_lock(&space_info->lock);
4996                 if (list_empty(&space_info->tickets)) {
4997                         space_info->flush = 0;
4998                         spin_unlock(&space_info->lock);
4999                         return;
5000                 }
5001                 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info->fs_root,
5002                                                               space_info);
5003                 ticket = list_first_entry(&space_info->tickets,
5004                                           struct reserve_ticket, list);
5005                 if (last_tickets_id == space_info->tickets_id) {
5006                         flush_state++;
5007                 } else {
5008                         last_tickets_id = space_info->tickets_id;
5009                         flush_state = FLUSH_DELAYED_ITEMS_NR;
5010                         if (commit_cycles)
5011                                 commit_cycles--;
5012                 }
5013
5014                 if (flush_state > COMMIT_TRANS) {
5015                         commit_cycles++;
5016                         if (commit_cycles > 2) {
5017                                 wake_all_tickets(&space_info->tickets);
5018                                 space_info->flush = 0;
5019                         } else {
5020                                 flush_state = FLUSH_DELAYED_ITEMS_NR;
5021                         }
5022                 }
5023                 spin_unlock(&space_info->lock);
5024         } while (flush_state <= COMMIT_TRANS);
5025 }
5026
5027 void btrfs_init_async_reclaim_work(struct work_struct *work)
5028 {
5029         INIT_WORK(work, btrfs_async_reclaim_metadata_space);
5030 }
5031
5032 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info,
5033                                             struct btrfs_space_info *space_info,
5034                                             struct reserve_ticket *ticket)
5035 {
5036         u64 to_reclaim;
5037         int flush_state = FLUSH_DELAYED_ITEMS_NR;
5038
5039         spin_lock(&space_info->lock);
5040         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info->fs_root,
5041                                                       space_info);
5042         if (!to_reclaim) {
5043                 spin_unlock(&space_info->lock);
5044                 return;
5045         }
5046         spin_unlock(&space_info->lock);
5047
5048         do {
5049                 flush_space(fs_info->fs_root, space_info, to_reclaim,
5050                             to_reclaim, flush_state);
5051                 flush_state++;
5052                 spin_lock(&space_info->lock);
5053                 if (ticket->bytes == 0) {
5054                         spin_unlock(&space_info->lock);
5055                         return;
5056                 }
5057                 spin_unlock(&space_info->lock);
5058
5059                 /*
5060                  * Priority flushers can't wait on delalloc without
5061                  * deadlocking.
5062                  */
5063                 if (flush_state == FLUSH_DELALLOC ||
5064                     flush_state == FLUSH_DELALLOC_WAIT)
5065                         flush_state = ALLOC_CHUNK;
5066         } while (flush_state < COMMIT_TRANS);
5067 }
5068
5069 static int wait_reserve_ticket(struct btrfs_fs_info *fs_info,
5070                                struct btrfs_space_info *space_info,
5071                                struct reserve_ticket *ticket, u64 orig_bytes)
5072
5073 {
5074         DEFINE_WAIT(wait);
5075         int ret = 0;
5076
5077         spin_lock(&space_info->lock);
5078         while (ticket->bytes > 0 && ticket->error == 0) {
5079                 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE);
5080                 if (ret) {
5081                         ret = -EINTR;
5082                         break;
5083                 }
5084                 spin_unlock(&space_info->lock);
5085
5086                 schedule();
5087
5088                 finish_wait(&ticket->wait, &wait);
5089                 spin_lock(&space_info->lock);
5090         }
5091         if (!ret)
5092                 ret = ticket->error;
5093         if (!list_empty(&ticket->list))
5094                 list_del_init(&ticket->list);
5095         if (ticket->bytes && ticket->bytes < orig_bytes) {
5096                 u64 num_bytes = orig_bytes - ticket->bytes;
5097                 space_info->bytes_may_use -= num_bytes;
5098                 trace_btrfs_space_reservation(fs_info, "space_info",
5099                                               space_info->flags, num_bytes, 0);
5100         }
5101         spin_unlock(&space_info->lock);
5102
5103         return ret;
5104 }
5105
5106 /**
5107  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5108  * @root - the root we're allocating for
5109  * @space_info - the space info we want to allocate from
5110  * @orig_bytes - the number of bytes we want
5111  * @flush - whether or not we can flush to make our reservation
5112  *
5113  * This will reserve orig_bytes number of bytes from the space info associated
5114  * with the block_rsv.  If there is not enough space it will make an attempt to
5115  * flush out space to make room.  It will do this by flushing delalloc if
5116  * possible or committing the transaction.  If flush is 0 then no attempts to
5117  * regain reservations will be made and this will fail if there is not enough
5118  * space already.
5119  */
5120 static int __reserve_metadata_bytes(struct btrfs_root *root,
5121                                     struct btrfs_space_info *space_info,
5122                                     u64 orig_bytes,
5123                                     enum btrfs_reserve_flush_enum flush)
5124 {
5125         struct btrfs_fs_info *fs_info = root->fs_info;
5126         struct reserve_ticket ticket;
5127         u64 used;
5128         int ret = 0;
5129
5130         ASSERT(orig_bytes);
5131         ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL);
5132
5133         spin_lock(&space_info->lock);
5134         ret = -ENOSPC;
5135         used = space_info->bytes_used + space_info->bytes_reserved +
5136                 space_info->bytes_pinned + space_info->bytes_readonly +
5137                 space_info->bytes_may_use;
5138
5139         /*
5140          * If we have enough space then hooray, make our reservation and carry
5141          * on.  If not see if we can overcommit, and if we can, hooray carry on.
5142          * If not things get more complicated.
5143          */
5144         if (used + orig_bytes <= space_info->total_bytes) {
5145                 space_info->bytes_may_use += orig_bytes;
5146                 trace_btrfs_space_reservation(fs_info, "space_info",
5147                                               space_info->flags, orig_bytes, 1);
5148                 ret = 0;
5149         } else if (can_overcommit(root, space_info, orig_bytes, flush)) {
5150                 space_info->bytes_may_use += orig_bytes;
5151                 trace_btrfs_space_reservation(fs_info, "space_info",
5152                                               space_info->flags, orig_bytes, 1);
5153                 ret = 0;
5154         }
5155
5156         /*
5157          * If we couldn't make a reservation then setup our reservation ticket
5158          * and kick the async worker if it's not already running.
5159          *
5160          * If we are a priority flusher then we just need to add our ticket to
5161          * the list and we will do our own flushing further down.
5162          */
5163         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
5164                 ticket.bytes = orig_bytes;
5165                 ticket.error = 0;
5166                 init_waitqueue_head(&ticket.wait);
5167                 if (flush == BTRFS_RESERVE_FLUSH_ALL) {
5168                         list_add_tail(&ticket.list, &space_info->tickets);
5169                         if (!space_info->flush) {
5170                                 space_info->flush = 1;
5171                                 trace_btrfs_trigger_flush(fs_info,
5172                                                           space_info->flags,
5173                                                           orig_bytes, flush,
5174                                                           "enospc");
5175                                 queue_work(system_unbound_wq,
5176                                            &root->fs_info->async_reclaim_work);
5177                         }
5178                 } else {
5179                         list_add_tail(&ticket.list,
5180                                       &space_info->priority_tickets);
5181                 }
5182         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
5183                 used += orig_bytes;
5184                 /*
5185                  * We will do the space reservation dance during log replay,
5186                  * which means we won't have fs_info->fs_root set, so don't do
5187                  * the async reclaim as we will panic.
5188                  */
5189                 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) &&
5190                     need_do_async_reclaim(space_info, root, used) &&
5191                     !work_busy(&fs_info->async_reclaim_work)) {
5192                         trace_btrfs_trigger_flush(fs_info, space_info->flags,
5193                                                   orig_bytes, flush, "preempt");
5194                         queue_work(system_unbound_wq,
5195                                    &fs_info->async_reclaim_work);
5196                 }
5197         }
5198         spin_unlock(&space_info->lock);
5199         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
5200                 return ret;
5201
5202         if (flush == BTRFS_RESERVE_FLUSH_ALL)
5203                 return wait_reserve_ticket(fs_info, space_info, &ticket,
5204                                            orig_bytes);
5205
5206         ret = 0;
5207         priority_reclaim_metadata_space(fs_info, space_info, &ticket);
5208         spin_lock(&space_info->lock);
5209         if (ticket.bytes) {
5210                 if (ticket.bytes < orig_bytes) {
5211                         u64 num_bytes = orig_bytes - ticket.bytes;
5212                         space_info->bytes_may_use -= num_bytes;
5213                         trace_btrfs_space_reservation(fs_info, "space_info",
5214                                                       space_info->flags,
5215                                                       num_bytes, 0);
5216
5217                 }
5218                 list_del_init(&ticket.list);
5219                 ret = -ENOSPC;
5220         }
5221         spin_unlock(&space_info->lock);
5222         ASSERT(list_empty(&ticket.list));
5223         return ret;
5224 }
5225
5226 /**
5227  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5228  * @root - the root we're allocating for
5229  * @block_rsv - the block_rsv we're allocating for
5230  * @orig_bytes - the number of bytes we want
5231  * @flush - whether or not we can flush to make our reservation
5232  *
5233  * This will reserve orgi_bytes number of bytes from the space info associated
5234  * with the block_rsv.  If there is not enough space it will make an attempt to
5235  * flush out space to make room.  It will do this by flushing delalloc if
5236  * possible or committing the transaction.  If flush is 0 then no attempts to
5237  * regain reservations will be made and this will fail if there is not enough
5238  * space already.
5239  */
5240 static int reserve_metadata_bytes(struct btrfs_root *root,
5241                                   struct btrfs_block_rsv *block_rsv,
5242                                   u64 orig_bytes,
5243                                   enum btrfs_reserve_flush_enum flush)
5244 {
5245         struct btrfs_fs_info *fs_info = root->fs_info;
5246         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5247         int ret;
5248
5249         ret = __reserve_metadata_bytes(root, block_rsv->space_info, orig_bytes,
5250                                        flush);
5251         if (ret == -ENOSPC &&
5252             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
5253                 if (block_rsv != global_rsv &&
5254                     !block_rsv_use_bytes(global_rsv, orig_bytes))
5255                         ret = 0;
5256         }
5257         if (ret == -ENOSPC)
5258                 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
5259                                               block_rsv->space_info->flags,
5260                                               orig_bytes, 1);
5261         return ret;
5262 }
5263
5264 static struct btrfs_block_rsv *get_block_rsv(
5265                                         const struct btrfs_trans_handle *trans,
5266                                         const struct btrfs_root *root)
5267 {
5268         struct btrfs_fs_info *fs_info = root->fs_info;
5269         struct btrfs_block_rsv *block_rsv = NULL;
5270
5271         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
5272             (root == fs_info->csum_root && trans->adding_csums) ||
5273             (root == fs_info->uuid_root))
5274                 block_rsv = trans->block_rsv;
5275
5276         if (!block_rsv)
5277                 block_rsv = root->block_rsv;
5278
5279         if (!block_rsv)
5280                 block_rsv = &fs_info->empty_block_rsv;
5281
5282         return block_rsv;
5283 }
5284
5285 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
5286                                u64 num_bytes)
5287 {
5288         int ret = -ENOSPC;
5289         spin_lock(&block_rsv->lock);
5290         if (block_rsv->reserved >= num_bytes) {
5291                 block_rsv->reserved -= num_bytes;
5292                 if (block_rsv->reserved < block_rsv->size)
5293                         block_rsv->full = 0;
5294                 ret = 0;
5295         }
5296         spin_unlock(&block_rsv->lock);
5297         return ret;
5298 }
5299
5300 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
5301                                 u64 num_bytes, int update_size)
5302 {
5303         spin_lock(&block_rsv->lock);
5304         block_rsv->reserved += num_bytes;
5305         if (update_size)
5306                 block_rsv->size += num_bytes;
5307         else if (block_rsv->reserved >= block_rsv->size)
5308                 block_rsv->full = 1;
5309         spin_unlock(&block_rsv->lock);
5310 }
5311
5312 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
5313                              struct btrfs_block_rsv *dest, u64 num_bytes,
5314                              int min_factor)
5315 {
5316         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5317         u64 min_bytes;
5318
5319         if (global_rsv->space_info != dest->space_info)
5320                 return -ENOSPC;
5321
5322         spin_lock(&global_rsv->lock);
5323         min_bytes = div_factor(global_rsv->size, min_factor);
5324         if (global_rsv->reserved < min_bytes + num_bytes) {
5325                 spin_unlock(&global_rsv->lock);
5326                 return -ENOSPC;
5327         }
5328         global_rsv->reserved -= num_bytes;
5329         if (global_rsv->reserved < global_rsv->size)
5330                 global_rsv->full = 0;
5331         spin_unlock(&global_rsv->lock);
5332
5333         block_rsv_add_bytes(dest, num_bytes, 1);
5334         return 0;
5335 }
5336
5337 /*
5338  * This is for space we already have accounted in space_info->bytes_may_use, so
5339  * basically when we're returning space from block_rsv's.
5340  */
5341 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
5342                                      struct btrfs_space_info *space_info,
5343                                      u64 num_bytes)
5344 {
5345         struct reserve_ticket *ticket;
5346         struct list_head *head;
5347         u64 used;
5348         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH;
5349         bool check_overcommit = false;
5350
5351         spin_lock(&space_info->lock);
5352         head = &space_info->priority_tickets;
5353
5354         /*
5355          * If we are over our limit then we need to check and see if we can
5356          * overcommit, and if we can't then we just need to free up our space
5357          * and not satisfy any requests.
5358          */
5359         used = space_info->bytes_used + space_info->bytes_reserved +
5360                 space_info->bytes_pinned + space_info->bytes_readonly +
5361                 space_info->bytes_may_use;
5362         if (used - num_bytes >= space_info->total_bytes)
5363                 check_overcommit = true;
5364 again:
5365         while (!list_empty(head) && num_bytes) {
5366                 ticket = list_first_entry(head, struct reserve_ticket,
5367                                           list);
5368                 /*
5369                  * We use 0 bytes because this space is already reserved, so
5370                  * adding the ticket space would be a double count.
5371                  */
5372                 if (check_overcommit &&
5373                     !can_overcommit(fs_info->extent_root, space_info, 0,
5374                                     flush))
5375                         break;
5376                 if (num_bytes >= ticket->bytes) {
5377                         list_del_init(&ticket->list);
5378                         num_bytes -= ticket->bytes;
5379                         ticket->bytes = 0;
5380                         space_info->tickets_id++;
5381                         wake_up(&ticket->wait);
5382                 } else {
5383                         ticket->bytes -= num_bytes;
5384                         num_bytes = 0;
5385                 }
5386         }
5387
5388         if (num_bytes && head == &space_info->priority_tickets) {
5389                 head = &space_info->tickets;
5390                 flush = BTRFS_RESERVE_FLUSH_ALL;
5391                 goto again;
5392         }
5393         space_info->bytes_may_use -= num_bytes;
5394         trace_btrfs_space_reservation(fs_info, "space_info",
5395                                       space_info->flags, num_bytes, 0);
5396         spin_unlock(&space_info->lock);
5397 }
5398
5399 /*
5400  * This is for newly allocated space that isn't accounted in
5401  * space_info->bytes_may_use yet.  So if we allocate a chunk or unpin an extent
5402  * we use this helper.
5403  */
5404 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
5405                                      struct btrfs_space_info *space_info,
5406                                      u64 num_bytes)
5407 {
5408         struct reserve_ticket *ticket;
5409         struct list_head *head = &space_info->priority_tickets;
5410
5411 again:
5412         while (!list_empty(head) && num_bytes) {
5413                 ticket = list_first_entry(head, struct reserve_ticket,
5414                                           list);
5415                 if (num_bytes >= ticket->bytes) {
5416                         trace_btrfs_space_reservation(fs_info, "space_info",
5417                                                       space_info->flags,
5418                                                       ticket->bytes, 1);
5419                         list_del_init(&ticket->list);
5420                         num_bytes -= ticket->bytes;
5421                         space_info->bytes_may_use += ticket->bytes;
5422                         ticket->bytes = 0;
5423                         space_info->tickets_id++;
5424                         wake_up(&ticket->wait);
5425                 } else {
5426                         trace_btrfs_space_reservation(fs_info, "space_info",
5427                                                       space_info->flags,
5428                                                       num_bytes, 1);
5429                         space_info->bytes_may_use += num_bytes;
5430                         ticket->bytes -= num_bytes;
5431                         num_bytes = 0;
5432                 }
5433         }
5434
5435         if (num_bytes && head == &space_info->priority_tickets) {
5436                 head = &space_info->tickets;
5437                 goto again;
5438         }
5439 }
5440
5441 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
5442                                     struct btrfs_block_rsv *block_rsv,
5443                                     struct btrfs_block_rsv *dest, u64 num_bytes)
5444 {
5445         struct btrfs_space_info *space_info = block_rsv->space_info;
5446
5447         spin_lock(&block_rsv->lock);
5448         if (num_bytes == (u64)-1)
5449                 num_bytes = block_rsv->size;
5450         block_rsv->size -= num_bytes;
5451         if (block_rsv->reserved >= block_rsv->size) {
5452                 num_bytes = block_rsv->reserved - block_rsv->size;
5453                 block_rsv->reserved = block_rsv->size;
5454                 block_rsv->full = 1;
5455         } else {
5456                 num_bytes = 0;
5457         }
5458         spin_unlock(&block_rsv->lock);
5459
5460         if (num_bytes > 0) {
5461                 if (dest) {
5462                         spin_lock(&dest->lock);
5463                         if (!dest->full) {
5464                                 u64 bytes_to_add;
5465
5466                                 bytes_to_add = dest->size - dest->reserved;
5467                                 bytes_to_add = min(num_bytes, bytes_to_add);
5468                                 dest->reserved += bytes_to_add;
5469                                 if (dest->reserved >= dest->size)
5470                                         dest->full = 1;
5471                                 num_bytes -= bytes_to_add;
5472                         }
5473                         spin_unlock(&dest->lock);
5474                 }
5475                 if (num_bytes)
5476                         space_info_add_old_bytes(fs_info, space_info,
5477                                                  num_bytes);
5478         }
5479 }
5480
5481 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src,
5482                             struct btrfs_block_rsv *dst, u64 num_bytes,
5483                             int update_size)
5484 {
5485         int ret;
5486
5487         ret = block_rsv_use_bytes(src, num_bytes);
5488         if (ret)
5489                 return ret;
5490
5491         block_rsv_add_bytes(dst, num_bytes, update_size);
5492         return 0;
5493 }
5494
5495 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
5496 {
5497         memset(rsv, 0, sizeof(*rsv));
5498         spin_lock_init(&rsv->lock);
5499         rsv->type = type;
5500 }
5501
5502 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_fs_info *fs_info,
5503                                               unsigned short type)
5504 {
5505         struct btrfs_block_rsv *block_rsv;
5506
5507         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
5508         if (!block_rsv)
5509                 return NULL;
5510
5511         btrfs_init_block_rsv(block_rsv, type);
5512         block_rsv->space_info = __find_space_info(fs_info,
5513                                                   BTRFS_BLOCK_GROUP_METADATA);
5514         return block_rsv;
5515 }
5516
5517 void btrfs_free_block_rsv(struct btrfs_fs_info *fs_info,
5518                           struct btrfs_block_rsv *rsv)
5519 {
5520         if (!rsv)
5521                 return;
5522         btrfs_block_rsv_release(fs_info, rsv, (u64)-1);
5523         kfree(rsv);
5524 }
5525
5526 void __btrfs_free_block_rsv(struct btrfs_block_rsv *rsv)
5527 {
5528         kfree(rsv);
5529 }
5530
5531 int btrfs_block_rsv_add(struct btrfs_root *root,
5532                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
5533                         enum btrfs_reserve_flush_enum flush)
5534 {
5535         int ret;
5536
5537         if (num_bytes == 0)
5538                 return 0;
5539
5540         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5541         if (!ret) {
5542                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
5543                 return 0;
5544         }
5545
5546         return ret;
5547 }
5548
5549 int btrfs_block_rsv_check(struct btrfs_block_rsv *block_rsv, int min_factor)
5550 {
5551         u64 num_bytes = 0;
5552         int ret = -ENOSPC;
5553
5554         if (!block_rsv)
5555                 return 0;
5556
5557         spin_lock(&block_rsv->lock);
5558         num_bytes = div_factor(block_rsv->size, min_factor);
5559         if (block_rsv->reserved >= num_bytes)
5560                 ret = 0;
5561         spin_unlock(&block_rsv->lock);
5562
5563         return ret;
5564 }
5565
5566 int btrfs_block_rsv_refill(struct btrfs_root *root,
5567                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
5568                            enum btrfs_reserve_flush_enum flush)
5569 {
5570         u64 num_bytes = 0;
5571         int ret = -ENOSPC;
5572
5573         if (!block_rsv)
5574                 return 0;
5575
5576         spin_lock(&block_rsv->lock);
5577         num_bytes = min_reserved;
5578         if (block_rsv->reserved >= num_bytes)
5579                 ret = 0;
5580         else
5581                 num_bytes -= block_rsv->reserved;
5582         spin_unlock(&block_rsv->lock);
5583
5584         if (!ret)
5585                 return 0;
5586
5587         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5588         if (!ret) {
5589                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
5590                 return 0;
5591         }
5592
5593         return ret;
5594 }
5595
5596 void btrfs_block_rsv_release(struct btrfs_fs_info *fs_info,
5597                              struct btrfs_block_rsv *block_rsv,
5598                              u64 num_bytes)
5599 {
5600         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5601
5602         if (global_rsv == block_rsv ||
5603             block_rsv->space_info != global_rsv->space_info)
5604                 global_rsv = NULL;
5605         block_rsv_release_bytes(fs_info, block_rsv, global_rsv, num_bytes);
5606 }
5607
5608 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
5609 {
5610         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
5611         struct btrfs_space_info *sinfo = block_rsv->space_info;
5612         u64 num_bytes;
5613
5614         /*
5615          * The global block rsv is based on the size of the extent tree, the
5616          * checksum tree and the root tree.  If the fs is empty we want to set
5617          * it to a minimal amount for safety.
5618          */
5619         num_bytes = btrfs_root_used(&fs_info->extent_root->root_item) +
5620                 btrfs_root_used(&fs_info->csum_root->root_item) +
5621                 btrfs_root_used(&fs_info->tree_root->root_item);
5622         num_bytes = max_t(u64, num_bytes, SZ_16M);
5623
5624         spin_lock(&sinfo->lock);
5625         spin_lock(&block_rsv->lock);
5626
5627         block_rsv->size = min_t(u64, num_bytes, SZ_512M);
5628
5629         if (block_rsv->reserved < block_rsv->size) {
5630                 num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
5631                         sinfo->bytes_reserved + sinfo->bytes_readonly +
5632                         sinfo->bytes_may_use;
5633                 if (sinfo->total_bytes > num_bytes) {
5634                         num_bytes = sinfo->total_bytes - num_bytes;
5635                         num_bytes = min(num_bytes,
5636                                         block_rsv->size - block_rsv->reserved);
5637                         block_rsv->reserved += num_bytes;
5638                         sinfo->bytes_may_use += num_bytes;
5639                         trace_btrfs_space_reservation(fs_info, "space_info",
5640                                                       sinfo->flags, num_bytes,
5641                                                       1);
5642                 }
5643         } else if (block_rsv->reserved > block_rsv->size) {
5644                 num_bytes = block_rsv->reserved - block_rsv->size;
5645                 sinfo->bytes_may_use -= num_bytes;
5646                 trace_btrfs_space_reservation(fs_info, "space_info",
5647                                       sinfo->flags, num_bytes, 0);
5648                 block_rsv->reserved = block_rsv->size;
5649         }
5650
5651         if (block_rsv->reserved == block_rsv->size)
5652                 block_rsv->full = 1;
5653         else
5654                 block_rsv->full = 0;
5655
5656         spin_unlock(&block_rsv->lock);
5657         spin_unlock(&sinfo->lock);
5658 }
5659
5660 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
5661 {
5662         struct btrfs_space_info *space_info;
5663
5664         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
5665         fs_info->chunk_block_rsv.space_info = space_info;
5666
5667         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5668         fs_info->global_block_rsv.space_info = space_info;
5669         fs_info->delalloc_block_rsv.space_info = space_info;
5670         fs_info->trans_block_rsv.space_info = space_info;
5671         fs_info->empty_block_rsv.space_info = space_info;
5672         fs_info->delayed_block_rsv.space_info = space_info;
5673
5674         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
5675         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
5676         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
5677         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
5678         if (fs_info->quota_root)
5679                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
5680         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
5681
5682         update_global_block_rsv(fs_info);
5683 }
5684
5685 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
5686 {
5687         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
5688                                 (u64)-1);
5689         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
5690         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
5691         WARN_ON(fs_info->trans_block_rsv.size > 0);
5692         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
5693         WARN_ON(fs_info->chunk_block_rsv.size > 0);
5694         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
5695         WARN_ON(fs_info->delayed_block_rsv.size > 0);
5696         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
5697 }
5698
5699 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
5700                                   struct btrfs_fs_info *fs_info)
5701 {
5702         if (!trans->block_rsv)
5703                 return;
5704
5705         if (!trans->bytes_reserved)
5706                 return;
5707
5708         trace_btrfs_space_reservation(fs_info, "transaction",
5709                                       trans->transid, trans->bytes_reserved, 0);
5710         btrfs_block_rsv_release(fs_info, trans->block_rsv,
5711                                 trans->bytes_reserved);
5712         trans->bytes_reserved = 0;
5713 }
5714
5715 /*
5716  * To be called after all the new block groups attached to the transaction
5717  * handle have been created (btrfs_create_pending_block_groups()).
5718  */
5719 void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans)
5720 {
5721         struct btrfs_fs_info *fs_info = trans->fs_info;
5722
5723         if (!trans->chunk_bytes_reserved)
5724                 return;
5725
5726         WARN_ON_ONCE(!list_empty(&trans->new_bgs));
5727
5728         block_rsv_release_bytes(fs_info, &fs_info->chunk_block_rsv, NULL,
5729                                 trans->chunk_bytes_reserved);
5730         trans->chunk_bytes_reserved = 0;
5731 }
5732
5733 /* Can only return 0 or -ENOSPC */
5734 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
5735                                   struct inode *inode)
5736 {
5737         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5738         struct btrfs_root *root = BTRFS_I(inode)->root;
5739         /*
5740          * We always use trans->block_rsv here as we will have reserved space
5741          * for our orphan when starting the transaction, using get_block_rsv()
5742          * here will sometimes make us choose the wrong block rsv as we could be
5743          * doing a reloc inode for a non refcounted root.
5744          */
5745         struct btrfs_block_rsv *src_rsv = trans->block_rsv;
5746         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
5747
5748         /*
5749          * We need to hold space in order to delete our orphan item once we've
5750          * added it, so this takes the reservation so we can release it later
5751          * when we are truly done with the orphan item.
5752          */
5753         u64 num_bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
5754
5755         trace_btrfs_space_reservation(fs_info, "orphan",
5756                                       btrfs_ino(inode), num_bytes, 1);
5757         return btrfs_block_rsv_migrate(src_rsv, dst_rsv, num_bytes, 1);
5758 }
5759
5760 void btrfs_orphan_release_metadata(struct inode *inode)
5761 {
5762         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5763         struct btrfs_root *root = BTRFS_I(inode)->root;
5764         u64 num_bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
5765
5766         trace_btrfs_space_reservation(fs_info, "orphan",
5767                                       btrfs_ino(inode), num_bytes, 0);
5768         btrfs_block_rsv_release(fs_info, root->orphan_block_rsv, num_bytes);
5769 }
5770
5771 /*
5772  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
5773  * root: the root of the parent directory
5774  * rsv: block reservation
5775  * items: the number of items that we need do reservation
5776  * qgroup_reserved: used to return the reserved size in qgroup
5777  *
5778  * This function is used to reserve the space for snapshot/subvolume
5779  * creation and deletion. Those operations are different with the
5780  * common file/directory operations, they change two fs/file trees
5781  * and root tree, the number of items that the qgroup reserves is
5782  * different with the free space reservation. So we can not use
5783  * the space reservation mechanism in start_transaction().
5784  */
5785 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
5786                                      struct btrfs_block_rsv *rsv,
5787                                      int items,
5788                                      u64 *qgroup_reserved,
5789                                      bool use_global_rsv)
5790 {
5791         u64 num_bytes;
5792         int ret;
5793         struct btrfs_fs_info *fs_info = root->fs_info;
5794         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5795
5796         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)) {
5797                 /* One for parent inode, two for dir entries */
5798                 num_bytes = 3 * fs_info->nodesize;
5799                 ret = btrfs_qgroup_reserve_meta(root, num_bytes);
5800                 if (ret)
5801                         return ret;
5802         } else {
5803                 num_bytes = 0;
5804         }
5805
5806         *qgroup_reserved = num_bytes;
5807
5808         num_bytes = btrfs_calc_trans_metadata_size(fs_info, items);
5809         rsv->space_info = __find_space_info(fs_info,
5810                                             BTRFS_BLOCK_GROUP_METADATA);
5811         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
5812                                   BTRFS_RESERVE_FLUSH_ALL);
5813
5814         if (ret == -ENOSPC && use_global_rsv)
5815                 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes, 1);
5816
5817         if (ret && *qgroup_reserved)
5818                 btrfs_qgroup_free_meta(root, *qgroup_reserved);
5819
5820         return ret;
5821 }
5822
5823 void btrfs_subvolume_release_metadata(struct btrfs_fs_info *fs_info,
5824                                       struct btrfs_block_rsv *rsv,
5825                                       u64 qgroup_reserved)
5826 {
5827         btrfs_block_rsv_release(fs_info, rsv, (u64)-1);
5828 }
5829
5830 /**
5831  * drop_outstanding_extent - drop an outstanding extent
5832  * @inode: the inode we're dropping the extent for
5833  * @num_bytes: the number of bytes we're releasing.
5834  *
5835  * This is called when we are freeing up an outstanding extent, either called
5836  * after an error or after an extent is written.  This will return the number of
5837  * reserved extents that need to be freed.  This must be called with
5838  * BTRFS_I(inode)->lock held.
5839  */
5840 static unsigned drop_outstanding_extent(struct inode *inode, u64 num_bytes)
5841 {
5842         unsigned drop_inode_space = 0;
5843         unsigned dropped_extents = 0;
5844         unsigned num_extents = 0;
5845
5846         num_extents = (unsigned)div64_u64(num_bytes +
5847                                           BTRFS_MAX_EXTENT_SIZE - 1,
5848                                           BTRFS_MAX_EXTENT_SIZE);
5849         ASSERT(num_extents);
5850         ASSERT(BTRFS_I(inode)->outstanding_extents >= num_extents);
5851         BTRFS_I(inode)->outstanding_extents -= num_extents;
5852
5853         if (BTRFS_I(inode)->outstanding_extents == 0 &&
5854             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5855                                &BTRFS_I(inode)->runtime_flags))
5856                 drop_inode_space = 1;
5857
5858         /*
5859          * If we have more or the same amount of outstanding extents than we have
5860          * reserved then we need to leave the reserved extents count alone.
5861          */
5862         if (BTRFS_I(inode)->outstanding_extents >=
5863             BTRFS_I(inode)->reserved_extents)
5864                 return drop_inode_space;
5865
5866         dropped_extents = BTRFS_I(inode)->reserved_extents -
5867                 BTRFS_I(inode)->outstanding_extents;
5868         BTRFS_I(inode)->reserved_extents -= dropped_extents;
5869         return dropped_extents + drop_inode_space;
5870 }
5871
5872 /**
5873  * calc_csum_metadata_size - return the amount of metadata space that must be
5874  *      reserved/freed for the given bytes.
5875  * @inode: the inode we're manipulating
5876  * @num_bytes: the number of bytes in question
5877  * @reserve: 1 if we are reserving space, 0 if we are freeing space
5878  *
5879  * This adjusts the number of csum_bytes in the inode and then returns the
5880  * correct amount of metadata that must either be reserved or freed.  We
5881  * calculate how many checksums we can fit into one leaf and then divide the
5882  * number of bytes that will need to be checksumed by this value to figure out
5883  * how many checksums will be required.  If we are adding bytes then the number
5884  * may go up and we will return the number of additional bytes that must be
5885  * reserved.  If it is going down we will return the number of bytes that must
5886  * be freed.
5887  *
5888  * This must be called with BTRFS_I(inode)->lock held.
5889  */
5890 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
5891                                    int reserve)
5892 {
5893         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5894         u64 old_csums, num_csums;
5895
5896         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
5897             BTRFS_I(inode)->csum_bytes == 0)
5898                 return 0;
5899
5900         old_csums = btrfs_csum_bytes_to_leaves(fs_info,
5901                                                BTRFS_I(inode)->csum_bytes);
5902         if (reserve)
5903                 BTRFS_I(inode)->csum_bytes += num_bytes;
5904         else
5905                 BTRFS_I(inode)->csum_bytes -= num_bytes;
5906         num_csums = btrfs_csum_bytes_to_leaves(fs_info,
5907                                                BTRFS_I(inode)->csum_bytes);
5908
5909         /* No change, no need to reserve more */
5910         if (old_csums == num_csums)
5911                 return 0;
5912
5913         if (reserve)
5914                 return btrfs_calc_trans_metadata_size(fs_info,
5915                                                       num_csums - old_csums);
5916
5917         return btrfs_calc_trans_metadata_size(fs_info, old_csums - num_csums);
5918 }
5919
5920 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
5921 {
5922         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5923         struct btrfs_root *root = BTRFS_I(inode)->root;
5924         struct btrfs_block_rsv *block_rsv = &fs_info->delalloc_block_rsv;
5925         u64 to_reserve = 0;
5926         u64 csum_bytes;
5927         unsigned nr_extents = 0;
5928         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
5929         int ret = 0;
5930         bool delalloc_lock = true;
5931         u64 to_free = 0;
5932         unsigned dropped;
5933         bool release_extra = false;
5934
5935         /* If we are a free space inode we need to not flush since we will be in
5936          * the middle of a transaction commit.  We also don't need the delalloc
5937          * mutex since we won't race with anybody.  We need this mostly to make
5938          * lockdep shut its filthy mouth.
5939          *
5940          * If we have a transaction open (can happen if we call truncate_block
5941          * from truncate), then we need FLUSH_LIMIT so we don't deadlock.
5942          */
5943         if (btrfs_is_free_space_inode(inode)) {
5944                 flush = BTRFS_RESERVE_NO_FLUSH;
5945                 delalloc_lock = false;
5946         } else if (current->journal_info) {
5947                 flush = BTRFS_RESERVE_FLUSH_LIMIT;
5948         }
5949
5950         if (flush != BTRFS_RESERVE_NO_FLUSH &&
5951             btrfs_transaction_in_commit(fs_info))
5952                 schedule_timeout(1);
5953
5954         if (delalloc_lock)
5955                 mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
5956
5957         num_bytes = ALIGN(num_bytes, fs_info->sectorsize);
5958
5959         spin_lock(&BTRFS_I(inode)->lock);
5960         nr_extents = (unsigned)div64_u64(num_bytes +
5961                                          BTRFS_MAX_EXTENT_SIZE - 1,
5962                                          BTRFS_MAX_EXTENT_SIZE);
5963         BTRFS_I(inode)->outstanding_extents += nr_extents;
5964
5965         nr_extents = 0;
5966         if (BTRFS_I(inode)->outstanding_extents >
5967             BTRFS_I(inode)->reserved_extents)
5968                 nr_extents += BTRFS_I(inode)->outstanding_extents -
5969                         BTRFS_I(inode)->reserved_extents;
5970
5971         /* We always want to reserve a slot for updating the inode. */
5972         to_reserve = btrfs_calc_trans_metadata_size(fs_info, nr_extents + 1);
5973         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
5974         csum_bytes = BTRFS_I(inode)->csum_bytes;
5975         spin_unlock(&BTRFS_I(inode)->lock);
5976
5977         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)) {
5978                 ret = btrfs_qgroup_reserve_meta(root,
5979                                 nr_extents * fs_info->nodesize);
5980                 if (ret)
5981                         goto out_fail;
5982         }
5983
5984         ret = btrfs_block_rsv_add(root, block_rsv, to_reserve, flush);
5985         if (unlikely(ret)) {
5986                 btrfs_qgroup_free_meta(root,
5987                                        nr_extents * fs_info->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(fs_info, 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(fs_info, "delalloc",
6005                                               btrfs_ino(inode), to_reserve, 1);
6006         if (release_extra)
6007                 btrfs_block_rsv_release(fs_info, block_rsv,
6008                                 btrfs_calc_trans_metadata_size(fs_info, 1));
6009         return 0;
6010
6011 out_fail:
6012         spin_lock(&BTRFS_I(inode)->lock);
6013         dropped = drop_outstanding_extent(inode, num_bytes);
6014         /*
6015          * If the inodes csum_bytes is the same as the original
6016          * csum_bytes then we know we haven't raced with any free()ers
6017          * so we can just reduce our inodes csum bytes and carry on.
6018          */
6019         if (BTRFS_I(inode)->csum_bytes == csum_bytes) {
6020                 calc_csum_metadata_size(inode, num_bytes, 0);
6021         } else {
6022                 u64 orig_csum_bytes = BTRFS_I(inode)->csum_bytes;
6023                 u64 bytes;
6024
6025                 /*
6026                  * This is tricky, but first we need to figure out how much we
6027                  * freed from any free-ers that occurred during this
6028                  * reservation, so we reset ->csum_bytes to the csum_bytes
6029                  * before we dropped our lock, and then call the free for the
6030                  * number of bytes that were freed while we were trying our
6031                  * reservation.
6032                  */
6033                 bytes = csum_bytes - BTRFS_I(inode)->csum_bytes;
6034                 BTRFS_I(inode)->csum_bytes = csum_bytes;
6035                 to_free = calc_csum_metadata_size(inode, bytes, 0);
6036
6037
6038                 /*
6039                  * Now we need to see how much we would have freed had we not
6040                  * been making this reservation and our ->csum_bytes were not
6041                  * artificially inflated.
6042                  */
6043                 BTRFS_I(inode)->csum_bytes = csum_bytes - num_bytes;
6044                 bytes = csum_bytes - orig_csum_bytes;
6045                 bytes = calc_csum_metadata_size(inode, bytes, 0);
6046
6047                 /*
6048                  * Now reset ->csum_bytes to what it should be.  If bytes is
6049                  * more than to_free then we would have freed more space had we
6050                  * not had an artificially high ->csum_bytes, so we need to free
6051                  * the remainder.  If bytes is the same or less then we don't
6052                  * need to do anything, the other free-ers did the correct
6053                  * thing.
6054                  */
6055                 BTRFS_I(inode)->csum_bytes = orig_csum_bytes - num_bytes;
6056                 if (bytes > to_free)
6057                         to_free = bytes - to_free;
6058                 else
6059                         to_free = 0;
6060         }
6061         spin_unlock(&BTRFS_I(inode)->lock);
6062         if (dropped)
6063                 to_free += btrfs_calc_trans_metadata_size(fs_info, dropped);
6064
6065         if (to_free) {
6066                 btrfs_block_rsv_release(fs_info, block_rsv, to_free);
6067                 trace_btrfs_space_reservation(fs_info, "delalloc",
6068                                               btrfs_ino(inode), to_free, 0);
6069         }
6070         if (delalloc_lock)
6071                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
6072         return ret;
6073 }
6074
6075 /**
6076  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
6077  * @inode: the inode to release the reservation for
6078  * @num_bytes: the number of bytes we're releasing
6079  *
6080  * This will release the metadata reservation for an inode.  This can be called
6081  * once we complete IO for a given set of bytes to release their metadata
6082  * reservations.
6083  */
6084 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
6085 {
6086         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6087         u64 to_free = 0;
6088         unsigned dropped;
6089
6090         num_bytes = ALIGN(num_bytes, fs_info->sectorsize);
6091         spin_lock(&BTRFS_I(inode)->lock);
6092         dropped = drop_outstanding_extent(inode, num_bytes);
6093
6094         if (num_bytes)
6095                 to_free = calc_csum_metadata_size(inode, num_bytes, 0);
6096         spin_unlock(&BTRFS_I(inode)->lock);
6097         if (dropped > 0)
6098                 to_free += btrfs_calc_trans_metadata_size(fs_info, dropped);
6099
6100         if (btrfs_is_testing(fs_info))
6101                 return;
6102
6103         trace_btrfs_space_reservation(fs_info, "delalloc",
6104                                       btrfs_ino(inode), to_free, 0);
6105
6106         btrfs_block_rsv_release(fs_info, &fs_info->delalloc_block_rsv, to_free);
6107 }
6108
6109 /**
6110  * btrfs_delalloc_reserve_space - reserve data and metadata space for
6111  * delalloc
6112  * @inode: inode we're writing to
6113  * @start: start range we are writing to
6114  * @len: how long the range we are writing to
6115  *
6116  * This will do the following things
6117  *
6118  * o reserve space in data space info for num bytes
6119  *   and reserve precious corresponding qgroup space
6120  *   (Done in check_data_free_space)
6121  *
6122  * o reserve space for metadata space, based on the number of outstanding
6123  *   extents and how much csums will be needed
6124  *   also reserve metadata space in a per root over-reserve method.
6125  * o add to the inodes->delalloc_bytes
6126  * o add it to the fs_info's delalloc inodes list.
6127  *   (Above 3 all done in delalloc_reserve_metadata)
6128  *
6129  * Return 0 for success
6130  * Return <0 for error(-ENOSPC or -EQUOT)
6131  */
6132 int btrfs_delalloc_reserve_space(struct inode *inode, u64 start, u64 len)
6133 {
6134         int ret;
6135
6136         ret = btrfs_check_data_free_space(inode, start, len);
6137         if (ret < 0)
6138                 return ret;
6139         ret = btrfs_delalloc_reserve_metadata(inode, len);
6140         if (ret < 0)
6141                 btrfs_free_reserved_data_space(inode, start, len);
6142         return ret;
6143 }
6144
6145 /**
6146  * btrfs_delalloc_release_space - release data and metadata space for delalloc
6147  * @inode: inode we're releasing space for
6148  * @start: start position of the space already reserved
6149  * @len: the len of the space already reserved
6150  *
6151  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
6152  * called in the case that we don't need the metadata AND data reservations
6153  * anymore.  So if there is an error or we insert an inline extent.
6154  *
6155  * This function will release the metadata space that was not used and will
6156  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
6157  * list if there are no delalloc bytes left.
6158  * Also it will handle the qgroup reserved space.
6159  */
6160 void btrfs_delalloc_release_space(struct inode *inode, u64 start, u64 len)
6161 {
6162         btrfs_delalloc_release_metadata(inode, len);
6163         btrfs_free_reserved_data_space(inode, start, len);
6164 }
6165
6166 static int update_block_group(struct btrfs_trans_handle *trans,
6167                               struct btrfs_fs_info *info, u64 bytenr,
6168                               u64 num_bytes, int alloc)
6169 {
6170         struct btrfs_block_group_cache *cache = NULL;
6171         u64 total = num_bytes;
6172         u64 old_val;
6173         u64 byte_in_group;
6174         int factor;
6175
6176         /* block accounting for super block */
6177         spin_lock(&info->delalloc_root_lock);
6178         old_val = btrfs_super_bytes_used(info->super_copy);
6179         if (alloc)
6180                 old_val += num_bytes;
6181         else
6182                 old_val -= num_bytes;
6183         btrfs_set_super_bytes_used(info->super_copy, old_val);
6184         spin_unlock(&info->delalloc_root_lock);
6185
6186         while (total) {
6187                 cache = btrfs_lookup_block_group(info, bytenr);
6188                 if (!cache)
6189                         return -ENOENT;
6190                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
6191                                     BTRFS_BLOCK_GROUP_RAID1 |
6192                                     BTRFS_BLOCK_GROUP_RAID10))
6193                         factor = 2;
6194                 else
6195                         factor = 1;
6196                 /*
6197                  * If this block group has free space cache written out, we
6198                  * need to make sure to load it if we are removing space.  This
6199                  * is because we need the unpinning stage to actually add the
6200                  * space back to the block group, otherwise we will leak space.
6201                  */
6202                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
6203                         cache_block_group(cache, 1);
6204
6205                 byte_in_group = bytenr - cache->key.objectid;
6206                 WARN_ON(byte_in_group > cache->key.offset);
6207
6208                 spin_lock(&cache->space_info->lock);
6209                 spin_lock(&cache->lock);
6210
6211                 if (btrfs_test_opt(info, SPACE_CACHE) &&
6212                     cache->disk_cache_state < BTRFS_DC_CLEAR)
6213                         cache->disk_cache_state = BTRFS_DC_CLEAR;
6214
6215                 old_val = btrfs_block_group_used(&cache->item);
6216                 num_bytes = min(total, cache->key.offset - byte_in_group);
6217                 if (alloc) {
6218                         old_val += num_bytes;
6219                         btrfs_set_block_group_used(&cache->item, old_val);
6220                         cache->reserved -= num_bytes;
6221                         cache->space_info->bytes_reserved -= num_bytes;
6222                         cache->space_info->bytes_used += num_bytes;
6223                         cache->space_info->disk_used += num_bytes * factor;
6224                         spin_unlock(&cache->lock);
6225                         spin_unlock(&cache->space_info->lock);
6226                 } else {
6227                         old_val -= num_bytes;
6228                         btrfs_set_block_group_used(&cache->item, old_val);
6229                         cache->pinned += num_bytes;
6230                         cache->space_info->bytes_pinned += num_bytes;
6231                         cache->space_info->bytes_used -= num_bytes;
6232                         cache->space_info->disk_used -= num_bytes * factor;
6233                         spin_unlock(&cache->lock);
6234                         spin_unlock(&cache->space_info->lock);
6235
6236                         trace_btrfs_space_reservation(info, "pinned",
6237                                                       cache->space_info->flags,
6238                                                       num_bytes, 1);
6239                         set_extent_dirty(info->pinned_extents,
6240                                          bytenr, bytenr + num_bytes - 1,
6241                                          GFP_NOFS | __GFP_NOFAIL);
6242                 }
6243
6244                 spin_lock(&trans->transaction->dirty_bgs_lock);
6245                 if (list_empty(&cache->dirty_list)) {
6246                         list_add_tail(&cache->dirty_list,
6247                                       &trans->transaction->dirty_bgs);
6248                                 trans->transaction->num_dirty_bgs++;
6249                         btrfs_get_block_group(cache);
6250                 }
6251                 spin_unlock(&trans->transaction->dirty_bgs_lock);
6252
6253                 /*
6254                  * No longer have used bytes in this block group, queue it for
6255                  * deletion. We do this after adding the block group to the
6256                  * dirty list to avoid races between cleaner kthread and space
6257                  * cache writeout.
6258                  */
6259                 if (!alloc && old_val == 0) {
6260                         spin_lock(&info->unused_bgs_lock);
6261                         if (list_empty(&cache->bg_list)) {
6262                                 btrfs_get_block_group(cache);
6263                                 list_add_tail(&cache->bg_list,
6264                                               &info->unused_bgs);
6265                         }
6266                         spin_unlock(&info->unused_bgs_lock);
6267                 }
6268
6269                 btrfs_put_block_group(cache);
6270                 total -= num_bytes;
6271                 bytenr += num_bytes;
6272         }
6273         return 0;
6274 }
6275
6276 static u64 first_logical_byte(struct btrfs_fs_info *fs_info, u64 search_start)
6277 {
6278         struct btrfs_block_group_cache *cache;
6279         u64 bytenr;
6280
6281         spin_lock(&fs_info->block_group_cache_lock);
6282         bytenr = fs_info->first_logical_byte;
6283         spin_unlock(&fs_info->block_group_cache_lock);
6284
6285         if (bytenr < (u64)-1)
6286                 return bytenr;
6287
6288         cache = btrfs_lookup_first_block_group(fs_info, search_start);
6289         if (!cache)
6290                 return 0;
6291
6292         bytenr = cache->key.objectid;
6293         btrfs_put_block_group(cache);
6294
6295         return bytenr;
6296 }
6297
6298 static int pin_down_extent(struct btrfs_fs_info *fs_info,
6299                            struct btrfs_block_group_cache *cache,
6300                            u64 bytenr, u64 num_bytes, int reserved)
6301 {
6302         spin_lock(&cache->space_info->lock);
6303         spin_lock(&cache->lock);
6304         cache->pinned += num_bytes;
6305         cache->space_info->bytes_pinned += num_bytes;
6306         if (reserved) {
6307                 cache->reserved -= num_bytes;
6308                 cache->space_info->bytes_reserved -= num_bytes;
6309         }
6310         spin_unlock(&cache->lock);
6311         spin_unlock(&cache->space_info->lock);
6312
6313         trace_btrfs_space_reservation(fs_info, "pinned",
6314                                       cache->space_info->flags, num_bytes, 1);
6315         set_extent_dirty(fs_info->pinned_extents, bytenr,
6316                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
6317         return 0;
6318 }
6319
6320 /*
6321  * this function must be called within transaction
6322  */
6323 int btrfs_pin_extent(struct btrfs_fs_info *fs_info,
6324                      u64 bytenr, u64 num_bytes, int reserved)
6325 {
6326         struct btrfs_block_group_cache *cache;
6327
6328         cache = btrfs_lookup_block_group(fs_info, bytenr);
6329         BUG_ON(!cache); /* Logic error */
6330
6331         pin_down_extent(fs_info, cache, bytenr, num_bytes, reserved);
6332
6333         btrfs_put_block_group(cache);
6334         return 0;
6335 }
6336
6337 /*
6338  * this function must be called within transaction
6339  */
6340 int btrfs_pin_extent_for_log_replay(struct btrfs_fs_info *fs_info,
6341                                     u64 bytenr, u64 num_bytes)
6342 {
6343         struct btrfs_block_group_cache *cache;
6344         int ret;
6345
6346         cache = btrfs_lookup_block_group(fs_info, bytenr);
6347         if (!cache)
6348                 return -EINVAL;
6349
6350         /*
6351          * pull in the free space cache (if any) so that our pin
6352          * removes the free space from the cache.  We have load_only set
6353          * to one because the slow code to read in the free extents does check
6354          * the pinned extents.
6355          */
6356         cache_block_group(cache, 1);
6357
6358         pin_down_extent(fs_info, cache, bytenr, num_bytes, 0);
6359
6360         /* remove us from the free space cache (if we're there at all) */
6361         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
6362         btrfs_put_block_group(cache);
6363         return ret;
6364 }
6365
6366 static int __exclude_logged_extent(struct btrfs_fs_info *fs_info,
6367                                    u64 start, u64 num_bytes)
6368 {
6369         int ret;
6370         struct btrfs_block_group_cache *block_group;
6371         struct btrfs_caching_control *caching_ctl;
6372
6373         block_group = btrfs_lookup_block_group(fs_info, start);
6374         if (!block_group)
6375                 return -EINVAL;
6376
6377         cache_block_group(block_group, 0);
6378         caching_ctl = get_caching_control(block_group);
6379
6380         if (!caching_ctl) {
6381                 /* Logic error */
6382                 BUG_ON(!block_group_cache_done(block_group));
6383                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6384         } else {
6385                 mutex_lock(&caching_ctl->mutex);
6386
6387                 if (start >= caching_ctl->progress) {
6388                         ret = add_excluded_extent(fs_info, start, num_bytes);
6389                 } else if (start + num_bytes <= caching_ctl->progress) {
6390                         ret = btrfs_remove_free_space(block_group,
6391                                                       start, num_bytes);
6392                 } else {
6393                         num_bytes = caching_ctl->progress - start;
6394                         ret = btrfs_remove_free_space(block_group,
6395                                                       start, num_bytes);
6396                         if (ret)
6397                                 goto out_lock;
6398
6399                         num_bytes = (start + num_bytes) -
6400                                 caching_ctl->progress;
6401                         start = caching_ctl->progress;
6402                         ret = add_excluded_extent(fs_info, start, num_bytes);
6403                 }
6404 out_lock:
6405                 mutex_unlock(&caching_ctl->mutex);
6406                 put_caching_control(caching_ctl);
6407         }
6408         btrfs_put_block_group(block_group);
6409         return ret;
6410 }
6411
6412 int btrfs_exclude_logged_extents(struct btrfs_fs_info *fs_info,
6413                                  struct extent_buffer *eb)
6414 {
6415         struct btrfs_file_extent_item *item;
6416         struct btrfs_key key;
6417         int found_type;
6418         int i;
6419
6420         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS))
6421                 return 0;
6422
6423         for (i = 0; i < btrfs_header_nritems(eb); i++) {
6424                 btrfs_item_key_to_cpu(eb, &key, i);
6425                 if (key.type != BTRFS_EXTENT_DATA_KEY)
6426                         continue;
6427                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
6428                 found_type = btrfs_file_extent_type(eb, item);
6429                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
6430                         continue;
6431                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
6432                         continue;
6433                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
6434                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
6435                 __exclude_logged_extent(fs_info, key.objectid, key.offset);
6436         }
6437
6438         return 0;
6439 }
6440
6441 static void
6442 btrfs_inc_block_group_reservations(struct btrfs_block_group_cache *bg)
6443 {
6444         atomic_inc(&bg->reservations);
6445 }
6446
6447 void btrfs_dec_block_group_reservations(struct btrfs_fs_info *fs_info,
6448                                         const u64 start)
6449 {
6450         struct btrfs_block_group_cache *bg;
6451
6452         bg = btrfs_lookup_block_group(fs_info, start);
6453         ASSERT(bg);
6454         if (atomic_dec_and_test(&bg->reservations))
6455                 wake_up_atomic_t(&bg->reservations);
6456         btrfs_put_block_group(bg);
6457 }
6458
6459 static int btrfs_wait_bg_reservations_atomic_t(atomic_t *a)
6460 {
6461         schedule();
6462         return 0;
6463 }
6464
6465 void btrfs_wait_block_group_reservations(struct btrfs_block_group_cache *bg)
6466 {
6467         struct btrfs_space_info *space_info = bg->space_info;
6468
6469         ASSERT(bg->ro);
6470
6471         if (!(bg->flags & BTRFS_BLOCK_GROUP_DATA))
6472                 return;
6473
6474         /*
6475          * Our block group is read only but before we set it to read only,
6476          * some task might have had allocated an extent from it already, but it
6477          * has not yet created a respective ordered extent (and added it to a
6478          * root's list of ordered extents).
6479          * Therefore wait for any task currently allocating extents, since the
6480          * block group's reservations counter is incremented while a read lock
6481          * on the groups' semaphore is held and decremented after releasing
6482          * the read access on that semaphore and creating the ordered extent.
6483          */
6484         down_write(&space_info->groups_sem);
6485         up_write(&space_info->groups_sem);
6486
6487         wait_on_atomic_t(&bg->reservations,
6488                          btrfs_wait_bg_reservations_atomic_t,
6489                          TASK_UNINTERRUPTIBLE);
6490 }
6491
6492 /**
6493  * btrfs_add_reserved_bytes - update the block_group and space info counters
6494  * @cache:      The cache we are manipulating
6495  * @ram_bytes:  The number of bytes of file content, and will be same to
6496  *              @num_bytes except for the compress path.
6497  * @num_bytes:  The number of bytes in question
6498  * @delalloc:   The blocks are allocated for the delalloc write
6499  *
6500  * This is called by the allocator when it reserves space. If this is a
6501  * reservation and the block group has become read only we cannot make the
6502  * reservation and return -EAGAIN, otherwise this function always succeeds.
6503  */
6504 static int btrfs_add_reserved_bytes(struct btrfs_block_group_cache *cache,
6505                                     u64 ram_bytes, u64 num_bytes, int delalloc)
6506 {
6507         struct btrfs_space_info *space_info = cache->space_info;
6508         int ret = 0;
6509
6510         spin_lock(&space_info->lock);
6511         spin_lock(&cache->lock);
6512         if (cache->ro) {
6513                 ret = -EAGAIN;
6514         } else {
6515                 cache->reserved += num_bytes;
6516                 space_info->bytes_reserved += num_bytes;
6517
6518                 trace_btrfs_space_reservation(cache->fs_info,
6519                                 "space_info", space_info->flags,
6520                                 ram_bytes, 0);
6521                 space_info->bytes_may_use -= ram_bytes;
6522                 if (delalloc)
6523                         cache->delalloc_bytes += num_bytes;
6524         }
6525         spin_unlock(&cache->lock);
6526         spin_unlock(&space_info->lock);
6527         return ret;
6528 }
6529
6530 /**
6531  * btrfs_free_reserved_bytes - update the block_group and space info counters
6532  * @cache:      The cache we are manipulating
6533  * @num_bytes:  The number of bytes in question
6534  * @delalloc:   The blocks are allocated for the delalloc write
6535  *
6536  * This is called by somebody who is freeing space that was never actually used
6537  * on disk.  For example if you reserve some space for a new leaf in transaction
6538  * A and before transaction A commits you free that leaf, you call this with
6539  * reserve set to 0 in order to clear the reservation.
6540  */
6541
6542 static int btrfs_free_reserved_bytes(struct btrfs_block_group_cache *cache,
6543                                      u64 num_bytes, int delalloc)
6544 {
6545         struct btrfs_space_info *space_info = cache->space_info;
6546         int ret = 0;
6547
6548         spin_lock(&space_info->lock);
6549         spin_lock(&cache->lock);
6550         if (cache->ro)
6551                 space_info->bytes_readonly += num_bytes;
6552         cache->reserved -= num_bytes;
6553         space_info->bytes_reserved -= num_bytes;
6554
6555         if (delalloc)
6556                 cache->delalloc_bytes -= num_bytes;
6557         spin_unlock(&cache->lock);
6558         spin_unlock(&space_info->lock);
6559         return ret;
6560 }
6561 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
6562                                 struct btrfs_fs_info *fs_info)
6563 {
6564         struct btrfs_caching_control *next;
6565         struct btrfs_caching_control *caching_ctl;
6566         struct btrfs_block_group_cache *cache;
6567
6568         down_write(&fs_info->commit_root_sem);
6569
6570         list_for_each_entry_safe(caching_ctl, next,
6571                                  &fs_info->caching_block_groups, list) {
6572                 cache = caching_ctl->block_group;
6573                 if (block_group_cache_done(cache)) {
6574                         cache->last_byte_to_unpin = (u64)-1;
6575                         list_del_init(&caching_ctl->list);
6576                         put_caching_control(caching_ctl);
6577                 } else {
6578                         cache->last_byte_to_unpin = caching_ctl->progress;
6579                 }
6580         }
6581
6582         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6583                 fs_info->pinned_extents = &fs_info->freed_extents[1];
6584         else
6585                 fs_info->pinned_extents = &fs_info->freed_extents[0];
6586
6587         up_write(&fs_info->commit_root_sem);
6588
6589         update_global_block_rsv(fs_info);
6590 }
6591
6592 /*
6593  * Returns the free cluster for the given space info and sets empty_cluster to
6594  * what it should be based on the mount options.
6595  */
6596 static struct btrfs_free_cluster *
6597 fetch_cluster_info(struct btrfs_fs_info *fs_info,
6598                    struct btrfs_space_info *space_info, u64 *empty_cluster)
6599 {
6600         struct btrfs_free_cluster *ret = NULL;
6601         bool ssd = btrfs_test_opt(fs_info, SSD);
6602
6603         *empty_cluster = 0;
6604         if (btrfs_mixed_space_info(space_info))
6605                 return ret;
6606
6607         if (ssd)
6608                 *empty_cluster = SZ_2M;
6609         if (space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
6610                 ret = &fs_info->meta_alloc_cluster;
6611                 if (!ssd)
6612                         *empty_cluster = SZ_64K;
6613         } else if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) && ssd) {
6614                 ret = &fs_info->data_alloc_cluster;
6615         }
6616
6617         return ret;
6618 }
6619
6620 static int unpin_extent_range(struct btrfs_fs_info *fs_info,
6621                               u64 start, u64 end,
6622                               const bool return_free_space)
6623 {
6624         struct btrfs_block_group_cache *cache = NULL;
6625         struct btrfs_space_info *space_info;
6626         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
6627         struct btrfs_free_cluster *cluster = NULL;
6628         u64 len;
6629         u64 total_unpinned = 0;
6630         u64 empty_cluster = 0;
6631         bool readonly;
6632
6633         while (start <= end) {
6634                 readonly = false;
6635                 if (!cache ||
6636                     start >= cache->key.objectid + cache->key.offset) {
6637                         if (cache)
6638                                 btrfs_put_block_group(cache);
6639                         total_unpinned = 0;
6640                         cache = btrfs_lookup_block_group(fs_info, start);
6641                         BUG_ON(!cache); /* Logic error */
6642
6643                         cluster = fetch_cluster_info(fs_info,
6644                                                      cache->space_info,
6645                                                      &empty_cluster);
6646                         empty_cluster <<= 1;
6647                 }
6648
6649                 len = cache->key.objectid + cache->key.offset - start;
6650                 len = min(len, end + 1 - start);
6651
6652                 if (start < cache->last_byte_to_unpin) {
6653                         len = min(len, cache->last_byte_to_unpin - start);
6654                         if (return_free_space)
6655                                 btrfs_add_free_space(cache, start, len);
6656                 }
6657
6658                 start += len;
6659                 total_unpinned += len;
6660                 space_info = cache->space_info;
6661
6662                 /*
6663                  * If this space cluster has been marked as fragmented and we've
6664                  * unpinned enough in this block group to potentially allow a
6665                  * cluster to be created inside of it go ahead and clear the
6666                  * fragmented check.
6667                  */
6668                 if (cluster && cluster->fragmented &&
6669                     total_unpinned > empty_cluster) {
6670                         spin_lock(&cluster->lock);
6671                         cluster->fragmented = 0;
6672                         spin_unlock(&cluster->lock);
6673                 }
6674
6675                 spin_lock(&space_info->lock);
6676                 spin_lock(&cache->lock);
6677                 cache->pinned -= len;
6678                 space_info->bytes_pinned -= len;
6679
6680                 trace_btrfs_space_reservation(fs_info, "pinned",
6681                                               space_info->flags, len, 0);
6682                 space_info->max_extent_size = 0;
6683                 percpu_counter_add(&space_info->total_bytes_pinned, -len);
6684                 if (cache->ro) {
6685                         space_info->bytes_readonly += len;
6686                         readonly = true;
6687                 }
6688                 spin_unlock(&cache->lock);
6689                 if (!readonly && return_free_space &&
6690                     global_rsv->space_info == space_info) {
6691                         u64 to_add = len;
6692                         WARN_ON(!return_free_space);
6693                         spin_lock(&global_rsv->lock);
6694                         if (!global_rsv->full) {
6695                                 to_add = min(len, global_rsv->size -
6696                                              global_rsv->reserved);
6697                                 global_rsv->reserved += to_add;
6698                                 space_info->bytes_may_use += to_add;
6699                                 if (global_rsv->reserved >= global_rsv->size)
6700                                         global_rsv->full = 1;
6701                                 trace_btrfs_space_reservation(fs_info,
6702                                                               "space_info",
6703                                                               space_info->flags,
6704                                                               to_add, 1);
6705                                 len -= to_add;
6706                         }
6707                         spin_unlock(&global_rsv->lock);
6708                         /* Add to any tickets we may have */
6709                         if (len)
6710                                 space_info_add_new_bytes(fs_info, space_info,
6711                                                          len);
6712                 }
6713                 spin_unlock(&space_info->lock);
6714         }
6715
6716         if (cache)
6717                 btrfs_put_block_group(cache);
6718         return 0;
6719 }
6720
6721 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
6722                                struct btrfs_fs_info *fs_info)
6723 {
6724         struct btrfs_block_group_cache *block_group, *tmp;
6725         struct list_head *deleted_bgs;
6726         struct extent_io_tree *unpin;
6727         u64 start;
6728         u64 end;
6729         int ret;
6730
6731         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6732                 unpin = &fs_info->freed_extents[1];
6733         else
6734                 unpin = &fs_info->freed_extents[0];
6735
6736         while (!trans->aborted) {
6737                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
6738                 ret = find_first_extent_bit(unpin, 0, &start, &end,
6739                                             EXTENT_DIRTY, NULL);
6740                 if (ret) {
6741                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6742                         break;
6743                 }
6744
6745                 if (btrfs_test_opt(fs_info, DISCARD))
6746                         ret = btrfs_discard_extent(fs_info, start,
6747                                                    end + 1 - start, NULL);
6748
6749                 clear_extent_dirty(unpin, start, end);
6750                 unpin_extent_range(fs_info, start, end, true);
6751                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6752                 cond_resched();
6753         }
6754
6755         /*
6756          * Transaction is finished.  We don't need the lock anymore.  We
6757          * do need to clean up the block groups in case of a transaction
6758          * abort.
6759          */
6760         deleted_bgs = &trans->transaction->deleted_bgs;
6761         list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
6762                 u64 trimmed = 0;
6763
6764                 ret = -EROFS;
6765                 if (!trans->aborted)
6766                         ret = btrfs_discard_extent(fs_info,
6767                                                    block_group->key.objectid,
6768                                                    block_group->key.offset,
6769                                                    &trimmed);
6770
6771                 list_del_init(&block_group->bg_list);
6772                 btrfs_put_block_group_trimming(block_group);
6773                 btrfs_put_block_group(block_group);
6774
6775                 if (ret) {
6776                         const char *errstr = btrfs_decode_error(ret);
6777                         btrfs_warn(fs_info,
6778                                    "Discard failed while removing blockgroup: errno=%d %s\n",
6779                                    ret, errstr);
6780                 }
6781         }
6782
6783         return 0;
6784 }
6785
6786 static void add_pinned_bytes(struct btrfs_fs_info *fs_info, u64 num_bytes,
6787                              u64 owner, u64 root_objectid)
6788 {
6789         struct btrfs_space_info *space_info;
6790         u64 flags;
6791
6792         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
6793                 if (root_objectid == BTRFS_CHUNK_TREE_OBJECTID)
6794                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
6795                 else
6796                         flags = BTRFS_BLOCK_GROUP_METADATA;
6797         } else {
6798                 flags = BTRFS_BLOCK_GROUP_DATA;
6799         }
6800
6801         space_info = __find_space_info(fs_info, flags);
6802         BUG_ON(!space_info); /* Logic bug */
6803         percpu_counter_add(&space_info->total_bytes_pinned, num_bytes);
6804 }
6805
6806
6807 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
6808                                 struct btrfs_fs_info *info,
6809                                 struct btrfs_delayed_ref_node *node, u64 parent,
6810                                 u64 root_objectid, u64 owner_objectid,
6811                                 u64 owner_offset, int refs_to_drop,
6812                                 struct btrfs_delayed_extent_op *extent_op)
6813 {
6814         struct btrfs_key key;
6815         struct btrfs_path *path;
6816         struct btrfs_root *extent_root = info->extent_root;
6817         struct extent_buffer *leaf;
6818         struct btrfs_extent_item *ei;
6819         struct btrfs_extent_inline_ref *iref;
6820         int ret;
6821         int is_data;
6822         int extent_slot = 0;
6823         int found_extent = 0;
6824         int num_to_del = 1;
6825         u32 item_size;
6826         u64 refs;
6827         u64 bytenr = node->bytenr;
6828         u64 num_bytes = node->num_bytes;
6829         int last_ref = 0;
6830         bool skinny_metadata = btrfs_fs_incompat(info, SKINNY_METADATA);
6831
6832         path = btrfs_alloc_path();
6833         if (!path)
6834                 return -ENOMEM;
6835
6836         path->reada = READA_FORWARD;
6837         path->leave_spinning = 1;
6838
6839         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
6840         BUG_ON(!is_data && refs_to_drop != 1);
6841
6842         if (is_data)
6843                 skinny_metadata = 0;
6844
6845         ret = lookup_extent_backref(trans, extent_root, path, &iref,
6846                                     bytenr, num_bytes, parent,
6847                                     root_objectid, owner_objectid,
6848                                     owner_offset);
6849         if (ret == 0) {
6850                 extent_slot = path->slots[0];
6851                 while (extent_slot >= 0) {
6852                         btrfs_item_key_to_cpu(path->nodes[0], &key,
6853                                               extent_slot);
6854                         if (key.objectid != bytenr)
6855                                 break;
6856                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
6857                             key.offset == num_bytes) {
6858                                 found_extent = 1;
6859                                 break;
6860                         }
6861                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
6862                             key.offset == owner_objectid) {
6863                                 found_extent = 1;
6864                                 break;
6865                         }
6866                         if (path->slots[0] - extent_slot > 5)
6867                                 break;
6868                         extent_slot--;
6869                 }
6870 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6871                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
6872                 if (found_extent && item_size < sizeof(*ei))
6873                         found_extent = 0;
6874 #endif
6875                 if (!found_extent) {
6876                         BUG_ON(iref);
6877                         ret = remove_extent_backref(trans, extent_root, path,
6878                                                     NULL, refs_to_drop,
6879                                                     is_data, &last_ref);
6880                         if (ret) {
6881                                 btrfs_abort_transaction(trans, ret);
6882                                 goto out;
6883                         }
6884                         btrfs_release_path(path);
6885                         path->leave_spinning = 1;
6886
6887                         key.objectid = bytenr;
6888                         key.type = BTRFS_EXTENT_ITEM_KEY;
6889                         key.offset = num_bytes;
6890
6891                         if (!is_data && skinny_metadata) {
6892                                 key.type = BTRFS_METADATA_ITEM_KEY;
6893                                 key.offset = owner_objectid;
6894                         }
6895
6896                         ret = btrfs_search_slot(trans, extent_root,
6897                                                 &key, path, -1, 1);
6898                         if (ret > 0 && skinny_metadata && path->slots[0]) {
6899                                 /*
6900                                  * Couldn't find our skinny metadata item,
6901                                  * see if we have ye olde extent item.
6902                                  */
6903                                 path->slots[0]--;
6904                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
6905                                                       path->slots[0]);
6906                                 if (key.objectid == bytenr &&
6907                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
6908                                     key.offset == num_bytes)
6909                                         ret = 0;
6910                         }
6911
6912                         if (ret > 0 && skinny_metadata) {
6913                                 skinny_metadata = false;
6914                                 key.objectid = bytenr;
6915                                 key.type = BTRFS_EXTENT_ITEM_KEY;
6916                                 key.offset = num_bytes;
6917                                 btrfs_release_path(path);
6918                                 ret = btrfs_search_slot(trans, extent_root,
6919                                                         &key, path, -1, 1);
6920                         }
6921
6922                         if (ret) {
6923                                 btrfs_err(info,
6924                                           "umm, got %d back from search, was looking for %llu",
6925                                           ret, bytenr);
6926                                 if (ret > 0)
6927                                         btrfs_print_leaf(info, path->nodes[0]);
6928                         }
6929                         if (ret < 0) {
6930                                 btrfs_abort_transaction(trans, ret);
6931                                 goto out;
6932                         }
6933                         extent_slot = path->slots[0];
6934                 }
6935         } else if (WARN_ON(ret == -ENOENT)) {
6936                 btrfs_print_leaf(info, path->nodes[0]);
6937                 btrfs_err(info,
6938                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
6939                         bytenr, parent, root_objectid, owner_objectid,
6940                         owner_offset);
6941                 btrfs_abort_transaction(trans, ret);
6942                 goto out;
6943         } else {
6944                 btrfs_abort_transaction(trans, ret);
6945                 goto out;
6946         }
6947
6948         leaf = path->nodes[0];
6949         item_size = btrfs_item_size_nr(leaf, extent_slot);
6950 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6951         if (item_size < sizeof(*ei)) {
6952                 BUG_ON(found_extent || extent_slot != path->slots[0]);
6953                 ret = convert_extent_item_v0(trans, extent_root, path,
6954                                              owner_objectid, 0);
6955                 if (ret < 0) {
6956                         btrfs_abort_transaction(trans, ret);
6957                         goto out;
6958                 }
6959
6960                 btrfs_release_path(path);
6961                 path->leave_spinning = 1;
6962
6963                 key.objectid = bytenr;
6964                 key.type = BTRFS_EXTENT_ITEM_KEY;
6965                 key.offset = num_bytes;
6966
6967                 ret = btrfs_search_slot(trans, extent_root, &key, path,
6968                                         -1, 1);
6969                 if (ret) {
6970                         btrfs_err(info,
6971                                   "umm, got %d back from search, was looking for %llu",
6972                                 ret, bytenr);
6973                         btrfs_print_leaf(info, path->nodes[0]);
6974                 }
6975                 if (ret < 0) {
6976                         btrfs_abort_transaction(trans, ret);
6977                         goto out;
6978                 }
6979
6980                 extent_slot = path->slots[0];
6981                 leaf = path->nodes[0];
6982                 item_size = btrfs_item_size_nr(leaf, extent_slot);
6983         }
6984 #endif
6985         BUG_ON(item_size < sizeof(*ei));
6986         ei = btrfs_item_ptr(leaf, extent_slot,
6987                             struct btrfs_extent_item);
6988         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
6989             key.type == BTRFS_EXTENT_ITEM_KEY) {
6990                 struct btrfs_tree_block_info *bi;
6991                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
6992                 bi = (struct btrfs_tree_block_info *)(ei + 1);
6993                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
6994         }
6995
6996         refs = btrfs_extent_refs(leaf, ei);
6997         if (refs < refs_to_drop) {
6998                 btrfs_err(info,
6999                           "trying to drop %d refs but we only have %Lu for bytenr %Lu",
7000                           refs_to_drop, refs, bytenr);
7001                 ret = -EINVAL;
7002                 btrfs_abort_transaction(trans, ret);
7003                 goto out;
7004         }
7005         refs -= refs_to_drop;
7006
7007         if (refs > 0) {
7008                 if (extent_op)
7009                         __run_delayed_extent_op(extent_op, leaf, ei);
7010                 /*
7011                  * In the case of inline back ref, reference count will
7012                  * be updated by remove_extent_backref
7013                  */
7014                 if (iref) {
7015                         BUG_ON(!found_extent);
7016                 } else {
7017                         btrfs_set_extent_refs(leaf, ei, refs);
7018                         btrfs_mark_buffer_dirty(leaf);
7019                 }
7020                 if (found_extent) {
7021                         ret = remove_extent_backref(trans, extent_root, path,
7022                                                     iref, refs_to_drop,
7023                                                     is_data, &last_ref);
7024                         if (ret) {
7025                                 btrfs_abort_transaction(trans, ret);
7026                                 goto out;
7027                         }
7028                 }
7029                 add_pinned_bytes(info, -num_bytes, owner_objectid,
7030                                  root_objectid);
7031         } else {
7032                 if (found_extent) {
7033                         BUG_ON(is_data && refs_to_drop !=
7034                                extent_data_ref_count(path, iref));
7035                         if (iref) {
7036                                 BUG_ON(path->slots[0] != extent_slot);
7037                         } else {
7038                                 BUG_ON(path->slots[0] != extent_slot + 1);
7039                                 path->slots[0] = extent_slot;
7040                                 num_to_del = 2;
7041                         }
7042                 }
7043
7044                 last_ref = 1;
7045                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
7046                                       num_to_del);
7047                 if (ret) {
7048                         btrfs_abort_transaction(trans, ret);
7049                         goto out;
7050                 }
7051                 btrfs_release_path(path);
7052
7053                 if (is_data) {
7054                         ret = btrfs_del_csums(trans, info, bytenr, num_bytes);
7055                         if (ret) {
7056                                 btrfs_abort_transaction(trans, ret);
7057                                 goto out;
7058                         }
7059                 }
7060
7061                 ret = add_to_free_space_tree(trans, info, bytenr, num_bytes);
7062                 if (ret) {
7063                         btrfs_abort_transaction(trans, ret);
7064                         goto out;
7065                 }
7066
7067                 ret = update_block_group(trans, info, bytenr, num_bytes, 0);
7068                 if (ret) {
7069                         btrfs_abort_transaction(trans, ret);
7070                         goto out;
7071                 }
7072         }
7073         btrfs_release_path(path);
7074
7075 out:
7076         btrfs_free_path(path);
7077         return ret;
7078 }
7079
7080 /*
7081  * when we free an block, it is possible (and likely) that we free the last
7082  * delayed ref for that extent as well.  This searches the delayed ref tree for
7083  * a given extent, and if there are no other delayed refs to be processed, it
7084  * removes it from the tree.
7085  */
7086 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
7087                                       u64 bytenr)
7088 {
7089         struct btrfs_delayed_ref_head *head;
7090         struct btrfs_delayed_ref_root *delayed_refs;
7091         int ret = 0;
7092
7093         delayed_refs = &trans->transaction->delayed_refs;
7094         spin_lock(&delayed_refs->lock);
7095         head = btrfs_find_delayed_ref_head(trans, bytenr);
7096         if (!head)
7097                 goto out_delayed_unlock;
7098
7099         spin_lock(&head->lock);
7100         if (!list_empty(&head->ref_list))
7101                 goto out;
7102
7103         if (head->extent_op) {
7104                 if (!head->must_insert_reserved)
7105                         goto out;
7106                 btrfs_free_delayed_extent_op(head->extent_op);
7107                 head->extent_op = NULL;
7108         }
7109
7110         /*
7111          * waiting for the lock here would deadlock.  If someone else has it
7112          * locked they are already in the process of dropping it anyway
7113          */
7114         if (!mutex_trylock(&head->mutex))
7115                 goto out;
7116
7117         /*
7118          * at this point we have a head with no other entries.  Go
7119          * ahead and process it.
7120          */
7121         head->node.in_tree = 0;
7122         rb_erase(&head->href_node, &delayed_refs->href_root);
7123
7124         atomic_dec(&delayed_refs->num_entries);
7125
7126         /*
7127          * we don't take a ref on the node because we're removing it from the
7128          * tree, so we just steal the ref the tree was holding.
7129          */
7130         delayed_refs->num_heads--;
7131         if (head->processing == 0)
7132                 delayed_refs->num_heads_ready--;
7133         head->processing = 0;
7134         spin_unlock(&head->lock);
7135         spin_unlock(&delayed_refs->lock);
7136
7137         BUG_ON(head->extent_op);
7138         if (head->must_insert_reserved)
7139                 ret = 1;
7140
7141         mutex_unlock(&head->mutex);
7142         btrfs_put_delayed_ref(&head->node);
7143         return ret;
7144 out:
7145         spin_unlock(&head->lock);
7146
7147 out_delayed_unlock:
7148         spin_unlock(&delayed_refs->lock);
7149         return 0;
7150 }
7151
7152 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
7153                            struct btrfs_root *root,
7154                            struct extent_buffer *buf,
7155                            u64 parent, int last_ref)
7156 {
7157         struct btrfs_fs_info *fs_info = root->fs_info;
7158         int pin = 1;
7159         int ret;
7160
7161         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7162                 ret = btrfs_add_delayed_tree_ref(fs_info, trans,
7163                                                  buf->start, buf->len,
7164                                                  parent,
7165                                                  root->root_key.objectid,
7166                                                  btrfs_header_level(buf),
7167                                                  BTRFS_DROP_DELAYED_REF, NULL);
7168                 BUG_ON(ret); /* -ENOMEM */
7169         }
7170
7171         if (!last_ref)
7172                 return;
7173
7174         if (btrfs_header_generation(buf) == trans->transid) {
7175                 struct btrfs_block_group_cache *cache;
7176
7177                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7178                         ret = check_ref_cleanup(trans, buf->start);
7179                         if (!ret)
7180                                 goto out;
7181                 }
7182
7183                 cache = btrfs_lookup_block_group(fs_info, buf->start);
7184
7185                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
7186                         pin_down_extent(fs_info, cache, buf->start,
7187                                         buf->len, 1);
7188                         btrfs_put_block_group(cache);
7189                         goto out;
7190                 }
7191
7192                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
7193
7194                 btrfs_add_free_space(cache, buf->start, buf->len);
7195                 btrfs_free_reserved_bytes(cache, buf->len, 0);
7196                 btrfs_put_block_group(cache);
7197                 trace_btrfs_reserved_extent_free(fs_info, buf->start, buf->len);
7198                 pin = 0;
7199         }
7200 out:
7201         if (pin)
7202                 add_pinned_bytes(fs_info, buf->len, btrfs_header_level(buf),
7203                                  root->root_key.objectid);
7204
7205         /*
7206          * Deleting the buffer, clear the corrupt flag since it doesn't matter
7207          * anymore.
7208          */
7209         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
7210 }
7211
7212 /* Can return -ENOMEM */
7213 int btrfs_free_extent(struct btrfs_trans_handle *trans,
7214                       struct btrfs_fs_info *fs_info,
7215                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
7216                       u64 owner, u64 offset)
7217 {
7218         int ret;
7219
7220         if (btrfs_is_testing(fs_info))
7221                 return 0;
7222
7223         add_pinned_bytes(fs_info, num_bytes, owner, root_objectid);
7224
7225         /*
7226          * tree log blocks never actually go into the extent allocation
7227          * tree, just update pinning info and exit early.
7228          */
7229         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
7230                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
7231                 /* unlocks the pinned mutex */
7232                 btrfs_pin_extent(fs_info, bytenr, num_bytes, 1);
7233                 ret = 0;
7234         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
7235                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
7236                                         num_bytes,
7237                                         parent, root_objectid, (int)owner,
7238                                         BTRFS_DROP_DELAYED_REF, NULL);
7239         } else {
7240                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
7241                                                 num_bytes,
7242                                                 parent, root_objectid, owner,
7243                                                 offset, 0,
7244                                                 BTRFS_DROP_DELAYED_REF, NULL);
7245         }
7246         return ret;
7247 }
7248
7249 /*
7250  * when we wait for progress in the block group caching, its because
7251  * our allocation attempt failed at least once.  So, we must sleep
7252  * and let some progress happen before we try again.
7253  *
7254  * This function will sleep at least once waiting for new free space to
7255  * show up, and then it will check the block group free space numbers
7256  * for our min num_bytes.  Another option is to have it go ahead
7257  * and look in the rbtree for a free extent of a given size, but this
7258  * is a good start.
7259  *
7260  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
7261  * any of the information in this block group.
7262  */
7263 static noinline void
7264 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
7265                                 u64 num_bytes)
7266 {
7267         struct btrfs_caching_control *caching_ctl;
7268
7269         caching_ctl = get_caching_control(cache);
7270         if (!caching_ctl)
7271                 return;
7272
7273         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
7274                    (cache->free_space_ctl->free_space >= num_bytes));
7275
7276         put_caching_control(caching_ctl);
7277 }
7278
7279 static noinline int
7280 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
7281 {
7282         struct btrfs_caching_control *caching_ctl;
7283         int ret = 0;
7284
7285         caching_ctl = get_caching_control(cache);
7286         if (!caching_ctl)
7287                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
7288
7289         wait_event(caching_ctl->wait, block_group_cache_done(cache));
7290         if (cache->cached == BTRFS_CACHE_ERROR)
7291                 ret = -EIO;
7292         put_caching_control(caching_ctl);
7293         return ret;
7294 }
7295
7296 int __get_raid_index(u64 flags)
7297 {
7298         if (flags & BTRFS_BLOCK_GROUP_RAID10)
7299                 return BTRFS_RAID_RAID10;
7300         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
7301                 return BTRFS_RAID_RAID1;
7302         else if (flags & BTRFS_BLOCK_GROUP_DUP)
7303                 return BTRFS_RAID_DUP;
7304         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
7305                 return BTRFS_RAID_RAID0;
7306         else if (flags & BTRFS_BLOCK_GROUP_RAID5)
7307                 return BTRFS_RAID_RAID5;
7308         else if (flags & BTRFS_BLOCK_GROUP_RAID6)
7309                 return BTRFS_RAID_RAID6;
7310
7311         return BTRFS_RAID_SINGLE; /* BTRFS_BLOCK_GROUP_SINGLE */
7312 }
7313
7314 int get_block_group_index(struct btrfs_block_group_cache *cache)
7315 {
7316         return __get_raid_index(cache->flags);
7317 }
7318
7319 static const char *btrfs_raid_type_names[BTRFS_NR_RAID_TYPES] = {
7320         [BTRFS_RAID_RAID10]     = "raid10",
7321         [BTRFS_RAID_RAID1]      = "raid1",
7322         [BTRFS_RAID_DUP]        = "dup",
7323         [BTRFS_RAID_RAID0]      = "raid0",
7324         [BTRFS_RAID_SINGLE]     = "single",
7325         [BTRFS_RAID_RAID5]      = "raid5",
7326         [BTRFS_RAID_RAID6]      = "raid6",
7327 };
7328
7329 static const char *get_raid_name(enum btrfs_raid_types type)
7330 {
7331         if (type >= BTRFS_NR_RAID_TYPES)
7332                 return NULL;
7333
7334         return btrfs_raid_type_names[type];
7335 }
7336
7337 enum btrfs_loop_type {
7338         LOOP_CACHING_NOWAIT = 0,
7339         LOOP_CACHING_WAIT = 1,
7340         LOOP_ALLOC_CHUNK = 2,
7341         LOOP_NO_EMPTY_SIZE = 3,
7342 };
7343
7344 static inline void
7345 btrfs_lock_block_group(struct btrfs_block_group_cache *cache,
7346                        int delalloc)
7347 {
7348         if (delalloc)
7349                 down_read(&cache->data_rwsem);
7350 }
7351
7352 static inline void
7353 btrfs_grab_block_group(struct btrfs_block_group_cache *cache,
7354                        int delalloc)
7355 {
7356         btrfs_get_block_group(cache);
7357         if (delalloc)
7358                 down_read(&cache->data_rwsem);
7359 }
7360
7361 static struct btrfs_block_group_cache *
7362 btrfs_lock_cluster(struct btrfs_block_group_cache *block_group,
7363                    struct btrfs_free_cluster *cluster,
7364                    int delalloc)
7365 {
7366         struct btrfs_block_group_cache *used_bg = NULL;
7367
7368         spin_lock(&cluster->refill_lock);
7369         while (1) {
7370                 used_bg = cluster->block_group;
7371                 if (!used_bg)
7372                         return NULL;
7373
7374                 if (used_bg == block_group)
7375                         return used_bg;
7376
7377                 btrfs_get_block_group(used_bg);
7378
7379                 if (!delalloc)
7380                         return used_bg;
7381
7382                 if (down_read_trylock(&used_bg->data_rwsem))
7383                         return used_bg;
7384
7385                 spin_unlock(&cluster->refill_lock);
7386
7387                 down_read(&used_bg->data_rwsem);
7388
7389                 spin_lock(&cluster->refill_lock);
7390                 if (used_bg == cluster->block_group)
7391                         return used_bg;
7392
7393                 up_read(&used_bg->data_rwsem);
7394                 btrfs_put_block_group(used_bg);
7395         }
7396 }
7397
7398 static inline void
7399 btrfs_release_block_group(struct btrfs_block_group_cache *cache,
7400                          int delalloc)
7401 {
7402         if (delalloc)
7403                 up_read(&cache->data_rwsem);
7404         btrfs_put_block_group(cache);
7405 }
7406
7407 /*
7408  * walks the btree of allocated extents and find a hole of a given size.
7409  * The key ins is changed to record the hole:
7410  * ins->objectid == start position
7411  * ins->flags = BTRFS_EXTENT_ITEM_KEY
7412  * ins->offset == the size of the hole.
7413  * Any available blocks before search_start are skipped.
7414  *
7415  * If there is no suitable free space, we will record the max size of
7416  * the free space extent currently.
7417  */
7418 static noinline int find_free_extent(struct btrfs_root *orig_root,
7419                                 u64 ram_bytes, u64 num_bytes, u64 empty_size,
7420                                 u64 hint_byte, struct btrfs_key *ins,
7421                                 u64 flags, int delalloc)
7422 {
7423         struct btrfs_fs_info *fs_info = orig_root->fs_info;
7424         int ret = 0;
7425         struct btrfs_root *root = fs_info->extent_root;
7426         struct btrfs_free_cluster *last_ptr = NULL;
7427         struct btrfs_block_group_cache *block_group = NULL;
7428         u64 search_start = 0;
7429         u64 max_extent_size = 0;
7430         u64 empty_cluster = 0;
7431         struct btrfs_space_info *space_info;
7432         int loop = 0;
7433         int index = __get_raid_index(flags);
7434         bool failed_cluster_refill = false;
7435         bool failed_alloc = false;
7436         bool use_cluster = true;
7437         bool have_caching_bg = false;
7438         bool orig_have_caching_bg = false;
7439         bool full_search = false;
7440
7441         WARN_ON(num_bytes < fs_info->sectorsize);
7442         ins->type = BTRFS_EXTENT_ITEM_KEY;
7443         ins->objectid = 0;
7444         ins->offset = 0;
7445
7446         trace_find_free_extent(fs_info, num_bytes, empty_size, flags);
7447
7448         space_info = __find_space_info(fs_info, flags);
7449         if (!space_info) {
7450                 btrfs_err(fs_info, "No space info for %llu", flags);
7451                 return -ENOSPC;
7452         }
7453
7454         /*
7455          * If our free space is heavily fragmented we may not be able to make
7456          * big contiguous allocations, so instead of doing the expensive search
7457          * for free space, simply return ENOSPC with our max_extent_size so we
7458          * can go ahead and search for a more manageable chunk.
7459          *
7460          * If our max_extent_size is large enough for our allocation simply
7461          * disable clustering since we will likely not be able to find enough
7462          * space to create a cluster and induce latency trying.
7463          */
7464         if (unlikely(space_info->max_extent_size)) {
7465                 spin_lock(&space_info->lock);
7466                 if (space_info->max_extent_size &&
7467                     num_bytes > space_info->max_extent_size) {
7468                         ins->offset = space_info->max_extent_size;
7469                         spin_unlock(&space_info->lock);
7470                         return -ENOSPC;
7471                 } else if (space_info->max_extent_size) {
7472                         use_cluster = false;
7473                 }
7474                 spin_unlock(&space_info->lock);
7475         }
7476
7477         last_ptr = fetch_cluster_info(fs_info, space_info, &empty_cluster);
7478         if (last_ptr) {
7479                 spin_lock(&last_ptr->lock);
7480                 if (last_ptr->block_group)
7481                         hint_byte = last_ptr->window_start;
7482                 if (last_ptr->fragmented) {
7483                         /*
7484                          * We still set window_start so we can keep track of the
7485                          * last place we found an allocation to try and save
7486                          * some time.
7487                          */
7488                         hint_byte = last_ptr->window_start;
7489                         use_cluster = false;
7490                 }
7491                 spin_unlock(&last_ptr->lock);
7492         }
7493
7494         search_start = max(search_start, first_logical_byte(fs_info, 0));
7495         search_start = max(search_start, hint_byte);
7496         if (search_start == hint_byte) {
7497                 block_group = btrfs_lookup_block_group(fs_info, search_start);
7498                 /*
7499                  * we don't want to use the block group if it doesn't match our
7500                  * allocation bits, or if its not cached.
7501                  *
7502                  * However if we are re-searching with an ideal block group
7503                  * picked out then we don't care that the block group is cached.
7504                  */
7505                 if (block_group && block_group_bits(block_group, flags) &&
7506                     block_group->cached != BTRFS_CACHE_NO) {
7507                         down_read(&space_info->groups_sem);
7508                         if (list_empty(&block_group->list) ||
7509                             block_group->ro) {
7510                                 /*
7511                                  * someone is removing this block group,
7512                                  * we can't jump into the have_block_group
7513                                  * target because our list pointers are not
7514                                  * valid
7515                                  */
7516                                 btrfs_put_block_group(block_group);
7517                                 up_read(&space_info->groups_sem);
7518                         } else {
7519                                 index = get_block_group_index(block_group);
7520                                 btrfs_lock_block_group(block_group, delalloc);
7521                                 goto have_block_group;
7522                         }
7523                 } else if (block_group) {
7524                         btrfs_put_block_group(block_group);
7525                 }
7526         }
7527 search:
7528         have_caching_bg = false;
7529         if (index == 0 || index == __get_raid_index(flags))
7530                 full_search = true;
7531         down_read(&space_info->groups_sem);
7532         list_for_each_entry(block_group, &space_info->block_groups[index],
7533                             list) {
7534                 u64 offset;
7535                 int cached;
7536
7537                 btrfs_grab_block_group(block_group, delalloc);
7538                 search_start = block_group->key.objectid;
7539
7540                 /*
7541                  * this can happen if we end up cycling through all the
7542                  * raid types, but we want to make sure we only allocate
7543                  * for the proper type.
7544                  */
7545                 if (!block_group_bits(block_group, flags)) {
7546                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
7547                                 BTRFS_BLOCK_GROUP_RAID1 |
7548                                 BTRFS_BLOCK_GROUP_RAID5 |
7549                                 BTRFS_BLOCK_GROUP_RAID6 |
7550                                 BTRFS_BLOCK_GROUP_RAID10;
7551
7552                         /*
7553                          * if they asked for extra copies and this block group
7554                          * doesn't provide them, bail.  This does allow us to
7555                          * fill raid0 from raid1.
7556                          */
7557                         if ((flags & extra) && !(block_group->flags & extra))
7558                                 goto loop;
7559                 }
7560
7561 have_block_group:
7562                 cached = block_group_cache_done(block_group);
7563                 if (unlikely(!cached)) {
7564                         have_caching_bg = true;
7565                         ret = cache_block_group(block_group, 0);
7566                         BUG_ON(ret < 0);
7567                         ret = 0;
7568                 }
7569
7570                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
7571                         goto loop;
7572                 if (unlikely(block_group->ro))
7573                         goto loop;
7574
7575                 /*
7576                  * Ok we want to try and use the cluster allocator, so
7577                  * lets look there
7578                  */
7579                 if (last_ptr && use_cluster) {
7580                         struct btrfs_block_group_cache *used_block_group;
7581                         unsigned long aligned_cluster;
7582                         /*
7583                          * the refill lock keeps out other
7584                          * people trying to start a new cluster
7585                          */
7586                         used_block_group = btrfs_lock_cluster(block_group,
7587                                                               last_ptr,
7588                                                               delalloc);
7589                         if (!used_block_group)
7590                                 goto refill_cluster;
7591
7592                         if (used_block_group != block_group &&
7593                             (used_block_group->ro ||
7594                              !block_group_bits(used_block_group, flags)))
7595                                 goto release_cluster;
7596
7597                         offset = btrfs_alloc_from_cluster(used_block_group,
7598                                                 last_ptr,
7599                                                 num_bytes,
7600                                                 used_block_group->key.objectid,
7601                                                 &max_extent_size);
7602                         if (offset) {
7603                                 /* we have a block, we're done */
7604                                 spin_unlock(&last_ptr->refill_lock);
7605                                 trace_btrfs_reserve_extent_cluster(fs_info,
7606                                                 used_block_group,
7607                                                 search_start, num_bytes);
7608                                 if (used_block_group != block_group) {
7609                                         btrfs_release_block_group(block_group,
7610                                                                   delalloc);
7611                                         block_group = used_block_group;
7612                                 }
7613                                 goto checks;
7614                         }
7615
7616                         WARN_ON(last_ptr->block_group != used_block_group);
7617 release_cluster:
7618                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
7619                          * set up a new clusters, so lets just skip it
7620                          * and let the allocator find whatever block
7621                          * it can find.  If we reach this point, we
7622                          * will have tried the cluster allocator
7623                          * plenty of times and not have found
7624                          * anything, so we are likely way too
7625                          * fragmented for the clustering stuff to find
7626                          * anything.
7627                          *
7628                          * However, if the cluster is taken from the
7629                          * current block group, release the cluster
7630                          * first, so that we stand a better chance of
7631                          * succeeding in the unclustered
7632                          * allocation.  */
7633                         if (loop >= LOOP_NO_EMPTY_SIZE &&
7634                             used_block_group != block_group) {
7635                                 spin_unlock(&last_ptr->refill_lock);
7636                                 btrfs_release_block_group(used_block_group,
7637                                                           delalloc);
7638                                 goto unclustered_alloc;
7639                         }
7640
7641                         /*
7642                          * this cluster didn't work out, free it and
7643                          * start over
7644                          */
7645                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7646
7647                         if (used_block_group != block_group)
7648                                 btrfs_release_block_group(used_block_group,
7649                                                           delalloc);
7650 refill_cluster:
7651                         if (loop >= LOOP_NO_EMPTY_SIZE) {
7652                                 spin_unlock(&last_ptr->refill_lock);
7653                                 goto unclustered_alloc;
7654                         }
7655
7656                         aligned_cluster = max_t(unsigned long,
7657                                                 empty_cluster + empty_size,
7658                                               block_group->full_stripe_len);
7659
7660                         /* allocate a cluster in this block group */
7661                         ret = btrfs_find_space_cluster(fs_info, block_group,
7662                                                        last_ptr, search_start,
7663                                                        num_bytes,
7664                                                        aligned_cluster);
7665                         if (ret == 0) {
7666                                 /*
7667                                  * now pull our allocation out of this
7668                                  * cluster
7669                                  */
7670                                 offset = btrfs_alloc_from_cluster(block_group,
7671                                                         last_ptr,
7672                                                         num_bytes,
7673                                                         search_start,
7674                                                         &max_extent_size);
7675                                 if (offset) {
7676                                         /* we found one, proceed */
7677                                         spin_unlock(&last_ptr->refill_lock);
7678                                         trace_btrfs_reserve_extent_cluster(fs_info,
7679                                                 block_group, search_start,
7680                                                 num_bytes);
7681                                         goto checks;
7682                                 }
7683                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
7684                                    && !failed_cluster_refill) {
7685                                 spin_unlock(&last_ptr->refill_lock);
7686
7687                                 failed_cluster_refill = true;
7688                                 wait_block_group_cache_progress(block_group,
7689                                        num_bytes + empty_cluster + empty_size);
7690                                 goto have_block_group;
7691                         }
7692
7693                         /*
7694                          * at this point we either didn't find a cluster
7695                          * or we weren't able to allocate a block from our
7696                          * cluster.  Free the cluster we've been trying
7697                          * to use, and go to the next block group
7698                          */
7699                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7700                         spin_unlock(&last_ptr->refill_lock);
7701                         goto loop;
7702                 }
7703
7704 unclustered_alloc:
7705                 /*
7706                  * We are doing an unclustered alloc, set the fragmented flag so
7707                  * we don't bother trying to setup a cluster again until we get
7708                  * more space.
7709                  */
7710                 if (unlikely(last_ptr)) {
7711                         spin_lock(&last_ptr->lock);
7712                         last_ptr->fragmented = 1;
7713                         spin_unlock(&last_ptr->lock);
7714                 }
7715                 spin_lock(&block_group->free_space_ctl->tree_lock);
7716                 if (cached &&
7717                     block_group->free_space_ctl->free_space <
7718                     num_bytes + empty_cluster + empty_size) {
7719                         if (block_group->free_space_ctl->free_space >
7720                             max_extent_size)
7721                                 max_extent_size =
7722                                         block_group->free_space_ctl->free_space;
7723                         spin_unlock(&block_group->free_space_ctl->tree_lock);
7724                         goto loop;
7725                 }
7726                 spin_unlock(&block_group->free_space_ctl->tree_lock);
7727
7728                 offset = btrfs_find_space_for_alloc(block_group, search_start,
7729                                                     num_bytes, empty_size,
7730                                                     &max_extent_size);
7731                 /*
7732                  * If we didn't find a chunk, and we haven't failed on this
7733                  * block group before, and this block group is in the middle of
7734                  * caching and we are ok with waiting, then go ahead and wait
7735                  * for progress to be made, and set failed_alloc to true.
7736                  *
7737                  * If failed_alloc is true then we've already waited on this
7738                  * block group once and should move on to the next block group.
7739                  */
7740                 if (!offset && !failed_alloc && !cached &&
7741                     loop > LOOP_CACHING_NOWAIT) {
7742                         wait_block_group_cache_progress(block_group,
7743                                                 num_bytes + empty_size);
7744                         failed_alloc = true;
7745                         goto have_block_group;
7746                 } else if (!offset) {
7747                         goto loop;
7748                 }
7749 checks:
7750                 search_start = ALIGN(offset, fs_info->stripesize);
7751
7752                 /* move on to the next group */
7753                 if (search_start + num_bytes >
7754                     block_group->key.objectid + block_group->key.offset) {
7755                         btrfs_add_free_space(block_group, offset, num_bytes);
7756                         goto loop;
7757                 }
7758
7759                 if (offset < search_start)
7760                         btrfs_add_free_space(block_group, offset,
7761                                              search_start - offset);
7762                 BUG_ON(offset > search_start);
7763
7764                 ret = btrfs_add_reserved_bytes(block_group, ram_bytes,
7765                                 num_bytes, delalloc);
7766                 if (ret == -EAGAIN) {
7767                         btrfs_add_free_space(block_group, offset, num_bytes);
7768                         goto loop;
7769                 }
7770                 btrfs_inc_block_group_reservations(block_group);
7771
7772                 /* we are all good, lets return */
7773                 ins->objectid = search_start;
7774                 ins->offset = num_bytes;
7775
7776                 trace_btrfs_reserve_extent(fs_info, block_group,
7777                                            search_start, num_bytes);
7778                 btrfs_release_block_group(block_group, delalloc);
7779                 break;
7780 loop:
7781                 failed_cluster_refill = false;
7782                 failed_alloc = false;
7783                 BUG_ON(index != get_block_group_index(block_group));
7784                 btrfs_release_block_group(block_group, delalloc);
7785         }
7786         up_read(&space_info->groups_sem);
7787
7788         if ((loop == LOOP_CACHING_NOWAIT) && have_caching_bg
7789                 && !orig_have_caching_bg)
7790                 orig_have_caching_bg = true;
7791
7792         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
7793                 goto search;
7794
7795         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
7796                 goto search;
7797
7798         /*
7799          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
7800          *                      caching kthreads as we move along
7801          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
7802          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
7803          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
7804          *                      again
7805          */
7806         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
7807                 index = 0;
7808                 if (loop == LOOP_CACHING_NOWAIT) {
7809                         /*
7810                          * We want to skip the LOOP_CACHING_WAIT step if we
7811                          * don't have any uncached bgs and we've already done a
7812                          * full search through.
7813                          */
7814                         if (orig_have_caching_bg || !full_search)
7815                                 loop = LOOP_CACHING_WAIT;
7816                         else
7817                                 loop = LOOP_ALLOC_CHUNK;
7818                 } else {
7819                         loop++;
7820                 }
7821
7822                 if (loop == LOOP_ALLOC_CHUNK) {
7823                         struct btrfs_trans_handle *trans;
7824                         int exist = 0;
7825
7826                         trans = current->journal_info;
7827                         if (trans)
7828                                 exist = 1;
7829                         else
7830                                 trans = btrfs_join_transaction(root);
7831
7832                         if (IS_ERR(trans)) {
7833                                 ret = PTR_ERR(trans);
7834                                 goto out;
7835                         }
7836
7837                         ret = do_chunk_alloc(trans, fs_info, flags,
7838                                              CHUNK_ALLOC_FORCE);
7839
7840                         /*
7841                          * If we can't allocate a new chunk we've already looped
7842                          * through at least once, move on to the NO_EMPTY_SIZE
7843                          * case.
7844                          */
7845                         if (ret == -ENOSPC)
7846                                 loop = LOOP_NO_EMPTY_SIZE;
7847
7848                         /*
7849                          * Do not bail out on ENOSPC since we
7850                          * can do more things.
7851                          */
7852                         if (ret < 0 && ret != -ENOSPC)
7853                                 btrfs_abort_transaction(trans, ret);
7854                         else
7855                                 ret = 0;
7856                         if (!exist)
7857                                 btrfs_end_transaction(trans);
7858                         if (ret)
7859                                 goto out;
7860                 }
7861
7862                 if (loop == LOOP_NO_EMPTY_SIZE) {
7863                         /*
7864                          * Don't loop again if we already have no empty_size and
7865                          * no empty_cluster.
7866                          */
7867                         if (empty_size == 0 &&
7868                             empty_cluster == 0) {
7869                                 ret = -ENOSPC;
7870                                 goto out;
7871                         }
7872                         empty_size = 0;
7873                         empty_cluster = 0;
7874                 }
7875
7876                 goto search;
7877         } else if (!ins->objectid) {
7878                 ret = -ENOSPC;
7879         } else if (ins->objectid) {
7880                 if (!use_cluster && last_ptr) {
7881                         spin_lock(&last_ptr->lock);
7882                         last_ptr->window_start = ins->objectid;
7883                         spin_unlock(&last_ptr->lock);
7884                 }
7885                 ret = 0;
7886         }
7887 out:
7888         if (ret == -ENOSPC) {
7889                 spin_lock(&space_info->lock);
7890                 space_info->max_extent_size = max_extent_size;
7891                 spin_unlock(&space_info->lock);
7892                 ins->offset = max_extent_size;
7893         }
7894         return ret;
7895 }
7896
7897 static void dump_space_info(struct btrfs_fs_info *fs_info,
7898                             struct btrfs_space_info *info, u64 bytes,
7899                             int dump_block_groups)
7900 {
7901         struct btrfs_block_group_cache *cache;
7902         int index = 0;
7903
7904         spin_lock(&info->lock);
7905         btrfs_info(fs_info, "space_info %llu has %llu free, is %sfull",
7906                    info->flags,
7907                    info->total_bytes - info->bytes_used - info->bytes_pinned -
7908                    info->bytes_reserved - info->bytes_readonly -
7909                    info->bytes_may_use, (info->full) ? "" : "not ");
7910         btrfs_info(fs_info,
7911                 "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu",
7912                 info->total_bytes, info->bytes_used, info->bytes_pinned,
7913                 info->bytes_reserved, info->bytes_may_use,
7914                 info->bytes_readonly);
7915         spin_unlock(&info->lock);
7916
7917         if (!dump_block_groups)
7918                 return;
7919
7920         down_read(&info->groups_sem);
7921 again:
7922         list_for_each_entry(cache, &info->block_groups[index], list) {
7923                 spin_lock(&cache->lock);
7924                 btrfs_info(fs_info,
7925                         "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s",
7926                         cache->key.objectid, cache->key.offset,
7927                         btrfs_block_group_used(&cache->item), cache->pinned,
7928                         cache->reserved, cache->ro ? "[readonly]" : "");
7929                 btrfs_dump_free_space(cache, bytes);
7930                 spin_unlock(&cache->lock);
7931         }
7932         if (++index < BTRFS_NR_RAID_TYPES)
7933                 goto again;
7934         up_read(&info->groups_sem);
7935 }
7936
7937 int btrfs_reserve_extent(struct btrfs_root *root, u64 ram_bytes,
7938                          u64 num_bytes, u64 min_alloc_size,
7939                          u64 empty_size, u64 hint_byte,
7940                          struct btrfs_key *ins, int is_data, int delalloc)
7941 {
7942         struct btrfs_fs_info *fs_info = root->fs_info;
7943         bool final_tried = num_bytes == min_alloc_size;
7944         u64 flags;
7945         int ret;
7946
7947         flags = btrfs_get_alloc_profile(root, is_data);
7948 again:
7949         WARN_ON(num_bytes < fs_info->sectorsize);
7950         ret = find_free_extent(root, ram_bytes, num_bytes, empty_size,
7951                                hint_byte, ins, flags, delalloc);
7952         if (!ret && !is_data) {
7953                 btrfs_dec_block_group_reservations(fs_info, ins->objectid);
7954         } else if (ret == -ENOSPC) {
7955                 if (!final_tried && ins->offset) {
7956                         num_bytes = min(num_bytes >> 1, ins->offset);
7957                         num_bytes = round_down(num_bytes,
7958                                                fs_info->sectorsize);
7959                         num_bytes = max(num_bytes, min_alloc_size);
7960                         ram_bytes = num_bytes;
7961                         if (num_bytes == min_alloc_size)
7962                                 final_tried = true;
7963                         goto again;
7964                 } else if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
7965                         struct btrfs_space_info *sinfo;
7966
7967                         sinfo = __find_space_info(fs_info, flags);
7968                         btrfs_err(fs_info,
7969                                   "allocation failed flags %llu, wanted %llu",
7970                                   flags, num_bytes);
7971                         if (sinfo)
7972                                 dump_space_info(fs_info, sinfo, num_bytes, 1);
7973                 }
7974         }
7975
7976         return ret;
7977 }
7978
7979 static int __btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
7980                                         u64 start, u64 len,
7981                                         int pin, int delalloc)
7982 {
7983         struct btrfs_block_group_cache *cache;
7984         int ret = 0;
7985
7986         cache = btrfs_lookup_block_group(fs_info, start);
7987         if (!cache) {
7988                 btrfs_err(fs_info, "Unable to find block group for %llu",
7989                           start);
7990                 return -ENOSPC;
7991         }
7992
7993         if (pin)
7994                 pin_down_extent(fs_info, cache, start, len, 1);
7995         else {
7996                 if (btrfs_test_opt(fs_info, DISCARD))
7997                         ret = btrfs_discard_extent(fs_info, start, len, NULL);
7998                 btrfs_add_free_space(cache, start, len);
7999                 btrfs_free_reserved_bytes(cache, len, delalloc);
8000                 trace_btrfs_reserved_extent_free(fs_info, start, len);
8001         }
8002
8003         btrfs_put_block_group(cache);
8004         return ret;
8005 }
8006
8007 int btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
8008                                u64 start, u64 len, int delalloc)
8009 {
8010         return __btrfs_free_reserved_extent(fs_info, start, len, 0, delalloc);
8011 }
8012
8013 int btrfs_free_and_pin_reserved_extent(struct btrfs_fs_info *fs_info,
8014                                        u64 start, u64 len)
8015 {
8016         return __btrfs_free_reserved_extent(fs_info, start, len, 1, 0);
8017 }
8018
8019 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8020                                       struct btrfs_fs_info *fs_info,
8021                                       u64 parent, u64 root_objectid,
8022                                       u64 flags, u64 owner, u64 offset,
8023                                       struct btrfs_key *ins, int ref_mod)
8024 {
8025         int ret;
8026         struct btrfs_extent_item *extent_item;
8027         struct btrfs_extent_inline_ref *iref;
8028         struct btrfs_path *path;
8029         struct extent_buffer *leaf;
8030         int type;
8031         u32 size;
8032
8033         if (parent > 0)
8034                 type = BTRFS_SHARED_DATA_REF_KEY;
8035         else
8036                 type = BTRFS_EXTENT_DATA_REF_KEY;
8037
8038         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
8039
8040         path = btrfs_alloc_path();
8041         if (!path)
8042                 return -ENOMEM;
8043
8044         path->leave_spinning = 1;
8045         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
8046                                       ins, size);
8047         if (ret) {
8048                 btrfs_free_path(path);
8049                 return ret;
8050         }
8051
8052         leaf = path->nodes[0];
8053         extent_item = btrfs_item_ptr(leaf, path->slots[0],
8054                                      struct btrfs_extent_item);
8055         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
8056         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8057         btrfs_set_extent_flags(leaf, extent_item,
8058                                flags | BTRFS_EXTENT_FLAG_DATA);
8059
8060         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8061         btrfs_set_extent_inline_ref_type(leaf, iref, type);
8062         if (parent > 0) {
8063                 struct btrfs_shared_data_ref *ref;
8064                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
8065                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
8066                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
8067         } else {
8068                 struct btrfs_extent_data_ref *ref;
8069                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
8070                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
8071                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
8072                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
8073                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
8074         }
8075
8076         btrfs_mark_buffer_dirty(path->nodes[0]);
8077         btrfs_free_path(path);
8078
8079         ret = remove_from_free_space_tree(trans, fs_info, ins->objectid,
8080                                           ins->offset);
8081         if (ret)
8082                 return ret;
8083
8084         ret = update_block_group(trans, fs_info, ins->objectid, ins->offset, 1);
8085         if (ret) { /* -ENOENT, logic error */
8086                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8087                         ins->objectid, ins->offset);
8088                 BUG();
8089         }
8090         trace_btrfs_reserved_extent_alloc(fs_info, ins->objectid, ins->offset);
8091         return ret;
8092 }
8093
8094 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
8095                                      struct btrfs_fs_info *fs_info,
8096                                      u64 parent, u64 root_objectid,
8097                                      u64 flags, struct btrfs_disk_key *key,
8098                                      int level, struct btrfs_key *ins)
8099 {
8100         int ret;
8101         struct btrfs_extent_item *extent_item;
8102         struct btrfs_tree_block_info *block_info;
8103         struct btrfs_extent_inline_ref *iref;
8104         struct btrfs_path *path;
8105         struct extent_buffer *leaf;
8106         u32 size = sizeof(*extent_item) + sizeof(*iref);
8107         u64 num_bytes = ins->offset;
8108         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
8109
8110         if (!skinny_metadata)
8111                 size += sizeof(*block_info);
8112
8113         path = btrfs_alloc_path();
8114         if (!path) {
8115                 btrfs_free_and_pin_reserved_extent(fs_info, ins->objectid,
8116                                                    fs_info->nodesize);
8117                 return -ENOMEM;
8118         }
8119
8120         path->leave_spinning = 1;
8121         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
8122                                       ins, size);
8123         if (ret) {
8124                 btrfs_free_path(path);
8125                 btrfs_free_and_pin_reserved_extent(fs_info, ins->objectid,
8126                                                    fs_info->nodesize);
8127                 return ret;
8128         }
8129
8130         leaf = path->nodes[0];
8131         extent_item = btrfs_item_ptr(leaf, path->slots[0],
8132                                      struct btrfs_extent_item);
8133         btrfs_set_extent_refs(leaf, extent_item, 1);
8134         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8135         btrfs_set_extent_flags(leaf, extent_item,
8136                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
8137
8138         if (skinny_metadata) {
8139                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8140                 num_bytes = fs_info->nodesize;
8141         } else {
8142                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
8143                 btrfs_set_tree_block_key(leaf, block_info, key);
8144                 btrfs_set_tree_block_level(leaf, block_info, level);
8145                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
8146         }
8147
8148         if (parent > 0) {
8149                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
8150                 btrfs_set_extent_inline_ref_type(leaf, iref,
8151                                                  BTRFS_SHARED_BLOCK_REF_KEY);
8152                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
8153         } else {
8154                 btrfs_set_extent_inline_ref_type(leaf, iref,
8155                                                  BTRFS_TREE_BLOCK_REF_KEY);
8156                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
8157         }
8158
8159         btrfs_mark_buffer_dirty(leaf);
8160         btrfs_free_path(path);
8161
8162         ret = remove_from_free_space_tree(trans, fs_info, ins->objectid,
8163                                           num_bytes);
8164         if (ret)
8165                 return ret;
8166
8167         ret = update_block_group(trans, fs_info, ins->objectid,
8168                                  fs_info->nodesize, 1);
8169         if (ret) { /* -ENOENT, logic error */
8170                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8171                         ins->objectid, ins->offset);
8172                 BUG();
8173         }
8174
8175         trace_btrfs_reserved_extent_alloc(fs_info, ins->objectid,
8176                                           fs_info->nodesize);
8177         return ret;
8178 }
8179
8180 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8181                                      u64 root_objectid, u64 owner,
8182                                      u64 offset, u64 ram_bytes,
8183                                      struct btrfs_key *ins)
8184 {
8185         struct btrfs_fs_info *fs_info = trans->fs_info;
8186         int ret;
8187
8188         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
8189
8190         ret = btrfs_add_delayed_data_ref(fs_info, trans, ins->objectid,
8191                                          ins->offset, 0,
8192                                          root_objectid, owner, offset,
8193                                          ram_bytes, BTRFS_ADD_DELAYED_EXTENT,
8194                                          NULL);
8195         return ret;
8196 }
8197
8198 /*
8199  * this is used by the tree logging recovery code.  It records that
8200  * an extent has been allocated and makes sure to clear the free
8201  * space cache bits as well
8202  */
8203 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
8204                                    struct btrfs_fs_info *fs_info,
8205                                    u64 root_objectid, u64 owner, u64 offset,
8206                                    struct btrfs_key *ins)
8207 {
8208         int ret;
8209         struct btrfs_block_group_cache *block_group;
8210         struct btrfs_space_info *space_info;
8211
8212         /*
8213          * Mixed block groups will exclude before processing the log so we only
8214          * need to do the exclude dance if this fs isn't mixed.
8215          */
8216         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS)) {
8217                 ret = __exclude_logged_extent(fs_info, ins->objectid,
8218                                               ins->offset);
8219                 if (ret)
8220                         return ret;
8221         }
8222
8223         block_group = btrfs_lookup_block_group(fs_info, ins->objectid);
8224         if (!block_group)
8225                 return -EINVAL;
8226
8227         space_info = block_group->space_info;
8228         spin_lock(&space_info->lock);
8229         spin_lock(&block_group->lock);
8230         space_info->bytes_reserved += ins->offset;
8231         block_group->reserved += ins->offset;
8232         spin_unlock(&block_group->lock);
8233         spin_unlock(&space_info->lock);
8234
8235         ret = alloc_reserved_file_extent(trans, fs_info, 0, root_objectid,
8236                                          0, owner, offset, ins, 1);
8237         btrfs_put_block_group(block_group);
8238         return ret;
8239 }
8240
8241 static struct extent_buffer *
8242 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
8243                       u64 bytenr, int level)
8244 {
8245         struct btrfs_fs_info *fs_info = root->fs_info;
8246         struct extent_buffer *buf;
8247
8248         buf = btrfs_find_create_tree_block(fs_info, bytenr);
8249         if (IS_ERR(buf))
8250                 return buf;
8251
8252         btrfs_set_header_generation(buf, trans->transid);
8253         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
8254         btrfs_tree_lock(buf);
8255         clean_tree_block(trans, fs_info, buf);
8256         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
8257
8258         btrfs_set_lock_blocking(buf);
8259         set_extent_buffer_uptodate(buf);
8260
8261         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
8262                 buf->log_index = root->log_transid % 2;
8263                 /*
8264                  * we allow two log transactions at a time, use different
8265                  * EXENT bit to differentiate dirty pages.
8266                  */
8267                 if (buf->log_index == 0)
8268                         set_extent_dirty(&root->dirty_log_pages, buf->start,
8269                                         buf->start + buf->len - 1, GFP_NOFS);
8270                 else
8271                         set_extent_new(&root->dirty_log_pages, buf->start,
8272                                         buf->start + buf->len - 1);
8273         } else {
8274                 buf->log_index = -1;
8275                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
8276                          buf->start + buf->len - 1, GFP_NOFS);
8277         }
8278         trans->dirty = true;
8279         /* this returns a buffer locked for blocking */
8280         return buf;
8281 }
8282
8283 static struct btrfs_block_rsv *
8284 use_block_rsv(struct btrfs_trans_handle *trans,
8285               struct btrfs_root *root, u32 blocksize)
8286 {
8287         struct btrfs_fs_info *fs_info = root->fs_info;
8288         struct btrfs_block_rsv *block_rsv;
8289         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
8290         int ret;
8291         bool global_updated = false;
8292
8293         block_rsv = get_block_rsv(trans, root);
8294
8295         if (unlikely(block_rsv->size == 0))
8296                 goto try_reserve;
8297 again:
8298         ret = block_rsv_use_bytes(block_rsv, blocksize);
8299         if (!ret)
8300                 return block_rsv;
8301
8302         if (block_rsv->failfast)
8303                 return ERR_PTR(ret);
8304
8305         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
8306                 global_updated = true;
8307                 update_global_block_rsv(fs_info);
8308                 goto again;
8309         }
8310
8311         if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
8312                 static DEFINE_RATELIMIT_STATE(_rs,
8313                                 DEFAULT_RATELIMIT_INTERVAL * 10,
8314                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
8315                 if (__ratelimit(&_rs))
8316                         WARN(1, KERN_DEBUG
8317                                 "BTRFS: block rsv returned %d\n", ret);
8318         }
8319 try_reserve:
8320         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
8321                                      BTRFS_RESERVE_NO_FLUSH);
8322         if (!ret)
8323                 return block_rsv;
8324         /*
8325          * If we couldn't reserve metadata bytes try and use some from
8326          * the global reserve if its space type is the same as the global
8327          * reservation.
8328          */
8329         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
8330             block_rsv->space_info == global_rsv->space_info) {
8331                 ret = block_rsv_use_bytes(global_rsv, blocksize);
8332                 if (!ret)
8333                         return global_rsv;
8334         }
8335         return ERR_PTR(ret);
8336 }
8337
8338 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
8339                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
8340 {
8341         block_rsv_add_bytes(block_rsv, blocksize, 0);
8342         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
8343 }
8344
8345 /*
8346  * finds a free extent and does all the dirty work required for allocation
8347  * returns the tree buffer or an ERR_PTR on error.
8348  */
8349 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
8350                                         struct btrfs_root *root,
8351                                         u64 parent, u64 root_objectid,
8352                                         struct btrfs_disk_key *key, int level,
8353                                         u64 hint, u64 empty_size)
8354 {
8355         struct btrfs_fs_info *fs_info = root->fs_info;
8356         struct btrfs_key ins;
8357         struct btrfs_block_rsv *block_rsv;
8358         struct extent_buffer *buf;
8359         struct btrfs_delayed_extent_op *extent_op;
8360         u64 flags = 0;
8361         int ret;
8362         u32 blocksize = fs_info->nodesize;
8363         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
8364
8365 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
8366         if (btrfs_is_testing(fs_info)) {
8367                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
8368                                             level);
8369                 if (!IS_ERR(buf))
8370                         root->alloc_bytenr += blocksize;
8371                 return buf;
8372         }
8373 #endif
8374
8375         block_rsv = use_block_rsv(trans, root, blocksize);
8376         if (IS_ERR(block_rsv))
8377                 return ERR_CAST(block_rsv);
8378
8379         ret = btrfs_reserve_extent(root, blocksize, blocksize, blocksize,
8380                                    empty_size, hint, &ins, 0, 0);
8381         if (ret)
8382                 goto out_unuse;
8383
8384         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level);
8385         if (IS_ERR(buf)) {
8386                 ret = PTR_ERR(buf);
8387                 goto out_free_reserved;
8388         }
8389
8390         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
8391                 if (parent == 0)
8392                         parent = ins.objectid;
8393                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
8394         } else
8395                 BUG_ON(parent > 0);
8396
8397         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
8398                 extent_op = btrfs_alloc_delayed_extent_op();
8399                 if (!extent_op) {
8400                         ret = -ENOMEM;
8401                         goto out_free_buf;
8402                 }
8403                 if (key)
8404                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
8405                 else
8406                         memset(&extent_op->key, 0, sizeof(extent_op->key));
8407                 extent_op->flags_to_set = flags;
8408                 extent_op->update_key = skinny_metadata ? false : true;
8409                 extent_op->update_flags = true;
8410                 extent_op->is_data = false;
8411                 extent_op->level = level;
8412
8413                 ret = btrfs_add_delayed_tree_ref(fs_info, trans,
8414                                                  ins.objectid, ins.offset,
8415                                                  parent, root_objectid, level,
8416                                                  BTRFS_ADD_DELAYED_EXTENT,
8417                                                  extent_op);
8418                 if (ret)
8419                         goto out_free_delayed;
8420         }
8421         return buf;
8422
8423 out_free_delayed:
8424         btrfs_free_delayed_extent_op(extent_op);
8425 out_free_buf:
8426         free_extent_buffer(buf);
8427 out_free_reserved:
8428         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 0);
8429 out_unuse:
8430         unuse_block_rsv(fs_info, block_rsv, blocksize);
8431         return ERR_PTR(ret);
8432 }
8433
8434 struct walk_control {
8435         u64 refs[BTRFS_MAX_LEVEL];
8436         u64 flags[BTRFS_MAX_LEVEL];
8437         struct btrfs_key update_progress;
8438         int stage;
8439         int level;
8440         int shared_level;
8441         int update_ref;
8442         int keep_locks;
8443         int reada_slot;
8444         int reada_count;
8445         int for_reloc;
8446 };
8447
8448 #define DROP_REFERENCE  1
8449 #define UPDATE_BACKREF  2
8450
8451 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
8452                                      struct btrfs_root *root,
8453                                      struct walk_control *wc,
8454                                      struct btrfs_path *path)
8455 {
8456         struct btrfs_fs_info *fs_info = root->fs_info;
8457         u64 bytenr;
8458         u64 generation;
8459         u64 refs;
8460         u64 flags;
8461         u32 nritems;
8462         struct btrfs_key key;
8463         struct extent_buffer *eb;
8464         int ret;
8465         int slot;
8466         int nread = 0;
8467
8468         if (path->slots[wc->level] < wc->reada_slot) {
8469                 wc->reada_count = wc->reada_count * 2 / 3;
8470                 wc->reada_count = max(wc->reada_count, 2);
8471         } else {
8472                 wc->reada_count = wc->reada_count * 3 / 2;
8473                 wc->reada_count = min_t(int, wc->reada_count,
8474                                         BTRFS_NODEPTRS_PER_BLOCK(fs_info));
8475         }
8476
8477         eb = path->nodes[wc->level];
8478         nritems = btrfs_header_nritems(eb);
8479
8480         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
8481                 if (nread >= wc->reada_count)
8482                         break;
8483
8484                 cond_resched();
8485                 bytenr = btrfs_node_blockptr(eb, slot);
8486                 generation = btrfs_node_ptr_generation(eb, slot);
8487
8488                 if (slot == path->slots[wc->level])
8489                         goto reada;
8490
8491                 if (wc->stage == UPDATE_BACKREF &&
8492                     generation <= root->root_key.offset)
8493                         continue;
8494
8495                 /* We don't lock the tree block, it's OK to be racy here */
8496                 ret = btrfs_lookup_extent_info(trans, fs_info, bytenr,
8497                                                wc->level - 1, 1, &refs,
8498                                                &flags);
8499                 /* We don't care about errors in readahead. */
8500                 if (ret < 0)
8501                         continue;
8502                 BUG_ON(refs == 0);
8503
8504                 if (wc->stage == DROP_REFERENCE) {
8505                         if (refs == 1)
8506                                 goto reada;
8507
8508                         if (wc->level == 1 &&
8509                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8510                                 continue;
8511                         if (!wc->update_ref ||
8512                             generation <= root->root_key.offset)
8513                                 continue;
8514                         btrfs_node_key_to_cpu(eb, &key, slot);
8515                         ret = btrfs_comp_cpu_keys(&key,
8516                                                   &wc->update_progress);
8517                         if (ret < 0)
8518                                 continue;
8519                 } else {
8520                         if (wc->level == 1 &&
8521                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8522                                 continue;
8523                 }
8524 reada:
8525                 readahead_tree_block(fs_info, bytenr);
8526                 nread++;
8527         }
8528         wc->reada_slot = slot;
8529 }
8530
8531 /*
8532  * helper to process tree block while walking down the tree.
8533  *
8534  * when wc->stage == UPDATE_BACKREF, this function updates
8535  * back refs for pointers in the block.
8536  *
8537  * NOTE: return value 1 means we should stop walking down.
8538  */
8539 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
8540                                    struct btrfs_root *root,
8541                                    struct btrfs_path *path,
8542                                    struct walk_control *wc, int lookup_info)
8543 {
8544         struct btrfs_fs_info *fs_info = root->fs_info;
8545         int level = wc->level;
8546         struct extent_buffer *eb = path->nodes[level];
8547         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8548         int ret;
8549
8550         if (wc->stage == UPDATE_BACKREF &&
8551             btrfs_header_owner(eb) != root->root_key.objectid)
8552                 return 1;
8553
8554         /*
8555          * when reference count of tree block is 1, it won't increase
8556          * again. once full backref flag is set, we never clear it.
8557          */
8558         if (lookup_info &&
8559             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
8560              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
8561                 BUG_ON(!path->locks[level]);
8562                 ret = btrfs_lookup_extent_info(trans, fs_info,
8563                                                eb->start, level, 1,
8564                                                &wc->refs[level],
8565                                                &wc->flags[level]);
8566                 BUG_ON(ret == -ENOMEM);
8567                 if (ret)
8568                         return ret;
8569                 BUG_ON(wc->refs[level] == 0);
8570         }
8571
8572         if (wc->stage == DROP_REFERENCE) {
8573                 if (wc->refs[level] > 1)
8574                         return 1;
8575
8576                 if (path->locks[level] && !wc->keep_locks) {
8577                         btrfs_tree_unlock_rw(eb, path->locks[level]);
8578                         path->locks[level] = 0;
8579                 }
8580                 return 0;
8581         }
8582
8583         /* wc->stage == UPDATE_BACKREF */
8584         if (!(wc->flags[level] & flag)) {
8585                 BUG_ON(!path->locks[level]);
8586                 ret = btrfs_inc_ref(trans, root, eb, 1);
8587                 BUG_ON(ret); /* -ENOMEM */
8588                 ret = btrfs_dec_ref(trans, root, eb, 0);
8589                 BUG_ON(ret); /* -ENOMEM */
8590                 ret = btrfs_set_disk_extent_flags(trans, fs_info, eb->start,
8591                                                   eb->len, flag,
8592                                                   btrfs_header_level(eb), 0);
8593                 BUG_ON(ret); /* -ENOMEM */
8594                 wc->flags[level] |= flag;
8595         }
8596
8597         /*
8598          * the block is shared by multiple trees, so it's not good to
8599          * keep the tree lock
8600          */
8601         if (path->locks[level] && level > 0) {
8602                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8603                 path->locks[level] = 0;
8604         }
8605         return 0;
8606 }
8607
8608 /*
8609  * helper to process tree block pointer.
8610  *
8611  * when wc->stage == DROP_REFERENCE, this function checks
8612  * reference count of the block pointed to. if the block
8613  * is shared and we need update back refs for the subtree
8614  * rooted at the block, this function changes wc->stage to
8615  * UPDATE_BACKREF. if the block is shared and there is no
8616  * need to update back, this function drops the reference
8617  * to the block.
8618  *
8619  * NOTE: return value 1 means we should stop walking down.
8620  */
8621 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
8622                                  struct btrfs_root *root,
8623                                  struct btrfs_path *path,
8624                                  struct walk_control *wc, int *lookup_info)
8625 {
8626         struct btrfs_fs_info *fs_info = root->fs_info;
8627         u64 bytenr;
8628         u64 generation;
8629         u64 parent;
8630         u32 blocksize;
8631         struct btrfs_key key;
8632         struct extent_buffer *next;
8633         int level = wc->level;
8634         int reada = 0;
8635         int ret = 0;
8636         bool need_account = false;
8637
8638         generation = btrfs_node_ptr_generation(path->nodes[level],
8639                                                path->slots[level]);
8640         /*
8641          * if the lower level block was created before the snapshot
8642          * was created, we know there is no need to update back refs
8643          * for the subtree
8644          */
8645         if (wc->stage == UPDATE_BACKREF &&
8646             generation <= root->root_key.offset) {
8647                 *lookup_info = 1;
8648                 return 1;
8649         }
8650
8651         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
8652         blocksize = fs_info->nodesize;
8653
8654         next = find_extent_buffer(fs_info, bytenr);
8655         if (!next) {
8656                 next = btrfs_find_create_tree_block(fs_info, bytenr);
8657                 if (IS_ERR(next))
8658                         return PTR_ERR(next);
8659
8660                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
8661                                                level - 1);
8662                 reada = 1;
8663         }
8664         btrfs_tree_lock(next);
8665         btrfs_set_lock_blocking(next);
8666
8667         ret = btrfs_lookup_extent_info(trans, fs_info, bytenr, level - 1, 1,
8668                                        &wc->refs[level - 1],
8669                                        &wc->flags[level - 1]);
8670         if (ret < 0)
8671                 goto out_unlock;
8672
8673         if (unlikely(wc->refs[level - 1] == 0)) {
8674                 btrfs_err(fs_info, "Missing references.");
8675                 ret = -EIO;
8676                 goto out_unlock;
8677         }
8678         *lookup_info = 0;
8679
8680         if (wc->stage == DROP_REFERENCE) {
8681                 if (wc->refs[level - 1] > 1) {
8682                         need_account = true;
8683                         if (level == 1 &&
8684                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8685                                 goto skip;
8686
8687                         if (!wc->update_ref ||
8688                             generation <= root->root_key.offset)
8689                                 goto skip;
8690
8691                         btrfs_node_key_to_cpu(path->nodes[level], &key,
8692                                               path->slots[level]);
8693                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
8694                         if (ret < 0)
8695                                 goto skip;
8696
8697                         wc->stage = UPDATE_BACKREF;
8698                         wc->shared_level = level - 1;
8699                 }
8700         } else {
8701                 if (level == 1 &&
8702                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8703                         goto skip;
8704         }
8705
8706         if (!btrfs_buffer_uptodate(next, generation, 0)) {
8707                 btrfs_tree_unlock(next);
8708                 free_extent_buffer(next);
8709                 next = NULL;
8710                 *lookup_info = 1;
8711         }
8712
8713         if (!next) {
8714                 if (reada && level == 1)
8715                         reada_walk_down(trans, root, wc, path);
8716                 next = read_tree_block(fs_info, bytenr, generation);
8717                 if (IS_ERR(next)) {
8718                         return PTR_ERR(next);
8719                 } else if (!extent_buffer_uptodate(next)) {
8720                         free_extent_buffer(next);
8721                         return -EIO;
8722                 }
8723                 btrfs_tree_lock(next);
8724                 btrfs_set_lock_blocking(next);
8725         }
8726
8727         level--;
8728         ASSERT(level == btrfs_header_level(next));
8729         if (level != btrfs_header_level(next)) {
8730                 btrfs_err(root->fs_info, "mismatched level");
8731                 ret = -EIO;
8732                 goto out_unlock;
8733         }
8734         path->nodes[level] = next;
8735         path->slots[level] = 0;
8736         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8737         wc->level = level;
8738         if (wc->level == 1)
8739                 wc->reada_slot = 0;
8740         return 0;
8741 skip:
8742         wc->refs[level - 1] = 0;
8743         wc->flags[level - 1] = 0;
8744         if (wc->stage == DROP_REFERENCE) {
8745                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
8746                         parent = path->nodes[level]->start;
8747                 } else {
8748                         ASSERT(root->root_key.objectid ==
8749                                btrfs_header_owner(path->nodes[level]));
8750                         if (root->root_key.objectid !=
8751                             btrfs_header_owner(path->nodes[level])) {
8752                                 btrfs_err(root->fs_info,
8753                                                 "mismatched block owner");
8754                                 ret = -EIO;
8755                                 goto out_unlock;
8756                         }
8757                         parent = 0;
8758                 }
8759
8760                 if (need_account) {
8761                         ret = btrfs_qgroup_trace_subtree(trans, root, next,
8762                                                          generation, level - 1);
8763                         if (ret) {
8764                                 btrfs_err_rl(fs_info,
8765                                              "Error %d accounting shared subtree. Quota is out of sync, rescan required.",
8766                                              ret);
8767                         }
8768                 }
8769                 ret = btrfs_free_extent(trans, fs_info, bytenr, blocksize,
8770                                         parent, root->root_key.objectid,
8771                                         level - 1, 0);
8772                 if (ret)
8773                         goto out_unlock;
8774         }
8775
8776         *lookup_info = 1;
8777         ret = 1;
8778
8779 out_unlock:
8780         btrfs_tree_unlock(next);
8781         free_extent_buffer(next);
8782
8783         return ret;
8784 }
8785
8786 /*
8787  * helper to process tree block while walking up the tree.
8788  *
8789  * when wc->stage == DROP_REFERENCE, this function drops
8790  * reference count on the block.
8791  *
8792  * when wc->stage == UPDATE_BACKREF, this function changes
8793  * wc->stage back to DROP_REFERENCE if we changed wc->stage
8794  * to UPDATE_BACKREF previously while processing the block.
8795  *
8796  * NOTE: return value 1 means we should stop walking up.
8797  */
8798 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
8799                                  struct btrfs_root *root,
8800                                  struct btrfs_path *path,
8801                                  struct walk_control *wc)
8802 {
8803         struct btrfs_fs_info *fs_info = root->fs_info;
8804         int ret;
8805         int level = wc->level;
8806         struct extent_buffer *eb = path->nodes[level];
8807         u64 parent = 0;
8808
8809         if (wc->stage == UPDATE_BACKREF) {
8810                 BUG_ON(wc->shared_level < level);
8811                 if (level < wc->shared_level)
8812                         goto out;
8813
8814                 ret = find_next_key(path, level + 1, &wc->update_progress);
8815                 if (ret > 0)
8816                         wc->update_ref = 0;
8817
8818                 wc->stage = DROP_REFERENCE;
8819                 wc->shared_level = -1;
8820                 path->slots[level] = 0;
8821
8822                 /*
8823                  * check reference count again if the block isn't locked.
8824                  * we should start walking down the tree again if reference
8825                  * count is one.
8826                  */
8827                 if (!path->locks[level]) {
8828                         BUG_ON(level == 0);
8829                         btrfs_tree_lock(eb);
8830                         btrfs_set_lock_blocking(eb);
8831                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8832
8833                         ret = btrfs_lookup_extent_info(trans, fs_info,
8834                                                        eb->start, level, 1,
8835                                                        &wc->refs[level],
8836                                                        &wc->flags[level]);
8837                         if (ret < 0) {
8838                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8839                                 path->locks[level] = 0;
8840                                 return ret;
8841                         }
8842                         BUG_ON(wc->refs[level] == 0);
8843                         if (wc->refs[level] == 1) {
8844                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8845                                 path->locks[level] = 0;
8846                                 return 1;
8847                         }
8848                 }
8849         }
8850
8851         /* wc->stage == DROP_REFERENCE */
8852         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
8853
8854         if (wc->refs[level] == 1) {
8855                 if (level == 0) {
8856                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8857                                 ret = btrfs_dec_ref(trans, root, eb, 1);
8858                         else
8859                                 ret = btrfs_dec_ref(trans, root, eb, 0);
8860                         BUG_ON(ret); /* -ENOMEM */
8861                         ret = btrfs_qgroup_trace_leaf_items(trans, fs_info, eb);
8862                         if (ret) {
8863                                 btrfs_err_rl(fs_info,
8864                                              "error %d accounting leaf items. Quota is out of sync, rescan required.",
8865                                              ret);
8866                         }
8867                 }
8868                 /* make block locked assertion in clean_tree_block happy */
8869                 if (!path->locks[level] &&
8870                     btrfs_header_generation(eb) == trans->transid) {
8871                         btrfs_tree_lock(eb);
8872                         btrfs_set_lock_blocking(eb);
8873                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8874                 }
8875                 clean_tree_block(trans, fs_info, eb);
8876         }
8877
8878         if (eb == root->node) {
8879                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8880                         parent = eb->start;
8881                 else
8882                         BUG_ON(root->root_key.objectid !=
8883                                btrfs_header_owner(eb));
8884         } else {
8885                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8886                         parent = path->nodes[level + 1]->start;
8887                 else
8888                         BUG_ON(root->root_key.objectid !=
8889                                btrfs_header_owner(path->nodes[level + 1]));
8890         }
8891
8892         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
8893 out:
8894         wc->refs[level] = 0;
8895         wc->flags[level] = 0;
8896         return 0;
8897 }
8898
8899 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
8900                                    struct btrfs_root *root,
8901                                    struct btrfs_path *path,
8902                                    struct walk_control *wc)
8903 {
8904         int level = wc->level;
8905         int lookup_info = 1;
8906         int ret;
8907
8908         while (level >= 0) {
8909                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
8910                 if (ret > 0)
8911                         break;
8912
8913                 if (level == 0)
8914                         break;
8915
8916                 if (path->slots[level] >=
8917                     btrfs_header_nritems(path->nodes[level]))
8918                         break;
8919
8920                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
8921                 if (ret > 0) {
8922                         path->slots[level]++;
8923                         continue;
8924                 } else if (ret < 0)
8925                         return ret;
8926                 level = wc->level;
8927         }
8928         return 0;
8929 }
8930
8931 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
8932                                  struct btrfs_root *root,
8933                                  struct btrfs_path *path,
8934                                  struct walk_control *wc, int max_level)
8935 {
8936         int level = wc->level;
8937         int ret;
8938
8939         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
8940         while (level < max_level && path->nodes[level]) {
8941                 wc->level = level;
8942                 if (path->slots[level] + 1 <
8943                     btrfs_header_nritems(path->nodes[level])) {
8944                         path->slots[level]++;
8945                         return 0;
8946                 } else {
8947                         ret = walk_up_proc(trans, root, path, wc);
8948                         if (ret > 0)
8949                                 return 0;
8950
8951                         if (path->locks[level]) {
8952                                 btrfs_tree_unlock_rw(path->nodes[level],
8953                                                      path->locks[level]);
8954                                 path->locks[level] = 0;
8955                         }
8956                         free_extent_buffer(path->nodes[level]);
8957                         path->nodes[level] = NULL;
8958                         level++;
8959                 }
8960         }
8961         return 1;
8962 }
8963
8964 /*
8965  * drop a subvolume tree.
8966  *
8967  * this function traverses the tree freeing any blocks that only
8968  * referenced by the tree.
8969  *
8970  * when a shared tree block is found. this function decreases its
8971  * reference count by one. if update_ref is true, this function
8972  * also make sure backrefs for the shared block and all lower level
8973  * blocks are properly updated.
8974  *
8975  * If called with for_reloc == 0, may exit early with -EAGAIN
8976  */
8977 int btrfs_drop_snapshot(struct btrfs_root *root,
8978                          struct btrfs_block_rsv *block_rsv, int update_ref,
8979                          int for_reloc)
8980 {
8981         struct btrfs_fs_info *fs_info = root->fs_info;
8982         struct btrfs_path *path;
8983         struct btrfs_trans_handle *trans;
8984         struct btrfs_root *tree_root = fs_info->tree_root;
8985         struct btrfs_root_item *root_item = &root->root_item;
8986         struct walk_control *wc;
8987         struct btrfs_key key;
8988         int err = 0;
8989         int ret;
8990         int level;
8991         bool root_dropped = false;
8992
8993         btrfs_debug(fs_info, "Drop subvolume %llu", root->objectid);
8994
8995         path = btrfs_alloc_path();
8996         if (!path) {
8997                 err = -ENOMEM;
8998                 goto out;
8999         }
9000
9001         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9002         if (!wc) {
9003                 btrfs_free_path(path);
9004                 err = -ENOMEM;
9005                 goto out;
9006         }
9007
9008         trans = btrfs_start_transaction(tree_root, 0);
9009         if (IS_ERR(trans)) {
9010                 err = PTR_ERR(trans);
9011                 goto out_free;
9012         }
9013
9014         if (block_rsv)
9015                 trans->block_rsv = block_rsv;
9016
9017         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
9018                 level = btrfs_header_level(root->node);
9019                 path->nodes[level] = btrfs_lock_root_node(root);
9020                 btrfs_set_lock_blocking(path->nodes[level]);
9021                 path->slots[level] = 0;
9022                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9023                 memset(&wc->update_progress, 0,
9024                        sizeof(wc->update_progress));
9025         } else {
9026                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
9027                 memcpy(&wc->update_progress, &key,
9028                        sizeof(wc->update_progress));
9029
9030                 level = root_item->drop_level;
9031                 BUG_ON(level == 0);
9032                 path->lowest_level = level;
9033                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
9034                 path->lowest_level = 0;
9035                 if (ret < 0) {
9036                         err = ret;
9037                         goto out_end_trans;
9038                 }
9039                 WARN_ON(ret > 0);
9040
9041                 /*
9042                  * unlock our path, this is safe because only this
9043                  * function is allowed to delete this snapshot
9044                  */
9045                 btrfs_unlock_up_safe(path, 0);
9046
9047                 level = btrfs_header_level(root->node);
9048                 while (1) {
9049                         btrfs_tree_lock(path->nodes[level]);
9050                         btrfs_set_lock_blocking(path->nodes[level]);
9051                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9052
9053                         ret = btrfs_lookup_extent_info(trans, fs_info,
9054                                                 path->nodes[level]->start,
9055                                                 level, 1, &wc->refs[level],
9056                                                 &wc->flags[level]);
9057                         if (ret < 0) {
9058                                 err = ret;
9059                                 goto out_end_trans;
9060                         }
9061                         BUG_ON(wc->refs[level] == 0);
9062
9063                         if (level == root_item->drop_level)
9064                                 break;
9065
9066                         btrfs_tree_unlock(path->nodes[level]);
9067                         path->locks[level] = 0;
9068                         WARN_ON(wc->refs[level] != 1);
9069                         level--;
9070                 }
9071         }
9072
9073         wc->level = level;
9074         wc->shared_level = -1;
9075         wc->stage = DROP_REFERENCE;
9076         wc->update_ref = update_ref;
9077         wc->keep_locks = 0;
9078         wc->for_reloc = for_reloc;
9079         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
9080
9081         while (1) {
9082
9083                 ret = walk_down_tree(trans, root, path, wc);
9084                 if (ret < 0) {
9085                         err = ret;
9086                         break;
9087                 }
9088
9089                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
9090                 if (ret < 0) {
9091                         err = ret;
9092                         break;
9093                 }
9094
9095                 if (ret > 0) {
9096                         BUG_ON(wc->stage != DROP_REFERENCE);
9097                         break;
9098                 }
9099
9100                 if (wc->stage == DROP_REFERENCE) {
9101                         level = wc->level;
9102                         btrfs_node_key(path->nodes[level],
9103                                        &root_item->drop_progress,
9104                                        path->slots[level]);
9105                         root_item->drop_level = level;
9106                 }
9107
9108                 BUG_ON(wc->level == 0);
9109                 if (btrfs_should_end_transaction(trans) ||
9110                     (!for_reloc && btrfs_need_cleaner_sleep(fs_info))) {
9111                         ret = btrfs_update_root(trans, tree_root,
9112                                                 &root->root_key,
9113                                                 root_item);
9114                         if (ret) {
9115                                 btrfs_abort_transaction(trans, ret);
9116                                 err = ret;
9117                                 goto out_end_trans;
9118                         }
9119
9120                         btrfs_end_transaction_throttle(trans);
9121                         if (!for_reloc && btrfs_need_cleaner_sleep(fs_info)) {
9122                                 btrfs_debug(fs_info,
9123                                             "drop snapshot early exit");
9124                                 err = -EAGAIN;
9125                                 goto out_free;
9126                         }
9127
9128                         trans = btrfs_start_transaction(tree_root, 0);
9129                         if (IS_ERR(trans)) {
9130                                 err = PTR_ERR(trans);
9131                                 goto out_free;
9132                         }
9133                         if (block_rsv)
9134                                 trans->block_rsv = block_rsv;
9135                 }
9136         }
9137         btrfs_release_path(path);
9138         if (err)
9139                 goto out_end_trans;
9140
9141         ret = btrfs_del_root(trans, tree_root, &root->root_key);
9142         if (ret) {
9143                 btrfs_abort_transaction(trans, ret);
9144                 goto out_end_trans;
9145         }
9146
9147         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
9148                 ret = btrfs_find_root(tree_root, &root->root_key, path,
9149                                       NULL, NULL);
9150                 if (ret < 0) {
9151                         btrfs_abort_transaction(trans, ret);
9152                         err = ret;
9153                         goto out_end_trans;
9154                 } else if (ret > 0) {
9155                         /* if we fail to delete the orphan item this time
9156                          * around, it'll get picked up the next time.
9157                          *
9158                          * The most common failure here is just -ENOENT.
9159                          */
9160                         btrfs_del_orphan_item(trans, tree_root,
9161                                               root->root_key.objectid);
9162                 }
9163         }
9164
9165         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
9166                 btrfs_add_dropped_root(trans, root);
9167         } else {
9168                 free_extent_buffer(root->node);
9169                 free_extent_buffer(root->commit_root);
9170                 btrfs_put_fs_root(root);
9171         }
9172         root_dropped = true;
9173 out_end_trans:
9174         btrfs_end_transaction_throttle(trans);
9175 out_free:
9176         kfree(wc);
9177         btrfs_free_path(path);
9178 out:
9179         /*
9180          * So if we need to stop dropping the snapshot for whatever reason we
9181          * need to make sure to add it back to the dead root list so that we
9182          * keep trying to do the work later.  This also cleans up roots if we
9183          * don't have it in the radix (like when we recover after a power fail
9184          * or unmount) so we don't leak memory.
9185          */
9186         if (!for_reloc && root_dropped == false)
9187                 btrfs_add_dead_root(root);
9188         if (err && err != -EAGAIN)
9189                 btrfs_handle_fs_error(fs_info, err, NULL);
9190         return err;
9191 }
9192
9193 /*
9194  * drop subtree rooted at tree block 'node'.
9195  *
9196  * NOTE: this function will unlock and release tree block 'node'
9197  * only used by relocation code
9198  */
9199 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
9200                         struct btrfs_root *root,
9201                         struct extent_buffer *node,
9202                         struct extent_buffer *parent)
9203 {
9204         struct btrfs_fs_info *fs_info = root->fs_info;
9205         struct btrfs_path *path;
9206         struct walk_control *wc;
9207         int level;
9208         int parent_level;
9209         int ret = 0;
9210         int wret;
9211
9212         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
9213
9214         path = btrfs_alloc_path();
9215         if (!path)
9216                 return -ENOMEM;
9217
9218         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9219         if (!wc) {
9220                 btrfs_free_path(path);
9221                 return -ENOMEM;
9222         }
9223
9224         btrfs_assert_tree_locked(parent);
9225         parent_level = btrfs_header_level(parent);
9226         extent_buffer_get(parent);
9227         path->nodes[parent_level] = parent;
9228         path->slots[parent_level] = btrfs_header_nritems(parent);
9229
9230         btrfs_assert_tree_locked(node);
9231         level = btrfs_header_level(node);
9232         path->nodes[level] = node;
9233         path->slots[level] = 0;
9234         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9235
9236         wc->refs[parent_level] = 1;
9237         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
9238         wc->level = level;
9239         wc->shared_level = -1;
9240         wc->stage = DROP_REFERENCE;
9241         wc->update_ref = 0;
9242         wc->keep_locks = 1;
9243         wc->for_reloc = 1;
9244         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
9245
9246         while (1) {
9247                 wret = walk_down_tree(trans, root, path, wc);
9248                 if (wret < 0) {
9249                         ret = wret;
9250                         break;
9251                 }
9252
9253                 wret = walk_up_tree(trans, root, path, wc, parent_level);
9254                 if (wret < 0)
9255                         ret = wret;
9256                 if (wret != 0)
9257                         break;
9258         }
9259
9260         kfree(wc);
9261         btrfs_free_path(path);
9262         return ret;
9263 }
9264
9265 static u64 update_block_group_flags(struct btrfs_fs_info *fs_info, u64 flags)
9266 {
9267         u64 num_devices;
9268         u64 stripped;
9269
9270         /*
9271          * if restripe for this chunk_type is on pick target profile and
9272          * return, otherwise do the usual balance
9273          */
9274         stripped = get_restripe_target(fs_info, flags);
9275         if (stripped)
9276                 return extended_to_chunk(stripped);
9277
9278         num_devices = fs_info->fs_devices->rw_devices;
9279
9280         stripped = BTRFS_BLOCK_GROUP_RAID0 |
9281                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
9282                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
9283
9284         if (num_devices == 1) {
9285                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9286                 stripped = flags & ~stripped;
9287
9288                 /* turn raid0 into single device chunks */
9289                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
9290                         return stripped;
9291
9292                 /* turn mirroring into duplication */
9293                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
9294                              BTRFS_BLOCK_GROUP_RAID10))
9295                         return stripped | BTRFS_BLOCK_GROUP_DUP;
9296         } else {
9297                 /* they already had raid on here, just return */
9298                 if (flags & stripped)
9299                         return flags;
9300
9301                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9302                 stripped = flags & ~stripped;
9303
9304                 /* switch duplicated blocks with raid1 */
9305                 if (flags & BTRFS_BLOCK_GROUP_DUP)
9306                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
9307
9308                 /* this is drive concat, leave it alone */
9309         }
9310
9311         return flags;
9312 }
9313
9314 static int inc_block_group_ro(struct btrfs_block_group_cache *cache, int force)
9315 {
9316         struct btrfs_space_info *sinfo = cache->space_info;
9317         u64 num_bytes;
9318         u64 min_allocable_bytes;
9319         int ret = -ENOSPC;
9320
9321         /*
9322          * We need some metadata space and system metadata space for
9323          * allocating chunks in some corner cases until we force to set
9324          * it to be readonly.
9325          */
9326         if ((sinfo->flags &
9327              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
9328             !force)
9329                 min_allocable_bytes = SZ_1M;
9330         else
9331                 min_allocable_bytes = 0;
9332
9333         spin_lock(&sinfo->lock);
9334         spin_lock(&cache->lock);
9335
9336         if (cache->ro) {
9337                 cache->ro++;
9338                 ret = 0;
9339                 goto out;
9340         }
9341
9342         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
9343                     cache->bytes_super - btrfs_block_group_used(&cache->item);
9344
9345         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
9346             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
9347             min_allocable_bytes <= sinfo->total_bytes) {
9348                 sinfo->bytes_readonly += num_bytes;
9349                 cache->ro++;
9350                 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
9351                 ret = 0;
9352         }
9353 out:
9354         spin_unlock(&cache->lock);
9355         spin_unlock(&sinfo->lock);
9356         return ret;
9357 }
9358
9359 int btrfs_inc_block_group_ro(struct btrfs_root *root,
9360                              struct btrfs_block_group_cache *cache)
9361
9362 {
9363         struct btrfs_fs_info *fs_info = root->fs_info;
9364         struct btrfs_trans_handle *trans;
9365         u64 alloc_flags;
9366         int ret;
9367
9368 again:
9369         trans = btrfs_join_transaction(root);
9370         if (IS_ERR(trans))
9371                 return PTR_ERR(trans);
9372
9373         /*
9374          * we're not allowed to set block groups readonly after the dirty
9375          * block groups cache has started writing.  If it already started,
9376          * back off and let this transaction commit
9377          */
9378         mutex_lock(&fs_info->ro_block_group_mutex);
9379         if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
9380                 u64 transid = trans->transid;
9381
9382                 mutex_unlock(&fs_info->ro_block_group_mutex);
9383                 btrfs_end_transaction(trans);
9384
9385                 ret = btrfs_wait_for_commit(fs_info, transid);
9386                 if (ret)
9387                         return ret;
9388                 goto again;
9389         }
9390
9391         /*
9392          * if we are changing raid levels, try to allocate a corresponding
9393          * block group with the new raid level.
9394          */
9395         alloc_flags = update_block_group_flags(fs_info, cache->flags);
9396         if (alloc_flags != cache->flags) {
9397                 ret = do_chunk_alloc(trans, fs_info, alloc_flags,
9398                                      CHUNK_ALLOC_FORCE);
9399                 /*
9400                  * ENOSPC is allowed here, we may have enough space
9401                  * already allocated at the new raid level to
9402                  * carry on
9403                  */
9404                 if (ret == -ENOSPC)
9405                         ret = 0;
9406                 if (ret < 0)
9407                         goto out;
9408         }
9409
9410         ret = inc_block_group_ro(cache, 0);
9411         if (!ret)
9412                 goto out;
9413         alloc_flags = get_alloc_profile(fs_info, cache->space_info->flags);
9414         ret = do_chunk_alloc(trans, fs_info, alloc_flags,
9415                              CHUNK_ALLOC_FORCE);
9416         if (ret < 0)
9417                 goto out;
9418         ret = inc_block_group_ro(cache, 0);
9419 out:
9420         if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
9421                 alloc_flags = update_block_group_flags(fs_info, cache->flags);
9422                 mutex_lock(&fs_info->chunk_mutex);
9423                 check_system_chunk(trans, fs_info, alloc_flags);
9424                 mutex_unlock(&fs_info->chunk_mutex);
9425         }
9426         mutex_unlock(&fs_info->ro_block_group_mutex);
9427
9428         btrfs_end_transaction(trans);
9429         return ret;
9430 }
9431
9432 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
9433                             struct btrfs_fs_info *fs_info, u64 type)
9434 {
9435         u64 alloc_flags = get_alloc_profile(fs_info, type);
9436
9437         return do_chunk_alloc(trans, fs_info, alloc_flags, CHUNK_ALLOC_FORCE);
9438 }
9439
9440 /*
9441  * helper to account the unused space of all the readonly block group in the
9442  * space_info. takes mirrors into account.
9443  */
9444 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
9445 {
9446         struct btrfs_block_group_cache *block_group;
9447         u64 free_bytes = 0;
9448         int factor;
9449
9450         /* It's df, we don't care if it's racy */
9451         if (list_empty(&sinfo->ro_bgs))
9452                 return 0;
9453
9454         spin_lock(&sinfo->lock);
9455         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
9456                 spin_lock(&block_group->lock);
9457
9458                 if (!block_group->ro) {
9459                         spin_unlock(&block_group->lock);
9460                         continue;
9461                 }
9462
9463                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
9464                                           BTRFS_BLOCK_GROUP_RAID10 |
9465                                           BTRFS_BLOCK_GROUP_DUP))
9466                         factor = 2;
9467                 else
9468                         factor = 1;
9469
9470                 free_bytes += (block_group->key.offset -
9471                                btrfs_block_group_used(&block_group->item)) *
9472                                factor;
9473
9474                 spin_unlock(&block_group->lock);
9475         }
9476         spin_unlock(&sinfo->lock);
9477
9478         return free_bytes;
9479 }
9480
9481 void btrfs_dec_block_group_ro(struct btrfs_block_group_cache *cache)
9482 {
9483         struct btrfs_space_info *sinfo = cache->space_info;
9484         u64 num_bytes;
9485
9486         BUG_ON(!cache->ro);
9487
9488         spin_lock(&sinfo->lock);
9489         spin_lock(&cache->lock);
9490         if (!--cache->ro) {
9491                 num_bytes = cache->key.offset - cache->reserved -
9492                             cache->pinned - cache->bytes_super -
9493                             btrfs_block_group_used(&cache->item);
9494                 sinfo->bytes_readonly -= num_bytes;
9495                 list_del_init(&cache->ro_list);
9496         }
9497         spin_unlock(&cache->lock);
9498         spin_unlock(&sinfo->lock);
9499 }
9500
9501 /*
9502  * checks to see if its even possible to relocate this block group.
9503  *
9504  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
9505  * ok to go ahead and try.
9506  */
9507 int btrfs_can_relocate(struct btrfs_fs_info *fs_info, u64 bytenr)
9508 {
9509         struct btrfs_root *root = fs_info->extent_root;
9510         struct btrfs_block_group_cache *block_group;
9511         struct btrfs_space_info *space_info;
9512         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
9513         struct btrfs_device *device;
9514         struct btrfs_trans_handle *trans;
9515         u64 min_free;
9516         u64 dev_min = 1;
9517         u64 dev_nr = 0;
9518         u64 target;
9519         int debug;
9520         int index;
9521         int full = 0;
9522         int ret = 0;
9523
9524         debug = btrfs_test_opt(fs_info, ENOSPC_DEBUG);
9525
9526         block_group = btrfs_lookup_block_group(fs_info, bytenr);
9527
9528         /* odd, couldn't find the block group, leave it alone */
9529         if (!block_group) {
9530                 if (debug)
9531                         btrfs_warn(fs_info,
9532                                    "can't find block group for bytenr %llu",
9533                                    bytenr);
9534                 return -1;
9535         }
9536
9537         min_free = btrfs_block_group_used(&block_group->item);
9538
9539         /* no bytes used, we're good */
9540         if (!min_free)
9541                 goto out;
9542
9543         space_info = block_group->space_info;
9544         spin_lock(&space_info->lock);
9545
9546         full = space_info->full;
9547
9548         /*
9549          * if this is the last block group we have in this space, we can't
9550          * relocate it unless we're able to allocate a new chunk below.
9551          *
9552          * Otherwise, we need to make sure we have room in the space to handle
9553          * all of the extents from this block group.  If we can, we're good
9554          */
9555         if ((space_info->total_bytes != block_group->key.offset) &&
9556             (space_info->bytes_used + space_info->bytes_reserved +
9557              space_info->bytes_pinned + space_info->bytes_readonly +
9558              min_free < space_info->total_bytes)) {
9559                 spin_unlock(&space_info->lock);
9560                 goto out;
9561         }
9562         spin_unlock(&space_info->lock);
9563
9564         /*
9565          * ok we don't have enough space, but maybe we have free space on our
9566          * devices to allocate new chunks for relocation, so loop through our
9567          * alloc devices and guess if we have enough space.  if this block
9568          * group is going to be restriped, run checks against the target
9569          * profile instead of the current one.
9570          */
9571         ret = -1;
9572
9573         /*
9574          * index:
9575          *      0: raid10
9576          *      1: raid1
9577          *      2: dup
9578          *      3: raid0
9579          *      4: single
9580          */
9581         target = get_restripe_target(fs_info, block_group->flags);
9582         if (target) {
9583                 index = __get_raid_index(extended_to_chunk(target));
9584         } else {
9585                 /*
9586                  * this is just a balance, so if we were marked as full
9587                  * we know there is no space for a new chunk
9588                  */
9589                 if (full) {
9590                         if (debug)
9591                                 btrfs_warn(fs_info,
9592                                            "no space to alloc new chunk for block group %llu",
9593                                            block_group->key.objectid);
9594                         goto out;
9595                 }
9596
9597                 index = get_block_group_index(block_group);
9598         }
9599
9600         if (index == BTRFS_RAID_RAID10) {
9601                 dev_min = 4;
9602                 /* Divide by 2 */
9603                 min_free >>= 1;
9604         } else if (index == BTRFS_RAID_RAID1) {
9605                 dev_min = 2;
9606         } else if (index == BTRFS_RAID_DUP) {
9607                 /* Multiply by 2 */
9608                 min_free <<= 1;
9609         } else if (index == BTRFS_RAID_RAID0) {
9610                 dev_min = fs_devices->rw_devices;
9611                 min_free = div64_u64(min_free, dev_min);
9612         }
9613
9614         /* We need to do this so that we can look at pending chunks */
9615         trans = btrfs_join_transaction(root);
9616         if (IS_ERR(trans)) {
9617                 ret = PTR_ERR(trans);
9618                 goto out;
9619         }
9620
9621         mutex_lock(&fs_info->chunk_mutex);
9622         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
9623                 u64 dev_offset;
9624
9625                 /*
9626                  * check to make sure we can actually find a chunk with enough
9627                  * space to fit our block group in.
9628                  */
9629                 if (device->total_bytes > device->bytes_used + min_free &&
9630                     !device->is_tgtdev_for_dev_replace) {
9631                         ret = find_free_dev_extent(trans, device, min_free,
9632                                                    &dev_offset, NULL);
9633                         if (!ret)
9634                                 dev_nr++;
9635
9636                         if (dev_nr >= dev_min)
9637                                 break;
9638
9639                         ret = -1;
9640                 }
9641         }
9642         if (debug && ret == -1)
9643                 btrfs_warn(fs_info,
9644                            "no space to allocate a new chunk for block group %llu",
9645                            block_group->key.objectid);
9646         mutex_unlock(&fs_info->chunk_mutex);
9647         btrfs_end_transaction(trans);
9648 out:
9649         btrfs_put_block_group(block_group);
9650         return ret;
9651 }
9652
9653 static int find_first_block_group(struct btrfs_fs_info *fs_info,
9654                                   struct btrfs_path *path,
9655                                   struct btrfs_key *key)
9656 {
9657         struct btrfs_root *root = fs_info->extent_root;
9658         int ret = 0;
9659         struct btrfs_key found_key;
9660         struct extent_buffer *leaf;
9661         int slot;
9662
9663         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
9664         if (ret < 0)
9665                 goto out;
9666
9667         while (1) {
9668                 slot = path->slots[0];
9669                 leaf = path->nodes[0];
9670                 if (slot >= btrfs_header_nritems(leaf)) {
9671                         ret = btrfs_next_leaf(root, path);
9672                         if (ret == 0)
9673                                 continue;
9674                         if (ret < 0)
9675                                 goto out;
9676                         break;
9677                 }
9678                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
9679
9680                 if (found_key.objectid >= key->objectid &&
9681                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
9682                         struct extent_map_tree *em_tree;
9683                         struct extent_map *em;
9684
9685                         em_tree = &root->fs_info->mapping_tree.map_tree;
9686                         read_lock(&em_tree->lock);
9687                         em = lookup_extent_mapping(em_tree, found_key.objectid,
9688                                                    found_key.offset);
9689                         read_unlock(&em_tree->lock);
9690                         if (!em) {
9691                                 btrfs_err(fs_info,
9692                         "logical %llu len %llu found bg but no related chunk",
9693                                           found_key.objectid, found_key.offset);
9694                                 ret = -ENOENT;
9695                         } else {
9696                                 ret = 0;
9697                         }
9698                         free_extent_map(em);
9699                         goto out;
9700                 }
9701                 path->slots[0]++;
9702         }
9703 out:
9704         return ret;
9705 }
9706
9707 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
9708 {
9709         struct btrfs_block_group_cache *block_group;
9710         u64 last = 0;
9711
9712         while (1) {
9713                 struct inode *inode;
9714
9715                 block_group = btrfs_lookup_first_block_group(info, last);
9716                 while (block_group) {
9717                         spin_lock(&block_group->lock);
9718                         if (block_group->iref)
9719                                 break;
9720                         spin_unlock(&block_group->lock);
9721                         block_group = next_block_group(info, block_group);
9722                 }
9723                 if (!block_group) {
9724                         if (last == 0)
9725                                 break;
9726                         last = 0;
9727                         continue;
9728                 }
9729
9730                 inode = block_group->inode;
9731                 block_group->iref = 0;
9732                 block_group->inode = NULL;
9733                 spin_unlock(&block_group->lock);
9734                 ASSERT(block_group->io_ctl.inode == NULL);
9735                 iput(inode);
9736                 last = block_group->key.objectid + block_group->key.offset;
9737                 btrfs_put_block_group(block_group);
9738         }
9739 }
9740
9741 int btrfs_free_block_groups(struct btrfs_fs_info *info)
9742 {
9743         struct btrfs_block_group_cache *block_group;
9744         struct btrfs_space_info *space_info;
9745         struct btrfs_caching_control *caching_ctl;
9746         struct rb_node *n;
9747
9748         down_write(&info->commit_root_sem);
9749         while (!list_empty(&info->caching_block_groups)) {
9750                 caching_ctl = list_entry(info->caching_block_groups.next,
9751                                          struct btrfs_caching_control, list);
9752                 list_del(&caching_ctl->list);
9753                 put_caching_control(caching_ctl);
9754         }
9755         up_write(&info->commit_root_sem);
9756
9757         spin_lock(&info->unused_bgs_lock);
9758         while (!list_empty(&info->unused_bgs)) {
9759                 block_group = list_first_entry(&info->unused_bgs,
9760                                                struct btrfs_block_group_cache,
9761                                                bg_list);
9762                 list_del_init(&block_group->bg_list);
9763                 btrfs_put_block_group(block_group);
9764         }
9765         spin_unlock(&info->unused_bgs_lock);
9766
9767         spin_lock(&info->block_group_cache_lock);
9768         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
9769                 block_group = rb_entry(n, struct btrfs_block_group_cache,
9770                                        cache_node);
9771                 rb_erase(&block_group->cache_node,
9772                          &info->block_group_cache_tree);
9773                 RB_CLEAR_NODE(&block_group->cache_node);
9774                 spin_unlock(&info->block_group_cache_lock);
9775
9776                 down_write(&block_group->space_info->groups_sem);
9777                 list_del(&block_group->list);
9778                 up_write(&block_group->space_info->groups_sem);
9779
9780                 if (block_group->cached == BTRFS_CACHE_STARTED)
9781                         wait_block_group_cache_done(block_group);
9782
9783                 /*
9784                  * We haven't cached this block group, which means we could
9785                  * possibly have excluded extents on this block group.
9786                  */
9787                 if (block_group->cached == BTRFS_CACHE_NO ||
9788                     block_group->cached == BTRFS_CACHE_ERROR)
9789                         free_excluded_extents(info, block_group);
9790
9791                 btrfs_remove_free_space_cache(block_group);
9792                 ASSERT(list_empty(&block_group->dirty_list));
9793                 ASSERT(list_empty(&block_group->io_list));
9794                 ASSERT(list_empty(&block_group->bg_list));
9795                 ASSERT(atomic_read(&block_group->count) == 1);
9796                 btrfs_put_block_group(block_group);
9797
9798                 spin_lock(&info->block_group_cache_lock);
9799         }
9800         spin_unlock(&info->block_group_cache_lock);
9801
9802         /* now that all the block groups are freed, go through and
9803          * free all the space_info structs.  This is only called during
9804          * the final stages of unmount, and so we know nobody is
9805          * using them.  We call synchronize_rcu() once before we start,
9806          * just to be on the safe side.
9807          */
9808         synchronize_rcu();
9809
9810         release_global_block_rsv(info);
9811
9812         while (!list_empty(&info->space_info)) {
9813                 int i;
9814
9815                 space_info = list_entry(info->space_info.next,
9816                                         struct btrfs_space_info,
9817                                         list);
9818
9819                 /*
9820                  * Do not hide this behind enospc_debug, this is actually
9821                  * important and indicates a real bug if this happens.
9822                  */
9823                 if (WARN_ON(space_info->bytes_pinned > 0 ||
9824                             space_info->bytes_reserved > 0 ||
9825                             space_info->bytes_may_use > 0))
9826                         dump_space_info(info, space_info, 0, 0);
9827                 list_del(&space_info->list);
9828                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
9829                         struct kobject *kobj;
9830                         kobj = space_info->block_group_kobjs[i];
9831                         space_info->block_group_kobjs[i] = NULL;
9832                         if (kobj) {
9833                                 kobject_del(kobj);
9834                                 kobject_put(kobj);
9835                         }
9836                 }
9837                 kobject_del(&space_info->kobj);
9838                 kobject_put(&space_info->kobj);
9839         }
9840         return 0;
9841 }
9842
9843 static void __link_block_group(struct btrfs_space_info *space_info,
9844                                struct btrfs_block_group_cache *cache)
9845 {
9846         int index = get_block_group_index(cache);
9847         bool first = false;
9848
9849         down_write(&space_info->groups_sem);
9850         if (list_empty(&space_info->block_groups[index]))
9851                 first = true;
9852         list_add_tail(&cache->list, &space_info->block_groups[index]);
9853         up_write(&space_info->groups_sem);
9854
9855         if (first) {
9856                 struct raid_kobject *rkobj;
9857                 int ret;
9858
9859                 rkobj = kzalloc(sizeof(*rkobj), GFP_NOFS);
9860                 if (!rkobj)
9861                         goto out_err;
9862                 rkobj->raid_type = index;
9863                 kobject_init(&rkobj->kobj, &btrfs_raid_ktype);
9864                 ret = kobject_add(&rkobj->kobj, &space_info->kobj,
9865                                   "%s", get_raid_name(index));
9866                 if (ret) {
9867                         kobject_put(&rkobj->kobj);
9868                         goto out_err;
9869                 }
9870                 space_info->block_group_kobjs[index] = &rkobj->kobj;
9871         }
9872
9873         return;
9874 out_err:
9875         btrfs_warn(cache->fs_info,
9876                    "failed to add kobject for block cache, ignoring");
9877 }
9878
9879 static struct btrfs_block_group_cache *
9880 btrfs_create_block_group_cache(struct btrfs_fs_info *fs_info,
9881                                u64 start, u64 size)
9882 {
9883         struct btrfs_block_group_cache *cache;
9884
9885         cache = kzalloc(sizeof(*cache), GFP_NOFS);
9886         if (!cache)
9887                 return NULL;
9888
9889         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
9890                                         GFP_NOFS);
9891         if (!cache->free_space_ctl) {
9892                 kfree(cache);
9893                 return NULL;
9894         }
9895
9896         cache->key.objectid = start;
9897         cache->key.offset = size;
9898         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
9899
9900         cache->sectorsize = fs_info->sectorsize;
9901         cache->fs_info = fs_info;
9902         cache->full_stripe_len = btrfs_full_stripe_len(fs_info,
9903                                                        &fs_info->mapping_tree,
9904                                                        start);
9905         set_free_space_tree_thresholds(cache);
9906
9907         atomic_set(&cache->count, 1);
9908         spin_lock_init(&cache->lock);
9909         init_rwsem(&cache->data_rwsem);
9910         INIT_LIST_HEAD(&cache->list);
9911         INIT_LIST_HEAD(&cache->cluster_list);
9912         INIT_LIST_HEAD(&cache->bg_list);
9913         INIT_LIST_HEAD(&cache->ro_list);
9914         INIT_LIST_HEAD(&cache->dirty_list);
9915         INIT_LIST_HEAD(&cache->io_list);
9916         btrfs_init_free_space_ctl(cache);
9917         atomic_set(&cache->trimming, 0);
9918         mutex_init(&cache->free_space_lock);
9919
9920         return cache;
9921 }
9922
9923 int btrfs_read_block_groups(struct btrfs_fs_info *info)
9924 {
9925         struct btrfs_path *path;
9926         int ret;
9927         struct btrfs_block_group_cache *cache;
9928         struct btrfs_space_info *space_info;
9929         struct btrfs_key key;
9930         struct btrfs_key found_key;
9931         struct extent_buffer *leaf;
9932         int need_clear = 0;
9933         u64 cache_gen;
9934         u64 feature;
9935         int mixed;
9936
9937         feature = btrfs_super_incompat_flags(info->super_copy);
9938         mixed = !!(feature & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS);
9939
9940         key.objectid = 0;
9941         key.offset = 0;
9942         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
9943         path = btrfs_alloc_path();
9944         if (!path)
9945                 return -ENOMEM;
9946         path->reada = READA_FORWARD;
9947
9948         cache_gen = btrfs_super_cache_generation(info->super_copy);
9949         if (btrfs_test_opt(info, SPACE_CACHE) &&
9950             btrfs_super_generation(info->super_copy) != cache_gen)
9951                 need_clear = 1;
9952         if (btrfs_test_opt(info, CLEAR_CACHE))
9953                 need_clear = 1;
9954
9955         while (1) {
9956                 ret = find_first_block_group(info, path, &key);
9957                 if (ret > 0)
9958                         break;
9959                 if (ret != 0)
9960                         goto error;
9961
9962                 leaf = path->nodes[0];
9963                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
9964
9965                 cache = btrfs_create_block_group_cache(info, found_key.objectid,
9966                                                        found_key.offset);
9967                 if (!cache) {
9968                         ret = -ENOMEM;
9969                         goto error;
9970                 }
9971
9972                 if (need_clear) {
9973                         /*
9974                          * When we mount with old space cache, we need to
9975                          * set BTRFS_DC_CLEAR and set dirty flag.
9976                          *
9977                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
9978                          *    truncate the old free space cache inode and
9979                          *    setup a new one.
9980                          * b) Setting 'dirty flag' makes sure that we flush
9981                          *    the new space cache info onto disk.
9982                          */
9983                         if (btrfs_test_opt(info, SPACE_CACHE))
9984                                 cache->disk_cache_state = BTRFS_DC_CLEAR;
9985                 }
9986
9987                 read_extent_buffer(leaf, &cache->item,
9988                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
9989                                    sizeof(cache->item));
9990                 cache->flags = btrfs_block_group_flags(&cache->item);
9991                 if (!mixed &&
9992                     ((cache->flags & BTRFS_BLOCK_GROUP_METADATA) &&
9993                     (cache->flags & BTRFS_BLOCK_GROUP_DATA))) {
9994                         btrfs_err(info,
9995 "bg %llu is a mixed block group but filesystem hasn't enabled mixed block groups",
9996                                   cache->key.objectid);
9997                         ret = -EINVAL;
9998                         goto error;
9999                 }
10000
10001                 key.objectid = found_key.objectid + found_key.offset;
10002                 btrfs_release_path(path);
10003
10004                 /*
10005                  * We need to exclude the super stripes now so that the space
10006                  * info has super bytes accounted for, otherwise we'll think
10007                  * we have more space than we actually do.
10008                  */
10009                 ret = exclude_super_stripes(info, cache);
10010                 if (ret) {
10011                         /*
10012                          * We may have excluded something, so call this just in
10013                          * case.
10014                          */
10015                         free_excluded_extents(info, cache);
10016                         btrfs_put_block_group(cache);
10017                         goto error;
10018                 }
10019
10020                 /*
10021                  * check for two cases, either we are full, and therefore
10022                  * don't need to bother with the caching work since we won't
10023                  * find any space, or we are empty, and we can just add all
10024                  * the space in and be done with it.  This saves us _alot_ of
10025                  * time, particularly in the full case.
10026                  */
10027                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
10028                         cache->last_byte_to_unpin = (u64)-1;
10029                         cache->cached = BTRFS_CACHE_FINISHED;
10030                         free_excluded_extents(info, cache);
10031                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10032                         cache->last_byte_to_unpin = (u64)-1;
10033                         cache->cached = BTRFS_CACHE_FINISHED;
10034                         add_new_free_space(cache, info,
10035                                            found_key.objectid,
10036                                            found_key.objectid +
10037                                            found_key.offset);
10038                         free_excluded_extents(info, cache);
10039                 }
10040
10041                 ret = btrfs_add_block_group_cache(info, cache);
10042                 if (ret) {
10043                         btrfs_remove_free_space_cache(cache);
10044                         btrfs_put_block_group(cache);
10045                         goto error;
10046                 }
10047
10048                 trace_btrfs_add_block_group(info, cache, 0);
10049                 ret = update_space_info(info, cache->flags, found_key.offset,
10050                                         btrfs_block_group_used(&cache->item),
10051                                         cache->bytes_super, &space_info);
10052                 if (ret) {
10053                         btrfs_remove_free_space_cache(cache);
10054                         spin_lock(&info->block_group_cache_lock);
10055                         rb_erase(&cache->cache_node,
10056                                  &info->block_group_cache_tree);
10057                         RB_CLEAR_NODE(&cache->cache_node);
10058                         spin_unlock(&info->block_group_cache_lock);
10059                         btrfs_put_block_group(cache);
10060                         goto error;
10061                 }
10062
10063                 cache->space_info = space_info;
10064
10065                 __link_block_group(space_info, cache);
10066
10067                 set_avail_alloc_bits(info, cache->flags);
10068                 if (btrfs_chunk_readonly(info, cache->key.objectid)) {
10069                         inc_block_group_ro(cache, 1);
10070                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10071                         spin_lock(&info->unused_bgs_lock);
10072                         /* Should always be true but just in case. */
10073                         if (list_empty(&cache->bg_list)) {
10074                                 btrfs_get_block_group(cache);
10075                                 list_add_tail(&cache->bg_list,
10076                                               &info->unused_bgs);
10077                         }
10078                         spin_unlock(&info->unused_bgs_lock);
10079                 }
10080         }
10081
10082         list_for_each_entry_rcu(space_info, &info->space_info, list) {
10083                 if (!(get_alloc_profile(info, space_info->flags) &
10084                       (BTRFS_BLOCK_GROUP_RAID10 |
10085                        BTRFS_BLOCK_GROUP_RAID1 |
10086                        BTRFS_BLOCK_GROUP_RAID5 |
10087                        BTRFS_BLOCK_GROUP_RAID6 |
10088                        BTRFS_BLOCK_GROUP_DUP)))
10089                         continue;
10090                 /*
10091                  * avoid allocating from un-mirrored block group if there are
10092                  * mirrored block groups.
10093                  */
10094                 list_for_each_entry(cache,
10095                                 &space_info->block_groups[BTRFS_RAID_RAID0],
10096                                 list)
10097                         inc_block_group_ro(cache, 1);
10098                 list_for_each_entry(cache,
10099                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
10100                                 list)
10101                         inc_block_group_ro(cache, 1);
10102         }
10103
10104         init_global_block_rsv(info);
10105         ret = 0;
10106 error:
10107         btrfs_free_path(path);
10108         return ret;
10109 }
10110
10111 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans,
10112                                        struct btrfs_fs_info *fs_info)
10113 {
10114         struct btrfs_block_group_cache *block_group, *tmp;
10115         struct btrfs_root *extent_root = fs_info->extent_root;
10116         struct btrfs_block_group_item item;
10117         struct btrfs_key key;
10118         int ret = 0;
10119         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
10120
10121         trans->can_flush_pending_bgs = false;
10122         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) {
10123                 if (ret)
10124                         goto next;
10125
10126                 spin_lock(&block_group->lock);
10127                 memcpy(&item, &block_group->item, sizeof(item));
10128                 memcpy(&key, &block_group->key, sizeof(key));
10129                 spin_unlock(&block_group->lock);
10130
10131                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
10132                                         sizeof(item));
10133                 if (ret)
10134                         btrfs_abort_transaction(trans, ret);
10135                 ret = btrfs_finish_chunk_alloc(trans, fs_info, key.objectid,
10136                                                key.offset);
10137                 if (ret)
10138                         btrfs_abort_transaction(trans, ret);
10139                 add_block_group_free_space(trans, fs_info, block_group);
10140                 /* already aborted the transaction if it failed. */
10141 next:
10142                 list_del_init(&block_group->bg_list);
10143         }
10144         trans->can_flush_pending_bgs = can_flush_pending_bgs;
10145 }
10146
10147 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
10148                            struct btrfs_fs_info *fs_info, u64 bytes_used,
10149                            u64 type, u64 chunk_objectid, u64 chunk_offset,
10150                            u64 size)
10151 {
10152         struct btrfs_block_group_cache *cache;
10153         int ret;
10154
10155         btrfs_set_log_full_commit(fs_info, trans);
10156
10157         cache = btrfs_create_block_group_cache(fs_info, chunk_offset, size);
10158         if (!cache)
10159                 return -ENOMEM;
10160
10161         btrfs_set_block_group_used(&cache->item, bytes_used);
10162         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
10163         btrfs_set_block_group_flags(&cache->item, type);
10164
10165         cache->flags = type;
10166         cache->last_byte_to_unpin = (u64)-1;
10167         cache->cached = BTRFS_CACHE_FINISHED;
10168         cache->needs_free_space = 1;
10169         ret = exclude_super_stripes(fs_info, cache);
10170         if (ret) {
10171                 /*
10172                  * We may have excluded something, so call this just in
10173                  * case.
10174                  */
10175                 free_excluded_extents(fs_info, cache);
10176                 btrfs_put_block_group(cache);
10177                 return ret;
10178         }
10179
10180         add_new_free_space(cache, fs_info, chunk_offset, chunk_offset + size);
10181
10182         free_excluded_extents(fs_info, cache);
10183
10184 #ifdef CONFIG_BTRFS_DEBUG
10185         if (btrfs_should_fragment_free_space(cache)) {
10186                 u64 new_bytes_used = size - bytes_used;
10187
10188                 bytes_used += new_bytes_used >> 1;
10189                 fragment_free_space(cache);
10190         }
10191 #endif
10192         /*
10193          * Call to ensure the corresponding space_info object is created and
10194          * assigned to our block group, but don't update its counters just yet.
10195          * We want our bg to be added to the rbtree with its ->space_info set.
10196          */
10197         ret = update_space_info(fs_info, cache->flags, 0, 0, 0,
10198                                 &cache->space_info);
10199         if (ret) {
10200                 btrfs_remove_free_space_cache(cache);
10201                 btrfs_put_block_group(cache);
10202                 return ret;
10203         }
10204
10205         ret = btrfs_add_block_group_cache(fs_info, cache);
10206         if (ret) {
10207                 btrfs_remove_free_space_cache(cache);
10208                 btrfs_put_block_group(cache);
10209                 return ret;
10210         }
10211
10212         /*
10213          * Now that our block group has its ->space_info set and is inserted in
10214          * the rbtree, update the space info's counters.
10215          */
10216         trace_btrfs_add_block_group(fs_info, cache, 1);
10217         ret = update_space_info(fs_info, cache->flags, size, bytes_used,
10218                                 cache->bytes_super, &cache->space_info);
10219         if (ret) {
10220                 btrfs_remove_free_space_cache(cache);
10221                 spin_lock(&fs_info->block_group_cache_lock);
10222                 rb_erase(&cache->cache_node,
10223                          &fs_info->block_group_cache_tree);
10224                 RB_CLEAR_NODE(&cache->cache_node);
10225                 spin_unlock(&fs_info->block_group_cache_lock);
10226                 btrfs_put_block_group(cache);
10227                 return ret;
10228         }
10229         update_global_block_rsv(fs_info);
10230
10231         __link_block_group(cache->space_info, cache);
10232
10233         list_add_tail(&cache->bg_list, &trans->new_bgs);
10234
10235         set_avail_alloc_bits(fs_info, type);
10236         return 0;
10237 }
10238
10239 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
10240 {
10241         u64 extra_flags = chunk_to_extended(flags) &
10242                                 BTRFS_EXTENDED_PROFILE_MASK;
10243
10244         write_seqlock(&fs_info->profiles_lock);
10245         if (flags & BTRFS_BLOCK_GROUP_DATA)
10246                 fs_info->avail_data_alloc_bits &= ~extra_flags;
10247         if (flags & BTRFS_BLOCK_GROUP_METADATA)
10248                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
10249         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
10250                 fs_info->avail_system_alloc_bits &= ~extra_flags;
10251         write_sequnlock(&fs_info->profiles_lock);
10252 }
10253
10254 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
10255                              struct btrfs_fs_info *fs_info, u64 group_start,
10256                              struct extent_map *em)
10257 {
10258         struct btrfs_root *root = fs_info->extent_root;
10259         struct btrfs_path *path;
10260         struct btrfs_block_group_cache *block_group;
10261         struct btrfs_free_cluster *cluster;
10262         struct btrfs_root *tree_root = fs_info->tree_root;
10263         struct btrfs_key key;
10264         struct inode *inode;
10265         struct kobject *kobj = NULL;
10266         int ret;
10267         int index;
10268         int factor;
10269         struct btrfs_caching_control *caching_ctl = NULL;
10270         bool remove_em;
10271
10272         block_group = btrfs_lookup_block_group(fs_info, group_start);
10273         BUG_ON(!block_group);
10274         BUG_ON(!block_group->ro);
10275
10276         /*
10277          * Free the reserved super bytes from this block group before
10278          * remove it.
10279          */
10280         free_excluded_extents(fs_info, block_group);
10281
10282         memcpy(&key, &block_group->key, sizeof(key));
10283         index = get_block_group_index(block_group);
10284         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
10285                                   BTRFS_BLOCK_GROUP_RAID1 |
10286                                   BTRFS_BLOCK_GROUP_RAID10))
10287                 factor = 2;
10288         else
10289                 factor = 1;
10290
10291         /* make sure this block group isn't part of an allocation cluster */
10292         cluster = &fs_info->data_alloc_cluster;
10293         spin_lock(&cluster->refill_lock);
10294         btrfs_return_cluster_to_free_space(block_group, cluster);
10295         spin_unlock(&cluster->refill_lock);
10296
10297         /*
10298          * make sure this block group isn't part of a metadata
10299          * allocation cluster
10300          */
10301         cluster = &fs_info->meta_alloc_cluster;
10302         spin_lock(&cluster->refill_lock);
10303         btrfs_return_cluster_to_free_space(block_group, cluster);
10304         spin_unlock(&cluster->refill_lock);
10305
10306         path = btrfs_alloc_path();
10307         if (!path) {
10308                 ret = -ENOMEM;
10309                 goto out;
10310         }
10311
10312         /*
10313          * get the inode first so any iput calls done for the io_list
10314          * aren't the final iput (no unlinks allowed now)
10315          */
10316         inode = lookup_free_space_inode(tree_root, block_group, path);
10317
10318         mutex_lock(&trans->transaction->cache_write_mutex);
10319         /*
10320          * make sure our free spache cache IO is done before remove the
10321          * free space inode
10322          */
10323         spin_lock(&trans->transaction->dirty_bgs_lock);
10324         if (!list_empty(&block_group->io_list)) {
10325                 list_del_init(&block_group->io_list);
10326
10327                 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
10328
10329                 spin_unlock(&trans->transaction->dirty_bgs_lock);
10330                 btrfs_wait_cache_io(trans, block_group, path);
10331                 btrfs_put_block_group(block_group);
10332                 spin_lock(&trans->transaction->dirty_bgs_lock);
10333         }
10334
10335         if (!list_empty(&block_group->dirty_list)) {
10336                 list_del_init(&block_group->dirty_list);
10337                 btrfs_put_block_group(block_group);
10338         }
10339         spin_unlock(&trans->transaction->dirty_bgs_lock);
10340         mutex_unlock(&trans->transaction->cache_write_mutex);
10341
10342         if (!IS_ERR(inode)) {
10343                 ret = btrfs_orphan_add(trans, inode);
10344                 if (ret) {
10345                         btrfs_add_delayed_iput(inode);
10346                         goto out;
10347                 }
10348                 clear_nlink(inode);
10349                 /* One for the block groups ref */
10350                 spin_lock(&block_group->lock);
10351                 if (block_group->iref) {
10352                         block_group->iref = 0;
10353                         block_group->inode = NULL;
10354                         spin_unlock(&block_group->lock);
10355                         iput(inode);
10356                 } else {
10357                         spin_unlock(&block_group->lock);
10358                 }
10359                 /* One for our lookup ref */
10360                 btrfs_add_delayed_iput(inode);
10361         }
10362
10363         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
10364         key.offset = block_group->key.objectid;
10365         key.type = 0;
10366
10367         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
10368         if (ret < 0)
10369                 goto out;
10370         if (ret > 0)
10371                 btrfs_release_path(path);
10372         if (ret == 0) {
10373                 ret = btrfs_del_item(trans, tree_root, path);
10374                 if (ret)
10375                         goto out;
10376                 btrfs_release_path(path);
10377         }
10378
10379         spin_lock(&fs_info->block_group_cache_lock);
10380         rb_erase(&block_group->cache_node,
10381                  &fs_info->block_group_cache_tree);
10382         RB_CLEAR_NODE(&block_group->cache_node);
10383
10384         if (fs_info->first_logical_byte == block_group->key.objectid)
10385                 fs_info->first_logical_byte = (u64)-1;
10386         spin_unlock(&fs_info->block_group_cache_lock);
10387
10388         down_write(&block_group->space_info->groups_sem);
10389         /*
10390          * we must use list_del_init so people can check to see if they
10391          * are still on the list after taking the semaphore
10392          */
10393         list_del_init(&block_group->list);
10394         if (list_empty(&block_group->space_info->block_groups[index])) {
10395                 kobj = block_group->space_info->block_group_kobjs[index];
10396                 block_group->space_info->block_group_kobjs[index] = NULL;
10397                 clear_avail_alloc_bits(fs_info, block_group->flags);
10398         }
10399         up_write(&block_group->space_info->groups_sem);
10400         if (kobj) {
10401                 kobject_del(kobj);
10402                 kobject_put(kobj);
10403         }
10404
10405         if (block_group->has_caching_ctl)
10406                 caching_ctl = get_caching_control(block_group);
10407         if (block_group->cached == BTRFS_CACHE_STARTED)
10408                 wait_block_group_cache_done(block_group);
10409         if (block_group->has_caching_ctl) {
10410                 down_write(&fs_info->commit_root_sem);
10411                 if (!caching_ctl) {
10412                         struct btrfs_caching_control *ctl;
10413
10414                         list_for_each_entry(ctl,
10415                                     &fs_info->caching_block_groups, list)
10416                                 if (ctl->block_group == block_group) {
10417                                         caching_ctl = ctl;
10418                                         atomic_inc(&caching_ctl->count);
10419                                         break;
10420                                 }
10421                 }
10422                 if (caching_ctl)
10423                         list_del_init(&caching_ctl->list);
10424                 up_write(&fs_info->commit_root_sem);
10425                 if (caching_ctl) {
10426                         /* Once for the caching bgs list and once for us. */
10427                         put_caching_control(caching_ctl);
10428                         put_caching_control(caching_ctl);
10429                 }
10430         }
10431
10432         spin_lock(&trans->transaction->dirty_bgs_lock);
10433         if (!list_empty(&block_group->dirty_list)) {
10434                 WARN_ON(1);
10435         }
10436         if (!list_empty(&block_group->io_list)) {
10437                 WARN_ON(1);
10438         }
10439         spin_unlock(&trans->transaction->dirty_bgs_lock);
10440         btrfs_remove_free_space_cache(block_group);
10441
10442         spin_lock(&block_group->space_info->lock);
10443         list_del_init(&block_group->ro_list);
10444
10445         if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
10446                 WARN_ON(block_group->space_info->total_bytes
10447                         < block_group->key.offset);
10448                 WARN_ON(block_group->space_info->bytes_readonly
10449                         < block_group->key.offset);
10450                 WARN_ON(block_group->space_info->disk_total
10451                         < block_group->key.offset * factor);
10452         }
10453         block_group->space_info->total_bytes -= block_group->key.offset;
10454         block_group->space_info->bytes_readonly -= block_group->key.offset;
10455         block_group->space_info->disk_total -= block_group->key.offset * factor;
10456
10457         spin_unlock(&block_group->space_info->lock);
10458
10459         memcpy(&key, &block_group->key, sizeof(key));
10460
10461         mutex_lock(&fs_info->chunk_mutex);
10462         if (!list_empty(&em->list)) {
10463                 /* We're in the transaction->pending_chunks list. */
10464                 free_extent_map(em);
10465         }
10466         spin_lock(&block_group->lock);
10467         block_group->removed = 1;
10468         /*
10469          * At this point trimming can't start on this block group, because we
10470          * removed the block group from the tree fs_info->block_group_cache_tree
10471          * so no one can't find it anymore and even if someone already got this
10472          * block group before we removed it from the rbtree, they have already
10473          * incremented block_group->trimming - if they didn't, they won't find
10474          * any free space entries because we already removed them all when we
10475          * called btrfs_remove_free_space_cache().
10476          *
10477          * And we must not remove the extent map from the fs_info->mapping_tree
10478          * to prevent the same logical address range and physical device space
10479          * ranges from being reused for a new block group. This is because our
10480          * fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
10481          * completely transactionless, so while it is trimming a range the
10482          * currently running transaction might finish and a new one start,
10483          * allowing for new block groups to be created that can reuse the same
10484          * physical device locations unless we take this special care.
10485          *
10486          * There may also be an implicit trim operation if the file system
10487          * is mounted with -odiscard. The same protections must remain
10488          * in place until the extents have been discarded completely when
10489          * the transaction commit has completed.
10490          */
10491         remove_em = (atomic_read(&block_group->trimming) == 0);
10492         /*
10493          * Make sure a trimmer task always sees the em in the pinned_chunks list
10494          * if it sees block_group->removed == 1 (needs to lock block_group->lock
10495          * before checking block_group->removed).
10496          */
10497         if (!remove_em) {
10498                 /*
10499                  * Our em might be in trans->transaction->pending_chunks which
10500                  * is protected by fs_info->chunk_mutex ([lock|unlock]_chunks),
10501                  * and so is the fs_info->pinned_chunks list.
10502                  *
10503                  * So at this point we must be holding the chunk_mutex to avoid
10504                  * any races with chunk allocation (more specifically at
10505                  * volumes.c:contains_pending_extent()), to ensure it always
10506                  * sees the em, either in the pending_chunks list or in the
10507                  * pinned_chunks list.
10508                  */
10509                 list_move_tail(&em->list, &fs_info->pinned_chunks);
10510         }
10511         spin_unlock(&block_group->lock);
10512
10513         if (remove_em) {
10514                 struct extent_map_tree *em_tree;
10515
10516                 em_tree = &fs_info->mapping_tree.map_tree;
10517                 write_lock(&em_tree->lock);
10518                 /*
10519                  * The em might be in the pending_chunks list, so make sure the
10520                  * chunk mutex is locked, since remove_extent_mapping() will
10521                  * delete us from that list.
10522                  */
10523                 remove_extent_mapping(em_tree, em);
10524                 write_unlock(&em_tree->lock);
10525                 /* once for the tree */
10526                 free_extent_map(em);
10527         }
10528
10529         mutex_unlock(&fs_info->chunk_mutex);
10530
10531         ret = remove_block_group_free_space(trans, fs_info, block_group);
10532         if (ret)
10533                 goto out;
10534
10535         btrfs_put_block_group(block_group);
10536         btrfs_put_block_group(block_group);
10537
10538         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
10539         if (ret > 0)
10540                 ret = -EIO;
10541         if (ret < 0)
10542                 goto out;
10543
10544         ret = btrfs_del_item(trans, root, path);
10545 out:
10546         btrfs_free_path(path);
10547         return ret;
10548 }
10549
10550 struct btrfs_trans_handle *
10551 btrfs_start_trans_remove_block_group(struct btrfs_fs_info *fs_info,
10552                                      const u64 chunk_offset)
10553 {
10554         struct extent_map_tree *em_tree = &fs_info->mapping_tree.map_tree;
10555         struct extent_map *em;
10556         struct map_lookup *map;
10557         unsigned int num_items;
10558
10559         read_lock(&em_tree->lock);
10560         em = lookup_extent_mapping(em_tree, chunk_offset, 1);
10561         read_unlock(&em_tree->lock);
10562         ASSERT(em && em->start == chunk_offset);
10563
10564         /*
10565          * We need to reserve 3 + N units from the metadata space info in order
10566          * to remove a block group (done at btrfs_remove_chunk() and at
10567          * btrfs_remove_block_group()), which are used for:
10568          *
10569          * 1 unit for adding the free space inode's orphan (located in the tree
10570          * of tree roots).
10571          * 1 unit for deleting the block group item (located in the extent
10572          * tree).
10573          * 1 unit for deleting the free space item (located in tree of tree
10574          * roots).
10575          * N units for deleting N device extent items corresponding to each
10576          * stripe (located in the device tree).
10577          *
10578          * In order to remove a block group we also need to reserve units in the
10579          * system space info in order to update the chunk tree (update one or
10580          * more device items and remove one chunk item), but this is done at
10581          * btrfs_remove_chunk() through a call to check_system_chunk().
10582          */
10583         map = em->map_lookup;
10584         num_items = 3 + map->num_stripes;
10585         free_extent_map(em);
10586
10587         return btrfs_start_transaction_fallback_global_rsv(fs_info->extent_root,
10588                                                            num_items, 1);
10589 }
10590
10591 /*
10592  * Process the unused_bgs list and remove any that don't have any allocated
10593  * space inside of them.
10594  */
10595 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
10596 {
10597         struct btrfs_block_group_cache *block_group;
10598         struct btrfs_space_info *space_info;
10599         struct btrfs_trans_handle *trans;
10600         int ret = 0;
10601
10602         if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
10603                 return;
10604
10605         spin_lock(&fs_info->unused_bgs_lock);
10606         while (!list_empty(&fs_info->unused_bgs)) {
10607                 u64 start, end;
10608                 int trimming;
10609
10610                 block_group = list_first_entry(&fs_info->unused_bgs,
10611                                                struct btrfs_block_group_cache,
10612                                                bg_list);
10613                 list_del_init(&block_group->bg_list);
10614
10615                 space_info = block_group->space_info;
10616
10617                 if (ret || btrfs_mixed_space_info(space_info)) {
10618                         btrfs_put_block_group(block_group);
10619                         continue;
10620                 }
10621                 spin_unlock(&fs_info->unused_bgs_lock);
10622
10623                 mutex_lock(&fs_info->delete_unused_bgs_mutex);
10624
10625                 /* Don't want to race with allocators so take the groups_sem */
10626                 down_write(&space_info->groups_sem);
10627                 spin_lock(&block_group->lock);
10628                 if (block_group->reserved ||
10629                     btrfs_block_group_used(&block_group->item) ||
10630                     block_group->ro ||
10631                     list_is_singular(&block_group->list)) {
10632                         /*
10633                          * We want to bail if we made new allocations or have
10634                          * outstanding allocations in this block group.  We do
10635                          * the ro check in case balance is currently acting on
10636                          * this block group.
10637                          */
10638                         spin_unlock(&block_group->lock);
10639                         up_write(&space_info->groups_sem);
10640                         goto next;
10641                 }
10642                 spin_unlock(&block_group->lock);
10643
10644                 /* We don't want to force the issue, only flip if it's ok. */
10645                 ret = inc_block_group_ro(block_group, 0);
10646                 up_write(&space_info->groups_sem);
10647                 if (ret < 0) {
10648                         ret = 0;
10649                         goto next;
10650                 }
10651
10652                 /*
10653                  * Want to do this before we do anything else so we can recover
10654                  * properly if we fail to join the transaction.
10655                  */
10656                 trans = btrfs_start_trans_remove_block_group(fs_info,
10657                                                      block_group->key.objectid);
10658                 if (IS_ERR(trans)) {
10659                         btrfs_dec_block_group_ro(block_group);
10660                         ret = PTR_ERR(trans);
10661                         goto next;
10662                 }
10663
10664                 /*
10665                  * We could have pending pinned extents for this block group,
10666                  * just delete them, we don't care about them anymore.
10667                  */
10668                 start = block_group->key.objectid;
10669                 end = start + block_group->key.offset - 1;
10670                 /*
10671                  * Hold the unused_bg_unpin_mutex lock to avoid racing with
10672                  * btrfs_finish_extent_commit(). If we are at transaction N,
10673                  * another task might be running finish_extent_commit() for the
10674                  * previous transaction N - 1, and have seen a range belonging
10675                  * to the block group in freed_extents[] before we were able to
10676                  * clear the whole block group range from freed_extents[]. This
10677                  * means that task can lookup for the block group after we
10678                  * unpinned it from freed_extents[] and removed it, leading to
10679                  * a BUG_ON() at btrfs_unpin_extent_range().
10680                  */
10681                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
10682                 ret = clear_extent_bits(&fs_info->freed_extents[0], start, end,
10683                                   EXTENT_DIRTY);
10684                 if (ret) {
10685                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10686                         btrfs_dec_block_group_ro(block_group);
10687                         goto end_trans;
10688                 }
10689                 ret = clear_extent_bits(&fs_info->freed_extents[1], start, end,
10690                                   EXTENT_DIRTY);
10691                 if (ret) {
10692                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10693                         btrfs_dec_block_group_ro(block_group);
10694                         goto end_trans;
10695                 }
10696                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10697
10698                 /* Reset pinned so btrfs_put_block_group doesn't complain */
10699                 spin_lock(&space_info->lock);
10700                 spin_lock(&block_group->lock);
10701
10702                 space_info->bytes_pinned -= block_group->pinned;
10703                 space_info->bytes_readonly += block_group->pinned;
10704                 percpu_counter_add(&space_info->total_bytes_pinned,
10705                                    -block_group->pinned);
10706                 block_group->pinned = 0;
10707
10708                 spin_unlock(&block_group->lock);
10709                 spin_unlock(&space_info->lock);
10710
10711                 /* DISCARD can flip during remount */
10712                 trimming = btrfs_test_opt(fs_info, DISCARD);
10713
10714                 /* Implicit trim during transaction commit. */
10715                 if (trimming)
10716                         btrfs_get_block_group_trimming(block_group);
10717
10718                 /*
10719                  * Btrfs_remove_chunk will abort the transaction if things go
10720                  * horribly wrong.
10721                  */
10722                 ret = btrfs_remove_chunk(trans, fs_info,
10723                                          block_group->key.objectid);
10724
10725                 if (ret) {
10726                         if (trimming)
10727                                 btrfs_put_block_group_trimming(block_group);
10728                         goto end_trans;
10729                 }
10730
10731                 /*
10732                  * If we're not mounted with -odiscard, we can just forget
10733                  * about this block group. Otherwise we'll need to wait
10734                  * until transaction commit to do the actual discard.
10735                  */
10736                 if (trimming) {
10737                         spin_lock(&fs_info->unused_bgs_lock);
10738                         /*
10739                          * A concurrent scrub might have added us to the list
10740                          * fs_info->unused_bgs, so use a list_move operation
10741                          * to add the block group to the deleted_bgs list.
10742                          */
10743                         list_move(&block_group->bg_list,
10744                                   &trans->transaction->deleted_bgs);
10745                         spin_unlock(&fs_info->unused_bgs_lock);
10746                         btrfs_get_block_group(block_group);
10747                 }
10748 end_trans:
10749                 btrfs_end_transaction(trans);
10750 next:
10751                 mutex_unlock(&fs_info->delete_unused_bgs_mutex);
10752                 btrfs_put_block_group(block_group);
10753                 spin_lock(&fs_info->unused_bgs_lock);
10754         }
10755         spin_unlock(&fs_info->unused_bgs_lock);
10756 }
10757
10758 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
10759 {
10760         struct btrfs_space_info *space_info;
10761         struct btrfs_super_block *disk_super;
10762         u64 features;
10763         u64 flags;
10764         int mixed = 0;
10765         int ret;
10766
10767         disk_super = fs_info->super_copy;
10768         if (!btrfs_super_root(disk_super))
10769                 return -EINVAL;
10770
10771         features = btrfs_super_incompat_flags(disk_super);
10772         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
10773                 mixed = 1;
10774
10775         flags = BTRFS_BLOCK_GROUP_SYSTEM;
10776         ret = update_space_info(fs_info, flags, 0, 0, 0, &space_info);
10777         if (ret)
10778                 goto out;
10779
10780         if (mixed) {
10781                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
10782                 ret = update_space_info(fs_info, flags, 0, 0, 0, &space_info);
10783         } else {
10784                 flags = BTRFS_BLOCK_GROUP_METADATA;
10785                 ret = update_space_info(fs_info, flags, 0, 0, 0, &space_info);
10786                 if (ret)
10787                         goto out;
10788
10789                 flags = BTRFS_BLOCK_GROUP_DATA;
10790                 ret = update_space_info(fs_info, flags, 0, 0, 0, &space_info);
10791         }
10792 out:
10793         return ret;
10794 }
10795
10796 int btrfs_error_unpin_extent_range(struct btrfs_fs_info *fs_info,
10797                                    u64 start, u64 end)
10798 {
10799         return unpin_extent_range(fs_info, start, end, false);
10800 }
10801
10802 /*
10803  * It used to be that old block groups would be left around forever.
10804  * Iterating over them would be enough to trim unused space.  Since we
10805  * now automatically remove them, we also need to iterate over unallocated
10806  * space.
10807  *
10808  * We don't want a transaction for this since the discard may take a
10809  * substantial amount of time.  We don't require that a transaction be
10810  * running, but we do need to take a running transaction into account
10811  * to ensure that we're not discarding chunks that were released in
10812  * the current transaction.
10813  *
10814  * Holding the chunks lock will prevent other threads from allocating
10815  * or releasing chunks, but it won't prevent a running transaction
10816  * from committing and releasing the memory that the pending chunks
10817  * list head uses.  For that, we need to take a reference to the
10818  * transaction.
10819  */
10820 static int btrfs_trim_free_extents(struct btrfs_device *device,
10821                                    u64 minlen, u64 *trimmed)
10822 {
10823         u64 start = 0, len = 0;
10824         int ret;
10825
10826         *trimmed = 0;
10827
10828         /* Not writeable = nothing to do. */
10829         if (!device->writeable)
10830                 return 0;
10831
10832         /* No free space = nothing to do. */
10833         if (device->total_bytes <= device->bytes_used)
10834                 return 0;
10835
10836         ret = 0;
10837
10838         while (1) {
10839                 struct btrfs_fs_info *fs_info = device->fs_info;
10840                 struct btrfs_transaction *trans;
10841                 u64 bytes;
10842
10843                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
10844                 if (ret)
10845                         return ret;
10846
10847                 down_read(&fs_info->commit_root_sem);
10848
10849                 spin_lock(&fs_info->trans_lock);
10850                 trans = fs_info->running_transaction;
10851                 if (trans)
10852                         atomic_inc(&trans->use_count);
10853                 spin_unlock(&fs_info->trans_lock);
10854
10855                 ret = find_free_dev_extent_start(trans, device, minlen, start,
10856                                                  &start, &len);
10857                 if (trans)
10858                         btrfs_put_transaction(trans);
10859
10860                 if (ret) {
10861                         up_read(&fs_info->commit_root_sem);
10862                         mutex_unlock(&fs_info->chunk_mutex);
10863                         if (ret == -ENOSPC)
10864                                 ret = 0;
10865                         break;
10866                 }
10867
10868                 ret = btrfs_issue_discard(device->bdev, start, len, &bytes);
10869                 up_read(&fs_info->commit_root_sem);
10870                 mutex_unlock(&fs_info->chunk_mutex);
10871
10872                 if (ret)
10873                         break;
10874
10875                 start += len;
10876                 *trimmed += bytes;
10877
10878                 if (fatal_signal_pending(current)) {
10879                         ret = -ERESTARTSYS;
10880                         break;
10881                 }
10882
10883                 cond_resched();
10884         }
10885
10886         return ret;
10887 }
10888
10889 int btrfs_trim_fs(struct btrfs_fs_info *fs_info, struct fstrim_range *range)
10890 {
10891         struct btrfs_block_group_cache *cache = NULL;
10892         struct btrfs_device *device;
10893         struct list_head *devices;
10894         u64 group_trimmed;
10895         u64 start;
10896         u64 end;
10897         u64 trimmed = 0;
10898         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
10899         int ret = 0;
10900
10901         /*
10902          * try to trim all FS space, our block group may start from non-zero.
10903          */
10904         if (range->len == total_bytes)
10905                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
10906         else
10907                 cache = btrfs_lookup_block_group(fs_info, range->start);
10908
10909         while (cache) {
10910                 if (cache->key.objectid >= (range->start + range->len)) {
10911                         btrfs_put_block_group(cache);
10912                         break;
10913                 }
10914
10915                 start = max(range->start, cache->key.objectid);
10916                 end = min(range->start + range->len,
10917                                 cache->key.objectid + cache->key.offset);
10918
10919                 if (end - start >= range->minlen) {
10920                         if (!block_group_cache_done(cache)) {
10921                                 ret = cache_block_group(cache, 0);
10922                                 if (ret) {
10923                                         btrfs_put_block_group(cache);
10924                                         break;
10925                                 }
10926                                 ret = wait_block_group_cache_done(cache);
10927                                 if (ret) {
10928                                         btrfs_put_block_group(cache);
10929                                         break;
10930                                 }
10931                         }
10932                         ret = btrfs_trim_block_group(cache,
10933                                                      &group_trimmed,
10934                                                      start,
10935                                                      end,
10936                                                      range->minlen);
10937
10938                         trimmed += group_trimmed;
10939                         if (ret) {
10940                                 btrfs_put_block_group(cache);
10941                                 break;
10942                         }
10943                 }
10944
10945                 cache = next_block_group(fs_info, cache);
10946         }
10947
10948         mutex_lock(&fs_info->fs_devices->device_list_mutex);
10949         devices = &fs_info->fs_devices->alloc_list;
10950         list_for_each_entry(device, devices, dev_alloc_list) {
10951                 ret = btrfs_trim_free_extents(device, range->minlen,
10952                                               &group_trimmed);
10953                 if (ret)
10954                         break;
10955
10956                 trimmed += group_trimmed;
10957         }
10958         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
10959
10960         range->len = trimmed;
10961         return ret;
10962 }
10963
10964 /*
10965  * btrfs_{start,end}_write_no_snapshoting() are similar to
10966  * mnt_{want,drop}_write(), they are used to prevent some tasks from writing
10967  * data into the page cache through nocow before the subvolume is snapshoted,
10968  * but flush the data into disk after the snapshot creation, or to prevent
10969  * operations while snapshoting is ongoing and that cause the snapshot to be
10970  * inconsistent (writes followed by expanding truncates for example).
10971  */
10972 void btrfs_end_write_no_snapshoting(struct btrfs_root *root)
10973 {
10974         percpu_counter_dec(&root->subv_writers->counter);
10975         /*
10976          * Make sure counter is updated before we wake up waiters.
10977          */
10978         smp_mb();
10979         if (waitqueue_active(&root->subv_writers->wait))
10980                 wake_up(&root->subv_writers->wait);
10981 }
10982
10983 int btrfs_start_write_no_snapshoting(struct btrfs_root *root)
10984 {
10985         if (atomic_read(&root->will_be_snapshoted))
10986                 return 0;
10987
10988         percpu_counter_inc(&root->subv_writers->counter);
10989         /*
10990          * Make sure counter is updated before we check for snapshot creation.
10991          */
10992         smp_mb();
10993         if (atomic_read(&root->will_be_snapshoted)) {
10994                 btrfs_end_write_no_snapshoting(root);
10995                 return 0;
10996         }
10997         return 1;
10998 }
10999
11000 static int wait_snapshoting_atomic_t(atomic_t *a)
11001 {
11002         schedule();
11003         return 0;
11004 }
11005
11006 void btrfs_wait_for_snapshot_creation(struct btrfs_root *root)
11007 {
11008         while (true) {
11009                 int ret;
11010
11011                 ret = btrfs_start_write_no_snapshoting(root);
11012                 if (ret)
11013                         break;
11014                 wait_on_atomic_t(&root->will_be_snapshoted,
11015                                  wait_snapshoting_atomic_t,
11016                                  TASK_UNINTERRUPTIBLE);
11017         }
11018 }