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/sched/signal.h>
  20 #include <linux/pagemap.h>
  21 #include <linux/writeback.h>
  22 #include <linux/blkdev.h>
  23 #include <linux/sort.h>
  24 #include <linux/rcupdate.h>
  25 #include <linux/kthread.h>
  26 #include <linux/slab.h>
  27 #include <linux/ratelimit.h>
  28 #include <linux/percpu_counter.h>
  29 #include <linux/lockdep.h>
  30 #include <linux/crc32c.h>
  31 #include "tree-log.h"
  32 #include "disk-io.h"
  33 #include "print-tree.h"
  34 #include "volumes.h"
  35 #include "raid56.h"
  36 #include "locking.h"
  37 #include "free-space-cache.h"
  38 #include "free-space-tree.h"
  39 #include "math.h"
  40 #include "sysfs.h"
  41 #include "qgroup.h"
  42 #include "ref-verify.h"
  43
  44 #undef SCRAMBLE_DELAYED_REFS
  45
  46 /*
  47  * control flags for do_chunk_alloc's force field
  48  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
  49  * if we really need one.
  50  *
  51  * CHUNK_ALLOC_LIMITED means to only try and allocate one
  52  * if we have very few chunks already allocated.  This is
  53  * used as part of the clustering code to help make sure
  54  * we have a good pool of storage to cluster in, without
  55  * filling the FS with empty chunks
  56  *
  57  * CHUNK_ALLOC_FORCE means it must try to allocate one
  58  *
  59  */
  60 enum {
  61         CHUNK_ALLOC_NO_FORCE = 0,
  62         CHUNK_ALLOC_LIMITED = 1,
  63         CHUNK_ALLOC_FORCE = 2,
  64 };
  65
  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 block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
  94                                u64 num_bytes);
  95 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
  96                                      struct btrfs_space_info *space_info,
  97                                      u64 num_bytes);
  98 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
  99                                      struct btrfs_space_info *space_info,
 100                                      u64 num_bytes);
 101
 102 static noinline int
 103 block_group_cache_done(struct btrfs_block_group_cache *cache)
 104 {
 105         smp_mb();
 106         return cache->cached == BTRFS_CACHE_FINISHED ||
 107                 cache->cached == BTRFS_CACHE_ERROR;
 108 }
 109
 110 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
 111 {
 112         return (cache->flags & bits) == bits;
 113 }
 114
 115 void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 116 {
 117         atomic_inc(&cache->count);
 118 }
 119
 120 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 121 {
 122         if (atomic_dec_and_test(&cache->count)) {
 123                 WARN_ON(cache->pinned > 0);
 124                 WARN_ON(cache->reserved > 0);
 125
 126                 /*
 127                  * If not empty, someone is still holding mutex of
 128                  * full_stripe_lock, which can only be released by caller.
 129                  * And it will definitely cause use-after-free when caller
 130                  * tries to release full stripe lock.
 131                  *
 132                  * No better way to resolve, but only to warn.
 133                  */
 134                 WARN_ON(!RB_EMPTY_ROOT(&cache->full_stripe_locks_root.root));
 135                 kfree(cache->free_space_ctl);
 136                 kfree(cache);
 137         }
 138 }
 139
 140 /*
 141  * this adds the block group to the fs_info rb tree for the block group
 142  * cache
 143  */
 144 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 145                                 struct btrfs_block_group_cache *block_group)
 146 {
 147         struct rb_node **p;
 148         struct rb_node *parent = NULL;
 149         struct btrfs_block_group_cache *cache;
 150
 151         spin_lock(&info->block_group_cache_lock);
 152         p = &info->block_group_cache_tree.rb_node;
 153
 154         while (*p) {
 155                 parent = *p;
 156                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 157                                  cache_node);
 158                 if (block_group->key.objectid < cache->key.objectid) {
 159                         p = &(*p)->rb_left;
 160                 } else if (block_group->key.objectid > cache->key.objectid) {
 161                         p = &(*p)->rb_right;
 162                 } else {
 163                         spin_unlock(&info->block_group_cache_lock);
 164                         return -EEXIST;
 165                 }
 166         }
 167
 168         rb_link_node(&block_group->cache_node, parent, p);
 169         rb_insert_color(&block_group->cache_node,
 170                         &info->block_group_cache_tree);
 171
 172         if (info->first_logical_byte > block_group->key.objectid)
 173                 info->first_logical_byte = block_group->key.objectid;
 174
 175         spin_unlock(&info->block_group_cache_lock);
 176
 177         return 0;
 178 }
 179
 180 /*
 181  * This will return the block group at or after bytenr if contains is 0, else
 182  * it will return the block group that contains the bytenr
 183  */
 184 static struct btrfs_block_group_cache *
 185 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 186                               int contains)
 187 {
 188         struct btrfs_block_group_cache *cache, *ret = NULL;
 189         struct rb_node *n;
 190         u64 end, start;
 191
 192         spin_lock(&info->block_group_cache_lock);
 193         n = info->block_group_cache_tree.rb_node;
 194
 195         while (n) {
 196                 cache = rb_entry(n, struct btrfs_block_group_cache,
 197                                  cache_node);
 198                 end = cache->key.objectid + cache->key.offset - 1;
 199                 start = cache->key.objectid;
 200
 201                 if (bytenr < start) {
 202                         if (!contains && (!ret || start < ret->key.objectid))
 203                                 ret = cache;
 204                         n = n->rb_left;
 205                 } else if (bytenr > start) {
 206                         if (contains && bytenr <= end) {
 207                                 ret = cache;
 208                                 break;
 209                         }
 210                         n = n->rb_right;
 211                 } else {
 212                         ret = cache;
 213                         break;
 214                 }
 215         }
 216         if (ret) {
 217                 btrfs_get_block_group(ret);
 218                 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
 219                         info->first_logical_byte = ret->key.objectid;
 220         }
 221         spin_unlock(&info->block_group_cache_lock);
 222
 223         return ret;
 224 }
 225
 226 static int add_excluded_extent(struct btrfs_fs_info *fs_info,
 227                                u64 start, u64 num_bytes)
 228 {
 229         u64 end = start + num_bytes - 1;
 230         set_extent_bits(&fs_info->freed_extents[0],
 231                         start, end, EXTENT_UPTODATE);
 232         set_extent_bits(&fs_info->freed_extents[1],
 233                         start, end, EXTENT_UPTODATE);
 234         return 0;
 235 }
 236
 237 static void free_excluded_extents(struct btrfs_fs_info *fs_info,
 238                                   struct btrfs_block_group_cache *cache)
 239 {
 240         u64 start, end;
 241
 242         start = cache->key.objectid;
 243         end = start + cache->key.offset - 1;
 244
 245         clear_extent_bits(&fs_info->freed_extents[0],
 246                           start, end, EXTENT_UPTODATE);
 247         clear_extent_bits(&fs_info->freed_extents[1],
 248                           start, end, EXTENT_UPTODATE);
 249 }
 250
 251 static int exclude_super_stripes(struct btrfs_fs_info *fs_info,
 252                                  struct btrfs_block_group_cache *cache)
 253 {
 254         u64 bytenr;
 255         u64 *logical;
 256         int stripe_len;
 257         int i, nr, ret;
 258
 259         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 260                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 261                 cache->bytes_super += stripe_len;
 262                 ret = add_excluded_extent(fs_info, cache->key.objectid,
 263                                           stripe_len);
 264                 if (ret)
 265                         return ret;
 266         }
 267
 268         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 269                 bytenr = btrfs_sb_offset(i);
 270                 ret = btrfs_rmap_block(fs_info, cache->key.objectid,
 271                                        bytenr, 0, &logical, &nr, &stripe_len);
 272                 if (ret)
 273                         return ret;
 274
 275                 while (nr--) {
 276                         u64 start, len;
 277
 278                         if (logical[nr] > cache->key.objectid +
 279                             cache->key.offset)
 280                                 continue;
 281
 282                         if (logical[nr] + stripe_len <= cache->key.objectid)
 283                                 continue;
 284
 285                         start = logical[nr];
 286                         if (start < cache->key.objectid) {
 287                                 start = cache->key.objectid;
 288                                 len = (logical[nr] + stripe_len) - start;
 289                         } else {
 290                                 len = min_t(u64, stripe_len,
 291                                             cache->key.objectid +
 292                                             cache->key.offset - start);
 293                         }
 294
 295                         cache->bytes_super += len;
 296                         ret = add_excluded_extent(fs_info, start, len);
 297                         if (ret) {
 298                                 kfree(logical);
 299                                 return ret;
 300                         }
 301                 }
 302
 303                 kfree(logical);
 304         }
 305         return 0;
 306 }
 307
 308 static struct btrfs_caching_control *
 309 get_caching_control(struct btrfs_block_group_cache *cache)
 310 {
 311         struct btrfs_caching_control *ctl;
 312
 313         spin_lock(&cache->lock);
 314         if (!cache->caching_ctl) {
 315                 spin_unlock(&cache->lock);
 316                 return NULL;
 317         }
 318
 319         ctl = cache->caching_ctl;
 320         refcount_inc(&ctl->count);
 321         spin_unlock(&cache->lock);
 322         return ctl;
 323 }
 324
 325 static void put_caching_control(struct btrfs_caching_control *ctl)
 326 {
 327         if (refcount_dec_and_test(&ctl->count))
 328                 kfree(ctl);
 329 }
 330
 331 #ifdef CONFIG_BTRFS_DEBUG
 332 static void fragment_free_space(struct btrfs_block_group_cache *block_group)
 333 {
 334         struct btrfs_fs_info *fs_info = block_group->fs_info;
 335         u64 start = block_group->key.objectid;
 336         u64 len = block_group->key.offset;
 337         u64 chunk = block_group->flags & BTRFS_BLOCK_GROUP_METADATA ?
 338                 fs_info->nodesize : fs_info->sectorsize;
 339         u64 step = chunk << 1;
 340
 341         while (len > chunk) {
 342                 btrfs_remove_free_space(block_group, start, chunk);
 343                 start += step;
 344                 if (len < step)
 345                         len = 0;
 346                 else
 347                         len -= step;
 348         }
 349 }
 350 #endif
 351
 352 /*
 353  * this is only called by cache_block_group, since we could have freed extents
 354  * we need to check the pinned_extents for any extents that can't be used yet
 355  * since their free space will be released as soon as the transaction commits.
 356  */
 357 u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 358                        struct btrfs_fs_info *info, u64 start, u64 end)
 359 {
 360         u64 extent_start, extent_end, size, total_added = 0;
 361         int ret;
 362
 363         while (start < end) {
 364                 ret = find_first_extent_bit(info->pinned_extents, start,
 365                                             &extent_start, &extent_end,
 366                                             EXTENT_DIRTY | EXTENT_UPTODATE,
 367                                             NULL);
 368                 if (ret)
 369                         break;
 370
 371                 if (extent_start <= start) {
 372                         start = extent_end + 1;
 373                 } else if (extent_start > start && extent_start < end) {
 374                         size = extent_start - start;
 375                         total_added += size;
 376                         ret = btrfs_add_free_space(block_group, start,
 377                                                    size);
 378                         BUG_ON(ret); /* -ENOMEM or logic error */
 379                         start = extent_end + 1;
 380                 } else {
 381                         break;
 382                 }
 383         }
 384
 385         if (start < end) {
 386                 size = end - start;
 387                 total_added += size;
 388                 ret = btrfs_add_free_space(block_group, start, size);
 389                 BUG_ON(ret); /* -ENOMEM or logic error */
 390         }
 391
 392         return total_added;
 393 }
 394
 395 static int load_extent_tree_free(struct btrfs_caching_control *caching_ctl)
 396 {
 397         struct btrfs_block_group_cache *block_group = caching_ctl->block_group;
 398         struct btrfs_fs_info *fs_info = block_group->fs_info;
 399         struct btrfs_root *extent_root = fs_info->extent_root;
 400         struct btrfs_path *path;
 401         struct extent_buffer *leaf;
 402         struct btrfs_key key;
 403         u64 total_found = 0;
 404         u64 last = 0;
 405         u32 nritems;
 406         int ret;
 407         bool wakeup = true;
 408
 409         path = btrfs_alloc_path();
 410         if (!path)
 411                 return -ENOMEM;
 412
 413         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 414
 415 #ifdef CONFIG_BTRFS_DEBUG
 416         /*
 417          * If we're fragmenting we don't want to make anybody think we can
 418          * allocate from this block group until we've had a chance to fragment
 419          * the free space.
 420          */
 421         if (btrfs_should_fragment_free_space(block_group))
 422                 wakeup = false;
 423 #endif
 424         /*
 425          * We don't want to deadlock with somebody trying to allocate a new
 426          * extent for the extent root while also trying to search the extent
 427          * root to add free space.  So we skip locking and search the commit
 428          * root, since its read-only
 429          */
 430         path->skip_locking = 1;
 431         path->search_commit_root = 1;
 432         path->reada = READA_FORWARD;
 433
 434         key.objectid = last;
 435         key.offset = 0;
 436         key.type = BTRFS_EXTENT_ITEM_KEY;
 437
 438 next:
 439         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 440         if (ret < 0)
 441                 goto out;
 442
 443         leaf = path->nodes[0];
 444         nritems = btrfs_header_nritems(leaf);
 445
 446         while (1) {
 447                 if (btrfs_fs_closing(fs_info) > 1) {
 448                         last = (u64)-1;
 449                         break;
 450                 }
 451
 452                 if (path->slots[0] < nritems) {
 453                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 454                 } else {
 455                         ret = find_next_key(path, 0, &key);
 456                         if (ret)
 457                                 break;
 458
 459                         if (need_resched() ||
 460                             rwsem_is_contended(&fs_info->commit_root_sem)) {
 461                                 if (wakeup)
 462                                         caching_ctl->progress = last;
 463                                 btrfs_release_path(path);
 464                                 up_read(&fs_info->commit_root_sem);
 465                                 mutex_unlock(&caching_ctl->mutex);
 466                                 cond_resched();
 467                                 mutex_lock(&caching_ctl->mutex);
 468                                 down_read(&fs_info->commit_root_sem);
 469                                 goto next;
 470                         }
 471
 472                         ret = btrfs_next_leaf(extent_root, path);
 473                         if (ret < 0)
 474                                 goto out;
 475                         if (ret)
 476                                 break;
 477                         leaf = path->nodes[0];
 478                         nritems = btrfs_header_nritems(leaf);
 479                         continue;
 480                 }
 481
 482                 if (key.objectid < last) {
 483                         key.objectid = last;
 484                         key.offset = 0;
 485                         key.type = BTRFS_EXTENT_ITEM_KEY;
 486
 487                         if (wakeup)
 488                                 caching_ctl->progress = last;
 489                         btrfs_release_path(path);
 490                         goto next;
 491                 }
 492
 493                 if (key.objectid < block_group->key.objectid) {
 494                         path->slots[0]++;
 495                         continue;
 496                 }
 497
 498                 if (key.objectid >= block_group->key.objectid +
 499                     block_group->key.offset)
 500                         break;
 501
 502                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 503                     key.type == BTRFS_METADATA_ITEM_KEY) {
 504                         total_found += add_new_free_space(block_group,
 505                                                           fs_info, last,
 506                                                           key.objectid);
 507                         if (key.type == BTRFS_METADATA_ITEM_KEY)
 508                                 last = key.objectid +
 509                                         fs_info->nodesize;
 510                         else
 511                                 last = key.objectid + key.offset;
 512
 513                         if (total_found > CACHING_CTL_WAKE_UP) {
 514                                 total_found = 0;
 515                                 if (wakeup)
 516                                         wake_up(&caching_ctl->wait);
 517                         }
 518                 }
 519                 path->slots[0]++;
 520         }
 521         ret = 0;
 522
 523         total_found += add_new_free_space(block_group, fs_info, last,
 524                                           block_group->key.objectid +
 525                                           block_group->key.offset);
 526         caching_ctl->progress = (u64)-1;
 527
 528 out:
 529         btrfs_free_path(path);
 530         return ret;
 531 }
 532
 533 static noinline void caching_thread(struct btrfs_work *work)
 534 {
 535         struct btrfs_block_group_cache *block_group;
 536         struct btrfs_fs_info *fs_info;
 537         struct btrfs_caching_control *caching_ctl;
 538         int ret;
 539
 540         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 541         block_group = caching_ctl->block_group;
 542         fs_info = block_group->fs_info;
 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         refcount_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                 refcount_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 (btrfs_test_opt(fs_info, 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         refcount_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 static void add_pinned_bytes(struct btrfs_fs_info *fs_info, s64 num_bytes,
 759                              u64 owner, u64 root_objectid)
 760 {
 761         struct btrfs_space_info *space_info;
 762         u64 flags;
 763
 764         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
 765                 if (root_objectid == BTRFS_CHUNK_TREE_OBJECTID)
 766                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
 767                 else
 768                         flags = BTRFS_BLOCK_GROUP_METADATA;
 769         } else {
 770                 flags = BTRFS_BLOCK_GROUP_DATA;
 771         }
 772
 773         space_info = __find_space_info(fs_info, flags);
 774         ASSERT(space_info);
 775         percpu_counter_add(&space_info->total_bytes_pinned, num_bytes);
 776 }
 777
 778 /*
 779  * after adding space to the filesystem, we need to clear the full flags
 780  * on all the space infos.
 781  */
 782 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 783 {
 784         struct list_head *head = &info->space_info;
 785         struct btrfs_space_info *found;
 786
 787         rcu_read_lock();
 788         list_for_each_entry_rcu(found, head, list)
 789                 found->full = 0;
 790         rcu_read_unlock();
 791 }
 792
 793 /* simple helper to search for an existing data extent at a given offset */
 794 int btrfs_lookup_data_extent(struct btrfs_fs_info *fs_info, u64 start, u64 len)
 795 {
 796         int ret;
 797         struct btrfs_key key;
 798         struct btrfs_path *path;
 799
 800         path = btrfs_alloc_path();
 801         if (!path)
 802                 return -ENOMEM;
 803
 804         key.objectid = start;
 805         key.offset = len;
 806         key.type = BTRFS_EXTENT_ITEM_KEY;
 807         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
 808         btrfs_free_path(path);
 809         return ret;
 810 }
 811
 812 /*
 813  * helper function to lookup reference count and flags of a tree block.
 814  *
 815  * the head node for delayed ref is used to store the sum of all the
 816  * reference count modifications queued up in the rbtree. the head
 817  * node may also store the extent flags to set. This way you can check
 818  * to see what the reference count and extent flags would be if all of
 819  * the delayed refs are not processed.
 820  */
 821 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 822                              struct btrfs_fs_info *fs_info, u64 bytenr,
 823                              u64 offset, int metadata, u64 *refs, u64 *flags)
 824 {
 825         struct btrfs_delayed_ref_head *head;
 826         struct btrfs_delayed_ref_root *delayed_refs;
 827         struct btrfs_path *path;
 828         struct btrfs_extent_item *ei;
 829         struct extent_buffer *leaf;
 830         struct btrfs_key key;
 831         u32 item_size;
 832         u64 num_refs;
 833         u64 extent_flags;
 834         int ret;
 835
 836         /*
 837          * If we don't have skinny metadata, don't bother doing anything
 838          * different
 839          */
 840         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA)) {
 841                 offset = fs_info->nodesize;
 842                 metadata = 0;
 843         }
 844
 845         path = btrfs_alloc_path();
 846         if (!path)
 847                 return -ENOMEM;
 848
 849         if (!trans) {
 850                 path->skip_locking = 1;
 851                 path->search_commit_root = 1;
 852         }
 853
 854 search_again:
 855         key.objectid = bytenr;
 856         key.offset = offset;
 857         if (metadata)
 858                 key.type = BTRFS_METADATA_ITEM_KEY;
 859         else
 860                 key.type = BTRFS_EXTENT_ITEM_KEY;
 861
 862         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 0);
 863         if (ret < 0)
 864                 goto out_free;
 865
 866         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 867                 if (path->slots[0]) {
 868                         path->slots[0]--;
 869                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 870                                               path->slots[0]);
 871                         if (key.objectid == bytenr &&
 872                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 873                             key.offset == fs_info->nodesize)
 874                                 ret = 0;
 875                 }
 876         }
 877
 878         if (ret == 0) {
 879                 leaf = path->nodes[0];
 880                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 881                 if (item_size >= sizeof(*ei)) {
 882                         ei = btrfs_item_ptr(leaf, path->slots[0],
 883                                             struct btrfs_extent_item);
 884                         num_refs = btrfs_extent_refs(leaf, ei);
 885                         extent_flags = btrfs_extent_flags(leaf, ei);
 886                 } else {
 887 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 888                         struct btrfs_extent_item_v0 *ei0;
 889                         BUG_ON(item_size != sizeof(*ei0));
 890                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 891                                              struct btrfs_extent_item_v0);
 892                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
 893                         /* FIXME: this isn't correct for data */
 894                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 895 #else
 896                         BUG();
 897 #endif
 898                 }
 899                 BUG_ON(num_refs == 0);
 900         } else {
 901                 num_refs = 0;
 902                 extent_flags = 0;
 903                 ret = 0;
 904         }
 905
 906         if (!trans)
 907                 goto out;
 908
 909         delayed_refs = &trans->transaction->delayed_refs;
 910         spin_lock(&delayed_refs->lock);
 911         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
 912         if (head) {
 913                 if (!mutex_trylock(&head->mutex)) {
 914                         refcount_inc(&head->refs);
 915                         spin_unlock(&delayed_refs->lock);
 916
 917                         btrfs_release_path(path);
 918
 919                         /*
 920                          * Mutex was contended, block until it's released and try
 921                          * again
 922                          */
 923                         mutex_lock(&head->mutex);
 924                         mutex_unlock(&head->mutex);
 925                         btrfs_put_delayed_ref_head(head);
 926                         goto search_again;
 927                 }
 928                 spin_lock(&head->lock);
 929                 if (head->extent_op && head->extent_op->update_flags)
 930                         extent_flags |= head->extent_op->flags_to_set;
 931                 else
 932                         BUG_ON(num_refs == 0);
 933
 934                 num_refs += head->ref_mod;
 935                 spin_unlock(&head->lock);
 936                 mutex_unlock(&head->mutex);
 937         }
 938         spin_unlock(&delayed_refs->lock);
 939 out:
 940         WARN_ON(num_refs == 0);
 941         if (refs)
 942                 *refs = num_refs;
 943         if (flags)
 944                 *flags = extent_flags;
 945 out_free:
 946         btrfs_free_path(path);
 947         return ret;
 948 }
 949
 950 /*
 951  * Back reference rules.  Back refs have three main goals:
 952  *
 953  * 1) differentiate between all holders of references to an extent so that
 954  *    when a reference is dropped we can make sure it was a valid reference
 955  *    before freeing the extent.
 956  *
 957  * 2) Provide enough information to quickly find the holders of an extent
 958  *    if we notice a given block is corrupted or bad.
 959  *
 960  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 961  *    maintenance.  This is actually the same as #2, but with a slightly
 962  *    different use case.
 963  *
 964  * There are two kinds of back refs. The implicit back refs is optimized
 965  * for pointers in non-shared tree blocks. For a given pointer in a block,
 966  * back refs of this kind provide information about the block's owner tree
 967  * and the pointer's key. These information allow us to find the block by
 968  * b-tree searching. The full back refs is for pointers in tree blocks not
 969  * referenced by their owner trees. The location of tree block is recorded
 970  * in the back refs. Actually the full back refs is generic, and can be
 971  * used in all cases the implicit back refs is used. The major shortcoming
 972  * of the full back refs is its overhead. Every time a tree block gets
 973  * COWed, we have to update back refs entry for all pointers in it.
 974  *
 975  * For a newly allocated tree block, we use implicit back refs for
 976  * pointers in it. This means most tree related operations only involve
 977  * implicit back refs. For a tree block created in old transaction, the
 978  * only way to drop a reference to it is COW it. So we can detect the
 979  * event that tree block loses its owner tree's reference and do the
 980  * back refs conversion.
 981  *
 982  * When a tree block is COWed through a tree, there are four cases:
 983  *
 984  * The reference count of the block is one and the tree is the block's
 985  * owner tree. Nothing to do in this case.
 986  *
 987  * The reference count of the block is one and the tree is not the
 988  * block's owner tree. In this case, full back refs is used for pointers
 989  * in the block. Remove these full back refs, add implicit back refs for
 990  * every pointers in the new block.
 991  *
 992  * The reference count of the block is greater than one and the tree is
 993  * the block's owner tree. In this case, implicit back refs is used for
 994  * pointers in the block. Add full back refs for every pointers in the
 995  * block, increase lower level extents' reference counts. The original
 996  * implicit back refs are entailed to the new block.
 997  *
 998  * The reference count of the block is greater than one and the tree is
 999  * not the block's owner tree. Add implicit back refs for every pointer in
1000  * the new block, increase lower level extents' reference count.
1001  *
1002  * Back Reference Key composing:
1003  *
1004  * The key objectid corresponds to the first byte in the extent,
1005  * The key type is used to differentiate between types of back refs.
1006  * There are different meanings of the key offset for different types
1007  * of back refs.
1008  *
1009  * File extents can be referenced by:
1010  *
1011  * - multiple snapshots, subvolumes, or different generations in one subvol
1012  * - different files inside a single subvolume
1013  * - different offsets inside a file (bookend extents in file.c)
1014  *
1015  * The extent ref structure for the implicit back refs has fields for:
1016  *
1017  * - Objectid of the subvolume root
1018  * - objectid of the file holding the reference
1019  * - original offset in the file
1020  * - how many bookend extents
1021  *
1022  * The key offset for the implicit back refs is hash of the first
1023  * three fields.
1024  *
1025  * The extent ref structure for the full back refs has field for:
1026  *
1027  * - number of pointers in the tree leaf
1028  *
1029  * The key offset for the implicit back refs is the first byte of
1030  * the tree leaf
1031  *
1032  * When a file extent is allocated, The implicit back refs is used.
1033  * the fields are filled in:
1034  *
1035  *     (root_key.objectid, inode objectid, offset in file, 1)
1036  *
1037  * When a file extent is removed file truncation, we find the
1038  * corresponding implicit back refs and check the following fields:
1039  *
1040  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
1041  *
1042  * Btree extents can be referenced by:
1043  *
1044  * - Different subvolumes
1045  *
1046  * Both the implicit back refs and the full back refs for tree blocks
1047  * only consist of key. The key offset for the implicit back refs is
1048  * objectid of block's owner tree. The key offset for the full back refs
1049  * is the first byte of parent block.
1050  *
1051  * When implicit back refs is used, information about the lowest key and
1052  * level of the tree block are required. These information are stored in
1053  * tree block info structure.
1054  */
1055
1056 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1057 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
1058                                   struct btrfs_fs_info *fs_info,
1059                                   struct btrfs_path *path,
1060                                   u64 owner, u32 extra_size)
1061 {
1062         struct btrfs_root *root = fs_info->extent_root;
1063         struct btrfs_extent_item *item;
1064         struct btrfs_extent_item_v0 *ei0;
1065         struct btrfs_extent_ref_v0 *ref0;
1066         struct btrfs_tree_block_info *bi;
1067         struct extent_buffer *leaf;
1068         struct btrfs_key key;
1069         struct btrfs_key found_key;
1070         u32 new_size = sizeof(*item);
1071         u64 refs;
1072         int ret;
1073
1074         leaf = path->nodes[0];
1075         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
1076
1077         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1078         ei0 = btrfs_item_ptr(leaf, path->slots[0],
1079                              struct btrfs_extent_item_v0);
1080         refs = btrfs_extent_refs_v0(leaf, ei0);
1081
1082         if (owner == (u64)-1) {
1083                 while (1) {
1084                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1085                                 ret = btrfs_next_leaf(root, path);
1086                                 if (ret < 0)
1087                                         return ret;
1088                                 BUG_ON(ret > 0); /* Corruption */
1089                                 leaf = path->nodes[0];
1090                         }
1091                         btrfs_item_key_to_cpu(leaf, &found_key,
1092                                               path->slots[0]);
1093                         BUG_ON(key.objectid != found_key.objectid);
1094                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
1095                                 path->slots[0]++;
1096                                 continue;
1097                         }
1098                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1099                                               struct btrfs_extent_ref_v0);
1100                         owner = btrfs_ref_objectid_v0(leaf, ref0);
1101                         break;
1102                 }
1103         }
1104         btrfs_release_path(path);
1105
1106         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1107                 new_size += sizeof(*bi);
1108
1109         new_size -= sizeof(*ei0);
1110         ret = btrfs_search_slot(trans, root, &key, path,
1111                                 new_size + extra_size, 1);
1112         if (ret < 0)
1113                 return ret;
1114         BUG_ON(ret); /* Corruption */
1115
1116         btrfs_extend_item(fs_info, path, new_size);
1117
1118         leaf = path->nodes[0];
1119         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1120         btrfs_set_extent_refs(leaf, item, refs);
1121         /* FIXME: get real generation */
1122         btrfs_set_extent_generation(leaf, item, 0);
1123         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1124                 btrfs_set_extent_flags(leaf, item,
1125                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1126                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1127                 bi = (struct btrfs_tree_block_info *)(item + 1);
1128                 /* FIXME: get first key of the block */
1129                 memzero_extent_buffer(leaf, (unsigned long)bi, sizeof(*bi));
1130                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1131         } else {
1132                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1133         }
1134         btrfs_mark_buffer_dirty(leaf);
1135         return 0;
1136 }
1137 #endif
1138
1139 /*
1140  * is_data == BTRFS_REF_TYPE_BLOCK, tree block type is required,
1141  * is_data == BTRFS_REF_TYPE_DATA, data type is requried,
1142  * is_data == BTRFS_REF_TYPE_ANY, either type is OK.
1143  */
1144 int btrfs_get_extent_inline_ref_type(const struct extent_buffer *eb,
1145                                      struct btrfs_extent_inline_ref *iref,
1146                                      enum btrfs_inline_ref_type is_data)
1147 {
1148         int type = btrfs_extent_inline_ref_type(eb, iref);
1149         u64 offset = btrfs_extent_inline_ref_offset(eb, iref);
1150
1151         if (type == BTRFS_TREE_BLOCK_REF_KEY ||
1152             type == BTRFS_SHARED_BLOCK_REF_KEY ||
1153             type == BTRFS_SHARED_DATA_REF_KEY ||
1154             type == BTRFS_EXTENT_DATA_REF_KEY) {
1155                 if (is_data == BTRFS_REF_TYPE_BLOCK) {
1156                         if (type == BTRFS_TREE_BLOCK_REF_KEY)
1157                                 return type;
1158                         if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1159                                 ASSERT(eb->fs_info);
1160                                 /*
1161                                  * Every shared one has parent tree
1162                                  * block, which must be aligned to
1163                                  * nodesize.
1164                                  */
1165                                 if (offset &&
1166                                     IS_ALIGNED(offset, eb->fs_info->nodesize))
1167                                         return type;
1168                         }
1169                 } else if (is_data == BTRFS_REF_TYPE_DATA) {
1170                         if (type == BTRFS_EXTENT_DATA_REF_KEY)
1171                                 return type;
1172                         if (type == BTRFS_SHARED_DATA_REF_KEY) {
1173                                 ASSERT(eb->fs_info);
1174                                 /*
1175                                  * Every shared one has parent tree
1176                                  * block, which must be aligned to
1177                                  * nodesize.
1178                                  */
1179                                 if (offset &&
1180                                     IS_ALIGNED(offset, eb->fs_info->nodesize))
1181                                         return type;
1182                         }
1183                 } else {
1184                         ASSERT(is_data == BTRFS_REF_TYPE_ANY);
1185                         return type;
1186                 }
1187         }
1188
1189         btrfs_print_leaf((struct extent_buffer *)eb);
1190         btrfs_err(eb->fs_info, "eb %llu invalid extent inline ref type %d",
1191                   eb->start, type);
1192         WARN_ON(1);
1193
1194         return BTRFS_REF_TYPE_INVALID;
1195 }
1196
1197 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1198 {
1199         u32 high_crc = ~(u32)0;
1200         u32 low_crc = ~(u32)0;
1201         __le64 lenum;
1202
1203         lenum = cpu_to_le64(root_objectid);
1204         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
1205         lenum = cpu_to_le64(owner);
1206         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1207         lenum = cpu_to_le64(offset);
1208         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1209
1210         return ((u64)high_crc << 31) ^ (u64)low_crc;
1211 }
1212
1213 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1214                                      struct btrfs_extent_data_ref *ref)
1215 {
1216         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1217                                     btrfs_extent_data_ref_objectid(leaf, ref),
1218                                     btrfs_extent_data_ref_offset(leaf, ref));
1219 }
1220
1221 static int match_extent_data_ref(struct extent_buffer *leaf,
1222                                  struct btrfs_extent_data_ref *ref,
1223                                  u64 root_objectid, u64 owner, u64 offset)
1224 {
1225         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1226             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1227             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1228                 return 0;
1229         return 1;
1230 }
1231
1232 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1233                                            struct btrfs_fs_info *fs_info,
1234                                            struct btrfs_path *path,
1235                                            u64 bytenr, u64 parent,
1236                                            u64 root_objectid,
1237                                            u64 owner, u64 offset)
1238 {
1239         struct btrfs_root *root = fs_info->extent_root;
1240         struct btrfs_key key;
1241         struct btrfs_extent_data_ref *ref;
1242         struct extent_buffer *leaf;
1243         u32 nritems;
1244         int ret;
1245         int recow;
1246         int err = -ENOENT;
1247
1248         key.objectid = bytenr;
1249         if (parent) {
1250                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1251                 key.offset = parent;
1252         } else {
1253                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1254                 key.offset = hash_extent_data_ref(root_objectid,
1255                                                   owner, offset);
1256         }
1257 again:
1258         recow = 0;
1259         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1260         if (ret < 0) {
1261                 err = ret;
1262                 goto fail;
1263         }
1264
1265         if (parent) {
1266                 if (!ret)
1267                         return 0;
1268 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1269                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1270                 btrfs_release_path(path);
1271                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1272                 if (ret < 0) {
1273                         err = ret;
1274                         goto fail;
1275                 }
1276                 if (!ret)
1277                         return 0;
1278 #endif
1279                 goto fail;
1280         }
1281
1282         leaf = path->nodes[0];
1283         nritems = btrfs_header_nritems(leaf);
1284         while (1) {
1285                 if (path->slots[0] >= nritems) {
1286                         ret = btrfs_next_leaf(root, path);
1287                         if (ret < 0)
1288                                 err = ret;
1289                         if (ret)
1290                                 goto fail;
1291
1292                         leaf = path->nodes[0];
1293                         nritems = btrfs_header_nritems(leaf);
1294                         recow = 1;
1295                 }
1296
1297                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1298                 if (key.objectid != bytenr ||
1299                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1300                         goto fail;
1301
1302                 ref = btrfs_item_ptr(leaf, path->slots[0],
1303                                      struct btrfs_extent_data_ref);
1304
1305                 if (match_extent_data_ref(leaf, ref, root_objectid,
1306                                           owner, offset)) {
1307                         if (recow) {
1308                                 btrfs_release_path(path);
1309                                 goto again;
1310                         }
1311                         err = 0;
1312                         break;
1313                 }
1314                 path->slots[0]++;
1315         }
1316 fail:
1317         return err;
1318 }
1319
1320 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1321                                            struct btrfs_fs_info *fs_info,
1322                                            struct btrfs_path *path,
1323                                            u64 bytenr, u64 parent,
1324                                            u64 root_objectid, u64 owner,
1325                                            u64 offset, int refs_to_add)
1326 {
1327         struct btrfs_root *root = fs_info->extent_root;
1328         struct btrfs_key key;
1329         struct extent_buffer *leaf;
1330         u32 size;
1331         u32 num_refs;
1332         int ret;
1333
1334         key.objectid = bytenr;
1335         if (parent) {
1336                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1337                 key.offset = parent;
1338                 size = sizeof(struct btrfs_shared_data_ref);
1339         } else {
1340                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1341                 key.offset = hash_extent_data_ref(root_objectid,
1342                                                   owner, offset);
1343                 size = sizeof(struct btrfs_extent_data_ref);
1344         }
1345
1346         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1347         if (ret && ret != -EEXIST)
1348                 goto fail;
1349
1350         leaf = path->nodes[0];
1351         if (parent) {
1352                 struct btrfs_shared_data_ref *ref;
1353                 ref = btrfs_item_ptr(leaf, path->slots[0],
1354                                      struct btrfs_shared_data_ref);
1355                 if (ret == 0) {
1356                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1357                 } else {
1358                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1359                         num_refs += refs_to_add;
1360                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1361                 }
1362         } else {
1363                 struct btrfs_extent_data_ref *ref;
1364                 while (ret == -EEXIST) {
1365                         ref = btrfs_item_ptr(leaf, path->slots[0],
1366                                              struct btrfs_extent_data_ref);
1367                         if (match_extent_data_ref(leaf, ref, root_objectid,
1368                                                   owner, offset))
1369                                 break;
1370                         btrfs_release_path(path);
1371                         key.offset++;
1372                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1373                                                       size);
1374                         if (ret && ret != -EEXIST)
1375                                 goto fail;
1376
1377                         leaf = path->nodes[0];
1378                 }
1379                 ref = btrfs_item_ptr(leaf, path->slots[0],
1380                                      struct btrfs_extent_data_ref);
1381                 if (ret == 0) {
1382                         btrfs_set_extent_data_ref_root(leaf, ref,
1383                                                        root_objectid);
1384                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1385                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1386                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1387                 } else {
1388                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1389                         num_refs += refs_to_add;
1390                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1391                 }
1392         }
1393         btrfs_mark_buffer_dirty(leaf);
1394         ret = 0;
1395 fail:
1396         btrfs_release_path(path);
1397         return ret;
1398 }
1399
1400 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1401                                            struct btrfs_fs_info *fs_info,
1402                                            struct btrfs_path *path,
1403                                            int refs_to_drop, int *last_ref)
1404 {
1405         struct btrfs_key key;
1406         struct btrfs_extent_data_ref *ref1 = NULL;
1407         struct btrfs_shared_data_ref *ref2 = NULL;
1408         struct extent_buffer *leaf;
1409         u32 num_refs = 0;
1410         int ret = 0;
1411
1412         leaf = path->nodes[0];
1413         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1414
1415         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1416                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1417                                       struct btrfs_extent_data_ref);
1418                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1419         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1420                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1421                                       struct btrfs_shared_data_ref);
1422                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1423 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1424         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1425                 struct btrfs_extent_ref_v0 *ref0;
1426                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1427                                       struct btrfs_extent_ref_v0);
1428                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1429 #endif
1430         } else {
1431                 BUG();
1432         }
1433
1434         BUG_ON(num_refs < refs_to_drop);
1435         num_refs -= refs_to_drop;
1436
1437         if (num_refs == 0) {
1438                 ret = btrfs_del_item(trans, fs_info->extent_root, path);
1439                 *last_ref = 1;
1440         } else {
1441                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1442                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1443                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1444                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1445 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1446                 else {
1447                         struct btrfs_extent_ref_v0 *ref0;
1448                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1449                                         struct btrfs_extent_ref_v0);
1450                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1451                 }
1452 #endif
1453                 btrfs_mark_buffer_dirty(leaf);
1454         }
1455         return ret;
1456 }
1457
1458 static noinline u32 extent_data_ref_count(struct btrfs_path *path,
1459                                           struct btrfs_extent_inline_ref *iref)
1460 {
1461         struct btrfs_key key;
1462         struct extent_buffer *leaf;
1463         struct btrfs_extent_data_ref *ref1;
1464         struct btrfs_shared_data_ref *ref2;
1465         u32 num_refs = 0;
1466         int type;
1467
1468         leaf = path->nodes[0];
1469         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1470         if (iref) {
1471                 /*
1472                  * If type is invalid, we should have bailed out earlier than
1473                  * this call.
1474                  */
1475                 type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
1476                 ASSERT(type != BTRFS_REF_TYPE_INVALID);
1477                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1478                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1479                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1480                 } else {
1481                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1482                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1483                 }
1484         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1485                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1486                                       struct btrfs_extent_data_ref);
1487                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1488         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1489                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1490                                       struct btrfs_shared_data_ref);
1491                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1492 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1493         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1494                 struct btrfs_extent_ref_v0 *ref0;
1495                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1496                                       struct btrfs_extent_ref_v0);
1497                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1498 #endif
1499         } else {
1500                 WARN_ON(1);
1501         }
1502         return num_refs;
1503 }
1504
1505 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1506                                           struct btrfs_fs_info *fs_info,
1507                                           struct btrfs_path *path,
1508                                           u64 bytenr, u64 parent,
1509                                           u64 root_objectid)
1510 {
1511         struct btrfs_root *root = fs_info->extent_root;
1512         struct btrfs_key key;
1513         int ret;
1514
1515         key.objectid = bytenr;
1516         if (parent) {
1517                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1518                 key.offset = parent;
1519         } else {
1520                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1521                 key.offset = root_objectid;
1522         }
1523
1524         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1525         if (ret > 0)
1526                 ret = -ENOENT;
1527 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1528         if (ret == -ENOENT && parent) {
1529                 btrfs_release_path(path);
1530                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1531                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1532                 if (ret > 0)
1533                         ret = -ENOENT;
1534         }
1535 #endif
1536         return ret;
1537 }
1538
1539 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1540                                           struct btrfs_fs_info *fs_info,
1541                                           struct btrfs_path *path,
1542                                           u64 bytenr, u64 parent,
1543                                           u64 root_objectid)
1544 {
1545         struct btrfs_key key;
1546         int ret;
1547
1548         key.objectid = bytenr;
1549         if (parent) {
1550                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1551                 key.offset = parent;
1552         } else {
1553                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1554                 key.offset = root_objectid;
1555         }
1556
1557         ret = btrfs_insert_empty_item(trans, fs_info->extent_root,
1558                                       path, &key, 0);
1559         btrfs_release_path(path);
1560         return ret;
1561 }
1562
1563 static inline int extent_ref_type(u64 parent, u64 owner)
1564 {
1565         int type;
1566         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1567                 if (parent > 0)
1568                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1569                 else
1570                         type = BTRFS_TREE_BLOCK_REF_KEY;
1571         } else {
1572                 if (parent > 0)
1573                         type = BTRFS_SHARED_DATA_REF_KEY;
1574                 else
1575                         type = BTRFS_EXTENT_DATA_REF_KEY;
1576         }
1577         return type;
1578 }
1579
1580 static int find_next_key(struct btrfs_path *path, int level,
1581                          struct btrfs_key *key)
1582
1583 {
1584         for (; level < BTRFS_MAX_LEVEL; level++) {
1585                 if (!path->nodes[level])
1586                         break;
1587                 if (path->slots[level] + 1 >=
1588                     btrfs_header_nritems(path->nodes[level]))
1589                         continue;
1590                 if (level == 0)
1591                         btrfs_item_key_to_cpu(path->nodes[level], key,
1592                                               path->slots[level] + 1);
1593                 else
1594                         btrfs_node_key_to_cpu(path->nodes[level], key,
1595                                               path->slots[level] + 1);
1596                 return 0;
1597         }
1598         return 1;
1599 }
1600
1601 /*
1602  * look for inline back ref. if back ref is found, *ref_ret is set
1603  * to the address of inline back ref, and 0 is returned.
1604  *
1605  * if back ref isn't found, *ref_ret is set to the address where it
1606  * should be inserted, and -ENOENT is returned.
1607  *
1608  * if insert is true and there are too many inline back refs, the path
1609  * points to the extent item, and -EAGAIN is returned.
1610  *
1611  * NOTE: inline back refs are ordered in the same way that back ref
1612  *       items in the tree are ordered.
1613  */
1614 static noinline_for_stack
1615 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1616                                  struct btrfs_fs_info *fs_info,
1617                                  struct btrfs_path *path,
1618                                  struct btrfs_extent_inline_ref **ref_ret,
1619                                  u64 bytenr, u64 num_bytes,
1620                                  u64 parent, u64 root_objectid,
1621                                  u64 owner, u64 offset, int insert)
1622 {
1623         struct btrfs_root *root = fs_info->extent_root;
1624         struct btrfs_key key;
1625         struct extent_buffer *leaf;
1626         struct btrfs_extent_item *ei;
1627         struct btrfs_extent_inline_ref *iref;
1628         u64 flags;
1629         u64 item_size;
1630         unsigned long ptr;
1631         unsigned long end;
1632         int extra_size;
1633         int type;
1634         int want;
1635         int ret;
1636         int err = 0;
1637         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
1638         int needed;
1639
1640         key.objectid = bytenr;
1641         key.type = BTRFS_EXTENT_ITEM_KEY;
1642         key.offset = num_bytes;
1643
1644         want = extent_ref_type(parent, owner);
1645         if (insert) {
1646                 extra_size = btrfs_extent_inline_ref_size(want);
1647                 path->keep_locks = 1;
1648         } else
1649                 extra_size = -1;
1650
1651         /*
1652          * Owner is our parent level, so we can just add one to get the level
1653          * for the block we are interested in.
1654          */
1655         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
1656                 key.type = BTRFS_METADATA_ITEM_KEY;
1657                 key.offset = owner;
1658         }
1659
1660 again:
1661         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1662         if (ret < 0) {
1663                 err = ret;
1664                 goto out;
1665         }
1666
1667         /*
1668          * We may be a newly converted file system which still has the old fat
1669          * extent entries for metadata, so try and see if we have one of those.
1670          */
1671         if (ret > 0 && skinny_metadata) {
1672                 skinny_metadata = false;
1673                 if (path->slots[0]) {
1674                         path->slots[0]--;
1675                         btrfs_item_key_to_cpu(path->nodes[0], &key,
1676                                               path->slots[0]);
1677                         if (key.objectid == bytenr &&
1678                             key.type == BTRFS_EXTENT_ITEM_KEY &&
1679                             key.offset == num_bytes)
1680                                 ret = 0;
1681                 }
1682                 if (ret) {
1683                         key.objectid = bytenr;
1684                         key.type = BTRFS_EXTENT_ITEM_KEY;
1685                         key.offset = num_bytes;
1686                         btrfs_release_path(path);
1687                         goto again;
1688                 }
1689         }
1690
1691         if (ret && !insert) {
1692                 err = -ENOENT;
1693                 goto out;
1694         } else if (WARN_ON(ret)) {
1695                 err = -EIO;
1696                 goto out;
1697         }
1698
1699         leaf = path->nodes[0];
1700         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1701 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1702         if (item_size < sizeof(*ei)) {
1703                 if (!insert) {
1704                         err = -ENOENT;
1705                         goto out;
1706                 }
1707                 ret = convert_extent_item_v0(trans, fs_info, path, owner,
1708                                              extra_size);
1709                 if (ret < 0) {
1710                         err = ret;
1711                         goto out;
1712                 }
1713                 leaf = path->nodes[0];
1714                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1715         }
1716 #endif
1717         BUG_ON(item_size < sizeof(*ei));
1718
1719         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1720         flags = btrfs_extent_flags(leaf, ei);
1721
1722         ptr = (unsigned long)(ei + 1);
1723         end = (unsigned long)ei + item_size;
1724
1725         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
1726                 ptr += sizeof(struct btrfs_tree_block_info);
1727                 BUG_ON(ptr > end);
1728         }
1729
1730         if (owner >= BTRFS_FIRST_FREE_OBJECTID)
1731                 needed = BTRFS_REF_TYPE_DATA;
1732         else
1733                 needed = BTRFS_REF_TYPE_BLOCK;
1734
1735         err = -ENOENT;
1736         while (1) {
1737                 if (ptr >= end) {
1738                         WARN_ON(ptr > end);
1739                         break;
1740                 }
1741                 iref = (struct btrfs_extent_inline_ref *)ptr;
1742                 type = btrfs_get_extent_inline_ref_type(leaf, iref, needed);
1743                 if (type == BTRFS_REF_TYPE_INVALID) {
1744                         err = -EINVAL;
1745                         goto out;
1746                 }
1747
1748                 if (want < type)
1749                         break;
1750                 if (want > type) {
1751                         ptr += btrfs_extent_inline_ref_size(type);
1752                         continue;
1753                 }
1754
1755                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1756                         struct btrfs_extent_data_ref *dref;
1757                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1758                         if (match_extent_data_ref(leaf, dref, root_objectid,
1759                                                   owner, offset)) {
1760                                 err = 0;
1761                                 break;
1762                         }
1763                         if (hash_extent_data_ref_item(leaf, dref) <
1764                             hash_extent_data_ref(root_objectid, owner, offset))
1765                                 break;
1766                 } else {
1767                         u64 ref_offset;
1768                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1769                         if (parent > 0) {
1770                                 if (parent == ref_offset) {
1771                                         err = 0;
1772                                         break;
1773                                 }
1774                                 if (ref_offset < parent)
1775                                         break;
1776                         } else {
1777                                 if (root_objectid == ref_offset) {
1778                                         err = 0;
1779                                         break;
1780                                 }
1781                                 if (ref_offset < root_objectid)
1782                                         break;
1783                         }
1784                 }
1785                 ptr += btrfs_extent_inline_ref_size(type);
1786         }
1787         if (err == -ENOENT && insert) {
1788                 if (item_size + extra_size >=
1789                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1790                         err = -EAGAIN;
1791                         goto out;
1792                 }
1793                 /*
1794                  * To add new inline back ref, we have to make sure
1795                  * there is no corresponding back ref item.
1796                  * For simplicity, we just do not add new inline back
1797                  * ref if there is any kind of item for this block
1798                  */
1799                 if (find_next_key(path, 0, &key) == 0 &&
1800                     key.objectid == bytenr &&
1801                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1802                         err = -EAGAIN;
1803                         goto out;
1804                 }
1805         }
1806         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1807 out:
1808         if (insert) {
1809                 path->keep_locks = 0;
1810                 btrfs_unlock_up_safe(path, 1);
1811         }
1812         return err;
1813 }
1814
1815 /*
1816  * helper to add new inline back ref
1817  */
1818 static noinline_for_stack
1819 void setup_inline_extent_backref(struct btrfs_fs_info *fs_info,
1820                                  struct btrfs_path *path,
1821                                  struct btrfs_extent_inline_ref *iref,
1822                                  u64 parent, u64 root_objectid,
1823                                  u64 owner, u64 offset, int refs_to_add,
1824                                  struct btrfs_delayed_extent_op *extent_op)
1825 {
1826         struct extent_buffer *leaf;
1827         struct btrfs_extent_item *ei;
1828         unsigned long ptr;
1829         unsigned long end;
1830         unsigned long item_offset;
1831         u64 refs;
1832         int size;
1833         int type;
1834
1835         leaf = path->nodes[0];
1836         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1837         item_offset = (unsigned long)iref - (unsigned long)ei;
1838
1839         type = extent_ref_type(parent, owner);
1840         size = btrfs_extent_inline_ref_size(type);
1841
1842         btrfs_extend_item(fs_info, path, size);
1843
1844         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1845         refs = btrfs_extent_refs(leaf, ei);
1846         refs += refs_to_add;
1847         btrfs_set_extent_refs(leaf, ei, refs);
1848         if (extent_op)
1849                 __run_delayed_extent_op(extent_op, leaf, ei);
1850
1851         ptr = (unsigned long)ei + item_offset;
1852         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1853         if (ptr < end - size)
1854                 memmove_extent_buffer(leaf, ptr + size, ptr,
1855                                       end - size - ptr);
1856
1857         iref = (struct btrfs_extent_inline_ref *)ptr;
1858         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1859         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1860                 struct btrfs_extent_data_ref *dref;
1861                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1862                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1863                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1864                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1865                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1866         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1867                 struct btrfs_shared_data_ref *sref;
1868                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1869                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1870                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1871         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1872                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1873         } else {
1874                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1875         }
1876         btrfs_mark_buffer_dirty(leaf);
1877 }
1878
1879 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1880                                  struct btrfs_fs_info *fs_info,
1881                                  struct btrfs_path *path,
1882                                  struct btrfs_extent_inline_ref **ref_ret,
1883                                  u64 bytenr, u64 num_bytes, u64 parent,
1884                                  u64 root_objectid, u64 owner, u64 offset)
1885 {
1886         int ret;
1887
1888         ret = lookup_inline_extent_backref(trans, fs_info, path, ref_ret,
1889                                            bytenr, num_bytes, parent,
1890                                            root_objectid, owner, offset, 0);
1891         if (ret != -ENOENT)
1892                 return ret;
1893
1894         btrfs_release_path(path);
1895         *ref_ret = NULL;
1896
1897         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1898                 ret = lookup_tree_block_ref(trans, fs_info, path, bytenr,
1899                                             parent, root_objectid);
1900         } else {
1901                 ret = lookup_extent_data_ref(trans, fs_info, path, bytenr,
1902                                              parent, root_objectid, owner,
1903                                              offset);
1904         }
1905         return ret;
1906 }
1907
1908 /*
1909  * helper to update/remove inline back ref
1910  */
1911 static noinline_for_stack
1912 void update_inline_extent_backref(struct btrfs_fs_info *fs_info,
1913                                   struct btrfs_path *path,
1914                                   struct btrfs_extent_inline_ref *iref,
1915                                   int refs_to_mod,
1916                                   struct btrfs_delayed_extent_op *extent_op,
1917                                   int *last_ref)
1918 {
1919         struct extent_buffer *leaf;
1920         struct btrfs_extent_item *ei;
1921         struct btrfs_extent_data_ref *dref = NULL;
1922         struct btrfs_shared_data_ref *sref = NULL;
1923         unsigned long ptr;
1924         unsigned long end;
1925         u32 item_size;
1926         int size;
1927         int type;
1928         u64 refs;
1929
1930         leaf = path->nodes[0];
1931         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1932         refs = btrfs_extent_refs(leaf, ei);
1933         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1934         refs += refs_to_mod;
1935         btrfs_set_extent_refs(leaf, ei, refs);
1936         if (extent_op)
1937                 __run_delayed_extent_op(extent_op, leaf, ei);
1938
1939         /*
1940          * If type is invalid, we should have bailed out after
1941          * lookup_inline_extent_backref().
1942          */
1943         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_ANY);
1944         ASSERT(type != BTRFS_REF_TYPE_INVALID);
1945
1946         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1947                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1948                 refs = btrfs_extent_data_ref_count(leaf, dref);
1949         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1950                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1951                 refs = btrfs_shared_data_ref_count(leaf, sref);
1952         } else {
1953                 refs = 1;
1954                 BUG_ON(refs_to_mod != -1);
1955         }
1956
1957         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1958         refs += refs_to_mod;
1959
1960         if (refs > 0) {
1961                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1962                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1963                 else
1964                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1965         } else {
1966                 *last_ref = 1;
1967                 size =  btrfs_extent_inline_ref_size(type);
1968                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1969                 ptr = (unsigned long)iref;
1970                 end = (unsigned long)ei + item_size;
1971                 if (ptr + size < end)
1972                         memmove_extent_buffer(leaf, ptr, ptr + size,
1973                                               end - ptr - size);
1974                 item_size -= size;
1975                 btrfs_truncate_item(fs_info, path, item_size, 1);
1976         }
1977         btrfs_mark_buffer_dirty(leaf);
1978 }
1979
1980 static noinline_for_stack
1981 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1982                                  struct btrfs_fs_info *fs_info,
1983                                  struct btrfs_path *path,
1984                                  u64 bytenr, u64 num_bytes, u64 parent,
1985                                  u64 root_objectid, u64 owner,
1986                                  u64 offset, int refs_to_add,
1987                                  struct btrfs_delayed_extent_op *extent_op)
1988 {
1989         struct btrfs_extent_inline_ref *iref;
1990         int ret;
1991
1992         ret = lookup_inline_extent_backref(trans, fs_info, path, &iref,
1993                                            bytenr, num_bytes, parent,
1994                                            root_objectid, owner, offset, 1);
1995         if (ret == 0) {
1996                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1997                 update_inline_extent_backref(fs_info, path, iref,
1998                                              refs_to_add, extent_op, NULL);
1999         } else if (ret == -ENOENT) {
2000                 setup_inline_extent_backref(fs_info, path, iref, parent,
2001                                             root_objectid, owner, offset,
2002                                             refs_to_add, extent_op);
2003                 ret = 0;
2004         }
2005         return ret;
2006 }
2007
2008 static int insert_extent_backref(struct btrfs_trans_handle *trans,
2009                                  struct btrfs_fs_info *fs_info,
2010                                  struct btrfs_path *path,
2011                                  u64 bytenr, u64 parent, u64 root_objectid,
2012                                  u64 owner, u64 offset, int refs_to_add)
2013 {
2014         int ret;
2015         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
2016                 BUG_ON(refs_to_add != 1);
2017                 ret = insert_tree_block_ref(trans, fs_info, path, bytenr,
2018                                             parent, root_objectid);
2019         } else {
2020                 ret = insert_extent_data_ref(trans, fs_info, path, bytenr,
2021                                              parent, root_objectid,
2022                                              owner, offset, refs_to_add);
2023         }
2024         return ret;
2025 }
2026
2027 static int remove_extent_backref(struct btrfs_trans_handle *trans,
2028                                  struct btrfs_fs_info *fs_info,
2029                                  struct btrfs_path *path,
2030                                  struct btrfs_extent_inline_ref *iref,
2031                                  int refs_to_drop, int is_data, int *last_ref)
2032 {
2033         int ret = 0;
2034
2035         BUG_ON(!is_data && refs_to_drop != 1);
2036         if (iref) {
2037                 update_inline_extent_backref(fs_info, path, iref,
2038                                              -refs_to_drop, NULL, last_ref);
2039         } else if (is_data) {
2040                 ret = remove_extent_data_ref(trans, fs_info, path, refs_to_drop,
2041                                              last_ref);
2042         } else {
2043                 *last_ref = 1;
2044                 ret = btrfs_del_item(trans, fs_info->extent_root, path);
2045         }
2046         return ret;
2047 }
2048
2049 #define in_range(b, first, len)        ((b) >= (first) && (b) < (first) + (len))
2050 static int btrfs_issue_discard(struct block_device *bdev, u64 start, u64 len,
2051                                u64 *discarded_bytes)
2052 {
2053         int j, ret = 0;
2054         u64 bytes_left, end;
2055         u64 aligned_start = ALIGN(start, 1 << 9);
2056
2057         if (WARN_ON(start != aligned_start)) {
2058                 len -= aligned_start - start;
2059                 len = round_down(len, 1 << 9);
2060                 start = aligned_start;
2061         }
2062
2063         *discarded_bytes = 0;
2064
2065         if (!len)
2066                 return 0;
2067
2068         end = start + len;
2069         bytes_left = len;
2070
2071         /* Skip any superblocks on this device. */
2072         for (j = 0; j < BTRFS_SUPER_MIRROR_MAX; j++) {
2073                 u64 sb_start = btrfs_sb_offset(j);
2074                 u64 sb_end = sb_start + BTRFS_SUPER_INFO_SIZE;
2075                 u64 size = sb_start - start;
2076
2077                 if (!in_range(sb_start, start, bytes_left) &&
2078                     !in_range(sb_end, start, bytes_left) &&
2079                     !in_range(start, sb_start, BTRFS_SUPER_INFO_SIZE))
2080                         continue;
2081
2082                 /*
2083                  * Superblock spans beginning of range.  Adjust start and
2084                  * try again.
2085                  */
2086                 if (sb_start <= start) {
2087                         start += sb_end - start;
2088                         if (start > end) {
2089                                 bytes_left = 0;
2090                                 break;
2091                         }
2092                         bytes_left = end - start;
2093                         continue;
2094                 }
2095
2096                 if (size) {
2097                         ret = blkdev_issue_discard(bdev, start >> 9, size >> 9,
2098                                                    GFP_NOFS, 0);
2099                         if (!ret)
2100                                 *discarded_bytes += size;
2101                         else if (ret != -EOPNOTSUPP)
2102                                 return ret;
2103                 }
2104
2105                 start = sb_end;
2106                 if (start > end) {
2107                         bytes_left = 0;
2108                         break;
2109                 }
2110                 bytes_left = end - start;
2111         }
2112
2113         if (bytes_left) {
2114                 ret = blkdev_issue_discard(bdev, start >> 9, bytes_left >> 9,
2115                                            GFP_NOFS, 0);
2116                 if (!ret)
2117                         *discarded_bytes += bytes_left;
2118         }
2119         return ret;
2120 }
2121
2122 int btrfs_discard_extent(struct btrfs_fs_info *fs_info, u64 bytenr,
2123                          u64 num_bytes, u64 *actual_bytes)
2124 {
2125         int ret;
2126         u64 discarded_bytes = 0;
2127         struct btrfs_bio *bbio = NULL;
2128
2129
2130         /*
2131          * Avoid races with device replace and make sure our bbio has devices
2132          * associated to its stripes that don't go away while we are discarding.
2133          */
2134         btrfs_bio_counter_inc_blocked(fs_info);
2135         /* Tell the block device(s) that the sectors can be discarded */
2136         ret = btrfs_map_block(fs_info, BTRFS_MAP_DISCARD, bytenr, &num_bytes,
2137                               &bbio, 0);
2138         /* Error condition is -ENOMEM */
2139         if (!ret) {
2140                 struct btrfs_bio_stripe *stripe = bbio->stripes;
2141                 int i;
2142
2143
2144                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
2145                         u64 bytes;
2146                         struct request_queue *req_q;
2147
2148                         if (!stripe->dev->bdev) {
2149                                 ASSERT(btrfs_test_opt(fs_info, DEGRADED));
2150                                 continue;
2151                         }
2152                         req_q = bdev_get_queue(stripe->dev->bdev);
2153                         if (!blk_queue_discard(req_q))
2154                                 continue;
2155
2156                         ret = btrfs_issue_discard(stripe->dev->bdev,
2157                                                   stripe->physical,
2158                                                   stripe->length,
2159                                                   &bytes);
2160                         if (!ret)
2161                                 discarded_bytes += bytes;
2162                         else if (ret != -EOPNOTSUPP)
2163                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
2164
2165                         /*
2166                          * Just in case we get back EOPNOTSUPP for some reason,
2167                          * just ignore the return value so we don't screw up
2168                          * people calling discard_extent.
2169                          */
2170                         ret = 0;
2171                 }
2172                 btrfs_put_bbio(bbio);
2173         }
2174         btrfs_bio_counter_dec(fs_info);
2175
2176         if (actual_bytes)
2177                 *actual_bytes = discarded_bytes;
2178
2179
2180         if (ret == -EOPNOTSUPP)
2181                 ret = 0;
2182         return ret;
2183 }
2184
2185 /* Can return -ENOMEM */
2186 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2187                          struct btrfs_root *root,
2188                          u64 bytenr, u64 num_bytes, u64 parent,
2189                          u64 root_objectid, u64 owner, u64 offset)
2190 {
2191         struct btrfs_fs_info *fs_info = root->fs_info;
2192         int old_ref_mod, new_ref_mod;
2193         int ret;
2194
2195         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
2196                root_objectid == BTRFS_TREE_LOG_OBJECTID);
2197
2198         btrfs_ref_tree_mod(root, bytenr, num_bytes, parent, root_objectid,
2199                            owner, offset, BTRFS_ADD_DELAYED_REF);
2200
2201         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
2202                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
2203                                                  num_bytes, parent,
2204                                                  root_objectid, (int)owner,
2205                                                  BTRFS_ADD_DELAYED_REF, NULL,
2206                                                  &old_ref_mod, &new_ref_mod);
2207         } else {
2208                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
2209                                                  num_bytes, parent,
2210                                                  root_objectid, owner, offset,
2211                                                  0, BTRFS_ADD_DELAYED_REF,
2212                                                  &old_ref_mod, &new_ref_mod);
2213         }
2214
2215         if (ret == 0 && old_ref_mod < 0 && new_ref_mod >= 0)
2216                 add_pinned_bytes(fs_info, -num_bytes, owner, root_objectid);
2217
2218         return ret;
2219 }
2220
2221 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2222                                   struct btrfs_fs_info *fs_info,
2223                                   struct btrfs_delayed_ref_node *node,
2224                                   u64 parent, u64 root_objectid,
2225                                   u64 owner, u64 offset, int refs_to_add,
2226                                   struct btrfs_delayed_extent_op *extent_op)
2227 {
2228         struct btrfs_path *path;
2229         struct extent_buffer *leaf;
2230         struct btrfs_extent_item *item;
2231         struct btrfs_key key;
2232         u64 bytenr = node->bytenr;
2233         u64 num_bytes = node->num_bytes;
2234         u64 refs;
2235         int ret;
2236
2237         path = btrfs_alloc_path();
2238         if (!path)
2239                 return -ENOMEM;
2240
2241         path->reada = READA_FORWARD;
2242         path->leave_spinning = 1;
2243         /* this will setup the path even if it fails to insert the back ref */
2244         ret = insert_inline_extent_backref(trans, fs_info, path, bytenr,
2245                                            num_bytes, parent, root_objectid,
2246                                            owner, offset,
2247                                            refs_to_add, extent_op);
2248         if ((ret < 0 && ret != -EAGAIN) || !ret)
2249                 goto out;
2250
2251         /*
2252          * Ok we had -EAGAIN which means we didn't have space to insert and
2253          * inline extent ref, so just update the reference count and add a
2254          * normal backref.
2255          */
2256         leaf = path->nodes[0];
2257         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2258         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2259         refs = btrfs_extent_refs(leaf, item);
2260         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
2261         if (extent_op)
2262                 __run_delayed_extent_op(extent_op, leaf, item);
2263
2264         btrfs_mark_buffer_dirty(leaf);
2265         btrfs_release_path(path);
2266
2267         path->reada = READA_FORWARD;
2268         path->leave_spinning = 1;
2269         /* now insert the actual backref */
2270         ret = insert_extent_backref(trans, fs_info, path, bytenr, parent,
2271                                     root_objectid, owner, offset, refs_to_add);
2272         if (ret)
2273                 btrfs_abort_transaction(trans, ret);
2274 out:
2275         btrfs_free_path(path);
2276         return ret;
2277 }
2278
2279 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
2280                                 struct btrfs_fs_info *fs_info,
2281                                 struct btrfs_delayed_ref_node *node,
2282                                 struct btrfs_delayed_extent_op *extent_op,
2283                                 int insert_reserved)
2284 {
2285         int ret = 0;
2286         struct btrfs_delayed_data_ref *ref;
2287         struct btrfs_key ins;
2288         u64 parent = 0;
2289         u64 ref_root = 0;
2290         u64 flags = 0;
2291
2292         ins.objectid = node->bytenr;
2293         ins.offset = node->num_bytes;
2294         ins.type = BTRFS_EXTENT_ITEM_KEY;
2295
2296         ref = btrfs_delayed_node_to_data_ref(node);
2297         trace_run_delayed_data_ref(fs_info, node, ref, node->action);
2298
2299         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2300                 parent = ref->parent;
2301         ref_root = ref->root;
2302
2303         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2304                 if (extent_op)
2305                         flags |= extent_op->flags_to_set;
2306                 ret = alloc_reserved_file_extent(trans, fs_info,
2307                                                  parent, ref_root, flags,
2308                                                  ref->objectid, ref->offset,
2309                                                  &ins, node->ref_mod);
2310         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2311                 ret = __btrfs_inc_extent_ref(trans, fs_info, node, parent,
2312                                              ref_root, ref->objectid,
2313                                              ref->offset, node->ref_mod,
2314                                              extent_op);
2315         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2316                 ret = __btrfs_free_extent(trans, fs_info, node, parent,
2317                                           ref_root, ref->objectid,
2318                                           ref->offset, node->ref_mod,
2319                                           extent_op);
2320         } else {
2321                 BUG();
2322         }
2323         return ret;
2324 }
2325
2326 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2327                                     struct extent_buffer *leaf,
2328                                     struct btrfs_extent_item *ei)
2329 {
2330         u64 flags = btrfs_extent_flags(leaf, ei);
2331         if (extent_op->update_flags) {
2332                 flags |= extent_op->flags_to_set;
2333                 btrfs_set_extent_flags(leaf, ei, flags);
2334         }
2335
2336         if (extent_op->update_key) {
2337                 struct btrfs_tree_block_info *bi;
2338                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2339                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2340                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2341         }
2342 }
2343
2344 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2345                                  struct btrfs_fs_info *fs_info,
2346                                  struct btrfs_delayed_ref_head *head,
2347                                  struct btrfs_delayed_extent_op *extent_op)
2348 {
2349         struct btrfs_key key;
2350         struct btrfs_path *path;
2351         struct btrfs_extent_item *ei;
2352         struct extent_buffer *leaf;
2353         u32 item_size;
2354         int ret;
2355         int err = 0;
2356         int metadata = !extent_op->is_data;
2357
2358         if (trans->aborted)
2359                 return 0;
2360
2361         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2362                 metadata = 0;
2363
2364         path = btrfs_alloc_path();
2365         if (!path)
2366                 return -ENOMEM;
2367
2368         key.objectid = head->bytenr;
2369
2370         if (metadata) {
2371                 key.type = BTRFS_METADATA_ITEM_KEY;
2372                 key.offset = extent_op->level;
2373         } else {
2374                 key.type = BTRFS_EXTENT_ITEM_KEY;
2375                 key.offset = head->num_bytes;
2376         }
2377
2378 again:
2379         path->reada = READA_FORWARD;
2380         path->leave_spinning = 1;
2381         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 1);
2382         if (ret < 0) {
2383                 err = ret;
2384                 goto out;
2385         }
2386         if (ret > 0) {
2387                 if (metadata) {
2388                         if (path->slots[0] > 0) {
2389                                 path->slots[0]--;
2390                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
2391                                                       path->slots[0]);
2392                                 if (key.objectid == head->bytenr &&
2393                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
2394                                     key.offset == head->num_bytes)
2395                                         ret = 0;
2396                         }
2397                         if (ret > 0) {
2398                                 btrfs_release_path(path);
2399                                 metadata = 0;
2400
2401                                 key.objectid = head->bytenr;
2402                                 key.offset = head->num_bytes;
2403                                 key.type = BTRFS_EXTENT_ITEM_KEY;
2404                                 goto again;
2405                         }
2406                 } else {
2407                         err = -EIO;
2408                         goto out;
2409                 }
2410         }
2411
2412         leaf = path->nodes[0];
2413         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2414 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2415         if (item_size < sizeof(*ei)) {
2416                 ret = convert_extent_item_v0(trans, fs_info, path, (u64)-1, 0);
2417                 if (ret < 0) {
2418                         err = ret;
2419                         goto out;
2420                 }
2421                 leaf = path->nodes[0];
2422                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2423         }
2424 #endif
2425         BUG_ON(item_size < sizeof(*ei));
2426         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2427         __run_delayed_extent_op(extent_op, leaf, ei);
2428
2429         btrfs_mark_buffer_dirty(leaf);
2430 out:
2431         btrfs_free_path(path);
2432         return err;
2433 }
2434
2435 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2436                                 struct btrfs_fs_info *fs_info,
2437                                 struct btrfs_delayed_ref_node *node,
2438                                 struct btrfs_delayed_extent_op *extent_op,
2439                                 int insert_reserved)
2440 {
2441         int ret = 0;
2442         struct btrfs_delayed_tree_ref *ref;
2443         struct btrfs_key ins;
2444         u64 parent = 0;
2445         u64 ref_root = 0;
2446         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
2447
2448         ref = btrfs_delayed_node_to_tree_ref(node);
2449         trace_run_delayed_tree_ref(fs_info, node, ref, node->action);
2450
2451         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2452                 parent = ref->parent;
2453         ref_root = ref->root;
2454
2455         ins.objectid = node->bytenr;
2456         if (skinny_metadata) {
2457                 ins.offset = ref->level;
2458                 ins.type = BTRFS_METADATA_ITEM_KEY;
2459         } else {
2460                 ins.offset = node->num_bytes;
2461                 ins.type = BTRFS_EXTENT_ITEM_KEY;
2462         }
2463
2464         if (node->ref_mod != 1) {
2465                 btrfs_err(fs_info,
2466         "btree block(%llu) has %d references rather than 1: action %d ref_root %llu parent %llu",
2467                           node->bytenr, node->ref_mod, node->action, ref_root,
2468                           parent);
2469                 return -EIO;
2470         }
2471         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2472                 BUG_ON(!extent_op || !extent_op->update_flags);
2473                 ret = alloc_reserved_tree_block(trans, fs_info,
2474                                                 parent, ref_root,
2475                                                 extent_op->flags_to_set,
2476                                                 &extent_op->key,
2477                                                 ref->level, &ins);
2478         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2479                 ret = __btrfs_inc_extent_ref(trans, fs_info, node,
2480                                              parent, ref_root,
2481                                              ref->level, 0, 1,
2482                                              extent_op);
2483         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2484                 ret = __btrfs_free_extent(trans, fs_info, node,
2485                                           parent, ref_root,
2486                                           ref->level, 0, 1, extent_op);
2487         } else {
2488                 BUG();
2489         }
2490         return ret;
2491 }
2492
2493 /* helper function to actually process a single delayed ref entry */
2494 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2495                                struct btrfs_fs_info *fs_info,
2496                                struct btrfs_delayed_ref_node *node,
2497                                struct btrfs_delayed_extent_op *extent_op,
2498                                int insert_reserved)
2499 {
2500         int ret = 0;
2501
2502         if (trans->aborted) {
2503                 if (insert_reserved)
2504                         btrfs_pin_extent(fs_info, node->bytenr,
2505                                          node->num_bytes, 1);
2506                 return 0;
2507         }
2508
2509         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2510             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2511                 ret = run_delayed_tree_ref(trans, fs_info, node, extent_op,
2512                                            insert_reserved);
2513         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2514                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2515                 ret = run_delayed_data_ref(trans, fs_info, node, extent_op,
2516                                            insert_reserved);
2517         else
2518                 BUG();
2519         return ret;
2520 }
2521
2522 static inline struct btrfs_delayed_ref_node *
2523 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2524 {
2525         struct btrfs_delayed_ref_node *ref;
2526
2527         if (RB_EMPTY_ROOT(&head->ref_tree))
2528                 return NULL;
2529
2530         /*
2531          * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
2532          * This is to prevent a ref count from going down to zero, which deletes
2533          * the extent item from the extent tree, when there still are references
2534          * to add, which would fail because they would not find the extent item.
2535          */
2536         if (!list_empty(&head->ref_add_list))
2537                 return list_first_entry(&head->ref_add_list,
2538                                 struct btrfs_delayed_ref_node, add_list);
2539
2540         ref = rb_entry(rb_first(&head->ref_tree),
2541                        struct btrfs_delayed_ref_node, ref_node);
2542         ASSERT(list_empty(&ref->add_list));
2543         return ref;
2544 }
2545
2546 static void unselect_delayed_ref_head(struct btrfs_delayed_ref_root *delayed_refs,
2547                                       struct btrfs_delayed_ref_head *head)
2548 {
2549         spin_lock(&delayed_refs->lock);
2550         head->processing = 0;
2551         delayed_refs->num_heads_ready++;
2552         spin_unlock(&delayed_refs->lock);
2553         btrfs_delayed_ref_unlock(head);
2554 }
2555
2556 static int cleanup_extent_op(struct btrfs_trans_handle *trans,
2557                              struct btrfs_fs_info *fs_info,
2558                              struct btrfs_delayed_ref_head *head)
2559 {
2560         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
2561         int ret;
2562
2563         if (!extent_op)
2564                 return 0;
2565         head->extent_op = NULL;
2566         if (head->must_insert_reserved) {
2567                 btrfs_free_delayed_extent_op(extent_op);
2568                 return 0;
2569         }
2570         spin_unlock(&head->lock);
2571         ret = run_delayed_extent_op(trans, fs_info, head, extent_op);
2572         btrfs_free_delayed_extent_op(extent_op);
2573         return ret ? ret : 1;
2574 }
2575
2576 static int cleanup_ref_head(struct btrfs_trans_handle *trans,
2577                             struct btrfs_fs_info *fs_info,
2578                             struct btrfs_delayed_ref_head *head)
2579 {
2580         struct btrfs_delayed_ref_root *delayed_refs;
2581         int ret;
2582
2583         delayed_refs = &trans->transaction->delayed_refs;
2584
2585         ret = cleanup_extent_op(trans, fs_info, head);
2586         if (ret < 0) {
2587                 unselect_delayed_ref_head(delayed_refs, head);
2588                 btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
2589                 return ret;
2590         } else if (ret) {
2591                 return ret;
2592         }
2593
2594         /*
2595          * Need to drop our head ref lock and re-acquire the delayed ref lock
2596          * and then re-check to make sure nobody got added.
2597          */
2598         spin_unlock(&head->lock);
2599         spin_lock(&delayed_refs->lock);
2600         spin_lock(&head->lock);
2601         if (!RB_EMPTY_ROOT(&head->ref_tree) || head->extent_op) {
2602                 spin_unlock(&head->lock);
2603                 spin_unlock(&delayed_refs->lock);
2604                 return 1;
2605         }
2606         delayed_refs->num_heads--;
2607         rb_erase(&head->href_node, &delayed_refs->href_root);
2608         RB_CLEAR_NODE(&head->href_node);
2609         spin_unlock(&delayed_refs->lock);
2610         spin_unlock(&head->lock);
2611         atomic_dec(&delayed_refs->num_entries);
2612
2613         trace_run_delayed_ref_head(fs_info, head, 0);
2614
2615         if (head->total_ref_mod < 0) {
2616                 struct btrfs_block_group_cache *cache;
2617
2618                 cache = btrfs_lookup_block_group(fs_info, head->bytenr);
2619                 ASSERT(cache);
2620                 percpu_counter_add(&cache->space_info->total_bytes_pinned,
2621                                    -head->num_bytes);
2622                 btrfs_put_block_group(cache);
2623
2624                 if (head->is_data) {
2625                         spin_lock(&delayed_refs->lock);
2626                         delayed_refs->pending_csums -= head->num_bytes;
2627                         spin_unlock(&delayed_refs->lock);
2628                 }
2629         }
2630
2631         if (head->must_insert_reserved) {
2632                 btrfs_pin_extent(fs_info, head->bytenr,
2633                                  head->num_bytes, 1);
2634                 if (head->is_data) {
2635                         ret = btrfs_del_csums(trans, fs_info, head->bytenr,
2636                                               head->num_bytes);
2637                 }
2638         }
2639
2640         /* Also free its reserved qgroup space */
2641         btrfs_qgroup_free_delayed_ref(fs_info, head->qgroup_ref_root,
2642                                       head->qgroup_reserved);
2643         btrfs_delayed_ref_unlock(head);
2644         btrfs_put_delayed_ref_head(head);
2645         return 0;
2646 }
2647
2648 /*
2649  * Returns 0 on success or if called with an already aborted transaction.
2650  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2651  */
2652 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2653                                              unsigned long nr)
2654 {
2655         struct btrfs_fs_info *fs_info = trans->fs_info;
2656         struct btrfs_delayed_ref_root *delayed_refs;
2657         struct btrfs_delayed_ref_node *ref;
2658         struct btrfs_delayed_ref_head *locked_ref = NULL;
2659         struct btrfs_delayed_extent_op *extent_op;
2660         ktime_t start = ktime_get();
2661         int ret;
2662         unsigned long count = 0;
2663         unsigned long actual_count = 0;
2664         int must_insert_reserved = 0;
2665
2666         delayed_refs = &trans->transaction->delayed_refs;
2667         while (1) {
2668                 if (!locked_ref) {
2669                         if (count >= nr)
2670                                 break;
2671
2672                         spin_lock(&delayed_refs->lock);
2673                         locked_ref = btrfs_select_ref_head(trans);
2674                         if (!locked_ref) {
2675                                 spin_unlock(&delayed_refs->lock);
2676                                 break;
2677                         }
2678
2679                         /* grab the lock that says we are going to process
2680                          * all the refs for this head */
2681                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2682                         spin_unlock(&delayed_refs->lock);
2683                         /*
2684                          * we may have dropped the spin lock to get the head
2685                          * mutex lock, and that might have given someone else
2686                          * time to free the head.  If that's true, it has been
2687                          * removed from our list and we can move on.
2688                          */
2689                         if (ret == -EAGAIN) {
2690                                 locked_ref = NULL;
2691                                 count++;
2692                                 continue;
2693                         }
2694                 }
2695
2696                 /*
2697                  * We need to try and merge add/drops of the same ref since we
2698                  * can run into issues with relocate dropping the implicit ref
2699                  * and then it being added back again before the drop can
2700                  * finish.  If we merged anything we need to re-loop so we can
2701                  * get a good ref.
2702                  * Or we can get node references of the same type that weren't
2703                  * merged when created due to bumps in the tree mod seq, and
2704                  * we need to merge them to prevent adding an inline extent
2705                  * backref before dropping it (triggering a BUG_ON at
2706                  * insert_inline_extent_backref()).
2707                  */
2708                 spin_lock(&locked_ref->lock);
2709                 btrfs_merge_delayed_refs(trans, fs_info, delayed_refs,
2710                                          locked_ref);
2711
2712                 /*
2713                  * locked_ref is the head node, so we have to go one
2714                  * node back for any delayed ref updates
2715                  */
2716                 ref = select_delayed_ref(locked_ref);
2717
2718                 if (ref && ref->seq &&
2719                     btrfs_check_delayed_seq(fs_info, delayed_refs, ref->seq)) {
2720                         spin_unlock(&locked_ref->lock);
2721                         unselect_delayed_ref_head(delayed_refs, locked_ref);
2722                         locked_ref = NULL;
2723                         cond_resched();
2724                         count++;
2725                         continue;
2726                 }
2727
2728                 /*
2729                  * We're done processing refs in this ref_head, clean everything
2730                  * up and move on to the next ref_head.
2731                  */
2732                 if (!ref) {
2733                         ret = cleanup_ref_head(trans, fs_info, locked_ref);
2734                         if (ret > 0 ) {
2735                                 /* We dropped our lock, we need to loop. */
2736                                 ret = 0;
2737                                 continue;
2738                         } else if (ret) {
2739                                 return ret;
2740                         }
2741                         locked_ref = NULL;
2742                         count++;
2743                         continue;
2744                 }
2745
2746                 actual_count++;
2747                 ref->in_tree = 0;
2748                 rb_erase(&ref->ref_node, &locked_ref->ref_tree);
2749                 RB_CLEAR_NODE(&ref->ref_node);
2750                 if (!list_empty(&ref->add_list))
2751                         list_del(&ref->add_list);
2752                 /*
2753                  * When we play the delayed ref, also correct the ref_mod on
2754                  * head
2755                  */
2756                 switch (ref->action) {
2757                 case BTRFS_ADD_DELAYED_REF:
2758                 case BTRFS_ADD_DELAYED_EXTENT:
2759                         locked_ref->ref_mod -= ref->ref_mod;
2760                         break;
2761                 case BTRFS_DROP_DELAYED_REF:
2762                         locked_ref->ref_mod += ref->ref_mod;
2763                         break;
2764                 default:
2765                         WARN_ON(1);
2766                 }
2767                 atomic_dec(&delayed_refs->num_entries);
2768
2769                 /*
2770                  * Record the must-insert_reserved flag before we drop the spin
2771                  * lock.
2772                  */
2773                 must_insert_reserved = locked_ref->must_insert_reserved;
2774                 locked_ref->must_insert_reserved = 0;
2775
2776                 extent_op = locked_ref->extent_op;
2777                 locked_ref->extent_op = NULL;
2778                 spin_unlock(&locked_ref->lock);
2779
2780                 ret = run_one_delayed_ref(trans, fs_info, ref, extent_op,
2781                                           must_insert_reserved);
2782
2783                 btrfs_free_delayed_extent_op(extent_op);
2784                 if (ret) {
2785                         unselect_delayed_ref_head(delayed_refs, locked_ref);
2786                         btrfs_put_delayed_ref(ref);
2787                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d",
2788                                     ret);
2789                         return ret;
2790                 }
2791
2792                 btrfs_put_delayed_ref(ref);
2793                 count++;
2794                 cond_resched();
2795         }
2796
2797         /*
2798          * We don't want to include ref heads since we can have empty ref heads
2799          * and those will drastically skew our runtime down since we just do
2800          * accounting, no actual extent tree updates.
2801          */
2802         if (actual_count > 0) {
2803                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2804                 u64 avg;
2805
2806                 /*
2807                  * We weigh the current average higher than our current runtime
2808                  * to avoid large swings in the average.
2809                  */
2810                 spin_lock(&delayed_refs->lock);
2811                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2812                 fs_info->avg_delayed_ref_runtime = avg >> 2;    /* div by 4 */
2813                 spin_unlock(&delayed_refs->lock);
2814         }
2815         return 0;
2816 }
2817
2818 #ifdef SCRAMBLE_DELAYED_REFS
2819 /*
2820  * Normally delayed refs get processed in ascending bytenr order. This
2821  * correlates in most cases to the order added. To expose dependencies on this
2822  * order, we start to process the tree in the middle instead of the beginning
2823  */
2824 static u64 find_middle(struct rb_root *root)
2825 {
2826         struct rb_node *n = root->rb_node;
2827         struct btrfs_delayed_ref_node *entry;
2828         int alt = 1;
2829         u64 middle;
2830         u64 first = 0, last = 0;
2831
2832         n = rb_first(root);
2833         if (n) {
2834                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2835                 first = entry->bytenr;
2836         }
2837         n = rb_last(root);
2838         if (n) {
2839                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2840                 last = entry->bytenr;
2841         }
2842         n = root->rb_node;
2843
2844         while (n) {
2845                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2846                 WARN_ON(!entry->in_tree);
2847
2848                 middle = entry->bytenr;
2849
2850                 if (alt)
2851                         n = n->rb_left;
2852                 else
2853                         n = n->rb_right;
2854
2855                 alt = 1 - alt;
2856         }
2857         return middle;
2858 }
2859 #endif
2860
2861 static inline u64 heads_to_leaves(struct btrfs_fs_info *fs_info, u64 heads)
2862 {
2863         u64 num_bytes;
2864
2865         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2866                              sizeof(struct btrfs_extent_inline_ref));
2867         if (!btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2868                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2869
2870         /*
2871          * We don't ever fill up leaves all the way so multiply by 2 just to be
2872          * closer to what we're really going to want to use.
2873          */
2874         return div_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(fs_info));
2875 }
2876
2877 /*
2878  * Takes the number of bytes to be csumm'ed and figures out how many leaves it
2879  * would require to store the csums for that many bytes.
2880  */
2881 u64 btrfs_csum_bytes_to_leaves(struct btrfs_fs_info *fs_info, u64 csum_bytes)
2882 {
2883         u64 csum_size;
2884         u64 num_csums_per_leaf;
2885         u64 num_csums;
2886
2887         csum_size = BTRFS_MAX_ITEM_SIZE(fs_info);
2888         num_csums_per_leaf = div64_u64(csum_size,
2889                         (u64)btrfs_super_csum_size(fs_info->super_copy));
2890         num_csums = div64_u64(csum_bytes, fs_info->sectorsize);
2891         num_csums += num_csums_per_leaf - 1;
2892         num_csums = div64_u64(num_csums, num_csums_per_leaf);
2893         return num_csums;
2894 }
2895
2896 int btrfs_check_space_for_delayed_refs(struct btrfs_trans_handle *trans,
2897                                        struct btrfs_fs_info *fs_info)
2898 {
2899         struct btrfs_block_rsv *global_rsv;
2900         u64 num_heads = trans->transaction->delayed_refs.num_heads_ready;
2901         u64 csum_bytes = trans->transaction->delayed_refs.pending_csums;
2902         unsigned int num_dirty_bgs = trans->transaction->num_dirty_bgs;
2903         u64 num_bytes, num_dirty_bgs_bytes;
2904         int ret = 0;
2905
2906         num_bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
2907         num_heads = heads_to_leaves(fs_info, num_heads);
2908         if (num_heads > 1)
2909                 num_bytes += (num_heads - 1) * fs_info->nodesize;
2910         num_bytes <<= 1;
2911         num_bytes += btrfs_csum_bytes_to_leaves(fs_info, csum_bytes) *
2912                                                         fs_info->nodesize;
2913         num_dirty_bgs_bytes = btrfs_calc_trans_metadata_size(fs_info,
2914                                                              num_dirty_bgs);
2915         global_rsv = &fs_info->global_block_rsv;
2916
2917         /*
2918          * If we can't allocate any more chunks lets make sure we have _lots_ of
2919          * wiggle room since running delayed refs can create more delayed refs.
2920          */
2921         if (global_rsv->space_info->full) {
2922                 num_dirty_bgs_bytes <<= 1;
2923                 num_bytes <<= 1;
2924         }
2925
2926         spin_lock(&global_rsv->lock);
2927         if (global_rsv->reserved <= num_bytes + num_dirty_bgs_bytes)
2928                 ret = 1;
2929         spin_unlock(&global_rsv->lock);
2930         return ret;
2931 }
2932
2933 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans,
2934                                        struct btrfs_fs_info *fs_info)
2935 {
2936         u64 num_entries =
2937                 atomic_read(&trans->transaction->delayed_refs.num_entries);
2938         u64 avg_runtime;
2939         u64 val;
2940
2941         smp_mb();
2942         avg_runtime = fs_info->avg_delayed_ref_runtime;
2943         val = num_entries * avg_runtime;
2944         if (val >= NSEC_PER_SEC)
2945                 return 1;
2946         if (val >= NSEC_PER_SEC / 2)
2947                 return 2;
2948
2949         return btrfs_check_space_for_delayed_refs(trans, fs_info);
2950 }
2951
2952 struct async_delayed_refs {
2953         struct btrfs_root *root;
2954         u64 transid;
2955         int count;
2956         int error;
2957         int sync;
2958         struct completion wait;
2959         struct btrfs_work work;
2960 };
2961
2962 static inline struct async_delayed_refs *
2963 to_async_delayed_refs(struct btrfs_work *work)
2964 {
2965         return container_of(work, struct async_delayed_refs, work);
2966 }
2967
2968 static void delayed_ref_async_start(struct btrfs_work *work)
2969 {
2970         struct async_delayed_refs *async = to_async_delayed_refs(work);
2971         struct btrfs_trans_handle *trans;
2972         struct btrfs_fs_info *fs_info = async->root->fs_info;
2973         int ret;
2974
2975         /* if the commit is already started, we don't need to wait here */
2976         if (btrfs_transaction_blocked(fs_info))
2977                 goto done;
2978
2979         trans = btrfs_join_transaction(async->root);
2980         if (IS_ERR(trans)) {
2981                 async->error = PTR_ERR(trans);
2982                 goto done;
2983         }
2984
2985         /*
2986          * trans->sync means that when we call end_transaction, we won't
2987          * wait on delayed refs
2988          */
2989         trans->sync = true;
2990
2991         /* Don't bother flushing if we got into a different transaction */
2992         if (trans->transid > async->transid)
2993                 goto end;
2994
2995         ret = btrfs_run_delayed_refs(trans, async->count);
2996         if (ret)
2997                 async->error = ret;
2998 end:
2999         ret = btrfs_end_transaction(trans);
3000         if (ret && !async->error)
3001                 async->error = ret;
3002 done:
3003         if (async->sync)
3004                 complete(&async->wait);
3005         else
3006                 kfree(async);
3007 }
3008
3009 int btrfs_async_run_delayed_refs(struct btrfs_fs_info *fs_info,
3010                                  unsigned long count, u64 transid, int wait)
3011 {
3012         struct async_delayed_refs *async;
3013         int ret;
3014
3015         async = kmalloc(sizeof(*async), GFP_NOFS);
3016         if (!async)
3017                 return -ENOMEM;
3018
3019         async->root = fs_info->tree_root;
3020         async->count = count;
3021         async->error = 0;
3022         async->transid = transid;
3023         if (wait)
3024                 async->sync = 1;
3025         else
3026                 async->sync = 0;
3027         init_completion(&async->wait);
3028
3029         btrfs_init_work(&async->work, btrfs_extent_refs_helper,
3030                         delayed_ref_async_start, NULL, NULL);
3031
3032         btrfs_queue_work(fs_info->extent_workers, &async->work);
3033
3034         if (wait) {
3035                 wait_for_completion(&async->wait);
3036                 ret = async->error;
3037                 kfree(async);
3038                 return ret;
3039         }
3040         return 0;
3041 }
3042
3043 /*
3044  * this starts processing the delayed reference count updates and
3045  * extent insertions we have queued up so far.  count can be
3046  * 0, which means to process everything in the tree at the start
3047  * of the run (but not newly added entries), or it can be some target
3048  * number you'd like to process.
3049  *
3050  * Returns 0 on success or if called with an aborted transaction
3051  * Returns <0 on error and aborts the transaction
3052  */
3053 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
3054                            unsigned long count)
3055 {
3056         struct btrfs_fs_info *fs_info = trans->fs_info;
3057         struct rb_node *node;
3058         struct btrfs_delayed_ref_root *delayed_refs;
3059         struct btrfs_delayed_ref_head *head;
3060         int ret;
3061         int run_all = count == (unsigned long)-1;
3062         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
3063
3064         /* We'll clean this up in btrfs_cleanup_transaction */
3065         if (trans->aborted)
3066                 return 0;
3067
3068         if (test_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags))
3069                 return 0;
3070
3071         delayed_refs = &trans->transaction->delayed_refs;
3072         if (count == 0)
3073                 count = atomic_read(&delayed_refs->num_entries) * 2;
3074
3075 again:
3076 #ifdef SCRAMBLE_DELAYED_REFS
3077         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
3078 #endif
3079         trans->can_flush_pending_bgs = false;
3080         ret = __btrfs_run_delayed_refs(trans, count);
3081         if (ret < 0) {
3082                 btrfs_abort_transaction(trans, ret);
3083                 return ret;
3084         }
3085
3086         if (run_all) {
3087                 if (!list_empty(&trans->new_bgs))
3088                         btrfs_create_pending_block_groups(trans);
3089
3090                 spin_lock(&delayed_refs->lock);
3091                 node = rb_first(&delayed_refs->href_root);
3092                 if (!node) {
3093                         spin_unlock(&delayed_refs->lock);
3094                         goto out;
3095                 }
3096                 head = rb_entry(node, struct btrfs_delayed_ref_head,
3097                                 href_node);
3098                 refcount_inc(&head->refs);
3099                 spin_unlock(&delayed_refs->lock);
3100
3101                 /* Mutex was contended, block until it's released and retry. */
3102                 mutex_lock(&head->mutex);
3103                 mutex_unlock(&head->mutex);
3104
3105                 btrfs_put_delayed_ref_head(head);
3106                 cond_resched();
3107                 goto again;
3108         }
3109 out:
3110         trans->can_flush_pending_bgs = can_flush_pending_bgs;
3111         return 0;
3112 }
3113
3114 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
3115                                 struct btrfs_fs_info *fs_info,
3116                                 u64 bytenr, u64 num_bytes, u64 flags,
3117                                 int level, int is_data)
3118 {
3119         struct btrfs_delayed_extent_op *extent_op;
3120         int ret;
3121
3122         extent_op = btrfs_alloc_delayed_extent_op();
3123         if (!extent_op)
3124                 return -ENOMEM;
3125
3126         extent_op->flags_to_set = flags;
3127         extent_op->update_flags = true;
3128         extent_op->update_key = false;
3129         extent_op->is_data = is_data ? true : false;
3130         extent_op->level = level;
3131
3132         ret = btrfs_add_delayed_extent_op(fs_info, trans, bytenr,
3133                                           num_bytes, extent_op);
3134         if (ret)
3135                 btrfs_free_delayed_extent_op(extent_op);
3136         return ret;
3137 }
3138
3139 static noinline int check_delayed_ref(struct btrfs_root *root,
3140                                       struct btrfs_path *path,
3141                                       u64 objectid, u64 offset, u64 bytenr)
3142 {
3143         struct btrfs_delayed_ref_head *head;
3144         struct btrfs_delayed_ref_node *ref;
3145         struct btrfs_delayed_data_ref *data_ref;
3146         struct btrfs_delayed_ref_root *delayed_refs;
3147         struct btrfs_transaction *cur_trans;
3148         struct rb_node *node;
3149         int ret = 0;
3150
3151         cur_trans = root->fs_info->running_transaction;
3152         if (!cur_trans)
3153                 return 0;
3154
3155         delayed_refs = &cur_trans->delayed_refs;
3156         spin_lock(&delayed_refs->lock);
3157         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
3158         if (!head) {
3159                 spin_unlock(&delayed_refs->lock);
3160                 return 0;
3161         }
3162
3163         if (!mutex_trylock(&head->mutex)) {
3164                 refcount_inc(&head->refs);
3165                 spin_unlock(&delayed_refs->lock);
3166
3167                 btrfs_release_path(path);
3168
3169                 /*
3170                  * Mutex was contended, block until it's released and let
3171                  * caller try again
3172                  */
3173                 mutex_lock(&head->mutex);
3174                 mutex_unlock(&head->mutex);
3175                 btrfs_put_delayed_ref_head(head);
3176                 return -EAGAIN;
3177         }
3178         spin_unlock(&delayed_refs->lock);
3179
3180         spin_lock(&head->lock);
3181         /*
3182          * XXX: We should replace this with a proper search function in the
3183          * future.
3184          */
3185         for (node = rb_first(&head->ref_tree); node; node = rb_next(node)) {
3186                 ref = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
3187                 /* If it's a shared ref we know a cross reference exists */
3188                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
3189                         ret = 1;
3190                         break;
3191                 }
3192
3193                 data_ref = btrfs_delayed_node_to_data_ref(ref);
3194
3195                 /*
3196                  * If our ref doesn't match the one we're currently looking at
3197                  * then we have a cross reference.
3198                  */
3199                 if (data_ref->root != root->root_key.objectid ||
3200                     data_ref->objectid != objectid ||
3201                     data_ref->offset != offset) {
3202                         ret = 1;
3203                         break;
3204                 }
3205         }
3206         spin_unlock(&head->lock);
3207         mutex_unlock(&head->mutex);
3208         return ret;
3209 }
3210
3211 static noinline int check_committed_ref(struct btrfs_root *root,
3212                                         struct btrfs_path *path,
3213                                         u64 objectid, u64 offset, u64 bytenr)
3214 {
3215         struct btrfs_fs_info *fs_info = root->fs_info;
3216         struct btrfs_root *extent_root = fs_info->extent_root;
3217         struct extent_buffer *leaf;
3218         struct btrfs_extent_data_ref *ref;
3219         struct btrfs_extent_inline_ref *iref;
3220         struct btrfs_extent_item *ei;
3221         struct btrfs_key key;
3222         u32 item_size;
3223         int type;
3224         int ret;
3225
3226         key.objectid = bytenr;
3227         key.offset = (u64)-1;
3228         key.type = BTRFS_EXTENT_ITEM_KEY;
3229
3230         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
3231         if (ret < 0)
3232                 goto out;
3233         BUG_ON(ret == 0); /* Corruption */
3234
3235         ret = -ENOENT;
3236         if (path->slots[0] == 0)
3237                 goto out;
3238
3239         path->slots[0]--;
3240         leaf = path->nodes[0];
3241         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3242
3243         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
3244                 goto out;
3245
3246         ret = 1;
3247         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
3248 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3249         if (item_size < sizeof(*ei)) {
3250                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3251                 goto out;
3252         }
3253 #endif
3254         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
3255
3256         if (item_size != sizeof(*ei) +
3257             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
3258                 goto out;
3259
3260         if (btrfs_extent_generation(leaf, ei) <=
3261             btrfs_root_last_snapshot(&root->root_item))
3262                 goto out;
3263
3264         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
3265
3266         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
3267         if (type != BTRFS_EXTENT_DATA_REF_KEY)
3268                 goto out;
3269
3270         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
3271         if (btrfs_extent_refs(leaf, ei) !=
3272             btrfs_extent_data_ref_count(leaf, ref) ||
3273             btrfs_extent_data_ref_root(leaf, ref) !=
3274             root->root_key.objectid ||
3275             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
3276             btrfs_extent_data_ref_offset(leaf, ref) != offset)
3277                 goto out;
3278
3279         ret = 0;
3280 out:
3281         return ret;
3282 }
3283
3284 int btrfs_cross_ref_exist(struct btrfs_root *root, u64 objectid, u64 offset,
3285                           u64 bytenr)
3286 {
3287         struct btrfs_path *path;
3288         int ret;
3289         int ret2;
3290
3291         path = btrfs_alloc_path();
3292         if (!path)
3293                 return -ENOENT;
3294
3295         do {
3296                 ret = check_committed_ref(root, path, objectid,
3297                                           offset, bytenr);
3298                 if (ret && ret != -ENOENT)
3299                         goto out;
3300
3301                 ret2 = check_delayed_ref(root, path, objectid,
3302                                          offset, bytenr);
3303         } while (ret2 == -EAGAIN);
3304
3305         if (ret2 && ret2 != -ENOENT) {
3306                 ret = ret2;
3307                 goto out;
3308         }
3309
3310         if (ret != -ENOENT || ret2 != -ENOENT)
3311                 ret = 0;
3312 out:
3313         btrfs_free_path(path);
3314         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
3315                 WARN_ON(ret > 0);
3316         return ret;
3317 }
3318
3319 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
3320                            struct btrfs_root *root,
3321                            struct extent_buffer *buf,
3322                            int full_backref, int inc)
3323 {
3324         struct btrfs_fs_info *fs_info = root->fs_info;
3325         u64 bytenr;
3326         u64 num_bytes;
3327         u64 parent;
3328         u64 ref_root;
3329         u32 nritems;
3330         struct btrfs_key key;
3331         struct btrfs_file_extent_item *fi;
3332         int i;
3333         int level;
3334         int ret = 0;
3335         int (*process_func)(struct btrfs_trans_handle *,
3336                             struct btrfs_root *,
3337                             u64, u64, u64, u64, u64, u64);
3338
3339
3340         if (btrfs_is_testing(fs_info))
3341                 return 0;
3342
3343         ref_root = btrfs_header_owner(buf);
3344         nritems = btrfs_header_nritems(buf);
3345         level = btrfs_header_level(buf);
3346
3347         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state) && level == 0)
3348                 return 0;
3349
3350         if (inc)
3351                 process_func = btrfs_inc_extent_ref;
3352         else
3353                 process_func = btrfs_free_extent;
3354
3355         if (full_backref)
3356                 parent = buf->start;
3357         else
3358                 parent = 0;
3359
3360         for (i = 0; i < nritems; i++) {
3361                 if (level == 0) {
3362                         btrfs_item_key_to_cpu(buf, &key, i);
3363                         if (key.type != BTRFS_EXTENT_DATA_KEY)
3364                                 continue;
3365                         fi = btrfs_item_ptr(buf, i,
3366                                             struct btrfs_file_extent_item);
3367                         if (btrfs_file_extent_type(buf, fi) ==
3368                             BTRFS_FILE_EXTENT_INLINE)
3369                                 continue;
3370                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3371                         if (bytenr == 0)
3372                                 continue;
3373
3374                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3375                         key.offset -= btrfs_file_extent_offset(buf, fi);
3376                         ret = process_func(trans, root, bytenr, num_bytes,
3377                                            parent, ref_root, key.objectid,
3378                                            key.offset);
3379                         if (ret)
3380                                 goto fail;
3381                 } else {
3382                         bytenr = btrfs_node_blockptr(buf, i);
3383                         num_bytes = fs_info->nodesize;
3384                         ret = process_func(trans, root, bytenr, num_bytes,
3385                                            parent, ref_root, level - 1, 0);
3386                         if (ret)
3387                                 goto fail;
3388                 }
3389         }
3390         return 0;
3391 fail:
3392         return ret;
3393 }
3394
3395 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3396                   struct extent_buffer *buf, int full_backref)
3397 {
3398         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
3399 }
3400
3401 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3402                   struct extent_buffer *buf, int full_backref)
3403 {
3404         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
3405 }
3406
3407 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3408                                  struct btrfs_fs_info *fs_info,
3409                                  struct btrfs_path *path,
3410                                  struct btrfs_block_group_cache *cache)
3411 {
3412         int ret;
3413         struct btrfs_root *extent_root = fs_info->extent_root;
3414         unsigned long bi;
3415         struct extent_buffer *leaf;
3416
3417         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3418         if (ret) {
3419                 if (ret > 0)
3420                         ret = -ENOENT;
3421                 goto fail;
3422         }
3423
3424         leaf = path->nodes[0];
3425         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3426         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3427         btrfs_mark_buffer_dirty(leaf);
3428 fail:
3429         btrfs_release_path(path);
3430         return ret;
3431
3432 }
3433
3434 static struct btrfs_block_group_cache *
3435 next_block_group(struct btrfs_fs_info *fs_info,
3436                  struct btrfs_block_group_cache *cache)
3437 {
3438         struct rb_node *node;
3439
3440         spin_lock(&fs_info->block_group_cache_lock);
3441
3442         /* If our block group was removed, we need a full search. */
3443         if (RB_EMPTY_NODE(&cache->cache_node)) {
3444                 const u64 next_bytenr = cache->key.objectid + cache->key.offset;
3445
3446                 spin_unlock(&fs_info->block_group_cache_lock);
3447                 btrfs_put_block_group(cache);
3448                 cache = btrfs_lookup_first_block_group(fs_info, next_bytenr); return cache;
3449         }
3450         node = rb_next(&cache->cache_node);
3451         btrfs_put_block_group(cache);
3452         if (node) {
3453                 cache = rb_entry(node, struct btrfs_block_group_cache,
3454                                  cache_node);
3455                 btrfs_get_block_group(cache);
3456         } else
3457                 cache = NULL;
3458         spin_unlock(&fs_info->block_group_cache_lock);
3459         return cache;
3460 }
3461
3462 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3463                             struct btrfs_trans_handle *trans,
3464                             struct btrfs_path *path)
3465 {
3466         struct btrfs_fs_info *fs_info = block_group->fs_info;
3467         struct btrfs_root *root = fs_info->tree_root;
3468         struct inode *inode = NULL;
3469         struct extent_changeset *data_reserved = NULL;
3470         u64 alloc_hint = 0;
3471         int dcs = BTRFS_DC_ERROR;
3472         u64 num_pages = 0;
3473         int retries = 0;
3474         int ret = 0;
3475
3476         /*
3477          * If this block group is smaller than 100 megs don't bother caching the
3478          * block group.
3479          */
3480         if (block_group->key.offset < (100 * SZ_1M)) {
3481                 spin_lock(&block_group->lock);
3482                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3483                 spin_unlock(&block_group->lock);
3484                 return 0;
3485         }
3486
3487         if (trans->aborted)
3488                 return 0;
3489 again:
3490         inode = lookup_free_space_inode(fs_info, block_group, path);
3491         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3492                 ret = PTR_ERR(inode);
3493                 btrfs_release_path(path);
3494                 goto out;
3495         }
3496
3497         if (IS_ERR(inode)) {
3498                 BUG_ON(retries);
3499                 retries++;
3500
3501                 if (block_group->ro)
3502                         goto out_free;
3503
3504                 ret = create_free_space_inode(fs_info, trans, block_group,
3505                                               path);
3506                 if (ret)
3507                         goto out_free;
3508                 goto again;
3509         }
3510
3511         /*
3512          * We want to set the generation to 0, that way if anything goes wrong
3513          * from here on out we know not to trust this cache when we load up next
3514          * time.
3515          */
3516         BTRFS_I(inode)->generation = 0;
3517         ret = btrfs_update_inode(trans, root, inode);
3518         if (ret) {
3519                 /*
3520                  * So theoretically we could recover from this, simply set the
3521                  * super cache generation to 0 so we know to invalidate the
3522                  * cache, but then we'd have to keep track of the block groups
3523                  * that fail this way so we know we _have_ to reset this cache
3524                  * before the next commit or risk reading stale cache.  So to
3525                  * limit our exposure to horrible edge cases lets just abort the
3526                  * transaction, this only happens in really bad situations
3527                  * anyway.
3528                  */
3529                 btrfs_abort_transaction(trans, ret);
3530                 goto out_put;
3531         }
3532         WARN_ON(ret);
3533
3534         /* We've already setup this transaction, go ahead and exit */
3535         if (block_group->cache_generation == trans->transid &&
3536             i_size_read(inode)) {
3537                 dcs = BTRFS_DC_SETUP;
3538                 goto out_put;
3539         }
3540
3541         if (i_size_read(inode) > 0) {
3542                 ret = btrfs_check_trunc_cache_free_space(fs_info,
3543                                         &fs_info->global_block_rsv);
3544                 if (ret)
3545                         goto out_put;
3546
3547                 ret = btrfs_truncate_free_space_cache(trans, NULL, inode);
3548                 if (ret)
3549                         goto out_put;
3550         }
3551
3552         spin_lock(&block_group->lock);
3553         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3554             !btrfs_test_opt(fs_info, SPACE_CACHE)) {
3555                 /*
3556                  * don't bother trying to write stuff out _if_
3557                  * a) we're not cached,
3558                  * b) we're with nospace_cache mount option,
3559                  * c) we're with v2 space_cache (FREE_SPACE_TREE).
3560                  */
3561                 dcs = BTRFS_DC_WRITTEN;
3562                 spin_unlock(&block_group->lock);
3563                 goto out_put;
3564         }
3565         spin_unlock(&block_group->lock);
3566
3567         /*
3568          * We hit an ENOSPC when setting up the cache in this transaction, just
3569          * skip doing the setup, we've already cleared the cache so we're safe.
3570          */
3571         if (test_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags)) {
3572                 ret = -ENOSPC;
3573                 goto out_put;
3574         }
3575
3576         /*
3577          * Try to preallocate enough space based on how big the block group is.
3578          * Keep in mind this has to include any pinned space which could end up
3579          * taking up quite a bit since it's not folded into the other space
3580          * cache.
3581          */
3582         num_pages = div_u64(block_group->key.offset, SZ_256M);
3583         if (!num_pages)
3584                 num_pages = 1;
3585
3586         num_pages *= 16;
3587         num_pages *= PAGE_SIZE;
3588
3589         ret = btrfs_check_data_free_space(inode, &data_reserved, 0, num_pages);
3590         if (ret)
3591                 goto out_put;
3592
3593         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3594                                               num_pages, num_pages,
3595                                               &alloc_hint);
3596         /*
3597          * Our cache requires contiguous chunks so that we don't modify a bunch
3598          * of metadata or split extents when writing the cache out, which means
3599          * we can enospc if we are heavily fragmented in addition to just normal
3600          * out of space conditions.  So if we hit this just skip setting up any
3601          * other block groups for this transaction, maybe we'll unpin enough
3602          * space the next time around.
3603          */
3604         if (!ret)
3605                 dcs = BTRFS_DC_SETUP;
3606         else if (ret == -ENOSPC)
3607                 set_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags);
3608
3609 out_put:
3610         iput(inode);
3611 out_free:
3612         btrfs_release_path(path);
3613 out:
3614         spin_lock(&block_group->lock);
3615         if (!ret && dcs == BTRFS_DC_SETUP)
3616                 block_group->cache_generation = trans->transid;
3617         block_group->disk_cache_state = dcs;
3618         spin_unlock(&block_group->lock);
3619
3620         extent_changeset_free(data_reserved);
3621         return ret;
3622 }
3623
3624 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans,
3625                             struct btrfs_fs_info *fs_info)
3626 {
3627         struct btrfs_block_group_cache *cache, *tmp;
3628         struct btrfs_transaction *cur_trans = trans->transaction;
3629         struct btrfs_path *path;
3630
3631         if (list_empty(&cur_trans->dirty_bgs) ||
3632             !btrfs_test_opt(fs_info, SPACE_CACHE))
3633                 return 0;
3634
3635         path = btrfs_alloc_path();
3636         if (!path)
3637                 return -ENOMEM;
3638
3639         /* Could add new block groups, use _safe just in case */
3640         list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3641                                  dirty_list) {
3642                 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3643                         cache_save_setup(cache, trans, path);
3644         }
3645
3646         btrfs_free_path(path);
3647         return 0;
3648 }
3649
3650 /*
3651  * transaction commit does final block group cache writeback during a
3652  * critical section where nothing is allowed to change the FS.  This is
3653  * required in order for the cache to actually match the block group,
3654  * but can introduce a lot of latency into the commit.
3655  *
3656  * So, btrfs_start_dirty_block_groups is here to kick off block group
3657  * cache IO.  There's a chance we'll have to redo some of it if the
3658  * block group changes again during the commit, but it greatly reduces
3659  * the commit latency by getting rid of the easy block groups while
3660  * we're still allowing others to join the commit.
3661  */
3662 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans)
3663 {
3664         struct btrfs_fs_info *fs_info = trans->fs_info;
3665         struct btrfs_block_group_cache *cache;
3666         struct btrfs_transaction *cur_trans = trans->transaction;
3667         int ret = 0;
3668         int should_put;
3669         struct btrfs_path *path = NULL;
3670         LIST_HEAD(dirty);
3671         struct list_head *io = &cur_trans->io_bgs;
3672         int num_started = 0;
3673         int loops = 0;
3674
3675         spin_lock(&cur_trans->dirty_bgs_lock);
3676         if (list_empty(&cur_trans->dirty_bgs)) {
3677                 spin_unlock(&cur_trans->dirty_bgs_lock);
3678                 return 0;
3679         }
3680         list_splice_init(&cur_trans->dirty_bgs, &dirty);
3681         spin_unlock(&cur_trans->dirty_bgs_lock);
3682
3683 again:
3684         /*
3685          * make sure all the block groups on our dirty list actually
3686          * exist
3687          */
3688         btrfs_create_pending_block_groups(trans);
3689
3690         if (!path) {
3691                 path = btrfs_alloc_path();
3692                 if (!path)
3693                         return -ENOMEM;
3694         }
3695
3696         /*
3697          * cache_write_mutex is here only to save us from balance or automatic
3698          * removal of empty block groups deleting this block group while we are
3699          * writing out the cache
3700          */
3701         mutex_lock(&trans->transaction->cache_write_mutex);
3702         while (!list_empty(&dirty)) {
3703                 cache = list_first_entry(&dirty,
3704                                          struct btrfs_block_group_cache,
3705                                          dirty_list);
3706                 /*
3707                  * this can happen if something re-dirties a block
3708                  * group that is already under IO.  Just wait for it to
3709                  * finish and then do it all again
3710                  */
3711                 if (!list_empty(&cache->io_list)) {
3712                         list_del_init(&cache->io_list);
3713                         btrfs_wait_cache_io(trans, cache, path);
3714                         btrfs_put_block_group(cache);
3715                 }
3716
3717
3718                 /*
3719                  * btrfs_wait_cache_io uses the cache->dirty_list to decide
3720                  * if it should update the cache_state.  Don't delete
3721                  * until after we wait.
3722                  *
3723                  * Since we're not running in the commit critical section
3724                  * we need the dirty_bgs_lock to protect from update_block_group
3725                  */
3726                 spin_lock(&cur_trans->dirty_bgs_lock);
3727                 list_del_init(&cache->dirty_list);
3728                 spin_unlock(&cur_trans->dirty_bgs_lock);
3729
3730                 should_put = 1;
3731
3732                 cache_save_setup(cache, trans, path);
3733
3734                 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
3735                         cache->io_ctl.inode = NULL;
3736                         ret = btrfs_write_out_cache(fs_info, trans,
3737                                                     cache, path);
3738                         if (ret == 0 && cache->io_ctl.inode) {
3739                                 num_started++;
3740                                 should_put = 0;
3741
3742                                 /*
3743                                  * The cache_write_mutex is protecting the
3744                                  * io_list, also refer to the definition of
3745                                  * btrfs_transaction::io_bgs for more details
3746                                  */
3747                                 list_add_tail(&cache->io_list, io);
3748                         } else {
3749                                 /*
3750                                  * if we failed to write the cache, the
3751                                  * generation will be bad and life goes on
3752                                  */
3753                                 ret = 0;
3754                         }
3755                 }
3756                 if (!ret) {
3757                         ret = write_one_cache_group(trans, fs_info,
3758                                                     path, cache);
3759                         /*
3760                          * Our block group might still be attached to the list
3761                          * of new block groups in the transaction handle of some
3762                          * other task (struct btrfs_trans_handle->new_bgs). This
3763                          * means its block group item isn't yet in the extent
3764                          * tree. If this happens ignore the error, as we will
3765                          * try again later in the critical section of the
3766                          * transaction commit.
3767                          */
3768                         if (ret == -ENOENT) {
3769                                 ret = 0;
3770                                 spin_lock(&cur_trans->dirty_bgs_lock);
3771                                 if (list_empty(&cache->dirty_list)) {
3772                                         list_add_tail(&cache->dirty_list,
3773                                                       &cur_trans->dirty_bgs);
3774                                         btrfs_get_block_group(cache);
3775                                 }
3776                                 spin_unlock(&cur_trans->dirty_bgs_lock);
3777                         } else if (ret) {
3778                                 btrfs_abort_transaction(trans, ret);
3779                         }
3780                 }
3781
3782                 /* if its not on the io list, we need to put the block group */
3783                 if (should_put)
3784                         btrfs_put_block_group(cache);
3785
3786                 if (ret)
3787                         break;
3788
3789                 /*
3790                  * Avoid blocking other tasks for too long. It might even save
3791                  * us from writing caches for block groups that are going to be
3792                  * removed.
3793                  */
3794                 mutex_unlock(&trans->transaction->cache_write_mutex);
3795                 mutex_lock(&trans->transaction->cache_write_mutex);
3796         }
3797         mutex_unlock(&trans->transaction->cache_write_mutex);
3798
3799         /*
3800          * go through delayed refs for all the stuff we've just kicked off
3801          * and then loop back (just once)
3802          */
3803         ret = btrfs_run_delayed_refs(trans, 0);
3804         if (!ret && loops == 0) {
3805                 loops++;
3806                 spin_lock(&cur_trans->dirty_bgs_lock);
3807                 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3808                 /*
3809                  * dirty_bgs_lock protects us from concurrent block group
3810                  * deletes too (not just cache_write_mutex).
3811                  */
3812                 if (!list_empty(&dirty)) {
3813                         spin_unlock(&cur_trans->dirty_bgs_lock);
3814                         goto again;
3815                 }
3816                 spin_unlock(&cur_trans->dirty_bgs_lock);
3817         } else if (ret < 0) {
3818                 btrfs_cleanup_dirty_bgs(cur_trans, fs_info);
3819         }
3820
3821         btrfs_free_path(path);
3822         return ret;
3823 }
3824
3825 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3826                                    struct btrfs_fs_info *fs_info)
3827 {
3828         struct btrfs_block_group_cache *cache;
3829         struct btrfs_transaction *cur_trans = trans->transaction;
3830         int ret = 0;
3831         int should_put;
3832         struct btrfs_path *path;
3833         struct list_head *io = &cur_trans->io_bgs;
3834         int num_started = 0;
3835
3836         path = btrfs_alloc_path();
3837         if (!path)
3838                 return -ENOMEM;
3839
3840         /*
3841          * Even though we are in the critical section of the transaction commit,
3842          * we can still have concurrent tasks adding elements to this
3843          * transaction's list of dirty block groups. These tasks correspond to
3844          * endio free space workers started when writeback finishes for a
3845          * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
3846          * allocate new block groups as a result of COWing nodes of the root
3847          * tree when updating the free space inode. The writeback for the space
3848          * caches is triggered by an earlier call to
3849          * btrfs_start_dirty_block_groups() and iterations of the following
3850          * loop.
3851          * Also we want to do the cache_save_setup first and then run the
3852          * delayed refs to make sure we have the best chance at doing this all
3853          * in one shot.
3854          */
3855         spin_lock(&cur_trans->dirty_bgs_lock);
3856         while (!list_empty(&cur_trans->dirty_bgs)) {
3857                 cache = list_first_entry(&cur_trans->dirty_bgs,
3858                                          struct btrfs_block_group_cache,
3859                                          dirty_list);
3860
3861                 /*
3862                  * this can happen if cache_save_setup re-dirties a block
3863                  * group that is already under IO.  Just wait for it to
3864                  * finish and then do it all again
3865                  */
3866                 if (!list_empty(&cache->io_list)) {
3867                         spin_unlock(&cur_trans->dirty_bgs_lock);
3868                         list_del_init(&cache->io_list);
3869                         btrfs_wait_cache_io(trans, cache, path);
3870                         btrfs_put_block_group(cache);
3871                         spin_lock(&cur_trans->dirty_bgs_lock);
3872                 }
3873
3874                 /*
3875                  * don't remove from the dirty list until after we've waited
3876                  * on any pending IO
3877                  */
3878                 list_del_init(&cache->dirty_list);
3879                 spin_unlock(&cur_trans->dirty_bgs_lock);
3880                 should_put = 1;
3881
3882                 cache_save_setup(cache, trans, path);
3883
3884                 if (!ret)
3885                         ret = btrfs_run_delayed_refs(trans,
3886                                                      (unsigned long) -1);
3887
3888                 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3889                         cache->io_ctl.inode = NULL;
3890                         ret = btrfs_write_out_cache(fs_info, trans,
3891                                                     cache, path);
3892                         if (ret == 0 && cache->io_ctl.inode) {
3893                                 num_started++;
3894                                 should_put = 0;
3895                                 list_add_tail(&cache->io_list, io);
3896                         } else {
3897                                 /*
3898                                  * if we failed to write the cache, the
3899                                  * generation will be bad and life goes on
3900                                  */
3901                                 ret = 0;
3902                         }
3903                 }
3904                 if (!ret) {
3905                         ret = write_one_cache_group(trans, fs_info,
3906                                                     path, cache);
3907                         /*
3908                          * One of the free space endio workers might have
3909                          * created a new block group while updating a free space
3910                          * cache's inode (at inode.c:btrfs_finish_ordered_io())
3911                          * and hasn't released its transaction handle yet, in
3912                          * which case the new block group is still attached to
3913                          * its transaction handle and its creation has not
3914                          * finished yet (no block group item in the extent tree
3915                          * yet, etc). If this is the case, wait for all free
3916                          * space endio workers to finish and retry. This is a
3917                          * a very rare case so no need for a more efficient and
3918                          * complex approach.
3919                          */
3920                         if (ret == -ENOENT) {
3921                                 wait_event(cur_trans->writer_wait,
3922                                    atomic_read(&cur_trans->num_writers) == 1);
3923                                 ret = write_one_cache_group(trans, fs_info,
3924                                                             path, cache);
3925                         }
3926                         if (ret)
3927                                 btrfs_abort_transaction(trans, ret);
3928                 }
3929
3930                 /* if its not on the io list, we need to put the block group */
3931                 if (should_put)
3932                         btrfs_put_block_group(cache);
3933                 spin_lock(&cur_trans->dirty_bgs_lock);
3934         }
3935         spin_unlock(&cur_trans->dirty_bgs_lock);
3936
3937         /*
3938          * Refer to the definition of io_bgs member for details why it's safe
3939          * to use it without any locking
3940          */
3941         while (!list_empty(io)) {
3942                 cache = list_first_entry(io, struct btrfs_block_group_cache,
3943                                          io_list);
3944                 list_del_init(&cache->io_list);
3945                 btrfs_wait_cache_io(trans, cache, path);
3946                 btrfs_put_block_group(cache);
3947         }
3948
3949         btrfs_free_path(path);
3950         return ret;
3951 }
3952
3953 int btrfs_extent_readonly(struct btrfs_fs_info *fs_info, u64 bytenr)
3954 {
3955         struct btrfs_block_group_cache *block_group;
3956         int readonly = 0;
3957
3958         block_group = btrfs_lookup_block_group(fs_info, bytenr);
3959         if (!block_group || block_group->ro)
3960                 readonly = 1;
3961         if (block_group)
3962                 btrfs_put_block_group(block_group);
3963         return readonly;
3964 }
3965
3966 bool btrfs_inc_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3967 {
3968         struct btrfs_block_group_cache *bg;
3969         bool ret = true;
3970
3971         bg = btrfs_lookup_block_group(fs_info, bytenr);
3972         if (!bg)
3973                 return false;
3974
3975         spin_lock(&bg->lock);
3976         if (bg->ro)
3977                 ret = false;
3978         else
3979                 atomic_inc(&bg->nocow_writers);
3980         spin_unlock(&bg->lock);
3981
3982         /* no put on block group, done by btrfs_dec_nocow_writers */
3983         if (!ret)
3984                 btrfs_put_block_group(bg);
3985
3986         return ret;
3987
3988 }
3989
3990 void btrfs_dec_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3991 {
3992         struct btrfs_block_group_cache *bg;
3993
3994         bg = btrfs_lookup_block_group(fs_info, bytenr);
3995         ASSERT(bg);
3996         if (atomic_dec_and_test(&bg->nocow_writers))
3997                 wake_up_atomic_t(&bg->nocow_writers);
3998         /*
3999          * Once for our lookup and once for the lookup done by a previous call
4000          * to btrfs_inc_nocow_writers()
4001          */
4002         btrfs_put_block_group(bg);
4003         btrfs_put_block_group(bg);
4004 }
4005
4006 void btrfs_wait_nocow_writers(struct btrfs_block_group_cache *bg)
4007 {
4008         wait_on_atomic_t(&bg->nocow_writers, atomic_t_wait,
4009                          TASK_UNINTERRUPTIBLE);
4010 }
4011
4012 static const char *alloc_name(u64 flags)
4013 {
4014         switch (flags) {
4015         case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA:
4016                 return "mixed";
4017         case BTRFS_BLOCK_GROUP_METADATA:
4018                 return "metadata";
4019         case BTRFS_BLOCK_GROUP_DATA:
4020                 return "data";
4021         case BTRFS_BLOCK_GROUP_SYSTEM:
4022                 return "system";
4023         default:
4024                 WARN_ON(1);
4025                 return "invalid-combination";
4026         };
4027 }
4028
4029 static int create_space_info(struct btrfs_fs_info *info, u64 flags,
4030                              struct btrfs_space_info **new)
4031 {
4032
4033         struct btrfs_space_info *space_info;
4034         int i;
4035         int ret;
4036
4037         space_info = kzalloc(sizeof(*space_info), GFP_NOFS);
4038         if (!space_info)
4039                 return -ENOMEM;
4040
4041         ret = percpu_counter_init(&space_info->total_bytes_pinned, 0,
4042                                  GFP_KERNEL);
4043         if (ret) {
4044                 kfree(space_info);
4045                 return ret;
4046         }
4047
4048         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
4049                 INIT_LIST_HEAD(&space_info->block_groups[i]);
4050         init_rwsem(&space_info->groups_sem);
4051         spin_lock_init(&space_info->lock);
4052         space_info->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
4053         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4054         init_waitqueue_head(&space_info->wait);
4055         INIT_LIST_HEAD(&space_info->ro_bgs);
4056         INIT_LIST_HEAD(&space_info->tickets);
4057         INIT_LIST_HEAD(&space_info->priority_tickets);
4058
4059         ret = kobject_init_and_add(&space_info->kobj, &space_info_ktype,
4060                                     info->space_info_kobj, "%s",
4061                                     alloc_name(space_info->flags));
4062         if (ret) {
4063                 percpu_counter_destroy(&space_info->total_bytes_pinned);
4064                 kfree(space_info);
4065                 return ret;
4066         }
4067
4068         *new = space_info;
4069         list_add_rcu(&space_info->list, &info->space_info);
4070         if (flags & BTRFS_BLOCK_GROUP_DATA)
4071                 info->data_sinfo = space_info;
4072
4073         return ret;
4074 }
4075
4076 static void update_space_info(struct btrfs_fs_info *info, u64 flags,
4077                              u64 total_bytes, u64 bytes_used,
4078                              u64 bytes_readonly,
4079                              struct btrfs_space_info **space_info)
4080 {
4081         struct btrfs_space_info *found;
4082         int factor;
4083
4084         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
4085                      BTRFS_BLOCK_GROUP_RAID10))
4086                 factor = 2;
4087         else
4088                 factor = 1;
4089
4090         found = __find_space_info(info, flags);
4091         ASSERT(found);
4092         spin_lock(&found->lock);
4093         found->total_bytes += total_bytes;
4094         found->disk_total += total_bytes * factor;
4095         found->bytes_used += bytes_used;
4096         found->disk_used += bytes_used * factor;
4097         found->bytes_readonly += bytes_readonly;
4098         if (total_bytes > 0)
4099                 found->full = 0;
4100         space_info_add_new_bytes(info, found, total_bytes -
4101                                  bytes_used - bytes_readonly);
4102         spin_unlock(&found->lock);
4103         *space_info = found;
4104 }
4105
4106 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
4107 {
4108         u64 extra_flags = chunk_to_extended(flags) &
4109                                 BTRFS_EXTENDED_PROFILE_MASK;
4110
4111         write_seqlock(&fs_info->profiles_lock);
4112         if (flags & BTRFS_BLOCK_GROUP_DATA)
4113                 fs_info->avail_data_alloc_bits |= extra_flags;
4114         if (flags & BTRFS_BLOCK_GROUP_METADATA)
4115                 fs_info->avail_metadata_alloc_bits |= extra_flags;
4116         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4117                 fs_info->avail_system_alloc_bits |= extra_flags;
4118         write_sequnlock(&fs_info->profiles_lock);
4119 }
4120
4121 /*
4122  * returns target flags in extended format or 0 if restripe for this
4123  * chunk_type is not in progress
4124  *
4125  * should be called with either volume_mutex or balance_lock held
4126  */
4127 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
4128 {
4129         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
4130         u64 target = 0;
4131
4132         if (!bctl)
4133                 return 0;
4134
4135         if (flags & BTRFS_BLOCK_GROUP_DATA &&
4136             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4137                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
4138         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
4139                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4140                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
4141         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
4142                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4143                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
4144         }
4145
4146         return target;
4147 }
4148
4149 /*
4150  * @flags: available profiles in extended format (see ctree.h)
4151  *
4152  * Returns reduced profile in chunk format.  If profile changing is in
4153  * progress (either running or paused) picks the target profile (if it's
4154  * already available), otherwise falls back to plain reducing.
4155  */
4156 static u64 btrfs_reduce_alloc_profile(struct btrfs_fs_info *fs_info, u64 flags)
4157 {
4158         u64 num_devices = fs_info->fs_devices->rw_devices;
4159         u64 target;
4160         u64 raid_type;
4161         u64 allowed = 0;
4162
4163         /*
4164          * see if restripe for this chunk_type is in progress, if so
4165          * try to reduce to the target profile
4166          */
4167         spin_lock(&fs_info->balance_lock);
4168         target = get_restripe_target(fs_info, flags);
4169         if (target) {
4170                 /* pick target profile only if it's already available */
4171                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
4172                         spin_unlock(&fs_info->balance_lock);
4173                         return extended_to_chunk(target);
4174                 }
4175         }
4176         spin_unlock(&fs_info->balance_lock);
4177
4178         /* First, mask out the RAID levels which aren't possible */
4179         for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
4180                 if (num_devices >= btrfs_raid_array[raid_type].devs_min)
4181                         allowed |= btrfs_raid_group[raid_type];
4182         }
4183         allowed &= flags;
4184
4185         if (allowed & BTRFS_BLOCK_GROUP_RAID6)
4186                 allowed = BTRFS_BLOCK_GROUP_RAID6;
4187         else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
4188                 allowed = BTRFS_BLOCK_GROUP_RAID5;
4189         else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
4190                 allowed = BTRFS_BLOCK_GROUP_RAID10;
4191         else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
4192                 allowed = BTRFS_BLOCK_GROUP_RAID1;
4193         else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
4194                 allowed = BTRFS_BLOCK_GROUP_RAID0;
4195
4196         flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
4197
4198         return extended_to_chunk(flags | allowed);
4199 }
4200
4201 static u64 get_alloc_profile(struct btrfs_fs_info *fs_info, u64 orig_flags)
4202 {
4203         unsigned seq;
4204         u64 flags;
4205
4206         do {
4207                 flags = orig_flags;
4208                 seq = read_seqbegin(&fs_info->profiles_lock);
4209
4210                 if (flags & BTRFS_BLOCK_GROUP_DATA)
4211                         flags |= fs_info->avail_data_alloc_bits;
4212                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4213                         flags |= fs_info->avail_system_alloc_bits;
4214                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
4215                         flags |= fs_info->avail_metadata_alloc_bits;
4216         } while (read_seqretry(&fs_info->profiles_lock, seq));
4217
4218         return btrfs_reduce_alloc_profile(fs_info, flags);
4219 }
4220
4221 static u64 get_alloc_profile_by_root(struct btrfs_root *root, int data)
4222 {
4223         struct btrfs_fs_info *fs_info = root->fs_info;
4224         u64 flags;
4225         u64 ret;
4226
4227         if (data)
4228                 flags = BTRFS_BLOCK_GROUP_DATA;
4229         else if (root == fs_info->chunk_root)
4230                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
4231         else
4232                 flags = BTRFS_BLOCK_GROUP_METADATA;
4233
4234         ret = get_alloc_profile(fs_info, flags);
4235         return ret;
4236 }
4237
4238 u64 btrfs_data_alloc_profile(struct btrfs_fs_info *fs_info)
4239 {
4240         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_DATA);
4241 }
4242
4243 u64 btrfs_metadata_alloc_profile(struct btrfs_fs_info *fs_info)
4244 {
4245         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4246 }
4247
4248 u64 btrfs_system_alloc_profile(struct btrfs_fs_info *fs_info)
4249 {
4250         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4251 }
4252
4253 static u64 btrfs_space_info_used(struct btrfs_space_info *s_info,
4254                                  bool may_use_included)
4255 {
4256         ASSERT(s_info);
4257         return s_info->bytes_used + s_info->bytes_reserved +
4258                 s_info->bytes_pinned + s_info->bytes_readonly +
4259                 (may_use_included ? s_info->bytes_may_use : 0);
4260 }
4261
4262 int btrfs_alloc_data_chunk_ondemand(struct btrfs_inode *inode, u64 bytes)
4263 {
4264         struct btrfs_root *root = inode->root;
4265         struct btrfs_fs_info *fs_info = root->fs_info;
4266         struct btrfs_space_info *data_sinfo = fs_info->data_sinfo;
4267         u64 used;
4268         int ret = 0;
4269         int need_commit = 2;
4270         int have_pinned_space;
4271
4272         /* make sure bytes are sectorsize aligned */
4273         bytes = ALIGN(bytes, fs_info->sectorsize);
4274
4275         if (btrfs_is_free_space_inode(inode)) {
4276                 need_commit = 0;
4277                 ASSERT(current->journal_info);
4278         }
4279
4280 again:
4281         /* make sure we have enough space to handle the data first */
4282         spin_lock(&data_sinfo->lock);
4283         used = btrfs_space_info_used(data_sinfo, true);
4284
4285         if (used + bytes > data_sinfo->total_bytes) {
4286                 struct btrfs_trans_handle *trans;
4287
4288                 /*
4289                  * if we don't have enough free bytes in this space then we need
4290                  * to alloc a new chunk.
4291                  */
4292                 if (!data_sinfo->full) {
4293                         u64 alloc_target;
4294
4295                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
4296                         spin_unlock(&data_sinfo->lock);
4297
4298                         alloc_target = btrfs_data_alloc_profile(fs_info);
4299                         /*
4300                          * It is ugly that we don't call nolock join
4301                          * transaction for the free space inode case here.
4302                          * But it is safe because we only do the data space
4303                          * reservation for the free space cache in the
4304                          * transaction context, the common join transaction
4305                          * just increase the counter of the current transaction
4306                          * handler, doesn't try to acquire the trans_lock of
4307                          * the fs.
4308                          */
4309                         trans = btrfs_join_transaction(root);
4310                         if (IS_ERR(trans))
4311                                 return PTR_ERR(trans);
4312
4313                         ret = do_chunk_alloc(trans, fs_info, alloc_target,
4314                                              CHUNK_ALLOC_NO_FORCE);
4315                         btrfs_end_transaction(trans);
4316                         if (ret < 0) {
4317                                 if (ret != -ENOSPC)
4318                                         return ret;
4319                                 else {
4320                                         have_pinned_space = 1;
4321                                         goto commit_trans;
4322                                 }
4323                         }
4324
4325                         goto again;
4326                 }
4327
4328                 /*
4329                  * If we don't have enough pinned space to deal with this
4330                  * allocation, and no removed chunk in current transaction,
4331                  * don't bother committing the transaction.
4332                  */
4333                 have_pinned_space = percpu_counter_compare(
4334                         &data_sinfo->total_bytes_pinned,
4335                         used + bytes - data_sinfo->total_bytes);
4336                 spin_unlock(&data_sinfo->lock);
4337
4338                 /* commit the current transaction and try again */
4339 commit_trans:
4340                 if (need_commit) {
4341                         need_commit--;
4342
4343                         if (need_commit > 0) {
4344                                 btrfs_start_delalloc_roots(fs_info, 0, -1);
4345                                 btrfs_wait_ordered_roots(fs_info, U64_MAX, 0,
4346                                                          (u64)-1);
4347                         }
4348
4349                         trans = btrfs_join_transaction(root);
4350                         if (IS_ERR(trans))
4351                                 return PTR_ERR(trans);
4352                         if (have_pinned_space >= 0 ||
4353                             test_bit(BTRFS_TRANS_HAVE_FREE_BGS,
4354                                      &trans->transaction->flags) ||
4355                             need_commit > 0) {
4356                                 ret = btrfs_commit_transaction(trans);
4357                                 if (ret)
4358                                         return ret;
4359                                 /*
4360                                  * The cleaner kthread might still be doing iput
4361                                  * operations. Wait for it to finish so that
4362                                  * more space is released.
4363                                  */
4364                                 mutex_lock(&fs_info->cleaner_delayed_iput_mutex);
4365                                 mutex_unlock(&fs_info->cleaner_delayed_iput_mutex);
4366                                 goto again;
4367                         } else {
4368                                 btrfs_end_transaction(trans);
4369                         }
4370                 }
4371
4372                 trace_btrfs_space_reservation(fs_info,
4373                                               "space_info:enospc",
4374                                               data_sinfo->flags, bytes, 1);
4375                 return -ENOSPC;
4376         }
4377         data_sinfo->bytes_may_use += bytes;
4378         trace_btrfs_space_reservation(fs_info, "space_info",
4379                                       data_sinfo->flags, bytes, 1);
4380         spin_unlock(&data_sinfo->lock);
4381
4382         return ret;
4383 }
4384
4385 int btrfs_check_data_free_space(struct inode *inode,
4386                         struct extent_changeset **reserved, u64 start, u64 len)
4387 {
4388         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4389         int ret;
4390
4391         /* align the range */
4392         len = round_up(start + len, fs_info->sectorsize) -
4393               round_down(start, fs_info->sectorsize);
4394         start = round_down(start, fs_info->sectorsize);
4395
4396         ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode), len);
4397         if (ret < 0)
4398                 return ret;
4399
4400         /* Use new btrfs_qgroup_reserve_data to reserve precious data space. */
4401         ret = btrfs_qgroup_reserve_data(inode, reserved, start, len);
4402         if (ret < 0)
4403                 btrfs_free_reserved_data_space_noquota(inode, start, len);
4404         else
4405                 ret = 0;
4406         return ret;
4407 }
4408
4409 /*
4410  * Called if we need to clear a data reservation for this inode
4411  * Normally in a error case.
4412  *
4413  * This one will *NOT* use accurate qgroup reserved space API, just for case
4414  * which we can't sleep and is sure it won't affect qgroup reserved space.
4415  * Like clear_bit_hook().
4416  */
4417 void btrfs_free_reserved_data_space_noquota(struct inode *inode, u64 start,
4418                                             u64 len)
4419 {
4420         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4421         struct btrfs_space_info *data_sinfo;
4422
4423         /* Make sure the range is aligned to sectorsize */
4424         len = round_up(start + len, fs_info->sectorsize) -
4425               round_down(start, fs_info->sectorsize);
4426         start = round_down(start, fs_info->sectorsize);
4427
4428         data_sinfo = fs_info->data_sinfo;
4429         spin_lock(&data_sinfo->lock);
4430         if (WARN_ON(data_sinfo->bytes_may_use < len))
4431                 data_sinfo->bytes_may_use = 0;
4432         else
4433                 data_sinfo->bytes_may_use -= len;
4434         trace_btrfs_space_reservation(fs_info, "space_info",
4435                                       data_sinfo->flags, len, 0);
4436         spin_unlock(&data_sinfo->lock);
4437 }
4438
4439 /*
4440  * Called if we need to clear a data reservation for this inode
4441  * Normally in a error case.
4442  *
4443  * This one will handle the per-inode data rsv map for accurate reserved
4444  * space framework.
4445  */
4446 void btrfs_free_reserved_data_space(struct inode *inode,
4447                         struct extent_changeset *reserved, u64 start, u64 len)
4448 {
4449         struct btrfs_root *root = BTRFS_I(inode)->root;
4450
4451         /* Make sure the range is aligned to sectorsize */
4452         len = round_up(start + len, root->fs_info->sectorsize) -
4453               round_down(start, root->fs_info->sectorsize);
4454         start = round_down(start, root->fs_info->sectorsize);
4455
4456         btrfs_free_reserved_data_space_noquota(inode, start, len);
4457         btrfs_qgroup_free_data(inode, reserved, start, len);
4458 }
4459
4460 static void force_metadata_allocation(struct btrfs_fs_info *info)
4461 {
4462         struct list_head *head = &info->space_info;
4463         struct btrfs_space_info *found;
4464
4465         rcu_read_lock();
4466         list_for_each_entry_rcu(found, head, list) {
4467                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
4468                         found->force_alloc = CHUNK_ALLOC_FORCE;
4469         }
4470         rcu_read_unlock();
4471 }
4472
4473 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
4474 {
4475         return (global->size << 1);
4476 }
4477
4478 static int should_alloc_chunk(struct btrfs_fs_info *fs_info,
4479                               struct btrfs_space_info *sinfo, int force)
4480 {
4481         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4482         u64 bytes_used = btrfs_space_info_used(sinfo, false);
4483         u64 thresh;
4484
4485         if (force == CHUNK_ALLOC_FORCE)
4486                 return 1;
4487
4488         /*
4489          * We need to take into account the global rsv because for all intents
4490          * and purposes it's used space.  Don't worry about locking the
4491          * global_rsv, it doesn't change except when the transaction commits.
4492          */
4493         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
4494                 bytes_used += calc_global_rsv_need_space(global_rsv);
4495
4496         /*
4497          * in limited mode, we want to have some free space up to
4498          * about 1% of the FS size.
4499          */
4500         if (force == CHUNK_ALLOC_LIMITED) {
4501                 thresh = btrfs_super_total_bytes(fs_info->super_copy);
4502                 thresh = max_t(u64, SZ_64M, div_factor_fine(thresh, 1));
4503
4504                 if (sinfo->total_bytes - bytes_used < thresh)
4505                         return 1;
4506         }
4507
4508         if (bytes_used + SZ_2M < div_factor(sinfo->total_bytes, 8))
4509                 return 0;
4510         return 1;
4511 }
4512
4513 static u64 get_profile_num_devs(struct btrfs_fs_info *fs_info, u64 type)
4514 {
4515         u64 num_dev;
4516
4517         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
4518                     BTRFS_BLOCK_GROUP_RAID0 |
4519                     BTRFS_BLOCK_GROUP_RAID5 |
4520                     BTRFS_BLOCK_GROUP_RAID6))
4521                 num_dev = fs_info->fs_devices->rw_devices;
4522         else if (type & BTRFS_BLOCK_GROUP_RAID1)
4523                 num_dev = 2;
4524         else
4525                 num_dev = 1;    /* DUP or single */
4526
4527         return num_dev;
4528 }
4529
4530 /*
4531  * If @is_allocation is true, reserve space in the system space info necessary
4532  * for allocating a chunk, otherwise if it's false, reserve space necessary for
4533  * removing a chunk.
4534  */
4535 void check_system_chunk(struct btrfs_trans_handle *trans,
4536                         struct btrfs_fs_info *fs_info, u64 type)
4537 {
4538         struct btrfs_space_info *info;
4539         u64 left;
4540         u64 thresh;
4541         int ret = 0;
4542         u64 num_devs;
4543
4544         /*
4545          * Needed because we can end up allocating a system chunk and for an
4546          * atomic and race free space reservation in the chunk block reserve.
4547          */
4548         lockdep_assert_held(&fs_info->chunk_mutex);
4549
4550         info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4551         spin_lock(&info->lock);
4552         left = info->total_bytes - btrfs_space_info_used(info, true);
4553         spin_unlock(&info->lock);
4554
4555         num_devs = get_profile_num_devs(fs_info, type);
4556
4557         /* num_devs device items to update and 1 chunk item to add or remove */
4558         thresh = btrfs_calc_trunc_metadata_size(fs_info, num_devs) +
4559                 btrfs_calc_trans_metadata_size(fs_info, 1);
4560
4561         if (left < thresh && btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
4562                 btrfs_info(fs_info, "left=%llu, need=%llu, flags=%llu",
4563                            left, thresh, type);
4564                 dump_space_info(fs_info, info, 0, 0);
4565         }
4566
4567         if (left < thresh) {
4568                 u64 flags = btrfs_system_alloc_profile(fs_info);
4569
4570                 /*
4571                  * Ignore failure to create system chunk. We might end up not
4572                  * needing it, as we might not need to COW all nodes/leafs from
4573                  * the paths we visit in the chunk tree (they were already COWed
4574                  * or created in the current transaction for example).
4575                  */
4576                 ret = btrfs_alloc_chunk(trans, fs_info, flags);
4577         }
4578
4579         if (!ret) {
4580                 ret = btrfs_block_rsv_add(fs_info->chunk_root,
4581                                           &fs_info->chunk_block_rsv,
4582                                           thresh, BTRFS_RESERVE_NO_FLUSH);
4583                 if (!ret)
4584                         trans->chunk_bytes_reserved += thresh;
4585         }
4586 }
4587
4588 /*
4589  * If force is CHUNK_ALLOC_FORCE:
4590  *    - return 1 if it successfully allocates a chunk,
4591  *    - return errors including -ENOSPC otherwise.
4592  * If force is NOT CHUNK_ALLOC_FORCE:
4593  *    - return 0 if it doesn't need to allocate a new chunk,
4594  *    - return 1 if it successfully allocates a chunk,
4595  *    - return errors including -ENOSPC otherwise.
4596  */
4597 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
4598                           struct btrfs_fs_info *fs_info, u64 flags, int force)
4599 {
4600         struct btrfs_space_info *space_info;
4601         int wait_for_alloc = 0;
4602         int ret = 0;
4603
4604         /* Don't re-enter if we're already allocating a chunk */
4605         if (trans->allocating_chunk)
4606                 return -ENOSPC;
4607
4608         space_info = __find_space_info(fs_info, flags);
4609         ASSERT(space_info);
4610
4611 again:
4612         spin_lock(&space_info->lock);
4613         if (force < space_info->force_alloc)
4614                 force = space_info->force_alloc;
4615         if (space_info->full) {
4616                 if (should_alloc_chunk(fs_info, space_info, force))
4617                         ret = -ENOSPC;
4618                 else
4619                         ret = 0;
4620                 spin_unlock(&space_info->lock);
4621                 return ret;
4622         }
4623
4624         if (!should_alloc_chunk(fs_info, space_info, force)) {
4625                 spin_unlock(&space_info->lock);
4626                 return 0;
4627         } else if (space_info->chunk_alloc) {
4628                 wait_for_alloc = 1;
4629         } else {
4630                 space_info->chunk_alloc = 1;
4631         }
4632
4633         spin_unlock(&space_info->lock);
4634
4635         mutex_lock(&fs_info->chunk_mutex);
4636
4637         /*
4638          * The chunk_mutex is held throughout the entirety of a chunk
4639          * allocation, so once we've acquired the chunk_mutex we know that the
4640          * other guy is done and we need to recheck and see if we should
4641          * allocate.
4642          */
4643         if (wait_for_alloc) {
4644                 mutex_unlock(&fs_info->chunk_mutex);
4645                 wait_for_alloc = 0;
4646                 goto again;
4647         }
4648
4649         trans->allocating_chunk = true;
4650
4651         /*
4652          * If we have mixed data/metadata chunks we want to make sure we keep
4653          * allocating mixed chunks instead of individual chunks.
4654          */
4655         if (btrfs_mixed_space_info(space_info))
4656                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4657
4658         /*
4659          * if we're doing a data chunk, go ahead and make sure that
4660          * we keep a reasonable number of metadata chunks allocated in the
4661          * FS as well.
4662          */
4663         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4664                 fs_info->data_chunk_allocations++;
4665                 if (!(fs_info->data_chunk_allocations %
4666                       fs_info->metadata_ratio))
4667                         force_metadata_allocation(fs_info);
4668         }
4669
4670         /*
4671          * Check if we have enough space in SYSTEM chunk because we may need
4672          * to update devices.
4673          */
4674         check_system_chunk(trans, fs_info, flags);
4675
4676         ret = btrfs_alloc_chunk(trans, fs_info, flags);
4677         trans->allocating_chunk = false;
4678
4679         spin_lock(&space_info->lock);
4680         if (ret < 0 && ret != -ENOSPC)
4681                 goto out;
4682         if (ret)
4683                 space_info->full = 1;
4684         else
4685                 ret = 1;
4686
4687         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4688 out:
4689         space_info->chunk_alloc = 0;
4690         spin_unlock(&space_info->lock);
4691         mutex_unlock(&fs_info->chunk_mutex);
4692         /*
4693          * When we allocate a new chunk we reserve space in the chunk block
4694          * reserve to make sure we can COW nodes/leafs in the chunk tree or
4695          * add new nodes/leafs to it if we end up needing to do it when
4696          * inserting the chunk item and updating device items as part of the
4697          * second phase of chunk allocation, performed by
4698          * btrfs_finish_chunk_alloc(). So make sure we don't accumulate a
4699          * large number of new block groups to create in our transaction
4700          * handle's new_bgs list to avoid exhausting the chunk block reserve
4701          * in extreme cases - like having a single transaction create many new
4702          * block groups when starting to write out the free space caches of all
4703          * the block groups that were made dirty during the lifetime of the
4704          * transaction.
4705          */
4706         if (trans->can_flush_pending_bgs &&
4707             trans->chunk_bytes_reserved >= (u64)SZ_2M) {
4708                 btrfs_create_pending_block_groups(trans);
4709                 btrfs_trans_release_chunk_metadata(trans);
4710         }
4711         return ret;
4712 }
4713
4714 static int can_overcommit(struct btrfs_fs_info *fs_info,
4715                           struct btrfs_space_info *space_info, u64 bytes,
4716                           enum btrfs_reserve_flush_enum flush,
4717                           bool system_chunk)
4718 {
4719         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4720         u64 profile;
4721         u64 space_size;
4722         u64 avail;
4723         u64 used;
4724
4725         /* Don't overcommit when in mixed mode. */
4726         if (space_info->flags & BTRFS_BLOCK_GROUP_DATA)
4727                 return 0;
4728
4729         if (system_chunk)
4730                 profile = btrfs_system_alloc_profile(fs_info);
4731         else
4732                 profile = btrfs_metadata_alloc_profile(fs_info);
4733
4734         used = btrfs_space_info_used(space_info, false);
4735
4736         /*
4737          * We only want to allow over committing if we have lots of actual space
4738          * free, but if we don't have enough space to handle the global reserve
4739          * space then we could end up having a real enospc problem when trying
4740          * to allocate a chunk or some other such important allocation.
4741          */
4742         spin_lock(&global_rsv->lock);
4743         space_size = calc_global_rsv_need_space(global_rsv);
4744         spin_unlock(&global_rsv->lock);
4745         if (used + space_size >= space_info->total_bytes)
4746                 return 0;
4747
4748         used += space_info->bytes_may_use;
4749
4750         avail = atomic64_read(&fs_info->free_chunk_space);
4751
4752         /*
4753          * If we have dup, raid1 or raid10 then only half of the free
4754          * space is actually useable.  For raid56, the space info used
4755          * doesn't include the parity drive, so we don't have to
4756          * change the math
4757          */
4758         if (profile & (BTRFS_BLOCK_GROUP_DUP |
4759                        BTRFS_BLOCK_GROUP_RAID1 |
4760                        BTRFS_BLOCK_GROUP_RAID10))
4761                 avail >>= 1;
4762
4763         /*
4764          * If we aren't flushing all things, let us overcommit up to
4765          * 1/2th of the space. If we can flush, don't let us overcommit
4766          * too much, let it overcommit up to 1/8 of the space.
4767          */
4768         if (flush == BTRFS_RESERVE_FLUSH_ALL)
4769                 avail >>= 3;
4770         else
4771                 avail >>= 1;
4772
4773         if (used + bytes < space_info->total_bytes + avail)
4774                 return 1;
4775         return 0;
4776 }
4777
4778 static void btrfs_writeback_inodes_sb_nr(struct btrfs_fs_info *fs_info,
4779                                          unsigned long nr_pages, int nr_items)
4780 {
4781         struct super_block *sb = fs_info->sb;
4782
4783         if (down_read_trylock(&sb->s_umount)) {
4784                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
4785                 up_read(&sb->s_umount);
4786         } else {
4787                 /*
4788                  * We needn't worry the filesystem going from r/w to r/o though
4789                  * we don't acquire ->s_umount mutex, because the filesystem
4790                  * should guarantee the delalloc inodes list be empty after
4791                  * the filesystem is readonly(all dirty pages are written to
4792                  * the disk).
4793                  */
4794                 btrfs_start_delalloc_roots(fs_info, 0, nr_items);
4795                 if (!current->journal_info)
4796                         btrfs_wait_ordered_roots(fs_info, nr_items, 0, (u64)-1);
4797         }
4798 }
4799
4800 static inline u64 calc_reclaim_items_nr(struct btrfs_fs_info *fs_info,
4801                                         u64 to_reclaim)
4802 {
4803         u64 bytes;
4804         u64 nr;
4805
4806         bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
4807         nr = div64_u64(to_reclaim, bytes);
4808         if (!nr)
4809                 nr = 1;
4810         return nr;
4811 }
4812
4813 #define EXTENT_SIZE_PER_ITEM    SZ_256K
4814
4815 /*
4816  * shrink metadata reservation for delalloc
4817  */
4818 static void shrink_delalloc(struct btrfs_fs_info *fs_info, u64 to_reclaim,
4819                             u64 orig, bool wait_ordered)
4820 {
4821         struct btrfs_space_info *space_info;
4822         struct btrfs_trans_handle *trans;
4823         u64 delalloc_bytes;
4824         u64 max_reclaim;
4825         u64 items;
4826         long time_left;
4827         unsigned long nr_pages;
4828         int loops;
4829
4830         /* Calc the number of the pages we need flush for space reservation */
4831         items = calc_reclaim_items_nr(fs_info, to_reclaim);
4832         to_reclaim = items * EXTENT_SIZE_PER_ITEM;
4833
4834         trans = (struct btrfs_trans_handle *)current->journal_info;
4835         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4836
4837         delalloc_bytes = percpu_counter_sum_positive(
4838                                                 &fs_info->delalloc_bytes);
4839         if (delalloc_bytes == 0) {
4840                 if (trans)
4841                         return;
4842                 if (wait_ordered)
4843                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
4844                 return;
4845         }
4846
4847         loops = 0;
4848         while (delalloc_bytes && loops < 3) {
4849                 max_reclaim = min(delalloc_bytes, to_reclaim);
4850                 nr_pages = max_reclaim >> PAGE_SHIFT;
4851                 btrfs_writeback_inodes_sb_nr(fs_info, nr_pages, items);
4852                 /*
4853                  * We need to wait for the async pages to actually start before
4854                  * we do anything.
4855                  */
4856                 max_reclaim = atomic_read(&fs_info->async_delalloc_pages);
4857                 if (!max_reclaim)
4858                         goto skip_async;
4859
4860                 if (max_reclaim <= nr_pages)
4861                         max_reclaim = 0;
4862                 else
4863                         max_reclaim -= nr_pages;
4864
4865                 wait_event(fs_info->async_submit_wait,
4866                            atomic_read(&fs_info->async_delalloc_pages) <=
4867                            (int)max_reclaim);
4868 skip_async:
4869                 spin_lock(&space_info->lock);
4870                 if (list_empty(&space_info->tickets) &&
4871                     list_empty(&space_info->priority_tickets)) {
4872                         spin_unlock(&space_info->lock);
4873                         break;
4874                 }
4875                 spin_unlock(&space_info->lock);
4876
4877                 loops++;
4878                 if (wait_ordered && !trans) {
4879                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
4880                 } else {
4881                         time_left = schedule_timeout_killable(1);
4882                         if (time_left)
4883                                 break;
4884                 }
4885                 delalloc_bytes = percpu_counter_sum_positive(
4886                                                 &fs_info->delalloc_bytes);
4887         }
4888 }
4889
4890 struct reserve_ticket {
4891         u64 bytes;
4892         int error;
4893         struct list_head list;
4894         wait_queue_head_t wait;
4895 };
4896
4897 /**
4898  * maybe_commit_transaction - possibly commit the transaction if its ok to
4899  * @root - the root we're allocating for
4900  * @bytes - the number of bytes we want to reserve
4901  * @force - force the commit
4902  *
4903  * This will check to make sure that committing the transaction will actually
4904  * get us somewhere and then commit the transaction if it does.  Otherwise it
4905  * will return -ENOSPC.
4906  */
4907 static int may_commit_transaction(struct btrfs_fs_info *fs_info,
4908                                   struct btrfs_space_info *space_info)
4909 {
4910         struct reserve_ticket *ticket = NULL;
4911         struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_block_rsv;
4912         struct btrfs_trans_handle *trans;
4913         u64 bytes;
4914
4915         trans = (struct btrfs_trans_handle *)current->journal_info;
4916         if (trans)
4917                 return -EAGAIN;
4918
4919         spin_lock(&space_info->lock);
4920         if (!list_empty(&space_info->priority_tickets))
4921                 ticket = list_first_entry(&space_info->priority_tickets,
4922                                           struct reserve_ticket, list);
4923         else if (!list_empty(&space_info->tickets))
4924                 ticket = list_first_entry(&space_info->tickets,
4925                                           struct reserve_ticket, list);
4926         bytes = (ticket) ? ticket->bytes : 0;
4927         spin_unlock(&space_info->lock);
4928
4929         if (!bytes)
4930                 return 0;
4931
4932         /* See if there is enough pinned space to make this reservation */
4933         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4934                                    bytes) >= 0)
4935                 goto commit;
4936
4937         /*
4938          * See if there is some space in the delayed insertion reservation for
4939          * this reservation.
4940          */
4941         if (space_info != delayed_rsv->space_info)
4942                 return -ENOSPC;
4943
4944         spin_lock(&delayed_rsv->lock);
4945         if (delayed_rsv->size > bytes)
4946                 bytes = 0;
4947         else
4948                 bytes -= delayed_rsv->size;
4949         spin_unlock(&delayed_rsv->lock);
4950
4951         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4952                                    bytes) < 0) {
4953                 return -ENOSPC;
4954         }
4955
4956 commit:
4957         trans = btrfs_join_transaction(fs_info->extent_root);
4958         if (IS_ERR(trans))
4959                 return -ENOSPC;
4960
4961         return btrfs_commit_transaction(trans);
4962 }
4963
4964 /*
4965  * Try to flush some data based on policy set by @state. This is only advisory
4966  * and may fail for various reasons. The caller is supposed to examine the
4967  * state of @space_info to detect the outcome.
4968  */
4969 static void flush_space(struct btrfs_fs_info *fs_info,
4970                        struct btrfs_space_info *space_info, u64 num_bytes,
4971                        int state)
4972 {
4973         struct btrfs_root *root = fs_info->extent_root;
4974         struct btrfs_trans_handle *trans;
4975         int nr;
4976         int ret = 0;
4977
4978         switch (state) {
4979         case FLUSH_DELAYED_ITEMS_NR:
4980         case FLUSH_DELAYED_ITEMS:
4981                 if (state == FLUSH_DELAYED_ITEMS_NR)
4982                         nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2;
4983                 else
4984                         nr = -1;
4985
4986                 trans = btrfs_join_transaction(root);
4987                 if (IS_ERR(trans)) {
4988                         ret = PTR_ERR(trans);
4989                         break;
4990                 }
4991                 ret = btrfs_run_delayed_items_nr(trans, nr);
4992                 btrfs_end_transaction(trans);
4993                 break;
4994         case FLUSH_DELALLOC:
4995         case FLUSH_DELALLOC_WAIT:
4996                 shrink_delalloc(fs_info, num_bytes * 2, num_bytes,
4997                                 state == FLUSH_DELALLOC_WAIT);
4998                 break;
4999         case ALLOC_CHUNK:
5000                 trans = btrfs_join_transaction(root);
5001                 if (IS_ERR(trans)) {
5002                         ret = PTR_ERR(trans);
5003                         break;
5004                 }
5005                 ret = do_chunk_alloc(trans, fs_info,
5006                                      btrfs_metadata_alloc_profile(fs_info),
5007                                      CHUNK_ALLOC_NO_FORCE);
5008                 btrfs_end_transaction(trans);
5009                 if (ret > 0 || ret == -ENOSPC)
5010                         ret = 0;
5011                 break;
5012         case COMMIT_TRANS:
5013                 ret = may_commit_transaction(fs_info, space_info);
5014                 break;
5015         default:
5016                 ret = -ENOSPC;
5017                 break;
5018         }
5019
5020         trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state,
5021                                 ret);
5022         return;
5023 }
5024
5025 static inline u64
5026 btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info,
5027                                  struct btrfs_space_info *space_info,
5028                                  bool system_chunk)
5029 {
5030         struct reserve_ticket *ticket;
5031         u64 used;
5032         u64 expected;
5033         u64 to_reclaim = 0;
5034
5035         list_for_each_entry(ticket, &space_info->tickets, list)
5036                 to_reclaim += ticket->bytes;
5037         list_for_each_entry(ticket, &space_info->priority_tickets, list)
5038                 to_reclaim += ticket->bytes;
5039         if (to_reclaim)
5040                 return to_reclaim;
5041
5042         to_reclaim = min_t(u64, num_online_cpus() * SZ_1M, SZ_16M);
5043         if (can_overcommit(fs_info, space_info, to_reclaim,
5044                            BTRFS_RESERVE_FLUSH_ALL, system_chunk))
5045                 return 0;
5046
5047         used = btrfs_space_info_used(space_info, true);
5048
5049         if (can_overcommit(fs_info, space_info, SZ_1M,
5050                            BTRFS_RESERVE_FLUSH_ALL, system_chunk))
5051                 expected = div_factor_fine(space_info->total_bytes, 95);
5052         else
5053                 expected = div_factor_fine(space_info->total_bytes, 90);
5054
5055         if (used > expected)
5056                 to_reclaim = used - expected;
5057         else
5058                 to_reclaim = 0;
5059         to_reclaim = min(to_reclaim, space_info->bytes_may_use +
5060                                      space_info->bytes_reserved);
5061         return to_reclaim;
5062 }
5063
5064 static inline int need_do_async_reclaim(struct btrfs_fs_info *fs_info,
5065                                         struct btrfs_space_info *space_info,
5066                                         u64 used, bool system_chunk)
5067 {
5068         u64 thresh = div_factor_fine(space_info->total_bytes, 98);
5069
5070         /* If we're just plain full then async reclaim just slows us down. */
5071         if ((space_info->bytes_used + space_info->bytes_reserved) >= thresh)
5072                 return 0;
5073
5074         if (!btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5075                                               system_chunk))
5076                 return 0;
5077
5078         return (used >= thresh && !btrfs_fs_closing(fs_info) &&
5079                 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
5080 }
5081
5082 static void wake_all_tickets(struct list_head *head)
5083 {
5084         struct reserve_ticket *ticket;
5085
5086         while (!list_empty(head)) {
5087                 ticket = list_first_entry(head, struct reserve_ticket, list);
5088                 list_del_init(&ticket->list);
5089                 ticket->error = -ENOSPC;
5090                 wake_up(&ticket->wait);
5091         }
5092 }
5093
5094 /*
5095  * This is for normal flushers, we can wait all goddamned day if we want to.  We
5096  * will loop and continuously try to flush as long as we are making progress.
5097  * We count progress as clearing off tickets each time we have to loop.
5098  */
5099 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
5100 {
5101         struct btrfs_fs_info *fs_info;
5102         struct btrfs_space_info *space_info;
5103         u64 to_reclaim;
5104         int flush_state;
5105         int commit_cycles = 0;
5106         u64 last_tickets_id;
5107
5108         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
5109         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5110
5111         spin_lock(&space_info->lock);
5112         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5113                                                       false);
5114         if (!to_reclaim) {
5115                 space_info->flush = 0;
5116                 spin_unlock(&space_info->lock);
5117                 return;
5118         }
5119         last_tickets_id = space_info->tickets_id;
5120         spin_unlock(&space_info->lock);
5121
5122         flush_state = FLUSH_DELAYED_ITEMS_NR;
5123         do {
5124                 flush_space(fs_info, space_info, to_reclaim, flush_state);
5125                 spin_lock(&space_info->lock);
5126                 if (list_empty(&space_info->tickets)) {
5127                         space_info->flush = 0;
5128                         spin_unlock(&space_info->lock);
5129                         return;
5130                 }
5131                 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info,
5132                                                               space_info,
5133                                                               false);
5134                 if (last_tickets_id == space_info->tickets_id) {
5135                         flush_state++;
5136                 } else {
5137                         last_tickets_id = space_info->tickets_id;
5138                         flush_state = FLUSH_DELAYED_ITEMS_NR;
5139                         if (commit_cycles)
5140                                 commit_cycles--;
5141                 }
5142
5143                 if (flush_state > COMMIT_TRANS) {
5144                         commit_cycles++;
5145                         if (commit_cycles > 2) {
5146                                 wake_all_tickets(&space_info->tickets);
5147                                 space_info->flush = 0;
5148                         } else {
5149                                 flush_state = FLUSH_DELAYED_ITEMS_NR;
5150                         }
5151                 }
5152                 spin_unlock(&space_info->lock);
5153         } while (flush_state <= COMMIT_TRANS);
5154 }
5155
5156 void btrfs_init_async_reclaim_work(struct work_struct *work)
5157 {
5158         INIT_WORK(work, btrfs_async_reclaim_metadata_space);
5159 }
5160
5161 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info,
5162                                             struct btrfs_space_info *space_info,
5163                                             struct reserve_ticket *ticket)
5164 {
5165         u64 to_reclaim;
5166         int flush_state = FLUSH_DELAYED_ITEMS_NR;
5167
5168         spin_lock(&space_info->lock);
5169         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5170                                                       false);
5171         if (!to_reclaim) {
5172                 spin_unlock(&space_info->lock);
5173                 return;
5174         }
5175         spin_unlock(&space_info->lock);
5176
5177         do {
5178                 flush_space(fs_info, space_info, to_reclaim, flush_state);
5179                 flush_state++;
5180                 spin_lock(&space_info->lock);
5181                 if (ticket->bytes == 0) {
5182                         spin_unlock(&space_info->lock);
5183                         return;
5184                 }
5185                 spin_unlock(&space_info->lock);
5186
5187                 /*
5188                  * Priority flushers can't wait on delalloc without
5189                  * deadlocking.
5190                  */
5191                 if (flush_state == FLUSH_DELALLOC ||
5192                     flush_state == FLUSH_DELALLOC_WAIT)
5193                         flush_state = ALLOC_CHUNK;
5194         } while (flush_state < COMMIT_TRANS);
5195 }
5196
5197 static int wait_reserve_ticket(struct btrfs_fs_info *fs_info,
5198                                struct btrfs_space_info *space_info,
5199                                struct reserve_ticket *ticket, u64 orig_bytes)
5200
5201 {
5202         DEFINE_WAIT(wait);
5203         int ret = 0;
5204
5205         spin_lock(&space_info->lock);
5206         while (ticket->bytes > 0 && ticket->error == 0) {
5207                 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE);
5208                 if (ret) {
5209                         ret = -EINTR;
5210                         break;
5211                 }
5212                 spin_unlock(&space_info->lock);
5213
5214                 schedule();
5215
5216                 finish_wait(&ticket->wait, &wait);
5217                 spin_lock(&space_info->lock);
5218         }
5219         if (!ret)
5220                 ret = ticket->error;
5221         if (!list_empty(&ticket->list))
5222                 list_del_init(&ticket->list);
5223         if (ticket->bytes && ticket->bytes < orig_bytes) {
5224                 u64 num_bytes = orig_bytes - ticket->bytes;
5225                 space_info->bytes_may_use -= num_bytes;
5226                 trace_btrfs_space_reservation(fs_info, "space_info",
5227                                               space_info->flags, num_bytes, 0);
5228         }
5229         spin_unlock(&space_info->lock);
5230
5231         return ret;
5232 }
5233
5234 /**
5235  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5236  * @root - the root we're allocating for
5237  * @space_info - the space info we want to allocate from
5238  * @orig_bytes - the number of bytes we want
5239  * @flush - whether or not we can flush to make our reservation
5240  *
5241  * This will reserve orig_bytes number of bytes from the space info associated
5242  * with the block_rsv.  If there is not enough space it will make an attempt to
5243  * flush out space to make room.  It will do this by flushing delalloc if
5244  * possible or committing the transaction.  If flush is 0 then no attempts to
5245  * regain reservations will be made and this will fail if there is not enough
5246  * space already.
5247  */
5248 static int __reserve_metadata_bytes(struct btrfs_fs_info *fs_info,
5249                                     struct btrfs_space_info *space_info,
5250                                     u64 orig_bytes,
5251                                     enum btrfs_reserve_flush_enum flush,
5252                                     bool system_chunk)
5253 {
5254         struct reserve_ticket ticket;
5255         u64 used;
5256         int ret = 0;
5257
5258         ASSERT(orig_bytes);
5259         ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL);
5260
5261         spin_lock(&space_info->lock);
5262         ret = -ENOSPC;
5263         used = btrfs_space_info_used(space_info, true);
5264
5265         /*
5266          * If we have enough space then hooray, make our reservation and carry
5267          * on.  If not see if we can overcommit, and if we can, hooray carry on.
5268          * If not things get more complicated.
5269          */
5270         if (used + orig_bytes <= space_info->total_bytes) {
5271                 space_info->bytes_may_use += orig_bytes;
5272                 trace_btrfs_space_reservation(fs_info, "space_info",
5273                                               space_info->flags, orig_bytes, 1);
5274                 ret = 0;
5275         } else if (can_overcommit(fs_info, space_info, orig_bytes, flush,
5276                                   system_chunk)) {
5277                 space_info->bytes_may_use += orig_bytes;
5278                 trace_btrfs_space_reservation(fs_info, "space_info",
5279                                               space_info->flags, orig_bytes, 1);
5280                 ret = 0;
5281         }
5282
5283         /*
5284          * If we couldn't make a reservation then setup our reservation ticket
5285          * and kick the async worker if it's not already running.
5286          *
5287          * If we are a priority flusher then we just need to add our ticket to
5288          * the list and we will do our own flushing further down.
5289          */
5290         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
5291                 ticket.bytes = orig_bytes;
5292                 ticket.error = 0;
5293                 init_waitqueue_head(&ticket.wait);
5294                 if (flush == BTRFS_RESERVE_FLUSH_ALL) {
5295                         list_add_tail(&ticket.list, &space_info->tickets);
5296                         if (!space_info->flush) {
5297                                 space_info->flush = 1;
5298                                 trace_btrfs_trigger_flush(fs_info,
5299                                                           space_info->flags,
5300                                                           orig_bytes, flush,
5301                                                           "enospc");
5302                                 queue_work(system_unbound_wq,
5303                                            &fs_info->async_reclaim_work);
5304                         }
5305                 } else {
5306                         list_add_tail(&ticket.list,
5307                                       &space_info->priority_tickets);
5308                 }
5309         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
5310                 used += orig_bytes;
5311                 /*
5312                  * We will do the space reservation dance during log replay,
5313                  * which means we won't have fs_info->fs_root set, so don't do
5314                  * the async reclaim as we will panic.
5315                  */
5316                 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) &&
5317                     need_do_async_reclaim(fs_info, space_info,
5318                                           used, system_chunk) &&
5319                     !work_busy(&fs_info->async_reclaim_work)) {
5320                         trace_btrfs_trigger_flush(fs_info, space_info->flags,
5321                                                   orig_bytes, flush, "preempt");
5322                         queue_work(system_unbound_wq,
5323                                    &fs_info->async_reclaim_work);
5324                 }
5325         }
5326         spin_unlock(&space_info->lock);
5327         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
5328                 return ret;
5329
5330         if (flush == BTRFS_RESERVE_FLUSH_ALL)
5331                 return wait_reserve_ticket(fs_info, space_info, &ticket,
5332                                            orig_bytes);
5333
5334         ret = 0;
5335         priority_reclaim_metadata_space(fs_info, space_info, &ticket);
5336         spin_lock(&space_info->lock);
5337         if (ticket.bytes) {
5338                 if (ticket.bytes < orig_bytes) {
5339                         u64 num_bytes = orig_bytes - ticket.bytes;
5340                         space_info->bytes_may_use -= num_bytes;
5341                         trace_btrfs_space_reservation(fs_info, "space_info",
5342                                                       space_info->flags,
5343                                                       num_bytes, 0);
5344
5345                 }
5346                 list_del_init(&ticket.list);
5347                 ret = -ENOSPC;
5348         }
5349         spin_unlock(&space_info->lock);
5350         ASSERT(list_empty(&ticket.list));
5351         return ret;
5352 }
5353
5354 /**
5355  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5356  * @root - the root we're allocating for
5357  * @block_rsv - the block_rsv we're allocating for
5358  * @orig_bytes - the number of bytes we want
5359  * @flush - whether or not we can flush to make our reservation
5360  *
5361  * This will reserve orgi_bytes number of bytes from the space info associated
5362  * with the block_rsv.  If there is not enough space it will make an attempt to
5363  * flush out space to make room.  It will do this by flushing delalloc if
5364  * possible or committing the transaction.  If flush is 0 then no attempts to
5365  * regain reservations will be made and this will fail if there is not enough
5366  * space already.
5367  */
5368 static int reserve_metadata_bytes(struct btrfs_root *root,
5369                                   struct btrfs_block_rsv *block_rsv,
5370                                   u64 orig_bytes,
5371                                   enum btrfs_reserve_flush_enum flush)
5372 {
5373         struct btrfs_fs_info *fs_info = root->fs_info;
5374         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5375         int ret;
5376         bool system_chunk = (root == fs_info->chunk_root);
5377
5378         ret = __reserve_metadata_bytes(fs_info, block_rsv->space_info,
5379                                        orig_bytes, flush, system_chunk);
5380         if (ret == -ENOSPC &&
5381             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
5382                 if (block_rsv != global_rsv &&
5383                     !block_rsv_use_bytes(global_rsv, orig_bytes))
5384                         ret = 0;
5385         }
5386         if (ret == -ENOSPC) {
5387                 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
5388                                               block_rsv->space_info->flags,
5389                                               orig_bytes, 1);
5390
5391                 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
5392                         dump_space_info(fs_info, block_rsv->space_info,
5393                                         orig_bytes, 0);
5394         }
5395         return ret;
5396 }
5397
5398 static struct btrfs_block_rsv *get_block_rsv(
5399                                         const struct btrfs_trans_handle *trans,
5400                                         const struct btrfs_root *root)
5401 {
5402         struct btrfs_fs_info *fs_info = root->fs_info;
5403         struct btrfs_block_rsv *block_rsv = NULL;
5404
5405         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
5406             (root == fs_info->csum_root && trans->adding_csums) ||
5407             (root == fs_info->uuid_root))
5408                 block_rsv = trans->block_rsv;
5409
5410         if (!block_rsv)
5411                 block_rsv = root->block_rsv;
5412
5413         if (!block_rsv)
5414                 block_rsv = &fs_info->empty_block_rsv;
5415
5416         return block_rsv;
5417 }
5418
5419 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
5420                                u64 num_bytes)
5421 {
5422         int ret = -ENOSPC;
5423         spin_lock(&block_rsv->lock);
5424         if (block_rsv->reserved >= num_bytes) {
5425                 block_rsv->reserved -= num_bytes;
5426                 if (block_rsv->reserved < block_rsv->size)
5427                         block_rsv->full = 0;
5428                 ret = 0;
5429         }
5430         spin_unlock(&block_rsv->lock);
5431         return ret;
5432 }
5433
5434 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
5435                                 u64 num_bytes, int update_size)
5436 {
5437         spin_lock(&block_rsv->lock);
5438         block_rsv->reserved += num_bytes;
5439         if (update_size)
5440                 block_rsv->size += num_bytes;
5441         else if (block_rsv->reserved >= block_rsv->size)
5442                 block_rsv->full = 1;
5443         spin_unlock(&block_rsv->lock);
5444 }
5445
5446 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
5447                              struct btrfs_block_rsv *dest, u64 num_bytes,
5448                              int min_factor)
5449 {
5450         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5451         u64 min_bytes;
5452
5453         if (global_rsv->space_info != dest->space_info)
5454                 return -ENOSPC;
5455
5456         spin_lock(&global_rsv->lock);
5457         min_bytes = div_factor(global_rsv->size, min_factor);
5458         if (global_rsv->reserved < min_bytes + num_bytes) {
5459                 spin_unlock(&global_rsv->lock);
5460                 return -ENOSPC;
5461         }
5462         global_rsv->reserved -= num_bytes;
5463         if (global_rsv->reserved < global_rsv->size)
5464                 global_rsv->full = 0;
5465         spin_unlock(&global_rsv->lock);
5466
5467         block_rsv_add_bytes(dest, num_bytes, 1);
5468         return 0;
5469 }
5470
5471 /*
5472  * This is for space we already have accounted in space_info->bytes_may_use, so
5473  * basically when we're returning space from block_rsv's.
5474  */
5475 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
5476                                      struct btrfs_space_info *space_info,
5477                                      u64 num_bytes)
5478 {
5479         struct reserve_ticket *ticket;
5480         struct list_head *head;
5481         u64 used;
5482         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH;
5483         bool check_overcommit = false;
5484
5485         spin_lock(&space_info->lock);
5486         head = &space_info->priority_tickets;
5487
5488         /*
5489          * If we are over our limit then we need to check and see if we can
5490          * overcommit, and if we can't then we just need to free up our space
5491          * and not satisfy any requests.
5492          */
5493         used = btrfs_space_info_used(space_info, true);
5494         if (used - num_bytes >= space_info->total_bytes)
5495                 check_overcommit = true;
5496 again:
5497         while (!list_empty(head) && num_bytes) {
5498                 ticket = list_first_entry(head, struct reserve_ticket,
5499                                           list);
5500                 /*
5501                  * We use 0 bytes because this space is already reserved, so
5502                  * adding the ticket space would be a double count.
5503                  */
5504                 if (check_overcommit &&
5505                     !can_overcommit(fs_info, space_info, 0, flush, false))
5506                         break;
5507                 if (num_bytes >= ticket->bytes) {
5508                         list_del_init(&ticket->list);
5509                         num_bytes -= ticket->bytes;
5510                         ticket->bytes = 0;
5511                         space_info->tickets_id++;
5512                         wake_up(&ticket->wait);
5513                 } else {
5514                         ticket->bytes -= num_bytes;
5515                         num_bytes = 0;
5516                 }
5517         }
5518
5519         if (num_bytes && head == &space_info->priority_tickets) {
5520                 head = &space_info->tickets;
5521                 flush = BTRFS_RESERVE_FLUSH_ALL;
5522                 goto again;
5523         }
5524         space_info->bytes_may_use -= num_bytes;
5525         trace_btrfs_space_reservation(fs_info, "space_info",
5526                                       space_info->flags, num_bytes, 0);
5527         spin_unlock(&space_info->lock);
5528 }
5529
5530 /*
5531  * This is for newly allocated space that isn't accounted in
5532  * space_info->bytes_may_use yet.  So if we allocate a chunk or unpin an extent
5533  * we use this helper.
5534  */
5535 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
5536                                      struct btrfs_space_info *space_info,
5537                                      u64 num_bytes)
5538 {
5539         struct reserve_ticket *ticket;
5540         struct list_head *head = &space_info->priority_tickets;
5541
5542 again:
5543         while (!list_empty(head) && num_bytes) {
5544                 ticket = list_first_entry(head, struct reserve_ticket,
5545                                           list);
5546                 if (num_bytes >= ticket->bytes) {
5547                         trace_btrfs_space_reservation(fs_info, "space_info",
5548                                                       space_info->flags,
5549                                                       ticket->bytes, 1);
5550                         list_del_init(&ticket->list);
5551                         num_bytes -= ticket->bytes;
5552                         space_info->bytes_may_use += ticket->bytes;
5553                         ticket->bytes = 0;
5554                         space_info->tickets_id++;
5555                         wake_up(&ticket->wait);
5556                 } else {
5557                         trace_btrfs_space_reservation(fs_info, "space_info",
5558                                                       space_info->flags,
5559                                                       num_bytes, 1);
5560                         space_info->bytes_may_use += num_bytes;
5561                         ticket->bytes -= num_bytes;
5562                         num_bytes = 0;
5563                 }
5564         }
5565
5566         if (num_bytes && head == &space_info->priority_tickets) {
5567                 head = &space_info->tickets;
5568                 goto again;
5569         }
5570 }
5571
5572 static u64 block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
5573                                     struct btrfs_block_rsv *block_rsv,
5574                                     struct btrfs_block_rsv *dest, u64 num_bytes)
5575 {
5576         struct btrfs_space_info *space_info = block_rsv->space_info;
5577         u64 ret;
5578
5579         spin_lock(&block_rsv->lock);
5580         if (num_bytes == (u64)-1)
5581                 num_bytes = block_rsv->size;
5582         block_rsv->size -= num_bytes;
5583         if (block_rsv->reserved >= block_rsv->size) {
5584                 num_bytes = block_rsv->reserved - block_rsv->size;
5585                 block_rsv->reserved = block_rsv->size;
5586                 block_rsv->full = 1;
5587         } else {
5588                 num_bytes = 0;
5589         }
5590         spin_unlock(&block_rsv->lock);
5591
5592         ret = num_bytes;
5593         if (num_bytes > 0) {
5594                 if (dest) {
5595                         spin_lock(&dest->lock);
5596                         if (!dest->full) {
5597                                 u64 bytes_to_add;
5598
5599                                 bytes_to_add = dest->size - dest->reserved;
5600                                 bytes_to_add = min(num_bytes, bytes_to_add);
5601                                 dest->reserved += bytes_to_add;
5602                                 if (dest->reserved >= dest->size)
5603                                         dest->full = 1;
5604                                 num_bytes -= bytes_to_add;
5605                         }
5606                         spin_unlock(&dest->lock);
5607                 }
5608                 if (num_bytes)
5609                         space_info_add_old_bytes(fs_info, space_info,
5610                                                  num_bytes);
5611         }
5612         return ret;
5613 }
5614
5615 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src,
5616                             struct btrfs_block_rsv *dst, u64 num_bytes,
5617                             int update_size)
5618 {
5619         int ret;
5620
5621         ret = block_rsv_use_bytes(src, num_bytes);
5622         if (ret)
5623                 return ret;
5624
5625         block_rsv_add_bytes(dst, num_bytes, update_size);
5626         return 0;
5627 }
5628
5629 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
5630 {
5631         memset(rsv, 0, sizeof(*rsv));
5632         spin_lock_init(&rsv->lock);
5633         rsv->type = type;
5634 }
5635
5636 void btrfs_init_metadata_block_rsv(struct btrfs_fs_info *fs_info,
5637                                    struct btrfs_block_rsv *rsv,
5638                                    unsigned short type)
5639 {
5640         btrfs_init_block_rsv(rsv, type);
5641         rsv->space_info = __find_space_info(fs_info,
5642                                             BTRFS_BLOCK_GROUP_METADATA);
5643 }
5644
5645 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_fs_info *fs_info,
5646                                               unsigned short type)
5647 {
5648         struct btrfs_block_rsv *block_rsv;
5649
5650         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
5651         if (!block_rsv)
5652                 return NULL;
5653
5654         btrfs_init_metadata_block_rsv(fs_info, block_rsv, type);
5655         return block_rsv;
5656 }
5657
5658 void btrfs_free_block_rsv(struct btrfs_fs_info *fs_info,
5659                           struct btrfs_block_rsv *rsv)
5660 {
5661         if (!rsv)
5662                 return;
5663         btrfs_block_rsv_release(fs_info, rsv, (u64)-1);
5664         kfree(rsv);
5665 }
5666
5667 void __btrfs_free_block_rsv(struct btrfs_block_rsv *rsv)
5668 {
5669         kfree(rsv);
5670 }
5671
5672 int btrfs_block_rsv_add(struct btrfs_root *root,
5673                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
5674                         enum btrfs_reserve_flush_enum flush)
5675 {
5676         int ret;
5677
5678         if (num_bytes == 0)
5679                 return 0;
5680
5681         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5682         if (!ret) {
5683                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
5684                 return 0;
5685         }
5686
5687         return ret;
5688 }
5689
5690 int btrfs_block_rsv_check(struct btrfs_block_rsv *block_rsv, int min_factor)
5691 {
5692         u64 num_bytes = 0;
5693         int ret = -ENOSPC;
5694
5695         if (!block_rsv)
5696                 return 0;
5697
5698         spin_lock(&block_rsv->lock);
5699         num_bytes = div_factor(block_rsv->size, min_factor);
5700         if (block_rsv->reserved >= num_bytes)
5701                 ret = 0;
5702         spin_unlock(&block_rsv->lock);
5703
5704         return ret;
5705 }
5706
5707 int btrfs_block_rsv_refill(struct btrfs_root *root,
5708                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
5709                            enum btrfs_reserve_flush_enum flush)
5710 {
5711         u64 num_bytes = 0;
5712         int ret = -ENOSPC;
5713
5714         if (!block_rsv)
5715                 return 0;
5716
5717         spin_lock(&block_rsv->lock);
5718         num_bytes = min_reserved;
5719         if (block_rsv->reserved >= num_bytes)
5720                 ret = 0;
5721         else
5722                 num_bytes -= block_rsv->reserved;
5723         spin_unlock(&block_rsv->lock);
5724
5725         if (!ret)
5726                 return 0;
5727
5728         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5729         if (!ret) {
5730                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
5731                 return 0;
5732         }
5733
5734         return ret;
5735 }
5736
5737 /**
5738  * btrfs_inode_rsv_refill - refill the inode block rsv.
5739  * @inode - the inode we are refilling.
5740  * @flush - the flusing restriction.
5741  *
5742  * Essentially the same as btrfs_block_rsv_refill, except it uses the
5743  * block_rsv->size as the minimum size.  We'll either refill the missing amount
5744  * or return if we already have enough space.  This will also handle the resreve
5745  * tracepoint for the reserved amount.
5746  */
5747 static int btrfs_inode_rsv_refill(struct btrfs_inode *inode,
5748                                   enum btrfs_reserve_flush_enum flush)
5749 {
5750         struct btrfs_root *root = inode->root;
5751         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
5752         u64 num_bytes = 0;
5753         int ret = -ENOSPC;
5754
5755         spin_lock(&block_rsv->lock);
5756         if (block_rsv->reserved < block_rsv->size)
5757                 num_bytes = block_rsv->size - block_rsv->reserved;
5758         spin_unlock(&block_rsv->lock);
5759
5760         if (num_bytes == 0)
5761                 return 0;
5762
5763         ret = btrfs_qgroup_reserve_meta_prealloc(root, num_bytes, true);
5764         if (ret)
5765                 return ret;
5766         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5767         if (!ret) {
5768                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
5769                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5770                                               btrfs_ino(inode), num_bytes, 1);
5771         }
5772         return ret;
5773 }
5774
5775 /**
5776  * btrfs_inode_rsv_release - release any excessive reservation.
5777  * @inode - the inode we need to release from.
5778  * @qgroup_free - free or convert qgroup meta.
5779  *   Unlike normal operation, qgroup meta reservation needs to know if we are
5780  *   freeing qgroup reservation or just converting it into per-trans.  Normally
5781  *   @qgroup_free is true for error handling, and false for normal release.
5782  *
5783  * This is the same as btrfs_block_rsv_release, except that it handles the
5784  * tracepoint for the reservation.
5785  */
5786 static void btrfs_inode_rsv_release(struct btrfs_inode *inode, bool qgroup_free)
5787 {
5788         struct btrfs_fs_info *fs_info = inode->root->fs_info;
5789         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5790         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
5791         u64 released = 0;
5792
5793         /*
5794          * Since we statically set the block_rsv->size we just want to say we
5795          * are releasing 0 bytes, and then we'll just get the reservation over
5796          * the size free'd.
5797          */
5798         released = block_rsv_release_bytes(fs_info, block_rsv, global_rsv, 0);
5799         if (released > 0)
5800                 trace_btrfs_space_reservation(fs_info, "delalloc",
5801                                               btrfs_ino(inode), released, 0);
5802         if (qgroup_free)
5803                 btrfs_qgroup_free_meta_prealloc(inode->root, released);
5804         else
5805                 btrfs_qgroup_convert_reserved_meta(inode->root, released);
5806 }
5807
5808 void btrfs_block_rsv_release(struct btrfs_fs_info *fs_info,
5809                              struct btrfs_block_rsv *block_rsv,
5810                              u64 num_bytes)
5811 {
5812         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5813
5814         if (global_rsv == block_rsv ||
5815             block_rsv->space_info != global_rsv->space_info)
5816                 global_rsv = NULL;
5817         block_rsv_release_bytes(fs_info, block_rsv, global_rsv, num_bytes);
5818 }
5819
5820 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
5821 {
5822         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
5823         struct btrfs_space_info *sinfo = block_rsv->space_info;
5824         u64 num_bytes;
5825
5826         /*
5827          * The global block rsv is based on the size of the extent tree, the
5828          * checksum tree and the root tree.  If the fs is empty we want to set
5829          * it to a minimal amount for safety.
5830          */
5831         num_bytes = btrfs_root_used(&fs_info->extent_root->root_item) +
5832                 btrfs_root_used(&fs_info->csum_root->root_item) +
5833                 btrfs_root_used(&fs_info->tree_root->root_item);
5834         num_bytes = max_t(u64, num_bytes, SZ_16M);
5835
5836         spin_lock(&sinfo->lock);
5837         spin_lock(&block_rsv->lock);
5838
5839         block_rsv->size = min_t(u64, num_bytes, SZ_512M);
5840
5841         if (block_rsv->reserved < block_rsv->size) {
5842                 num_bytes = btrfs_space_info_used(sinfo, true);
5843                 if (sinfo->total_bytes > num_bytes) {
5844                         num_bytes = sinfo->total_bytes - num_bytes;
5845                         num_bytes = min(num_bytes,
5846                                         block_rsv->size - block_rsv->reserved);
5847                         block_rsv->reserved += num_bytes;
5848                         sinfo->bytes_may_use += num_bytes;
5849                         trace_btrfs_space_reservation(fs_info, "space_info",
5850                                                       sinfo->flags, num_bytes,
5851                                                       1);
5852                 }
5853         } else if (block_rsv->reserved > block_rsv->size) {
5854                 num_bytes = block_rsv->reserved - block_rsv->size;
5855                 sinfo->bytes_may_use -= num_bytes;
5856                 trace_btrfs_space_reservation(fs_info, "space_info",
5857                                       sinfo->flags, num_bytes, 0);
5858                 block_rsv->reserved = block_rsv->size;
5859         }
5860
5861         if (block_rsv->reserved == block_rsv->size)
5862                 block_rsv->full = 1;
5863         else
5864                 block_rsv->full = 0;
5865
5866         spin_unlock(&block_rsv->lock);
5867         spin_unlock(&sinfo->lock);
5868 }
5869
5870 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
5871 {
5872         struct btrfs_space_info *space_info;
5873
5874         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
5875         fs_info->chunk_block_rsv.space_info = space_info;
5876
5877         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5878         fs_info->global_block_rsv.space_info = space_info;
5879         fs_info->trans_block_rsv.space_info = space_info;
5880         fs_info->empty_block_rsv.space_info = space_info;
5881         fs_info->delayed_block_rsv.space_info = space_info;
5882
5883         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
5884         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
5885         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
5886         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
5887         if (fs_info->quota_root)
5888                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
5889         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
5890
5891         update_global_block_rsv(fs_info);
5892 }
5893
5894 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
5895 {
5896         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
5897                                 (u64)-1);
5898         WARN_ON(fs_info->trans_block_rsv.size > 0);
5899         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
5900         WARN_ON(fs_info->chunk_block_rsv.size > 0);
5901         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
5902         WARN_ON(fs_info->delayed_block_rsv.size > 0);
5903         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
5904 }
5905
5906
5907 /*
5908  * To be called after all the new block groups attached to the transaction
5909  * handle have been created (btrfs_create_pending_block_groups()).
5910  */
5911 void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans)
5912 {
5913         struct btrfs_fs_info *fs_info = trans->fs_info;
5914
5915         if (!trans->chunk_bytes_reserved)
5916                 return;
5917
5918         WARN_ON_ONCE(!list_empty(&trans->new_bgs));
5919
5920         block_rsv_release_bytes(fs_info, &fs_info->chunk_block_rsv, NULL,
5921                                 trans->chunk_bytes_reserved);
5922         trans->chunk_bytes_reserved = 0;
5923 }
5924
5925 /* Can only return 0 or -ENOSPC */
5926 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
5927                                   struct btrfs_inode *inode)
5928 {
5929         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
5930         struct btrfs_root *root = inode->root;
5931         /*
5932          * We always use trans->block_rsv here as we will have reserved space
5933          * for our orphan when starting the transaction, using get_block_rsv()
5934          * here will sometimes make us choose the wrong block rsv as we could be
5935          * doing a reloc inode for a non refcounted root.
5936          */
5937         struct btrfs_block_rsv *src_rsv = trans->block_rsv;
5938         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
5939
5940         /*
5941          * We need to hold space in order to delete our orphan item once we've
5942          * added it, so this takes the reservation so we can release it later
5943          * when we are truly done with the orphan item.
5944          */
5945         u64 num_bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
5946
5947         trace_btrfs_space_reservation(fs_info, "orphan", btrfs_ino(inode),
5948                         num_bytes, 1);
5949         return btrfs_block_rsv_migrate(src_rsv, dst_rsv, num_bytes, 1);
5950 }
5951
5952 void btrfs_orphan_release_metadata(struct btrfs_inode *inode)
5953 {
5954         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
5955         struct btrfs_root *root = inode->root;
5956         u64 num_bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
5957
5958         trace_btrfs_space_reservation(fs_info, "orphan", btrfs_ino(inode),
5959                         num_bytes, 0);
5960         btrfs_block_rsv_release(fs_info, root->orphan_block_rsv, num_bytes);
5961 }
5962
5963 /*
5964  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
5965  * root: the root of the parent directory
5966  * rsv: block reservation
5967  * items: the number of items that we need do reservation
5968  * qgroup_reserved: used to return the reserved size in qgroup
5969  *
5970  * This function is used to reserve the space for snapshot/subvolume
5971  * creation and deletion. Those operations are different with the
5972  * common file/directory operations, they change two fs/file trees
5973  * and root tree, the number of items that the qgroup reserves is
5974  * different with the free space reservation. So we can not use
5975  * the space reservation mechanism in start_transaction().
5976  */
5977 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
5978                                      struct btrfs_block_rsv *rsv,
5979                                      int items,
5980                                      u64 *qgroup_reserved,
5981                                      bool use_global_rsv)
5982 {
5983         u64 num_bytes;
5984         int ret;
5985         struct btrfs_fs_info *fs_info = root->fs_info;
5986         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5987
5988         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)) {
5989                 /* One for parent inode, two for dir entries */
5990                 num_bytes = 3 * fs_info->nodesize;
5991                 ret = btrfs_qgroup_reserve_meta_prealloc(root, num_bytes, true);
5992                 if (ret)
5993                         return ret;
5994         } else {
5995                 num_bytes = 0;
5996         }
5997
5998         *qgroup_reserved = num_bytes;
5999
6000         num_bytes = btrfs_calc_trans_metadata_size(fs_info, items);
6001         rsv->space_info = __find_space_info(fs_info,
6002                                             BTRFS_BLOCK_GROUP_METADATA);
6003         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
6004                                   BTRFS_RESERVE_FLUSH_ALL);
6005
6006         if (ret == -ENOSPC && use_global_rsv)
6007                 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes, 1);
6008
6009         if (ret && *qgroup_reserved)
6010                 btrfs_qgroup_free_meta_prealloc(root, *qgroup_reserved);
6011
6012         return ret;
6013 }
6014
6015 void btrfs_subvolume_release_metadata(struct btrfs_fs_info *fs_info,
6016                                       struct btrfs_block_rsv *rsv)
6017 {
6018         btrfs_block_rsv_release(fs_info, rsv, (u64)-1);
6019 }
6020
6021 static void btrfs_calculate_inode_block_rsv_size(struct btrfs_fs_info *fs_info,
6022                                                  struct btrfs_inode *inode)
6023 {
6024         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
6025         u64 reserve_size = 0;
6026         u64 csum_leaves;
6027         unsigned outstanding_extents;
6028
6029         lockdep_assert_held(&inode->lock);
6030         outstanding_extents = inode->outstanding_extents;
6031         if (outstanding_extents)
6032                 reserve_size = btrfs_calc_trans_metadata_size(fs_info,
6033                                                 outstanding_extents + 1);
6034         csum_leaves = btrfs_csum_bytes_to_leaves(fs_info,
6035                                                  inode->csum_bytes);
6036         reserve_size += btrfs_calc_trans_metadata_size(fs_info,
6037                                                        csum_leaves);
6038
6039         spin_lock(&block_rsv->lock);
6040         block_rsv->size = reserve_size;
6041         spin_unlock(&block_rsv->lock);
6042 }
6043
6044 int btrfs_delalloc_reserve_metadata(struct btrfs_inode *inode, u64 num_bytes)
6045 {
6046         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
6047         unsigned nr_extents;
6048         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
6049         int ret = 0;
6050         bool delalloc_lock = true;
6051
6052         /* If we are a free space inode we need to not flush since we will be in
6053          * the middle of a transaction commit.  We also don't need the delalloc
6054          * mutex since we won't race with anybody.  We need this mostly to make
6055          * lockdep shut its filthy mouth.
6056          *
6057          * If we have a transaction open (can happen if we call truncate_block
6058          * from truncate), then we need FLUSH_LIMIT so we don't deadlock.
6059          */
6060         if (btrfs_is_free_space_inode(inode)) {
6061                 flush = BTRFS_RESERVE_NO_FLUSH;
6062                 delalloc_lock = false;
6063         } else {
6064                 if (current->journal_info)
6065                         flush = BTRFS_RESERVE_FLUSH_LIMIT;
6066
6067                 if (btrfs_transaction_in_commit(fs_info))
6068                         schedule_timeout(1);
6069         }
6070
6071         if (delalloc_lock)
6072                 mutex_lock(&inode->delalloc_mutex);
6073
6074         num_bytes = ALIGN(num_bytes, fs_info->sectorsize);
6075
6076         /* Add our new extents and calculate the new rsv size. */
6077         spin_lock(&inode->lock);
6078         nr_extents = count_max_extents(num_bytes);
6079         btrfs_mod_outstanding_extents(inode, nr_extents);
6080         inode->csum_bytes += num_bytes;
6081         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6082         spin_unlock(&inode->lock);
6083
6084         ret = btrfs_inode_rsv_refill(inode, flush);
6085         if (unlikely(ret))
6086                 goto out_fail;
6087
6088         if (delalloc_lock)
6089                 mutex_unlock(&inode->delalloc_mutex);
6090         return 0;
6091
6092 out_fail:
6093         spin_lock(&inode->lock);
6094         nr_extents = count_max_extents(num_bytes);
6095         btrfs_mod_outstanding_extents(inode, -nr_extents);
6096         inode->csum_bytes -= num_bytes;
6097         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6098         spin_unlock(&inode->lock);
6099
6100         btrfs_inode_rsv_release(inode, true);
6101         if (delalloc_lock)
6102                 mutex_unlock(&inode->delalloc_mutex);
6103         return ret;
6104 }
6105
6106 /**
6107  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
6108  * @inode: the inode to release the reservation for.
6109  * @num_bytes: the number of bytes we are releasing.
6110  * @qgroup_free: free qgroup reservation or convert it to per-trans reservation
6111  *
6112  * This will release the metadata reservation for an inode.  This can be called
6113  * once we complete IO for a given set of bytes to release their metadata
6114  * reservations, or on error for the same reason.
6115  */
6116 void btrfs_delalloc_release_metadata(struct btrfs_inode *inode, u64 num_bytes,
6117                                      bool qgroup_free)
6118 {
6119         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
6120
6121         num_bytes = ALIGN(num_bytes, fs_info->sectorsize);
6122         spin_lock(&inode->lock);
6123         inode->csum_bytes -= num_bytes;
6124         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6125         spin_unlock(&inode->lock);
6126
6127         if (btrfs_is_testing(fs_info))
6128                 return;
6129
6130         btrfs_inode_rsv_release(inode, qgroup_free);
6131 }
6132
6133 /**
6134  * btrfs_delalloc_release_extents - release our outstanding_extents
6135  * @inode: the inode to balance the reservation for.
6136  * @num_bytes: the number of bytes we originally reserved with
6137  * @qgroup_free: do we need to free qgroup meta reservation or convert them.
6138  *
6139  * When we reserve space we increase outstanding_extents for the extents we may
6140  * add.  Once we've set the range as delalloc or created our ordered extents we
6141  * have outstanding_extents to track the real usage, so we use this to free our
6142  * temporarily tracked outstanding_extents.  This _must_ be used in conjunction
6143  * with btrfs_delalloc_reserve_metadata.
6144  */
6145 void btrfs_delalloc_release_extents(struct btrfs_inode *inode, u64 num_bytes,
6146                                     bool qgroup_free)
6147 {
6148         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
6149         unsigned num_extents;
6150
6151         spin_lock(&inode->lock);
6152         num_extents = count_max_extents(num_bytes);
6153         btrfs_mod_outstanding_extents(inode, -num_extents);
6154         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6155         spin_unlock(&inode->lock);
6156
6157         if (btrfs_is_testing(fs_info))
6158                 return;
6159
6160         btrfs_inode_rsv_release(inode, qgroup_free);
6161 }
6162
6163 /**
6164  * btrfs_delalloc_reserve_space - reserve data and metadata space for
6165  * delalloc
6166  * @inode: inode we're writing to
6167  * @start: start range we are writing to
6168  * @len: how long the range we are writing to
6169  * @reserved: mandatory parameter, record actually reserved qgroup ranges of
6170  *            current reservation.
6171  *
6172  * This will do the following things
6173  *
6174  * o reserve space in data space info for num bytes
6175  *   and reserve precious corresponding qgroup space
6176  *   (Done in check_data_free_space)
6177  *
6178  * o reserve space for metadata space, based on the number of outstanding
6179  *   extents and how much csums will be needed
6180  *   also reserve metadata space in a per root over-reserve method.
6181  * o add to the inodes->delalloc_bytes
6182  * o add it to the fs_info's delalloc inodes list.
6183  *   (Above 3 all done in delalloc_reserve_metadata)
6184  *
6185  * Return 0 for success
6186  * Return <0 for error(-ENOSPC or -EQUOT)
6187  */
6188 int btrfs_delalloc_reserve_space(struct inode *inode,
6189                         struct extent_changeset **reserved, u64 start, u64 len)
6190 {
6191         int ret;
6192
6193         ret = btrfs_check_data_free_space(inode, reserved, start, len);
6194         if (ret < 0)
6195                 return ret;
6196         ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode), len);
6197         if (ret < 0)
6198                 btrfs_free_reserved_data_space(inode, *reserved, start, len);
6199         return ret;
6200 }
6201
6202 /**
6203  * btrfs_delalloc_release_space - release data and metadata space for delalloc
6204  * @inode: inode we're releasing space for
6205  * @start: start position of the space already reserved
6206  * @len: the len of the space already reserved
6207  * @release_bytes: the len of the space we consumed or didn't use
6208  *
6209  * This function will release the metadata space that was not used and will
6210  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
6211  * list if there are no delalloc bytes left.
6212  * Also it will handle the qgroup reserved space.
6213  */
6214 void btrfs_delalloc_release_space(struct inode *inode,
6215                                   struct extent_changeset *reserved,
6216                                   u64 start, u64 len, bool qgroup_free)
6217 {
6218         btrfs_delalloc_release_metadata(BTRFS_I(inode), len, qgroup_free);
6219         btrfs_free_reserved_data_space(inode, reserved, start, len);
6220 }
6221
6222 static int update_block_group(struct btrfs_trans_handle *trans,
6223                               struct btrfs_fs_info *info, u64 bytenr,
6224                               u64 num_bytes, int alloc)
6225 {
6226         struct btrfs_block_group_cache *cache = NULL;
6227         u64 total = num_bytes;
6228         u64 old_val;
6229         u64 byte_in_group;
6230         int factor;
6231
6232         /* block accounting for super block */
6233         spin_lock(&info->delalloc_root_lock);
6234         old_val = btrfs_super_bytes_used(info->super_copy);
6235         if (alloc)
6236                 old_val += num_bytes;
6237         else
6238                 old_val -= num_bytes;
6239         btrfs_set_super_bytes_used(info->super_copy, old_val);
6240         spin_unlock(&info->delalloc_root_lock);
6241
6242         while (total) {
6243                 cache = btrfs_lookup_block_group(info, bytenr);
6244                 if (!cache)
6245                         return -ENOENT;
6246                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
6247                                     BTRFS_BLOCK_GROUP_RAID1 |
6248                                     BTRFS_BLOCK_GROUP_RAID10))
6249                         factor = 2;
6250                 else
6251                         factor = 1;
6252                 /*
6253                  * If this block group has free space cache written out, we
6254                  * need to make sure to load it if we are removing space.  This
6255                  * is because we need the unpinning stage to actually add the
6256                  * space back to the block group, otherwise we will leak space.
6257                  */
6258                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
6259                         cache_block_group(cache, 1);
6260
6261                 byte_in_group = bytenr - cache->key.objectid;
6262                 WARN_ON(byte_in_group > cache->key.offset);
6263
6264                 spin_lock(&cache->space_info->lock);
6265                 spin_lock(&cache->lock);
6266
6267                 if (btrfs_test_opt(info, SPACE_CACHE) &&
6268                     cache->disk_cache_state < BTRFS_DC_CLEAR)
6269                         cache->disk_cache_state = BTRFS_DC_CLEAR;
6270
6271                 old_val = btrfs_block_group_used(&cache->item);
6272                 num_bytes = min(total, cache->key.offset - byte_in_group);
6273                 if (alloc) {
6274                         old_val += num_bytes;
6275                         btrfs_set_block_group_used(&cache->item, old_val);
6276                         cache->reserved -= num_bytes;
6277                         cache->space_info->bytes_reserved -= num_bytes;
6278                         cache->space_info->bytes_used += num_bytes;
6279                         cache->space_info->disk_used += num_bytes * factor;
6280                         spin_unlock(&cache->lock);
6281                         spin_unlock(&cache->space_info->lock);
6282                 } else {
6283                         old_val -= num_bytes;
6284                         btrfs_set_block_group_used(&cache->item, old_val);
6285                         cache->pinned += num_bytes;
6286                         cache->space_info->bytes_pinned += num_bytes;
6287                         cache->space_info->bytes_used -= num_bytes;
6288                         cache->space_info->disk_used -= num_bytes * factor;
6289                         spin_unlock(&cache->lock);
6290                         spin_unlock(&cache->space_info->lock);
6291
6292                         trace_btrfs_space_reservation(info, "pinned",
6293                                                       cache->space_info->flags,
6294                                                       num_bytes, 1);
6295                         percpu_counter_add(&cache->space_info->total_bytes_pinned,
6296                                            num_bytes);
6297                         set_extent_dirty(info->pinned_extents,
6298                                          bytenr, bytenr + num_bytes - 1,
6299                                          GFP_NOFS | __GFP_NOFAIL);
6300                 }
6301
6302                 spin_lock(&trans->transaction->dirty_bgs_lock);
6303                 if (list_empty(&cache->dirty_list)) {
6304                         list_add_tail(&cache->dirty_list,
6305                                       &trans->transaction->dirty_bgs);
6306                                 trans->transaction->num_dirty_bgs++;
6307                         btrfs_get_block_group(cache);
6308                 }
6309                 spin_unlock(&trans->transaction->dirty_bgs_lock);
6310
6311                 /*
6312                  * No longer have used bytes in this block group, queue it for
6313                  * deletion. We do this after adding the block group to the
6314                  * dirty list to avoid races between cleaner kthread and space
6315                  * cache writeout.
6316                  */
6317                 if (!alloc && old_val == 0) {
6318                         spin_lock(&info->unused_bgs_lock);
6319                         if (list_empty(&cache->bg_list)) {
6320                                 btrfs_get_block_group(cache);
6321                                 list_add_tail(&cache->bg_list,
6322                                               &info->unused_bgs);
6323                         }
6324                         spin_unlock(&info->unused_bgs_lock);
6325                 }
6326
6327                 btrfs_put_block_group(cache);
6328                 total -= num_bytes;
6329                 bytenr += num_bytes;
6330         }
6331         return 0;
6332 }
6333
6334 static u64 first_logical_byte(struct btrfs_fs_info *fs_info, u64 search_start)
6335 {
6336         struct btrfs_block_group_cache *cache;
6337         u64 bytenr;
6338
6339         spin_lock(&fs_info->block_group_cache_lock);
6340         bytenr = fs_info->first_logical_byte;
6341         spin_unlock(&fs_info->block_group_cache_lock);
6342
6343         if (bytenr < (u64)-1)
6344                 return bytenr;
6345
6346         cache = btrfs_lookup_first_block_group(fs_info, search_start);
6347         if (!cache)
6348                 return 0;
6349
6350         bytenr = cache->key.objectid;
6351         btrfs_put_block_group(cache);
6352
6353         return bytenr;
6354 }
6355
6356 static int pin_down_extent(struct btrfs_fs_info *fs_info,
6357                            struct btrfs_block_group_cache *cache,
6358                            u64 bytenr, u64 num_bytes, int reserved)
6359 {
6360         spin_lock(&cache->space_info->lock);
6361         spin_lock(&cache->lock);
6362         cache->pinned += num_bytes;
6363         cache->space_info->bytes_pinned += num_bytes;
6364         if (reserved) {
6365                 cache->reserved -= num_bytes;
6366                 cache->space_info->bytes_reserved -= num_bytes;
6367         }
6368         spin_unlock(&cache->lock);
6369         spin_unlock(&cache->space_info->lock);
6370
6371         trace_btrfs_space_reservation(fs_info, "pinned",
6372                                       cache->space_info->flags, num_bytes, 1);
6373         percpu_counter_add(&cache->space_info->total_bytes_pinned, num_bytes);
6374         set_extent_dirty(fs_info->pinned_extents, bytenr,
6375                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
6376         return 0;
6377 }
6378
6379 /*
6380  * this function must be called within transaction
6381  */
6382 int btrfs_pin_extent(struct btrfs_fs_info *fs_info,
6383                      u64 bytenr, u64 num_bytes, int reserved)
6384 {
6385         struct btrfs_block_group_cache *cache;
6386
6387         cache = btrfs_lookup_block_group(fs_info, bytenr);
6388         BUG_ON(!cache); /* Logic error */
6389
6390         pin_down_extent(fs_info, cache, bytenr, num_bytes, reserved);
6391
6392         btrfs_put_block_group(cache);
6393         return 0;
6394 }
6395
6396 /*
6397  * this function must be called within transaction
6398  */
6399 int btrfs_pin_extent_for_log_replay(struct btrfs_fs_info *fs_info,
6400                                     u64 bytenr, u64 num_bytes)
6401 {
6402         struct btrfs_block_group_cache *cache;
6403         int ret;
6404
6405         cache = btrfs_lookup_block_group(fs_info, bytenr);
6406         if (!cache)
6407                 return -EINVAL;
6408
6409         /*
6410          * pull in the free space cache (if any) so that our pin
6411          * removes the free space from the cache.  We have load_only set
6412          * to one because the slow code to read in the free extents does check
6413          * the pinned extents.
6414          */
6415         cache_block_group(cache, 1);
6416
6417         pin_down_extent(fs_info, cache, bytenr, num_bytes, 0);
6418
6419         /* remove us from the free space cache (if we're there at all) */
6420         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
6421         btrfs_put_block_group(cache);
6422         return ret;
6423 }
6424
6425 static int __exclude_logged_extent(struct btrfs_fs_info *fs_info,
6426                                    u64 start, u64 num_bytes)
6427 {
6428         int ret;
6429         struct btrfs_block_group_cache *block_group;
6430         struct btrfs_caching_control *caching_ctl;
6431
6432         block_group = btrfs_lookup_block_group(fs_info, start);
6433         if (!block_group)
6434                 return -EINVAL;
6435
6436         cache_block_group(block_group, 0);
6437         caching_ctl = get_caching_control(block_group);
6438
6439         if (!caching_ctl) {
6440                 /* Logic error */
6441                 BUG_ON(!block_group_cache_done(block_group));
6442                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6443         } else {
6444                 mutex_lock(&caching_ctl->mutex);
6445
6446                 if (start >= caching_ctl->progress) {
6447                         ret = add_excluded_extent(fs_info, start, num_bytes);
6448                 } else if (start + num_bytes <= caching_ctl->progress) {
6449                         ret = btrfs_remove_free_space(block_group,
6450                                                       start, num_bytes);
6451                 } else {
6452                         num_bytes = caching_ctl->progress - start;
6453                         ret = btrfs_remove_free_space(block_group,
6454                                                       start, num_bytes);
6455                         if (ret)
6456                                 goto out_lock;
6457
6458                         num_bytes = (start + num_bytes) -
6459                                 caching_ctl->progress;
6460                         start = caching_ctl->progress;
6461                         ret = add_excluded_extent(fs_info, start, num_bytes);
6462                 }
6463 out_lock:
6464                 mutex_unlock(&caching_ctl->mutex);
6465                 put_caching_control(caching_ctl);
6466         }
6467         btrfs_put_block_group(block_group);
6468         return ret;
6469 }
6470
6471 int btrfs_exclude_logged_extents(struct btrfs_fs_info *fs_info,
6472                                  struct extent_buffer *eb)
6473 {
6474         struct btrfs_file_extent_item *item;
6475         struct btrfs_key key;
6476         int found_type;
6477         int i;
6478
6479         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS))
6480                 return 0;
6481
6482         for (i = 0; i < btrfs_header_nritems(eb); i++) {
6483                 btrfs_item_key_to_cpu(eb, &key, i);
6484                 if (key.type != BTRFS_EXTENT_DATA_KEY)
6485                         continue;
6486                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
6487                 found_type = btrfs_file_extent_type(eb, item);
6488                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
6489                         continue;
6490                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
6491                         continue;
6492                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
6493                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
6494                 __exclude_logged_extent(fs_info, key.objectid, key.offset);
6495         }
6496
6497         return 0;
6498 }
6499
6500 static void
6501 btrfs_inc_block_group_reservations(struct btrfs_block_group_cache *bg)
6502 {
6503         atomic_inc(&bg->reservations);
6504 }
6505
6506 void btrfs_dec_block_group_reservations(struct btrfs_fs_info *fs_info,
6507                                         const u64 start)
6508 {
6509         struct btrfs_block_group_cache *bg;
6510
6511         bg = btrfs_lookup_block_group(fs_info, start);
6512         ASSERT(bg);
6513         if (atomic_dec_and_test(&bg->reservations))
6514                 wake_up_atomic_t(&bg->reservations);
6515         btrfs_put_block_group(bg);
6516 }
6517
6518 void btrfs_wait_block_group_reservations(struct btrfs_block_group_cache *bg)
6519 {
6520         struct btrfs_space_info *space_info = bg->space_info;
6521
6522         ASSERT(bg->ro);
6523
6524         if (!(bg->flags & BTRFS_BLOCK_GROUP_DATA))
6525                 return;
6526
6527         /*
6528          * Our block group is read only but before we set it to read only,
6529          * some task might have had allocated an extent from it already, but it
6530          * has not yet created a respective ordered extent (and added it to a
6531          * root's list of ordered extents).
6532          * Therefore wait for any task currently allocating extents, since the
6533          * block group's reservations counter is incremented while a read lock
6534          * on the groups' semaphore is held and decremented after releasing
6535          * the read access on that semaphore and creating the ordered extent.
6536          */
6537         down_write(&space_info->groups_sem);
6538         up_write(&space_info->groups_sem);
6539
6540         wait_on_atomic_t(&bg->reservations, atomic_t_wait,
6541                          TASK_UNINTERRUPTIBLE);
6542 }
6543
6544 /**
6545  * btrfs_add_reserved_bytes - update the block_group and space info counters
6546  * @cache:      The cache we are manipulating
6547  * @ram_bytes:  The number of bytes of file content, and will be same to
6548  *              @num_bytes except for the compress path.
6549  * @num_bytes:  The number of bytes in question
6550  * @delalloc:   The blocks are allocated for the delalloc write
6551  *
6552  * This is called by the allocator when it reserves space. If this is a
6553  * reservation and the block group has become read only we cannot make the
6554  * reservation and return -EAGAIN, otherwise this function always succeeds.
6555  */
6556 static int btrfs_add_reserved_bytes(struct btrfs_block_group_cache *cache,
6557                                     u64 ram_bytes, u64 num_bytes, int delalloc)
6558 {
6559         struct btrfs_space_info *space_info = cache->space_info;
6560         int ret = 0;
6561
6562         spin_lock(&space_info->lock);
6563         spin_lock(&cache->lock);
6564         if (cache->ro) {
6565                 ret = -EAGAIN;
6566         } else {
6567                 cache->reserved += num_bytes;
6568                 space_info->bytes_reserved += num_bytes;
6569
6570                 trace_btrfs_space_reservation(cache->fs_info,
6571                                 "space_info", space_info->flags,
6572                                 ram_bytes, 0);
6573                 space_info->bytes_may_use -= ram_bytes;
6574                 if (delalloc)
6575                         cache->delalloc_bytes += num_bytes;
6576         }
6577         spin_unlock(&cache->lock);
6578         spin_unlock(&space_info->lock);
6579         return ret;
6580 }
6581
6582 /**
6583  * btrfs_free_reserved_bytes - update the block_group and space info counters
6584  * @cache:      The cache we are manipulating
6585  * @num_bytes:  The number of bytes in question
6586  * @delalloc:   The blocks are allocated for the delalloc write
6587  *
6588  * This is called by somebody who is freeing space that was never actually used
6589  * on disk.  For example if you reserve some space for a new leaf in transaction
6590  * A and before transaction A commits you free that leaf, you call this with
6591  * reserve set to 0 in order to clear the reservation.
6592  */
6593
6594 static int btrfs_free_reserved_bytes(struct btrfs_block_group_cache *cache,
6595                                      u64 num_bytes, int delalloc)
6596 {
6597         struct btrfs_space_info *space_info = cache->space_info;
6598         int ret = 0;
6599
6600         spin_lock(&space_info->lock);
6601         spin_lock(&cache->lock);
6602         if (cache->ro)
6603                 space_info->bytes_readonly += num_bytes;
6604         cache->reserved -= num_bytes;
6605         space_info->bytes_reserved -= num_bytes;
6606
6607         if (delalloc)
6608                 cache->delalloc_bytes -= num_bytes;
6609         spin_unlock(&cache->lock);
6610         spin_unlock(&space_info->lock);
6611         return ret;
6612 }
6613 void btrfs_prepare_extent_commit(struct btrfs_fs_info *fs_info)
6614 {
6615         struct btrfs_caching_control *next;
6616         struct btrfs_caching_control *caching_ctl;
6617         struct btrfs_block_group_cache *cache;
6618
6619         down_write(&fs_info->commit_root_sem);
6620
6621         list_for_each_entry_safe(caching_ctl, next,
6622                                  &fs_info->caching_block_groups, list) {
6623                 cache = caching_ctl->block_group;
6624                 if (block_group_cache_done(cache)) {
6625                         cache->last_byte_to_unpin = (u64)-1;
6626                         list_del_init(&caching_ctl->list);
6627                         put_caching_control(caching_ctl);
6628                 } else {
6629                         cache->last_byte_to_unpin = caching_ctl->progress;
6630                 }
6631         }
6632
6633         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6634                 fs_info->pinned_extents = &fs_info->freed_extents[1];
6635         else
6636                 fs_info->pinned_extents = &fs_info->freed_extents[0];
6637
6638         up_write(&fs_info->commit_root_sem);
6639
6640         update_global_block_rsv(fs_info);
6641 }
6642
6643 /*
6644  * Returns the free cluster for the given space info and sets empty_cluster to
6645  * what it should be based on the mount options.
6646  */
6647 static struct btrfs_free_cluster *
6648 fetch_cluster_info(struct btrfs_fs_info *fs_info,
6649                    struct btrfs_space_info *space_info, u64 *empty_cluster)
6650 {
6651         struct btrfs_free_cluster *ret = NULL;
6652
6653         *empty_cluster = 0;
6654         if (btrfs_mixed_space_info(space_info))
6655                 return ret;
6656
6657         if (space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
6658                 ret = &fs_info->meta_alloc_cluster;
6659                 if (btrfs_test_opt(fs_info, SSD))
6660                         *empty_cluster = SZ_2M;
6661                 else
6662                         *empty_cluster = SZ_64K;
6663         } else if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) &&
6664                    btrfs_test_opt(fs_info, SSD_SPREAD)) {
6665                 *empty_cluster = SZ_2M;
6666                 ret = &fs_info->data_alloc_cluster;
6667         }
6668
6669         return ret;
6670 }
6671
6672 static int unpin_extent_range(struct btrfs_fs_info *fs_info,
6673                               u64 start, u64 end,
6674                               const bool return_free_space)
6675 {
6676         struct btrfs_block_group_cache *cache = NULL;
6677         struct btrfs_space_info *space_info;
6678         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
6679         struct btrfs_free_cluster *cluster = NULL;
6680         u64 len;
6681         u64 total_unpinned = 0;
6682         u64 empty_cluster = 0;
6683         bool readonly;
6684
6685         while (start <= end) {
6686                 readonly = false;
6687                 if (!cache ||
6688                     start >= cache->key.objectid + cache->key.offset) {
6689                         if (cache)
6690                                 btrfs_put_block_group(cache);
6691                         total_unpinned = 0;
6692                         cache = btrfs_lookup_block_group(fs_info, start);
6693                         BUG_ON(!cache); /* Logic error */
6694
6695                         cluster = fetch_cluster_info(fs_info,
6696                                                      cache->space_info,
6697                                                      &empty_cluster);
6698                         empty_cluster <<= 1;
6699                 }
6700
6701                 len = cache->key.objectid + cache->key.offset - start;
6702                 len = min(len, end + 1 - start);
6703
6704                 if (start < cache->last_byte_to_unpin) {
6705                         len = min(len, cache->last_byte_to_unpin - start);
6706                         if (return_free_space)
6707                                 btrfs_add_free_space(cache, start, len);
6708                 }
6709
6710                 start += len;
6711                 total_unpinned += len;
6712                 space_info = cache->space_info;
6713
6714                 /*
6715                  * If this space cluster has been marked as fragmented and we've
6716                  * unpinned enough in this block group to potentially allow a
6717                  * cluster to be created inside of it go ahead and clear the
6718                  * fragmented check.
6719                  */
6720                 if (cluster && cluster->fragmented &&
6721                     total_unpinned > empty_cluster) {
6722                         spin_lock(&cluster->lock);
6723                         cluster->fragmented = 0;
6724                         spin_unlock(&cluster->lock);
6725                 }
6726
6727                 spin_lock(&space_info->lock);
6728                 spin_lock(&cache->lock);
6729                 cache->pinned -= len;
6730                 space_info->bytes_pinned -= len;
6731
6732                 trace_btrfs_space_reservation(fs_info, "pinned",
6733                                               space_info->flags, len, 0);
6734                 space_info->max_extent_size = 0;
6735                 percpu_counter_add(&space_info->total_bytes_pinned, -len);
6736                 if (cache->ro) {
6737                         space_info->bytes_readonly += len;
6738                         readonly = true;
6739                 }
6740                 spin_unlock(&cache->lock);
6741                 if (!readonly && return_free_space &&
6742                     global_rsv->space_info == space_info) {
6743                         u64 to_add = len;
6744
6745                         spin_lock(&global_rsv->lock);
6746                         if (!global_rsv->full) {
6747                                 to_add = min(len, global_rsv->size -
6748                                              global_rsv->reserved);
6749                                 global_rsv->reserved += to_add;
6750                                 space_info->bytes_may_use += to_add;
6751                                 if (global_rsv->reserved >= global_rsv->size)
6752                                         global_rsv->full = 1;
6753                                 trace_btrfs_space_reservation(fs_info,
6754                                                               "space_info",
6755                                                               space_info->flags,
6756                                                               to_add, 1);
6757                                 len -= to_add;
6758                         }
6759                         spin_unlock(&global_rsv->lock);
6760                         /* Add to any tickets we may have */
6761                         if (len)
6762                                 space_info_add_new_bytes(fs_info, space_info,
6763                                                          len);
6764                 }
6765                 spin_unlock(&space_info->lock);
6766         }
6767
6768         if (cache)
6769                 btrfs_put_block_group(cache);
6770         return 0;
6771 }
6772
6773 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans)
6774 {
6775         struct btrfs_fs_info *fs_info = trans->fs_info;
6776         struct btrfs_block_group_cache *block_group, *tmp;
6777         struct list_head *deleted_bgs;
6778         struct extent_io_tree *unpin;
6779         u64 start;
6780         u64 end;
6781         int ret;
6782
6783         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6784                 unpin = &fs_info->freed_extents[1];
6785         else
6786                 unpin = &fs_info->freed_extents[0];
6787
6788         while (!trans->aborted) {
6789                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
6790                 ret = find_first_extent_bit(unpin, 0, &start, &end,
6791                                             EXTENT_DIRTY, NULL);
6792                 if (ret) {
6793                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6794                         break;
6795                 }
6796
6797                 if (btrfs_test_opt(fs_info, DISCARD))
6798                         ret = btrfs_discard_extent(fs_info, start,
6799                                                    end + 1 - start, NULL);
6800
6801                 clear_extent_dirty(unpin, start, end);
6802                 unpin_extent_range(fs_info, start, end, true);
6803                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6804                 cond_resched();
6805         }
6806
6807         /*
6808          * Transaction is finished.  We don't need the lock anymore.  We
6809          * do need to clean up the block groups in case of a transaction
6810          * abort.
6811          */
6812         deleted_bgs = &trans->transaction->deleted_bgs;
6813         list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
6814                 u64 trimmed = 0;
6815
6816                 ret = -EROFS;
6817                 if (!trans->aborted)
6818                         ret = btrfs_discard_extent(fs_info,
6819                                                    block_group->key.objectid,
6820                                                    block_group->key.offset,
6821                                                    &trimmed);
6822
6823                 list_del_init(&block_group->bg_list);
6824                 btrfs_put_block_group_trimming(block_group);
6825                 btrfs_put_block_group(block_group);
6826
6827                 if (ret) {
6828                         const char *errstr = btrfs_decode_error(ret);
6829                         btrfs_warn(fs_info,
6830                            "discard failed while removing blockgroup: errno=%d %s",
6831                                    ret, errstr);
6832                 }
6833         }
6834
6835         return 0;
6836 }
6837
6838 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
6839                                 struct btrfs_fs_info *info,
6840                                 struct btrfs_delayed_ref_node *node, u64 parent,
6841                                 u64 root_objectid, u64 owner_objectid,
6842                                 u64 owner_offset, int refs_to_drop,
6843                                 struct btrfs_delayed_extent_op *extent_op)
6844 {
6845         struct btrfs_key key;
6846         struct btrfs_path *path;
6847         struct btrfs_root *extent_root = info->extent_root;
6848         struct extent_buffer *leaf;
6849         struct btrfs_extent_item *ei;
6850         struct btrfs_extent_inline_ref *iref;
6851         int ret;
6852         int is_data;
6853         int extent_slot = 0;
6854         int found_extent = 0;
6855         int num_to_del = 1;
6856         u32 item_size;
6857         u64 refs;
6858         u64 bytenr = node->bytenr;
6859         u64 num_bytes = node->num_bytes;
6860         int last_ref = 0;
6861         bool skinny_metadata = btrfs_fs_incompat(info, SKINNY_METADATA);
6862
6863         path = btrfs_alloc_path();
6864         if (!path)
6865                 return -ENOMEM;
6866
6867         path->reada = READA_FORWARD;
6868         path->leave_spinning = 1;
6869
6870         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
6871         BUG_ON(!is_data && refs_to_drop != 1);
6872
6873         if (is_data)
6874                 skinny_metadata = false;
6875
6876         ret = lookup_extent_backref(trans, info, path, &iref,
6877                                     bytenr, num_bytes, parent,
6878                                     root_objectid, owner_objectid,
6879                                     owner_offset);
6880         if (ret == 0) {
6881                 extent_slot = path->slots[0];
6882                 while (extent_slot >= 0) {
6883                         btrfs_item_key_to_cpu(path->nodes[0], &key,
6884                                               extent_slot);
6885                         if (key.objectid != bytenr)
6886                                 break;
6887                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
6888                             key.offset == num_bytes) {
6889                                 found_extent = 1;
6890                                 break;
6891                         }
6892                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
6893                             key.offset == owner_objectid) {
6894                                 found_extent = 1;
6895                                 break;
6896                         }
6897                         if (path->slots[0] - extent_slot > 5)
6898                                 break;
6899                         extent_slot--;
6900                 }
6901 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6902                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
6903                 if (found_extent && item_size < sizeof(*ei))
6904                         found_extent = 0;
6905 #endif
6906                 if (!found_extent) {
6907                         BUG_ON(iref);
6908                         ret = remove_extent_backref(trans, info, path, NULL,
6909                                                     refs_to_drop,
6910                                                     is_data, &last_ref);
6911                         if (ret) {
6912                                 btrfs_abort_transaction(trans, ret);
6913                                 goto out;
6914                         }
6915                         btrfs_release_path(path);
6916                         path->leave_spinning = 1;
6917
6918                         key.objectid = bytenr;
6919                         key.type = BTRFS_EXTENT_ITEM_KEY;
6920                         key.offset = num_bytes;
6921
6922                         if (!is_data && skinny_metadata) {
6923                                 key.type = BTRFS_METADATA_ITEM_KEY;
6924                                 key.offset = owner_objectid;
6925                         }
6926
6927                         ret = btrfs_search_slot(trans, extent_root,
6928                                                 &key, path, -1, 1);
6929                         if (ret > 0 && skinny_metadata && path->slots[0]) {
6930                                 /*
6931                                  * Couldn't find our skinny metadata item,
6932                                  * see if we have ye olde extent item.
6933                                  */
6934                                 path->slots[0]--;
6935                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
6936                                                       path->slots[0]);
6937                                 if (key.objectid == bytenr &&
6938                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
6939                                     key.offset == num_bytes)
6940                                         ret = 0;
6941                         }
6942
6943                         if (ret > 0 && skinny_metadata) {
6944                                 skinny_metadata = false;
6945                                 key.objectid = bytenr;
6946                                 key.type = BTRFS_EXTENT_ITEM_KEY;
6947                                 key.offset = num_bytes;
6948                                 btrfs_release_path(path);
6949                                 ret = btrfs_search_slot(trans, extent_root,
6950                                                         &key, path, -1, 1);
6951                         }
6952
6953                         if (ret) {
6954                                 btrfs_err(info,
6955                                           "umm, got %d back from search, was looking for %llu",
6956                                           ret, bytenr);
6957                                 if (ret > 0)
6958                                         btrfs_print_leaf(path->nodes[0]);
6959                         }
6960                         if (ret < 0) {
6961                                 btrfs_abort_transaction(trans, ret);
6962                                 goto out;
6963                         }
6964                         extent_slot = path->slots[0];
6965                 }
6966         } else if (WARN_ON(ret == -ENOENT)) {
6967                 btrfs_print_leaf(path->nodes[0]);
6968                 btrfs_err(info,
6969                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
6970                         bytenr, parent, root_objectid, owner_objectid,
6971                         owner_offset);
6972                 btrfs_abort_transaction(trans, ret);
6973                 goto out;
6974         } else {
6975                 btrfs_abort_transaction(trans, ret);
6976                 goto out;
6977         }
6978
6979         leaf = path->nodes[0];
6980         item_size = btrfs_item_size_nr(leaf, extent_slot);
6981 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6982         if (item_size < sizeof(*ei)) {
6983                 BUG_ON(found_extent || extent_slot != path->slots[0]);
6984                 ret = convert_extent_item_v0(trans, info, path, owner_objectid,
6985                                              0);
6986                 if (ret < 0) {
6987                         btrfs_abort_transaction(trans, ret);
6988                         goto out;
6989                 }
6990
6991                 btrfs_release_path(path);
6992                 path->leave_spinning = 1;
6993
6994                 key.objectid = bytenr;
6995                 key.type = BTRFS_EXTENT_ITEM_KEY;
6996                 key.offset = num_bytes;
6997
6998                 ret = btrfs_search_slot(trans, extent_root, &key, path,
6999                                         -1, 1);
7000                 if (ret) {
7001                         btrfs_err(info,
7002                                   "umm, got %d back from search, was looking for %llu",
7003                                 ret, bytenr);
7004                         btrfs_print_leaf(path->nodes[0]);
7005                 }
7006                 if (ret < 0) {
7007                         btrfs_abort_transaction(trans, ret);
7008                         goto out;
7009                 }
7010
7011                 extent_slot = path->slots[0];
7012                 leaf = path->nodes[0];
7013                 item_size = btrfs_item_size_nr(leaf, extent_slot);
7014         }
7015 #endif
7016         BUG_ON(item_size < sizeof(*ei));
7017         ei = btrfs_item_ptr(leaf, extent_slot,
7018                             struct btrfs_extent_item);
7019         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
7020             key.type == BTRFS_EXTENT_ITEM_KEY) {
7021                 struct btrfs_tree_block_info *bi;
7022                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
7023                 bi = (struct btrfs_tree_block_info *)(ei + 1);
7024                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
7025         }
7026
7027         refs = btrfs_extent_refs(leaf, ei);
7028         if (refs < refs_to_drop) {
7029                 btrfs_err(info,
7030                           "trying to drop %d refs but we only have %Lu for bytenr %Lu",
7031                           refs_to_drop, refs, bytenr);
7032                 ret = -EINVAL;
7033                 btrfs_abort_transaction(trans, ret);
7034                 goto out;
7035         }
7036         refs -= refs_to_drop;
7037
7038         if (refs > 0) {
7039                 if (extent_op)
7040                         __run_delayed_extent_op(extent_op, leaf, ei);
7041                 /*
7042                  * In the case of inline back ref, reference count will
7043                  * be updated by remove_extent_backref
7044                  */
7045                 if (iref) {
7046                         BUG_ON(!found_extent);
7047                 } else {
7048                         btrfs_set_extent_refs(leaf, ei, refs);
7049                         btrfs_mark_buffer_dirty(leaf);
7050                 }
7051                 if (found_extent) {
7052                         ret = remove_extent_backref(trans, info, path,
7053                                                     iref, refs_to_drop,
7054                                                     is_data, &last_ref);
7055                         if (ret) {
7056                                 btrfs_abort_transaction(trans, ret);
7057                                 goto out;
7058                         }
7059                 }
7060         } else {
7061                 if (found_extent) {
7062                         BUG_ON(is_data && refs_to_drop !=
7063                                extent_data_ref_count(path, iref));
7064                         if (iref) {
7065                                 BUG_ON(path->slots[0] != extent_slot);
7066                         } else {
7067                                 BUG_ON(path->slots[0] != extent_slot + 1);
7068                                 path->slots[0] = extent_slot;
7069                                 num_to_del = 2;
7070                         }
7071                 }
7072
7073                 last_ref = 1;
7074                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
7075                                       num_to_del);
7076                 if (ret) {
7077                         btrfs_abort_transaction(trans, ret);
7078                         goto out;
7079                 }
7080                 btrfs_release_path(path);
7081
7082                 if (is_data) {
7083                         ret = btrfs_del_csums(trans, info, bytenr, num_bytes);
7084                         if (ret) {
7085                                 btrfs_abort_transaction(trans, ret);
7086                                 goto out;
7087                         }
7088                 }
7089
7090                 ret = add_to_free_space_tree(trans, info, bytenr, num_bytes);
7091                 if (ret) {
7092                         btrfs_abort_transaction(trans, ret);
7093                         goto out;
7094                 }
7095
7096                 ret = update_block_group(trans, info, bytenr, num_bytes, 0);
7097                 if (ret) {
7098                         btrfs_abort_transaction(trans, ret);
7099                         goto out;
7100                 }
7101         }
7102         btrfs_release_path(path);
7103
7104 out:
7105         btrfs_free_path(path);
7106         return ret;
7107 }
7108
7109 /*
7110  * when we free an block, it is possible (and likely) that we free the last
7111  * delayed ref for that extent as well.  This searches the delayed ref tree for
7112  * a given extent, and if there are no other delayed refs to be processed, it
7113  * removes it from the tree.
7114  */
7115 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
7116                                       u64 bytenr)
7117 {
7118         struct btrfs_delayed_ref_head *head;
7119         struct btrfs_delayed_ref_root *delayed_refs;
7120         int ret = 0;
7121
7122         delayed_refs = &trans->transaction->delayed_refs;
7123         spin_lock(&delayed_refs->lock);
7124         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
7125         if (!head)
7126                 goto out_delayed_unlock;
7127
7128         spin_lock(&head->lock);
7129         if (!RB_EMPTY_ROOT(&head->ref_tree))
7130                 goto out;
7131
7132         if (head->extent_op) {
7133                 if (!head->must_insert_reserved)
7134                         goto out;
7135                 btrfs_free_delayed_extent_op(head->extent_op);
7136                 head->extent_op = NULL;
7137         }
7138
7139         /*
7140          * waiting for the lock here would deadlock.  If someone else has it
7141          * locked they are already in the process of dropping it anyway
7142          */
7143         if (!mutex_trylock(&head->mutex))
7144                 goto out;
7145
7146         /*
7147          * at this point we have a head with no other entries.  Go
7148          * ahead and process it.
7149          */
7150         rb_erase(&head->href_node, &delayed_refs->href_root);
7151         RB_CLEAR_NODE(&head->href_node);
7152         atomic_dec(&delayed_refs->num_entries);
7153
7154         /*
7155          * we don't take a ref on the node because we're removing it from the
7156          * tree, so we just steal the ref the tree was holding.
7157          */
7158         delayed_refs->num_heads--;
7159         if (head->processing == 0)
7160                 delayed_refs->num_heads_ready--;
7161         head->processing = 0;
7162         spin_unlock(&head->lock);
7163         spin_unlock(&delayed_refs->lock);
7164
7165         BUG_ON(head->extent_op);
7166         if (head->must_insert_reserved)
7167                 ret = 1;
7168
7169         mutex_unlock(&head->mutex);
7170         btrfs_put_delayed_ref_head(head);
7171         return ret;
7172 out:
7173         spin_unlock(&head->lock);
7174
7175 out_delayed_unlock:
7176         spin_unlock(&delayed_refs->lock);
7177         return 0;
7178 }
7179
7180 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
7181                            struct btrfs_root *root,
7182                            struct extent_buffer *buf,
7183                            u64 parent, int last_ref)
7184 {
7185         struct btrfs_fs_info *fs_info = root->fs_info;
7186         int pin = 1;
7187         int ret;
7188
7189         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7190                 int old_ref_mod, new_ref_mod;
7191
7192                 btrfs_ref_tree_mod(root, buf->start, buf->len, parent,
7193                                    root->root_key.objectid,
7194                                    btrfs_header_level(buf), 0,
7195                                    BTRFS_DROP_DELAYED_REF);
7196                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, buf->start,
7197                                                  buf->len, parent,
7198                                                  root->root_key.objectid,
7199                                                  btrfs_header_level(buf),
7200                                                  BTRFS_DROP_DELAYED_REF, NULL,
7201                                                  &old_ref_mod, &new_ref_mod);
7202                 BUG_ON(ret); /* -ENOMEM */
7203                 pin = old_ref_mod >= 0 && new_ref_mod < 0;
7204         }
7205
7206         if (last_ref && btrfs_header_generation(buf) == trans->transid) {
7207                 struct btrfs_block_group_cache *cache;
7208
7209                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7210                         ret = check_ref_cleanup(trans, buf->start);
7211                         if (!ret)
7212                                 goto out;
7213                 }
7214
7215                 pin = 0;
7216                 cache = btrfs_lookup_block_group(fs_info, buf->start);
7217
7218                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
7219                         pin_down_extent(fs_info, cache, buf->start,
7220                                         buf->len, 1);
7221                         btrfs_put_block_group(cache);
7222                         goto out;
7223                 }
7224
7225                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
7226
7227                 btrfs_add_free_space(cache, buf->start, buf->len);
7228                 btrfs_free_reserved_bytes(cache, buf->len, 0);
7229                 btrfs_put_block_group(cache);
7230                 trace_btrfs_reserved_extent_free(fs_info, buf->start, buf->len);
7231         }
7232 out:
7233         if (pin)
7234                 add_pinned_bytes(fs_info, buf->len, btrfs_header_level(buf),
7235                                  root->root_key.objectid);
7236
7237         if (last_ref) {
7238                 /*
7239                  * Deleting the buffer, clear the corrupt flag since it doesn't
7240                  * matter anymore.
7241                  */
7242                 clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
7243         }
7244 }
7245
7246 /* Can return -ENOMEM */
7247 int btrfs_free_extent(struct btrfs_trans_handle *trans,
7248                       struct btrfs_root *root,
7249                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
7250                       u64 owner, u64 offset)
7251 {
7252         struct btrfs_fs_info *fs_info = root->fs_info;
7253         int old_ref_mod, new_ref_mod;
7254         int ret;
7255
7256         if (btrfs_is_testing(fs_info))
7257                 return 0;
7258
7259         if (root_objectid != BTRFS_TREE_LOG_OBJECTID)
7260                 btrfs_ref_tree_mod(root, bytenr, num_bytes, parent,
7261                                    root_objectid, owner, offset,
7262                                    BTRFS_DROP_DELAYED_REF);
7263
7264         /*
7265          * tree log blocks never actually go into the extent allocation
7266          * tree, just update pinning info and exit early.
7267          */
7268         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
7269                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
7270                 /* unlocks the pinned mutex */
7271                 btrfs_pin_extent(fs_info, bytenr, num_bytes, 1);
7272                 old_ref_mod = new_ref_mod = 0;
7273                 ret = 0;
7274         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
7275                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
7276                                                  num_bytes, parent,
7277                                                  root_objectid, (int)owner,
7278                                                  BTRFS_DROP_DELAYED_REF, NULL,
7279                                                  &old_ref_mod, &new_ref_mod);
7280         } else {
7281                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
7282                                                  num_bytes, parent,
7283                                                  root_objectid, owner, offset,
7284                                                  0, BTRFS_DROP_DELAYED_REF,
7285                                                  &old_ref_mod, &new_ref_mod);
7286         }
7287
7288         if (ret == 0 && old_ref_mod >= 0 && new_ref_mod < 0)
7289                 add_pinned_bytes(fs_info, num_bytes, owner, root_objectid);
7290
7291         return ret;
7292 }
7293
7294 /*
7295  * when we wait for progress in the block group caching, its because
7296  * our allocation attempt failed at least once.  So, we must sleep
7297  * and let some progress happen before we try again.
7298  *
7299  * This function will sleep at least once waiting for new free space to
7300  * show up, and then it will check the block group free space numbers
7301  * for our min num_bytes.  Another option is to have it go ahead
7302  * and look in the rbtree for a free extent of a given size, but this
7303  * is a good start.
7304  *
7305  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
7306  * any of the information in this block group.
7307  */
7308 static noinline void
7309 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
7310                                 u64 num_bytes)
7311 {
7312         struct btrfs_caching_control *caching_ctl;
7313
7314         caching_ctl = get_caching_control(cache);
7315         if (!caching_ctl)
7316                 return;
7317
7318         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
7319                    (cache->free_space_ctl->free_space >= num_bytes));
7320
7321         put_caching_control(caching_ctl);
7322 }
7323
7324 static noinline int
7325 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
7326 {
7327         struct btrfs_caching_control *caching_ctl;
7328         int ret = 0;
7329
7330         caching_ctl = get_caching_control(cache);
7331         if (!caching_ctl)
7332                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
7333
7334         wait_event(caching_ctl->wait, block_group_cache_done(cache));
7335         if (cache->cached == BTRFS_CACHE_ERROR)
7336                 ret = -EIO;
7337         put_caching_control(caching_ctl);
7338         return ret;
7339 }
7340
7341 static const char *btrfs_raid_type_names[BTRFS_NR_RAID_TYPES] = {
7342         [BTRFS_RAID_RAID10]     = "raid10",
7343         [BTRFS_RAID_RAID1]      = "raid1",
7344         [BTRFS_RAID_DUP]        = "dup",
7345         [BTRFS_RAID_RAID0]      = "raid0",
7346         [BTRFS_RAID_SINGLE]     = "single",
7347         [BTRFS_RAID_RAID5]      = "raid5",
7348         [BTRFS_RAID_RAID6]      = "raid6",
7349 };
7350
7351 static const char *get_raid_name(enum btrfs_raid_types type)
7352 {
7353         if (type >= BTRFS_NR_RAID_TYPES)
7354                 return NULL;
7355
7356         return btrfs_raid_type_names[type];
7357 }
7358
7359 enum btrfs_loop_type {
7360         LOOP_CACHING_NOWAIT = 0,
7361         LOOP_CACHING_WAIT = 1,
7362         LOOP_ALLOC_CHUNK = 2,
7363         LOOP_NO_EMPTY_SIZE = 3,
7364 };
7365
7366 static inline void
7367 btrfs_lock_block_group(struct btrfs_block_group_cache *cache,
7368                        int delalloc)
7369 {
7370         if (delalloc)
7371                 down_read(&cache->data_rwsem);
7372 }
7373
7374 static inline void
7375 btrfs_grab_block_group(struct btrfs_block_group_cache *cache,
7376                        int delalloc)
7377 {
7378         btrfs_get_block_group(cache);
7379         if (delalloc)
7380                 down_read(&cache->data_rwsem);
7381 }
7382
7383 static struct btrfs_block_group_cache *
7384 btrfs_lock_cluster(struct btrfs_block_group_cache *block_group,
7385                    struct btrfs_free_cluster *cluster,
7386                    int delalloc)
7387 {
7388         struct btrfs_block_group_cache *used_bg = NULL;
7389
7390         spin_lock(&cluster->refill_lock);
7391         while (1) {
7392                 used_bg = cluster->block_group;
7393                 if (!used_bg)
7394                         return NULL;
7395
7396                 if (used_bg == block_group)
7397                         return used_bg;
7398
7399                 btrfs_get_block_group(used_bg);
7400
7401                 if (!delalloc)
7402                         return used_bg;
7403
7404                 if (down_read_trylock(&used_bg->data_rwsem))
7405                         return used_bg;
7406
7407                 spin_unlock(&cluster->refill_lock);
7408
7409                 /* We should only have one-level nested. */
7410                 down_read_nested(&used_bg->data_rwsem, SINGLE_DEPTH_NESTING);
7411
7412                 spin_lock(&cluster->refill_lock);
7413                 if (used_bg == cluster->block_group)
7414                         return used_bg;
7415
7416                 up_read(&used_bg->data_rwsem);
7417                 btrfs_put_block_group(used_bg);
7418         }
7419 }
7420
7421 static inline void
7422 btrfs_release_block_group(struct btrfs_block_group_cache *cache,
7423                          int delalloc)
7424 {
7425         if (delalloc)
7426                 up_read(&cache->data_rwsem);
7427         btrfs_put_block_group(cache);
7428 }
7429
7430 /*
7431  * walks the btree of allocated extents and find a hole of a given size.
7432  * The key ins is changed to record the hole:
7433  * ins->objectid == start position
7434  * ins->flags = BTRFS_EXTENT_ITEM_KEY
7435  * ins->offset == the size of the hole.
7436  * Any available blocks before search_start are skipped.
7437  *
7438  * If there is no suitable free space, we will record the max size of
7439  * the free space extent currently.
7440  */
7441 static noinline int find_free_extent(struct btrfs_fs_info *fs_info,
7442                                 u64 ram_bytes, u64 num_bytes, u64 empty_size,
7443                                 u64 hint_byte, struct btrfs_key *ins,
7444                                 u64 flags, int delalloc)
7445 {
7446         int ret = 0;
7447         struct btrfs_root *root = fs_info->extent_root;
7448         struct btrfs_free_cluster *last_ptr = NULL;
7449         struct btrfs_block_group_cache *block_group = NULL;
7450         u64 search_start = 0;
7451         u64 max_extent_size = 0;
7452         u64 empty_cluster = 0;
7453         struct btrfs_space_info *space_info;
7454         int loop = 0;
7455         int index = btrfs_bg_flags_to_raid_index(flags);
7456         bool failed_cluster_refill = false;
7457         bool failed_alloc = false;
7458         bool use_cluster = true;
7459         bool have_caching_bg = false;
7460         bool orig_have_caching_bg = false;
7461         bool full_search = false;
7462
7463         WARN_ON(num_bytes < fs_info->sectorsize);
7464         ins->type = BTRFS_EXTENT_ITEM_KEY;
7465         ins->objectid = 0;
7466         ins->offset = 0;
7467
7468         trace_find_free_extent(fs_info, num_bytes, empty_size, flags);
7469
7470         space_info = __find_space_info(fs_info, flags);
7471         if (!space_info) {
7472                 btrfs_err(fs_info, "No space info for %llu", flags);
7473                 return -ENOSPC;
7474         }
7475
7476         /*
7477          * If our free space is heavily fragmented we may not be able to make
7478          * big contiguous allocations, so instead of doing the expensive search
7479          * for free space, simply return ENOSPC with our max_extent_size so we
7480          * can go ahead and search for a more manageable chunk.
7481          *
7482          * If our max_extent_size is large enough for our allocation simply
7483          * disable clustering since we will likely not be able to find enough
7484          * space to create a cluster and induce latency trying.
7485          */
7486         if (unlikely(space_info->max_extent_size)) {
7487                 spin_lock(&space_info->lock);
7488                 if (space_info->max_extent_size &&
7489                     num_bytes > space_info->max_extent_size) {
7490                         ins->offset = space_info->max_extent_size;
7491                         spin_unlock(&space_info->lock);
7492                         return -ENOSPC;
7493                 } else if (space_info->max_extent_size) {
7494                         use_cluster = false;
7495                 }
7496                 spin_unlock(&space_info->lock);
7497         }
7498
7499         last_ptr = fetch_cluster_info(fs_info, space_info, &empty_cluster);
7500         if (last_ptr) {
7501                 spin_lock(&last_ptr->lock);
7502                 if (last_ptr->block_group)
7503                         hint_byte = last_ptr->window_start;
7504                 if (last_ptr->fragmented) {
7505                         /*
7506                          * We still set window_start so we can keep track of the
7507                          * last place we found an allocation to try and save
7508                          * some time.
7509                          */
7510                         hint_byte = last_ptr->window_start;
7511                         use_cluster = false;
7512                 }
7513                 spin_unlock(&last_ptr->lock);
7514         }
7515
7516         search_start = max(search_start, first_logical_byte(fs_info, 0));
7517         search_start = max(search_start, hint_byte);
7518         if (search_start == hint_byte) {
7519                 block_group = btrfs_lookup_block_group(fs_info, search_start);
7520                 /*
7521                  * we don't want to use the block group if it doesn't match our
7522                  * allocation bits, or if its not cached.
7523                  *
7524                  * However if we are re-searching with an ideal block group
7525                  * picked out then we don't care that the block group is cached.
7526                  */
7527                 if (block_group && block_group_bits(block_group, flags) &&
7528                     block_group->cached != BTRFS_CACHE_NO) {
7529                         down_read(&space_info->groups_sem);
7530                         if (list_empty(&block_group->list) ||
7531                             block_group->ro) {
7532                                 /*
7533                                  * someone is removing this block group,
7534                                  * we can't jump into the have_block_group
7535                                  * target because our list pointers are not
7536                                  * valid
7537                                  */
7538                                 btrfs_put_block_group(block_group);
7539                                 up_read(&space_info->groups_sem);
7540                         } else {
7541                                 index = btrfs_bg_flags_to_raid_index(
7542                                                 block_group->flags);
7543                                 btrfs_lock_block_group(block_group, delalloc);
7544                                 goto have_block_group;
7545                         }
7546                 } else if (block_group) {
7547                         btrfs_put_block_group(block_group);
7548                 }
7549         }
7550 search:
7551         have_caching_bg = false;
7552         if (index == 0 || index == btrfs_bg_flags_to_raid_index(flags))
7553                 full_search = true;
7554         down_read(&space_info->groups_sem);
7555         list_for_each_entry(block_group, &space_info->block_groups[index],
7556                             list) {
7557                 u64 offset;
7558                 int cached;
7559
7560                 /* If the block group is read-only, we can skip it entirely. */
7561                 if (unlikely(block_group->ro))
7562                         continue;
7563
7564                 btrfs_grab_block_group(block_group, delalloc);
7565                 search_start = block_group->key.objectid;
7566
7567                 /*
7568                  * this can happen if we end up cycling through all the
7569                  * raid types, but we want to make sure we only allocate
7570                  * for the proper type.
7571                  */
7572                 if (!block_group_bits(block_group, flags)) {
7573                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
7574                                 BTRFS_BLOCK_GROUP_RAID1 |
7575                                 BTRFS_BLOCK_GROUP_RAID5 |
7576                                 BTRFS_BLOCK_GROUP_RAID6 |
7577                                 BTRFS_BLOCK_GROUP_RAID10;
7578
7579                         /*
7580                          * if they asked for extra copies and this block group
7581                          * doesn't provide them, bail.  This does allow us to
7582                          * fill raid0 from raid1.
7583                          */
7584                         if ((flags & extra) && !(block_group->flags & extra))
7585                                 goto loop;
7586                 }
7587
7588 have_block_group:
7589                 cached = block_group_cache_done(block_group);
7590                 if (unlikely(!cached)) {
7591                         have_caching_bg = true;
7592                         ret = cache_block_group(block_group, 0);
7593                         BUG_ON(ret < 0);
7594                         ret = 0;
7595                 }
7596
7597                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
7598                         goto loop;
7599
7600                 /*
7601                  * Ok we want to try and use the cluster allocator, so
7602                  * lets look there
7603                  */
7604                 if (last_ptr && use_cluster) {
7605                         struct btrfs_block_group_cache *used_block_group;
7606                         unsigned long aligned_cluster;
7607                         /*
7608                          * the refill lock keeps out other
7609                          * people trying to start a new cluster
7610                          */
7611                         used_block_group = btrfs_lock_cluster(block_group,
7612                                                               last_ptr,
7613                                                               delalloc);
7614                         if (!used_block_group)
7615                                 goto refill_cluster;
7616
7617                         if (used_block_group != block_group &&
7618                             (used_block_group->ro ||
7619                              !block_group_bits(used_block_group, flags)))
7620                                 goto release_cluster;
7621
7622                         offset = btrfs_alloc_from_cluster(used_block_group,
7623                                                 last_ptr,
7624                                                 num_bytes,
7625                                                 used_block_group->key.objectid,
7626                                                 &max_extent_size);
7627                         if (offset) {
7628                                 /* we have a block, we're done */
7629                                 spin_unlock(&last_ptr->refill_lock);
7630                                 trace_btrfs_reserve_extent_cluster(fs_info,
7631                                                 used_block_group,
7632                                                 search_start, num_bytes);
7633                                 if (used_block_group != block_group) {
7634                                         btrfs_release_block_group(block_group,
7635                                                                   delalloc);
7636                                         block_group = used_block_group;
7637                                 }
7638                                 goto checks;
7639                         }
7640
7641                         WARN_ON(last_ptr->block_group != used_block_group);
7642 release_cluster:
7643                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
7644                          * set up a new clusters, so lets just skip it
7645                          * and let the allocator find whatever block
7646                          * it can find.  If we reach this point, we
7647                          * will have tried the cluster allocator
7648                          * plenty of times and not have found
7649                          * anything, so we are likely way too
7650                          * fragmented for the clustering stuff to find
7651                          * anything.
7652                          *
7653                          * However, if the cluster is taken from the
7654                          * current block group, release the cluster
7655                          * first, so that we stand a better chance of
7656                          * succeeding in the unclustered
7657                          * allocation.  */
7658                         if (loop >= LOOP_NO_EMPTY_SIZE &&
7659                             used_block_group != block_group) {
7660                                 spin_unlock(&last_ptr->refill_lock);
7661                                 btrfs_release_block_group(used_block_group,
7662                                                           delalloc);
7663                                 goto unclustered_alloc;
7664                         }
7665
7666                         /*
7667                          * this cluster didn't work out, free it and
7668                          * start over
7669                          */
7670                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7671
7672                         if (used_block_group != block_group)
7673                                 btrfs_release_block_group(used_block_group,
7674                                                           delalloc);
7675 refill_cluster:
7676                         if (loop >= LOOP_NO_EMPTY_SIZE) {
7677                                 spin_unlock(&last_ptr->refill_lock);
7678                                 goto unclustered_alloc;
7679                         }
7680
7681                         aligned_cluster = max_t(unsigned long,
7682                                                 empty_cluster + empty_size,
7683                                               block_group->full_stripe_len);
7684
7685                         /* allocate a cluster in this block group */
7686                         ret = btrfs_find_space_cluster(fs_info, block_group,
7687                                                        last_ptr, search_start,
7688                                                        num_bytes,
7689                                                        aligned_cluster);
7690                         if (ret == 0) {
7691                                 /*
7692                                  * now pull our allocation out of this
7693                                  * cluster
7694                                  */
7695                                 offset = btrfs_alloc_from_cluster(block_group,
7696                                                         last_ptr,
7697                                                         num_bytes,
7698                                                         search_start,
7699                                                         &max_extent_size);
7700                                 if (offset) {
7701                                         /* we found one, proceed */
7702                                         spin_unlock(&last_ptr->refill_lock);
7703                                         trace_btrfs_reserve_extent_cluster(fs_info,
7704                                                 block_group, search_start,
7705                                                 num_bytes);
7706                                         goto checks;
7707                                 }
7708                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
7709                                    && !failed_cluster_refill) {
7710                                 spin_unlock(&last_ptr->refill_lock);
7711
7712                                 failed_cluster_refill = true;
7713                                 wait_block_group_cache_progress(block_group,
7714                                        num_bytes + empty_cluster + empty_size);
7715                                 goto have_block_group;
7716                         }
7717
7718                         /*
7719                          * at this point we either didn't find a cluster
7720                          * or we weren't able to allocate a block from our
7721                          * cluster.  Free the cluster we've been trying
7722                          * to use, and go to the next block group
7723                          */
7724                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7725                         spin_unlock(&last_ptr->refill_lock);
7726                         goto loop;
7727                 }
7728
7729 unclustered_alloc:
7730                 /*
7731                  * We are doing an unclustered alloc, set the fragmented flag so
7732                  * we don't bother trying to setup a cluster again until we get
7733                  * more space.
7734                  */
7735                 if (unlikely(last_ptr)) {
7736                         spin_lock(&last_ptr->lock);
7737                         last_ptr->fragmented = 1;
7738                         spin_unlock(&last_ptr->lock);
7739                 }
7740                 if (cached) {
7741                         struct btrfs_free_space_ctl *ctl =
7742                                 block_group->free_space_ctl;
7743
7744                         spin_lock(&ctl->tree_lock);
7745                         if (ctl->free_space <
7746                             num_bytes + empty_cluster + empty_size) {
7747                                 if (ctl->free_space > max_extent_size)
7748                                         max_extent_size = ctl->free_space;
7749                                 spin_unlock(&ctl->tree_lock);
7750                                 goto loop;
7751                         }
7752                         spin_unlock(&ctl->tree_lock);
7753                 }
7754
7755                 offset = btrfs_find_space_for_alloc(block_group, search_start,
7756                                                     num_bytes, empty_size,
7757                                                     &max_extent_size);
7758                 /*
7759                  * If we didn't find a chunk, and we haven't failed on this
7760                  * block group before, and this block group is in the middle of
7761                  * caching and we are ok with waiting, then go ahead and wait
7762                  * for progress to be made, and set failed_alloc to true.
7763                  *
7764                  * If failed_alloc is true then we've already waited on this
7765                  * block group once and should move on to the next block group.
7766                  */
7767                 if (!offset && !failed_alloc && !cached &&
7768                     loop > LOOP_CACHING_NOWAIT) {
7769                         wait_block_group_cache_progress(block_group,
7770                                                 num_bytes + empty_size);
7771                         failed_alloc = true;
7772                         goto have_block_group;
7773                 } else if (!offset) {
7774                         goto loop;
7775                 }
7776 checks:
7777                 search_start = ALIGN(offset, fs_info->stripesize);
7778
7779                 /* move on to the next group */
7780                 if (search_start + num_bytes >
7781                     block_group->key.objectid + block_group->key.offset) {
7782                         btrfs_add_free_space(block_group, offset, num_bytes);
7783                         goto loop;
7784                 }
7785
7786                 if (offset < search_start)
7787                         btrfs_add_free_space(block_group, offset,
7788                                              search_start - offset);
7789                 BUG_ON(offset > search_start);
7790
7791                 ret = btrfs_add_reserved_bytes(block_group, ram_bytes,
7792                                 num_bytes, delalloc);
7793                 if (ret == -EAGAIN) {
7794                         btrfs_add_free_space(block_group, offset, num_bytes);
7795                         goto loop;
7796                 }
7797                 btrfs_inc_block_group_reservations(block_group);
7798
7799                 /* we are all good, lets return */
7800                 ins->objectid = search_start;
7801                 ins->offset = num_bytes;
7802
7803                 trace_btrfs_reserve_extent(fs_info, block_group,
7804                                            search_start, num_bytes);
7805                 btrfs_release_block_group(block_group, delalloc);
7806                 break;
7807 loop:
7808                 failed_cluster_refill = false;
7809                 failed_alloc = false;
7810                 BUG_ON(btrfs_bg_flags_to_raid_index(block_group->flags) !=
7811                        index);
7812                 btrfs_release_block_group(block_group, delalloc);
7813                 cond_resched();
7814         }
7815         up_read(&space_info->groups_sem);
7816
7817         if ((loop == LOOP_CACHING_NOWAIT) && have_caching_bg
7818                 && !orig_have_caching_bg)
7819                 orig_have_caching_bg = true;
7820
7821         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
7822                 goto search;
7823
7824         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
7825                 goto search;
7826
7827         /*
7828          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
7829          *                      caching kthreads as we move along
7830          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
7831          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
7832          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
7833          *                      again
7834          */
7835         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
7836                 index = 0;
7837                 if (loop == LOOP_CACHING_NOWAIT) {
7838                         /*
7839                          * We want to skip the LOOP_CACHING_WAIT step if we
7840                          * don't have any uncached bgs and we've already done a
7841                          * full search through.
7842                          */
7843                         if (orig_have_caching_bg || !full_search)
7844                                 loop = LOOP_CACHING_WAIT;
7845                         else
7846                                 loop = LOOP_ALLOC_CHUNK;
7847                 } else {
7848                         loop++;
7849                 }
7850
7851                 if (loop == LOOP_ALLOC_CHUNK) {
7852                         struct btrfs_trans_handle *trans;
7853                         int exist = 0;
7854
7855                         trans = current->journal_info;
7856                         if (trans)
7857                                 exist = 1;
7858                         else
7859                                 trans = btrfs_join_transaction(root);
7860
7861                         if (IS_ERR(trans)) {
7862                                 ret = PTR_ERR(trans);
7863                                 goto out;
7864                         }
7865
7866                         ret = do_chunk_alloc(trans, fs_info, flags,
7867                                              CHUNK_ALLOC_FORCE);
7868
7869                         /*
7870                          * If we can't allocate a new chunk we've already looped
7871                          * through at least once, move on to the NO_EMPTY_SIZE
7872                          * case.
7873                          */
7874                         if (ret == -ENOSPC)
7875                                 loop = LOOP_NO_EMPTY_SIZE;
7876
7877                         /*
7878                          * Do not bail out on ENOSPC since we
7879                          * can do more things.
7880                          */
7881                         if (ret < 0 && ret != -ENOSPC)
7882                                 btrfs_abort_transaction(trans, ret);
7883                         else
7884                                 ret = 0;
7885                         if (!exist)
7886                                 btrfs_end_transaction(trans);
7887                         if (ret)
7888                                 goto out;
7889                 }
7890
7891                 if (loop == LOOP_NO_EMPTY_SIZE) {
7892                         /*
7893                          * Don't loop again if we already have no empty_size and
7894                          * no empty_cluster.
7895                          */
7896                         if (empty_size == 0 &&
7897                             empty_cluster == 0) {
7898                                 ret = -ENOSPC;
7899                                 goto out;
7900                         }
7901                         empty_size = 0;
7902                         empty_cluster = 0;
7903                 }
7904
7905                 goto search;
7906         } else if (!ins->objectid) {
7907                 ret = -ENOSPC;
7908         } else if (ins->objectid) {
7909                 if (!use_cluster && last_ptr) {
7910                         spin_lock(&last_ptr->lock);
7911                         last_ptr->window_start = ins->objectid;
7912                         spin_unlock(&last_ptr->lock);
7913                 }
7914                 ret = 0;
7915         }
7916 out:
7917         if (ret == -ENOSPC) {
7918                 spin_lock(&space_info->lock);
7919                 space_info->max_extent_size = max_extent_size;
7920                 spin_unlock(&space_info->lock);
7921                 ins->offset = max_extent_size;
7922         }
7923         return ret;
7924 }
7925
7926 static void dump_space_info(struct btrfs_fs_info *fs_info,
7927                             struct btrfs_space_info *info, u64 bytes,
7928                             int dump_block_groups)
7929 {
7930         struct btrfs_block_group_cache *cache;
7931         int index = 0;
7932
7933         spin_lock(&info->lock);
7934         btrfs_info(fs_info, "space_info %llu has %llu free, is %sfull",
7935                    info->flags,
7936                    info->total_bytes - btrfs_space_info_used(info, true),
7937                    info->full ? "" : "not ");
7938         btrfs_info(fs_info,
7939                 "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu",
7940                 info->total_bytes, info->bytes_used, info->bytes_pinned,
7941                 info->bytes_reserved, info->bytes_may_use,
7942                 info->bytes_readonly);
7943         spin_unlock(&info->lock);
7944
7945         if (!dump_block_groups)
7946                 return;
7947
7948         down_read(&info->groups_sem);
7949 again:
7950         list_for_each_entry(cache, &info->block_groups[index], list) {
7951                 spin_lock(&cache->lock);
7952                 btrfs_info(fs_info,
7953                         "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s",
7954                         cache->key.objectid, cache->key.offset,
7955                         btrfs_block_group_used(&cache->item), cache->pinned,
7956                         cache->reserved, cache->ro ? "[readonly]" : "");
7957                 btrfs_dump_free_space(cache, bytes);
7958                 spin_unlock(&cache->lock);
7959         }
7960         if (++index < BTRFS_NR_RAID_TYPES)
7961                 goto again;
7962         up_read(&info->groups_sem);
7963 }
7964
7965 /*
7966  * btrfs_reserve_extent - entry point to the extent allocator. Tries to find a
7967  *                        hole that is at least as big as @num_bytes.
7968  *
7969  * @root           -    The root that will contain this extent
7970  *
7971  * @ram_bytes      -    The amount of space in ram that @num_bytes take. This
7972  *                      is used for accounting purposes. This value differs
7973  *                      from @num_bytes only in the case of compressed extents.
7974  *
7975  * @num_bytes      -    Number of bytes to allocate on-disk.
7976  *
7977  * @min_alloc_size -    Indicates the minimum amount of space that the
7978  *                      allocator should try to satisfy. In some cases
7979  *                      @num_bytes may be larger than what is required and if
7980  *                      the filesystem is fragmented then allocation fails.
7981  *                      However, the presence of @min_alloc_size gives a
7982  *                      chance to try and satisfy the smaller allocation.
7983  *
7984  * @empty_size     -    A hint that you plan on doing more COW. This is the
7985  *                      size in bytes the allocator should try to find free
7986  *                      next to the block it returns.  This is just a hint and
7987  *                      may be ignored by the allocator.
7988  *
7989  * @hint_byte      -    Hint to the allocator to start searching above the byte
7990  *                      address passed. It might be ignored.
7991  *
7992  * @ins            -    This key is modified to record the found hole. It will
7993  *                      have the following values:
7994  *                      ins->objectid == start position
7995  *                      ins->flags = BTRFS_EXTENT_ITEM_KEY
7996  *                      ins->offset == the size of the hole.
7997  *
7998  * @is_data        -    Boolean flag indicating whether an extent is
7999  *                      allocated for data (true) or metadata (false)
8000  *
8001  * @delalloc       -    Boolean flag indicating whether this allocation is for
8002  *                      delalloc or not. If 'true' data_rwsem of block groups
8003  *                      is going to be acquired.
8004  *
8005  *
8006  * Returns 0 when an allocation succeeded or < 0 when an error occurred. In
8007  * case -ENOSPC is returned then @ins->offset will contain the size of the
8008  * largest available hole the allocator managed to find.
8009  */
8010 int btrfs_reserve_extent(struct btrfs_root *root, u64 ram_bytes,
8011                          u64 num_bytes, u64 min_alloc_size,
8012                          u64 empty_size, u64 hint_byte,
8013                          struct btrfs_key *ins, int is_data, int delalloc)
8014 {
8015         struct btrfs_fs_info *fs_info = root->fs_info;
8016         bool final_tried = num_bytes == min_alloc_size;
8017         u64 flags;
8018         int ret;
8019
8020         flags = get_alloc_profile_by_root(root, is_data);
8021 again:
8022         WARN_ON(num_bytes < fs_info->sectorsize);
8023         ret = find_free_extent(fs_info, ram_bytes, num_bytes, empty_size,
8024                                hint_byte, ins, flags, delalloc);
8025         if (!ret && !is_data) {
8026                 btrfs_dec_block_group_reservations(fs_info, ins->objectid);
8027         } else if (ret == -ENOSPC) {
8028                 if (!final_tried && ins->offset) {
8029                         num_bytes = min(num_bytes >> 1, ins->offset);
8030                         num_bytes = round_down(num_bytes,
8031                                                fs_info->sectorsize);
8032                         num_bytes = max(num_bytes, min_alloc_size);
8033                         ram_bytes = num_bytes;
8034                         if (num_bytes == min_alloc_size)
8035                                 final_tried = true;
8036                         goto again;
8037                 } else if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
8038                         struct btrfs_space_info *sinfo;
8039
8040                         sinfo = __find_space_info(fs_info, flags);
8041                         btrfs_err(fs_info,
8042                                   "allocation failed flags %llu, wanted %llu",
8043                                   flags, num_bytes);
8044                         if (sinfo)
8045                                 dump_space_info(fs_info, sinfo, num_bytes, 1);
8046                 }
8047         }
8048
8049         return ret;
8050 }
8051
8052 static int __btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
8053                                         u64 start, u64 len,
8054                                         int pin, int delalloc)
8055 {
8056         struct btrfs_block_group_cache *cache;
8057         int ret = 0;
8058
8059         cache = btrfs_lookup_block_group(fs_info, start);
8060         if (!cache) {
8061                 btrfs_err(fs_info, "Unable to find block group for %llu",
8062                           start);
8063                 return -ENOSPC;
8064         }
8065
8066         if (pin)
8067                 pin_down_extent(fs_info, cache, start, len, 1);
8068         else {
8069                 if (btrfs_test_opt(fs_info, DISCARD))
8070                         ret = btrfs_discard_extent(fs_info, start, len, NULL);
8071                 btrfs_add_free_space(cache, start, len);
8072                 btrfs_free_reserved_bytes(cache, len, delalloc);
8073                 trace_btrfs_reserved_extent_free(fs_info, start, len);
8074         }
8075
8076         btrfs_put_block_group(cache);
8077         return ret;
8078 }
8079
8080 int btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
8081                                u64 start, u64 len, int delalloc)
8082 {
8083         return __btrfs_free_reserved_extent(fs_info, start, len, 0, delalloc);
8084 }
8085
8086 int btrfs_free_and_pin_reserved_extent(struct btrfs_fs_info *fs_info,
8087                                        u64 start, u64 len)
8088 {
8089         return __btrfs_free_reserved_extent(fs_info, start, len, 1, 0);
8090 }
8091
8092 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8093                                       struct btrfs_fs_info *fs_info,
8094                                       u64 parent, u64 root_objectid,
8095                                       u64 flags, u64 owner, u64 offset,
8096                                       struct btrfs_key *ins, int ref_mod)
8097 {
8098         int ret;
8099         struct btrfs_extent_item *extent_item;
8100         struct btrfs_extent_inline_ref *iref;
8101         struct btrfs_path *path;
8102         struct extent_buffer *leaf;
8103         int type;
8104         u32 size;
8105
8106         if (parent > 0)
8107                 type = BTRFS_SHARED_DATA_REF_KEY;
8108         else
8109                 type = BTRFS_EXTENT_DATA_REF_KEY;
8110
8111         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
8112
8113         path = btrfs_alloc_path();
8114         if (!path)
8115                 return -ENOMEM;
8116
8117         path->leave_spinning = 1;
8118         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
8119                                       ins, size);
8120         if (ret) {
8121                 btrfs_free_path(path);
8122                 return ret;
8123         }
8124
8125         leaf = path->nodes[0];
8126         extent_item = btrfs_item_ptr(leaf, path->slots[0],
8127                                      struct btrfs_extent_item);
8128         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
8129         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8130         btrfs_set_extent_flags(leaf, extent_item,
8131                                flags | BTRFS_EXTENT_FLAG_DATA);
8132
8133         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8134         btrfs_set_extent_inline_ref_type(leaf, iref, type);
8135         if (parent > 0) {
8136                 struct btrfs_shared_data_ref *ref;
8137                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
8138                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
8139                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
8140         } else {
8141                 struct btrfs_extent_data_ref *ref;
8142                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
8143                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
8144                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
8145                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
8146                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
8147         }
8148
8149         btrfs_mark_buffer_dirty(path->nodes[0]);
8150         btrfs_free_path(path);
8151
8152         ret = remove_from_free_space_tree(trans, fs_info, ins->objectid,
8153                                           ins->offset);
8154         if (ret)
8155                 return ret;
8156
8157         ret = update_block_group(trans, fs_info, ins->objectid, ins->offset, 1);
8158         if (ret) { /* -ENOENT, logic error */
8159                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8160                         ins->objectid, ins->offset);
8161                 BUG();
8162         }
8163         trace_btrfs_reserved_extent_alloc(fs_info, ins->objectid, ins->offset);
8164         return ret;
8165 }
8166
8167 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
8168                                      struct btrfs_fs_info *fs_info,
8169                                      u64 parent, u64 root_objectid,
8170                                      u64 flags, struct btrfs_disk_key *key,
8171                                      int level, struct btrfs_key *ins)
8172 {
8173         int ret;
8174         struct btrfs_extent_item *extent_item;
8175         struct btrfs_tree_block_info *block_info;
8176         struct btrfs_extent_inline_ref *iref;
8177         struct btrfs_path *path;
8178         struct extent_buffer *leaf;
8179         u32 size = sizeof(*extent_item) + sizeof(*iref);
8180         u64 num_bytes = ins->offset;
8181         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
8182
8183         if (!skinny_metadata)
8184                 size += sizeof(*block_info);
8185
8186         path = btrfs_alloc_path();
8187         if (!path) {
8188                 btrfs_free_and_pin_reserved_extent(fs_info, ins->objectid,
8189                                                    fs_info->nodesize);
8190                 return -ENOMEM;
8191         }
8192
8193         path->leave_spinning = 1;
8194         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
8195                                       ins, size);
8196         if (ret) {
8197                 btrfs_free_path(path);
8198                 btrfs_free_and_pin_reserved_extent(fs_info, ins->objectid,
8199                                                    fs_info->nodesize);
8200                 return ret;
8201         }
8202
8203         leaf = path->nodes[0];
8204         extent_item = btrfs_item_ptr(leaf, path->slots[0],
8205                                      struct btrfs_extent_item);
8206         btrfs_set_extent_refs(leaf, extent_item, 1);
8207         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8208         btrfs_set_extent_flags(leaf, extent_item,
8209                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
8210
8211         if (skinny_metadata) {
8212                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8213                 num_bytes = fs_info->nodesize;
8214         } else {
8215                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
8216                 btrfs_set_tree_block_key(leaf, block_info, key);
8217                 btrfs_set_tree_block_level(leaf, block_info, level);
8218                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
8219         }
8220
8221         if (parent > 0) {
8222                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
8223                 btrfs_set_extent_inline_ref_type(leaf, iref,
8224                                                  BTRFS_SHARED_BLOCK_REF_KEY);
8225                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
8226         } else {
8227                 btrfs_set_extent_inline_ref_type(leaf, iref,
8228                                                  BTRFS_TREE_BLOCK_REF_KEY);
8229                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
8230         }
8231
8232         btrfs_mark_buffer_dirty(leaf);
8233         btrfs_free_path(path);
8234
8235         ret = remove_from_free_space_tree(trans, fs_info, ins->objectid,
8236                                           num_bytes);
8237         if (ret)
8238                 return ret;
8239
8240         ret = update_block_group(trans, fs_info, ins->objectid,
8241                                  fs_info->nodesize, 1);
8242         if (ret) { /* -ENOENT, logic error */
8243                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8244                         ins->objectid, ins->offset);
8245                 BUG();
8246         }
8247
8248         trace_btrfs_reserved_extent_alloc(fs_info, ins->objectid,
8249                                           fs_info->nodesize);
8250         return ret;
8251 }
8252
8253 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8254                                      struct btrfs_root *root, u64 owner,
8255                                      u64 offset, u64 ram_bytes,
8256                                      struct btrfs_key *ins)
8257 {
8258         struct btrfs_fs_info *fs_info = root->fs_info;
8259         int ret;
8260
8261         BUG_ON(root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
8262
8263         btrfs_ref_tree_mod(root, ins->objectid, ins->offset, 0,
8264                            root->root_key.objectid, owner, offset,
8265                            BTRFS_ADD_DELAYED_EXTENT);
8266
8267         ret = btrfs_add_delayed_data_ref(fs_info, trans, ins->objectid,
8268                                          ins->offset, 0,
8269                                          root->root_key.objectid, owner,
8270                                          offset, ram_bytes,
8271                                          BTRFS_ADD_DELAYED_EXTENT, NULL, NULL);
8272         return ret;
8273 }
8274
8275 /*
8276  * this is used by the tree logging recovery code.  It records that
8277  * an extent has been allocated and makes sure to clear the free
8278  * space cache bits as well
8279  */
8280 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
8281                                    struct btrfs_fs_info *fs_info,
8282                                    u64 root_objectid, u64 owner, u64 offset,
8283                                    struct btrfs_key *ins)
8284 {
8285         int ret;
8286         struct btrfs_block_group_cache *block_group;
8287         struct btrfs_space_info *space_info;
8288
8289         /*
8290          * Mixed block groups will exclude before processing the log so we only
8291          * need to do the exclude dance if this fs isn't mixed.
8292          */
8293         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS)) {
8294                 ret = __exclude_logged_extent(fs_info, ins->objectid,
8295                                               ins->offset);
8296                 if (ret)
8297                         return ret;
8298         }
8299
8300         block_group = btrfs_lookup_block_group(fs_info, ins->objectid);
8301         if (!block_group)
8302                 return -EINVAL;
8303
8304         space_info = block_group->space_info;
8305         spin_lock(&space_info->lock);
8306         spin_lock(&block_group->lock);
8307         space_info->bytes_reserved += ins->offset;
8308         block_group->reserved += ins->offset;
8309         spin_unlock(&block_group->lock);
8310         spin_unlock(&space_info->lock);
8311
8312         ret = alloc_reserved_file_extent(trans, fs_info, 0, root_objectid,
8313                                          0, owner, offset, ins, 1);
8314         btrfs_put_block_group(block_group);
8315         return ret;
8316 }
8317
8318 static struct extent_buffer *
8319 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
8320                       u64 bytenr, int level)
8321 {
8322         struct btrfs_fs_info *fs_info = root->fs_info;
8323         struct extent_buffer *buf;
8324
8325         buf = btrfs_find_create_tree_block(fs_info, bytenr);
8326         if (IS_ERR(buf))
8327                 return buf;
8328
8329         btrfs_set_header_generation(buf, trans->transid);
8330         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
8331         btrfs_tree_lock(buf);
8332         clean_tree_block(fs_info, buf);
8333         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
8334
8335         btrfs_set_lock_blocking(buf);
8336         set_extent_buffer_uptodate(buf);
8337
8338         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
8339                 buf->log_index = root->log_transid % 2;
8340                 /*
8341                  * we allow two log transactions at a time, use different
8342                  * EXENT bit to differentiate dirty pages.
8343                  */
8344                 if (buf->log_index == 0)
8345                         set_extent_dirty(&root->dirty_log_pages, buf->start,
8346                                         buf->start + buf->len - 1, GFP_NOFS);
8347                 else
8348                         set_extent_new(&root->dirty_log_pages, buf->start,
8349                                         buf->start + buf->len - 1);
8350         } else {
8351                 buf->log_index = -1;
8352                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
8353                          buf->start + buf->len - 1, GFP_NOFS);
8354         }
8355         trans->dirty = true;
8356         /* this returns a buffer locked for blocking */
8357         return buf;
8358 }
8359
8360 static struct btrfs_block_rsv *
8361 use_block_rsv(struct btrfs_trans_handle *trans,
8362               struct btrfs_root *root, u32 blocksize)
8363 {
8364         struct btrfs_fs_info *fs_info = root->fs_info;
8365         struct btrfs_block_rsv *block_rsv;
8366         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
8367         int ret;
8368         bool global_updated = false;
8369
8370         block_rsv = get_block_rsv(trans, root);
8371
8372         if (unlikely(block_rsv->size == 0))
8373                 goto try_reserve;
8374 again:
8375         ret = block_rsv_use_bytes(block_rsv, blocksize);
8376         if (!ret)
8377                 return block_rsv;
8378
8379         if (block_rsv->failfast)
8380                 return ERR_PTR(ret);
8381
8382         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
8383                 global_updated = true;
8384                 update_global_block_rsv(fs_info);
8385                 goto again;
8386         }
8387
8388         if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
8389                 static DEFINE_RATELIMIT_STATE(_rs,
8390                                 DEFAULT_RATELIMIT_INTERVAL * 10,
8391                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
8392                 if (__ratelimit(&_rs))
8393                         WARN(1, KERN_DEBUG
8394                                 "BTRFS: block rsv returned %d\n", ret);
8395         }
8396 try_reserve:
8397         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
8398                                      BTRFS_RESERVE_NO_FLUSH);
8399         if (!ret)
8400                 return block_rsv;
8401         /*
8402          * If we couldn't reserve metadata bytes try and use some from
8403          * the global reserve if its space type is the same as the global
8404          * reservation.
8405          */
8406         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
8407             block_rsv->space_info == global_rsv->space_info) {
8408                 ret = block_rsv_use_bytes(global_rsv, blocksize);
8409                 if (!ret)
8410                         return global_rsv;
8411         }
8412         return ERR_PTR(ret);
8413 }
8414
8415 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
8416                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
8417 {
8418         block_rsv_add_bytes(block_rsv, blocksize, 0);
8419         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
8420 }
8421
8422 /*
8423  * finds a free extent and does all the dirty work required for allocation
8424  * returns the tree buffer or an ERR_PTR on error.
8425  */
8426 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
8427                                              struct btrfs_root *root,
8428                                              u64 parent, u64 root_objectid,
8429                                              const struct btrfs_disk_key *key,
8430                                              int level, u64 hint,
8431                                              u64 empty_size)
8432 {
8433         struct btrfs_fs_info *fs_info = root->fs_info;
8434         struct btrfs_key ins;
8435         struct btrfs_block_rsv *block_rsv;
8436         struct extent_buffer *buf;
8437         struct btrfs_delayed_extent_op *extent_op;
8438         u64 flags = 0;
8439         int ret;
8440         u32 blocksize = fs_info->nodesize;
8441         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
8442
8443 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
8444         if (btrfs_is_testing(fs_info)) {
8445                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
8446                                             level);
8447                 if (!IS_ERR(buf))
8448                         root->alloc_bytenr += blocksize;
8449                 return buf;
8450         }
8451 #endif
8452
8453         block_rsv = use_block_rsv(trans, root, blocksize);
8454         if (IS_ERR(block_rsv))
8455                 return ERR_CAST(block_rsv);
8456
8457         ret = btrfs_reserve_extent(root, blocksize, blocksize, blocksize,
8458                                    empty_size, hint, &ins, 0, 0);
8459         if (ret)
8460                 goto out_unuse;
8461
8462         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level);
8463         if (IS_ERR(buf)) {
8464                 ret = PTR_ERR(buf);
8465                 goto out_free_reserved;
8466         }
8467
8468         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
8469                 if (parent == 0)
8470                         parent = ins.objectid;
8471                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
8472         } else
8473                 BUG_ON(parent > 0);
8474
8475         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
8476                 extent_op = btrfs_alloc_delayed_extent_op();
8477                 if (!extent_op) {
8478                         ret = -ENOMEM;
8479                         goto out_free_buf;
8480                 }
8481                 if (key)
8482                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
8483                 else
8484                         memset(&extent_op->key, 0, sizeof(extent_op->key));
8485                 extent_op->flags_to_set = flags;
8486                 extent_op->update_key = skinny_metadata ? false : true;
8487                 extent_op->update_flags = true;
8488                 extent_op->is_data = false;
8489                 extent_op->level = level;
8490
8491                 btrfs_ref_tree_mod(root, ins.objectid, ins.offset, parent,
8492                                    root_objectid, level, 0,
8493                                    BTRFS_ADD_DELAYED_EXTENT);
8494                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, ins.objectid,
8495                                                  ins.offset, parent,
8496                                                  root_objectid, level,
8497                                                  BTRFS_ADD_DELAYED_EXTENT,
8498                                                  extent_op, NULL, NULL);
8499                 if (ret)
8500                         goto out_free_delayed;
8501         }
8502         return buf;
8503
8504 out_free_delayed:
8505         btrfs_free_delayed_extent_op(extent_op);
8506 out_free_buf:
8507         free_extent_buffer(buf);
8508 out_free_reserved:
8509         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 0);
8510 out_unuse:
8511         unuse_block_rsv(fs_info, block_rsv, blocksize);
8512         return ERR_PTR(ret);
8513 }
8514
8515 struct walk_control {
8516         u64 refs[BTRFS_MAX_LEVEL];
8517         u64 flags[BTRFS_MAX_LEVEL];
8518         struct btrfs_key update_progress;
8519         int stage;
8520         int level;
8521         int shared_level;
8522         int update_ref;
8523         int keep_locks;
8524         int reada_slot;
8525         int reada_count;
8526         int for_reloc;
8527 };
8528
8529 #define DROP_REFERENCE  1
8530 #define UPDATE_BACKREF  2
8531
8532 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
8533                                      struct btrfs_root *root,
8534                                      struct walk_control *wc,
8535                                      struct btrfs_path *path)
8536 {
8537         struct btrfs_fs_info *fs_info = root->fs_info;
8538         u64 bytenr;
8539         u64 generation;
8540         u64 refs;
8541         u64 flags;
8542         u32 nritems;
8543         struct btrfs_key key;
8544         struct extent_buffer *eb;
8545         int ret;
8546         int slot;
8547         int nread = 0;
8548
8549         if (path->slots[wc->level] < wc->reada_slot) {
8550                 wc->reada_count = wc->reada_count * 2 / 3;
8551                 wc->reada_count = max(wc->reada_count, 2);
8552         } else {
8553                 wc->reada_count = wc->reada_count * 3 / 2;
8554                 wc->reada_count = min_t(int, wc->reada_count,
8555                                         BTRFS_NODEPTRS_PER_BLOCK(fs_info));
8556         }
8557
8558         eb = path->nodes[wc->level];
8559         nritems = btrfs_header_nritems(eb);
8560
8561         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
8562                 if (nread >= wc->reada_count)
8563                         break;
8564
8565                 cond_resched();
8566                 bytenr = btrfs_node_blockptr(eb, slot);
8567                 generation = btrfs_node_ptr_generation(eb, slot);
8568
8569                 if (slot == path->slots[wc->level])
8570                         goto reada;
8571
8572                 if (wc->stage == UPDATE_BACKREF &&
8573                     generation <= root->root_key.offset)
8574                         continue;
8575
8576                 /* We don't lock the tree block, it's OK to be racy here */
8577                 ret = btrfs_lookup_extent_info(trans, fs_info, bytenr,
8578                                                wc->level - 1, 1, &refs,
8579                                                &flags);
8580                 /* We don't care about errors in readahead. */
8581                 if (ret < 0)
8582                         continue;
8583                 BUG_ON(refs == 0);
8584
8585                 if (wc->stage == DROP_REFERENCE) {
8586                         if (refs == 1)
8587                                 goto reada;
8588
8589                         if (wc->level == 1 &&
8590                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8591                                 continue;
8592                         if (!wc->update_ref ||
8593                             generation <= root->root_key.offset)
8594                                 continue;
8595                         btrfs_node_key_to_cpu(eb, &key, slot);
8596                         ret = btrfs_comp_cpu_keys(&key,
8597                                                   &wc->update_progress);
8598                         if (ret < 0)
8599                                 continue;
8600                 } else {
8601                         if (wc->level == 1 &&
8602                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8603                                 continue;
8604                 }
8605 reada:
8606                 readahead_tree_block(fs_info, bytenr);
8607                 nread++;
8608         }
8609         wc->reada_slot = slot;
8610 }
8611
8612 /*
8613  * helper to process tree block while walking down the tree.
8614  *
8615  * when wc->stage == UPDATE_BACKREF, this function updates
8616  * back refs for pointers in the block.
8617  *
8618  * NOTE: return value 1 means we should stop walking down.
8619  */
8620 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
8621                                    struct btrfs_root *root,
8622                                    struct btrfs_path *path,
8623                                    struct walk_control *wc, int lookup_info)
8624 {
8625         struct btrfs_fs_info *fs_info = root->fs_info;
8626         int level = wc->level;
8627         struct extent_buffer *eb = path->nodes[level];
8628         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8629         int ret;
8630
8631         if (wc->stage == UPDATE_BACKREF &&
8632             btrfs_header_owner(eb) != root->root_key.objectid)
8633                 return 1;
8634
8635         /*
8636          * when reference count of tree block is 1, it won't increase
8637          * again. once full backref flag is set, we never clear it.
8638          */
8639         if (lookup_info &&
8640             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
8641              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
8642                 BUG_ON(!path->locks[level]);
8643                 ret = btrfs_lookup_extent_info(trans, fs_info,
8644                                                eb->start, level, 1,
8645                                                &wc->refs[level],
8646                                                &wc->flags[level]);
8647                 BUG_ON(ret == -ENOMEM);
8648                 if (ret)
8649                         return ret;
8650                 BUG_ON(wc->refs[level] == 0);
8651         }
8652
8653         if (wc->stage == DROP_REFERENCE) {
8654                 if (wc->refs[level] > 1)
8655                         return 1;
8656
8657                 if (path->locks[level] && !wc->keep_locks) {
8658                         btrfs_tree_unlock_rw(eb, path->locks[level]);
8659                         path->locks[level] = 0;
8660                 }
8661                 return 0;
8662         }
8663
8664         /* wc->stage == UPDATE_BACKREF */
8665         if (!(wc->flags[level] & flag)) {
8666                 BUG_ON(!path->locks[level]);
8667                 ret = btrfs_inc_ref(trans, root, eb, 1);
8668                 BUG_ON(ret); /* -ENOMEM */
8669                 ret = btrfs_dec_ref(trans, root, eb, 0);
8670                 BUG_ON(ret); /* -ENOMEM */
8671                 ret = btrfs_set_disk_extent_flags(trans, fs_info, eb->start,
8672                                                   eb->len, flag,
8673                                                   btrfs_header_level(eb), 0);
8674                 BUG_ON(ret); /* -ENOMEM */
8675                 wc->flags[level] |= flag;
8676         }
8677
8678         /*
8679          * the block is shared by multiple trees, so it's not good to
8680          * keep the tree lock
8681          */
8682         if (path->locks[level] && level > 0) {
8683                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8684                 path->locks[level] = 0;
8685         }
8686         return 0;
8687 }
8688
8689 /*
8690  * helper to process tree block pointer.
8691  *
8692  * when wc->stage == DROP_REFERENCE, this function checks
8693  * reference count of the block pointed to. if the block
8694  * is shared and we need update back refs for the subtree
8695  * rooted at the block, this function changes wc->stage to
8696  * UPDATE_BACKREF. if the block is shared and there is no
8697  * need to update back, this function drops the reference
8698  * to the block.
8699  *
8700  * NOTE: return value 1 means we should stop walking down.
8701  */
8702 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
8703                                  struct btrfs_root *root,
8704                                  struct btrfs_path *path,
8705                                  struct walk_control *wc, int *lookup_info)
8706 {
8707         struct btrfs_fs_info *fs_info = root->fs_info;
8708         u64 bytenr;
8709         u64 generation;
8710         u64 parent;
8711         u32 blocksize;
8712         struct btrfs_key key;
8713         struct btrfs_key first_key;
8714         struct extent_buffer *next;
8715         int level = wc->level;
8716         int reada = 0;
8717         int ret = 0;
8718         bool need_account = false;
8719
8720         generation = btrfs_node_ptr_generation(path->nodes[level],
8721                                                path->slots[level]);
8722         /*
8723          * if the lower level block was created before the snapshot
8724          * was created, we know there is no need to update back refs
8725          * for the subtree
8726          */
8727         if (wc->stage == UPDATE_BACKREF &&
8728             generation <= root->root_key.offset) {
8729                 *lookup_info = 1;
8730                 return 1;
8731         }
8732
8733         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
8734         btrfs_node_key_to_cpu(path->nodes[level], &first_key,
8735                               path->slots[level]);
8736         blocksize = fs_info->nodesize;
8737
8738         next = find_extent_buffer(fs_info, bytenr);
8739         if (!next) {
8740                 next = btrfs_find_create_tree_block(fs_info, bytenr);
8741                 if (IS_ERR(next))
8742                         return PTR_ERR(next);
8743
8744                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
8745                                                level - 1);
8746                 reada = 1;
8747         }
8748         btrfs_tree_lock(next);
8749         btrfs_set_lock_blocking(next);
8750
8751         ret = btrfs_lookup_extent_info(trans, fs_info, bytenr, level - 1, 1,
8752                                        &wc->refs[level - 1],
8753                                        &wc->flags[level - 1]);
8754         if (ret < 0)
8755                 goto out_unlock;
8756
8757         if (unlikely(wc->refs[level - 1] == 0)) {
8758                 btrfs_err(fs_info, "Missing references.");
8759                 ret = -EIO;
8760                 goto out_unlock;
8761         }
8762         *lookup_info = 0;
8763
8764         if (wc->stage == DROP_REFERENCE) {
8765                 if (wc->refs[level - 1] > 1) {
8766                         need_account = true;
8767                         if (level == 1 &&
8768                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8769                                 goto skip;
8770
8771                         if (!wc->update_ref ||
8772                             generation <= root->root_key.offset)
8773                                 goto skip;
8774
8775                         btrfs_node_key_to_cpu(path->nodes[level], &key,
8776                                               path->slots[level]);
8777                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
8778                         if (ret < 0)
8779                                 goto skip;
8780
8781                         wc->stage = UPDATE_BACKREF;
8782                         wc->shared_level = level - 1;
8783                 }
8784         } else {
8785                 if (level == 1 &&
8786                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8787                         goto skip;
8788         }
8789
8790         if (!btrfs_buffer_uptodate(next, generation, 0)) {
8791                 btrfs_tree_unlock(next);
8792                 free_extent_buffer(next);
8793                 next = NULL;
8794                 *lookup_info = 1;
8795         }
8796
8797         if (!next) {
8798                 if (reada && level == 1)
8799                         reada_walk_down(trans, root, wc, path);
8800                 next = read_tree_block(fs_info, bytenr, generation, level - 1,
8801                                        &first_key);
8802                 if (IS_ERR(next)) {
8803                         return PTR_ERR(next);
8804                 } else if (!extent_buffer_uptodate(next)) {
8805                         free_extent_buffer(next);
8806                         return -EIO;
8807                 }
8808                 btrfs_tree_lock(next);
8809                 btrfs_set_lock_blocking(next);
8810         }
8811
8812         level--;
8813         ASSERT(level == btrfs_header_level(next));
8814         if (level != btrfs_header_level(next)) {
8815                 btrfs_err(root->fs_info, "mismatched level");
8816                 ret = -EIO;
8817                 goto out_unlock;
8818         }
8819         path->nodes[level] = next;
8820         path->slots[level] = 0;
8821         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8822         wc->level = level;
8823         if (wc->level == 1)
8824                 wc->reada_slot = 0;
8825         return 0;
8826 skip:
8827         wc->refs[level - 1] = 0;
8828         wc->flags[level - 1] = 0;
8829         if (wc->stage == DROP_REFERENCE) {
8830                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
8831                         parent = path->nodes[level]->start;
8832                 } else {
8833                         ASSERT(root->root_key.objectid ==
8834                                btrfs_header_owner(path->nodes[level]));
8835                         if (root->root_key.objectid !=
8836                             btrfs_header_owner(path->nodes[level])) {
8837                                 btrfs_err(root->fs_info,
8838                                                 "mismatched block owner");
8839                                 ret = -EIO;
8840                                 goto out_unlock;
8841                         }
8842                         parent = 0;
8843                 }
8844
8845                 if (need_account) {
8846                         ret = btrfs_qgroup_trace_subtree(trans, root, next,
8847                                                          generation, level - 1);
8848                         if (ret) {
8849                                 btrfs_err_rl(fs_info,
8850                                              "Error %d accounting shared subtree. Quota is out of sync, rescan required.",
8851                                              ret);
8852                         }
8853                 }
8854                 ret = btrfs_free_extent(trans, root, bytenr, blocksize,
8855                                         parent, root->root_key.objectid,
8856                                         level - 1, 0);
8857                 if (ret)
8858                         goto out_unlock;
8859         }
8860
8861         *lookup_info = 1;
8862         ret = 1;
8863
8864 out_unlock:
8865         btrfs_tree_unlock(next);
8866         free_extent_buffer(next);
8867
8868         return ret;
8869 }
8870
8871 /*
8872  * helper to process tree block while walking up the tree.
8873  *
8874  * when wc->stage == DROP_REFERENCE, this function drops
8875  * reference count on the block.
8876  *
8877  * when wc->stage == UPDATE_BACKREF, this function changes
8878  * wc->stage back to DROP_REFERENCE if we changed wc->stage
8879  * to UPDATE_BACKREF previously while processing the block.
8880  *
8881  * NOTE: return value 1 means we should stop walking up.
8882  */
8883 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
8884                                  struct btrfs_root *root,
8885                                  struct btrfs_path *path,
8886                                  struct walk_control *wc)
8887 {
8888         struct btrfs_fs_info *fs_info = root->fs_info;
8889         int ret;
8890         int level = wc->level;
8891         struct extent_buffer *eb = path->nodes[level];
8892         u64 parent = 0;
8893
8894         if (wc->stage == UPDATE_BACKREF) {
8895                 BUG_ON(wc->shared_level < level);
8896                 if (level < wc->shared_level)
8897                         goto out;
8898
8899                 ret = find_next_key(path, level + 1, &wc->update_progress);
8900                 if (ret > 0)
8901                         wc->update_ref = 0;
8902
8903                 wc->stage = DROP_REFERENCE;
8904                 wc->shared_level = -1;
8905                 path->slots[level] = 0;
8906
8907                 /*
8908                  * check reference count again if the block isn't locked.
8909                  * we should start walking down the tree again if reference
8910                  * count is one.
8911                  */
8912                 if (!path->locks[level]) {
8913                         BUG_ON(level == 0);
8914                         btrfs_tree_lock(eb);
8915                         btrfs_set_lock_blocking(eb);
8916                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8917
8918                         ret = btrfs_lookup_extent_info(trans, fs_info,
8919                                                        eb->start, level, 1,
8920                                                        &wc->refs[level],
8921                                                        &wc->flags[level]);
8922                         if (ret < 0) {
8923                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8924                                 path->locks[level] = 0;
8925                                 return ret;
8926                         }
8927                         BUG_ON(wc->refs[level] == 0);
8928                         if (wc->refs[level] == 1) {
8929                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8930                                 path->locks[level] = 0;
8931                                 return 1;
8932                         }
8933                 }
8934         }
8935
8936         /* wc->stage == DROP_REFERENCE */
8937         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
8938
8939         if (wc->refs[level] == 1) {
8940                 if (level == 0) {
8941                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8942                                 ret = btrfs_dec_ref(trans, root, eb, 1);
8943                         else
8944                                 ret = btrfs_dec_ref(trans, root, eb, 0);
8945                         BUG_ON(ret); /* -ENOMEM */
8946                         ret = btrfs_qgroup_trace_leaf_items(trans, fs_info, eb);
8947                         if (ret) {
8948                                 btrfs_err_rl(fs_info,
8949                                              "error %d accounting leaf items. Quota is out of sync, rescan required.",
8950                                              ret);
8951                         }
8952                 }
8953                 /* make block locked assertion in clean_tree_block happy */
8954                 if (!path->locks[level] &&
8955                     btrfs_header_generation(eb) == trans->transid) {
8956                         btrfs_tree_lock(eb);
8957                         btrfs_set_lock_blocking(eb);
8958                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8959                 }
8960                 clean_tree_block(fs_info, eb);
8961         }
8962
8963         if (eb == root->node) {
8964                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8965                         parent = eb->start;
8966                 else
8967                         BUG_ON(root->root_key.objectid !=
8968                                btrfs_header_owner(eb));
8969         } else {
8970                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8971                         parent = path->nodes[level + 1]->start;
8972                 else
8973                         BUG_ON(root->root_key.objectid !=
8974                                btrfs_header_owner(path->nodes[level + 1]));
8975         }
8976
8977         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
8978 out:
8979         wc->refs[level] = 0;
8980         wc->flags[level] = 0;
8981         return 0;
8982 }
8983
8984 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
8985                                    struct btrfs_root *root,
8986                                    struct btrfs_path *path,
8987                                    struct walk_control *wc)
8988 {
8989         int level = wc->level;
8990         int lookup_info = 1;
8991         int ret;
8992
8993         while (level >= 0) {
8994                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
8995                 if (ret > 0)
8996                         break;
8997
8998                 if (level == 0)
8999                         break;
9000
9001                 if (path->slots[level] >=
9002                     btrfs_header_nritems(path->nodes[level]))
9003                         break;
9004
9005                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
9006                 if (ret > 0) {
9007                         path->slots[level]++;
9008                         continue;
9009                 } else if (ret < 0)
9010                         return ret;
9011                 level = wc->level;
9012         }
9013         return 0;
9014 }
9015
9016 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
9017                                  struct btrfs_root *root,
9018                                  struct btrfs_path *path,
9019                                  struct walk_control *wc, int max_level)
9020 {
9021         int level = wc->level;
9022         int ret;
9023
9024         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
9025         while (level < max_level && path->nodes[level]) {
9026                 wc->level = level;
9027                 if (path->slots[level] + 1 <
9028                     btrfs_header_nritems(path->nodes[level])) {
9029                         path->slots[level]++;
9030                         return 0;
9031                 } else {
9032                         ret = walk_up_proc(trans, root, path, wc);
9033                         if (ret > 0)
9034                                 return 0;
9035
9036                         if (path->locks[level]) {
9037                                 btrfs_tree_unlock_rw(path->nodes[level],
9038                                                      path->locks[level]);
9039                                 path->locks[level] = 0;
9040                         }
9041                         free_extent_buffer(path->nodes[level]);
9042                         path->nodes[level] = NULL;
9043                         level++;
9044                 }
9045         }
9046         return 1;
9047 }
9048
9049 /*
9050  * drop a subvolume tree.
9051  *
9052  * this function traverses the tree freeing any blocks that only
9053  * referenced by the tree.
9054  *
9055  * when a shared tree block is found. this function decreases its
9056  * reference count by one. if update_ref is true, this function
9057  * also make sure backrefs for the shared block and all lower level
9058  * blocks are properly updated.
9059  *
9060  * If called with for_reloc == 0, may exit early with -EAGAIN
9061  */
9062 int btrfs_drop_snapshot(struct btrfs_root *root,
9063                          struct btrfs_block_rsv *block_rsv, int update_ref,
9064                          int for_reloc)
9065 {
9066         struct btrfs_fs_info *fs_info = root->fs_info;
9067         struct btrfs_path *path;
9068         struct btrfs_trans_handle *trans;
9069         struct btrfs_root *tree_root = fs_info->tree_root;
9070         struct btrfs_root_item *root_item = &root->root_item;
9071         struct walk_control *wc;
9072         struct btrfs_key key;
9073         int err = 0;
9074         int ret;
9075         int level;
9076         bool root_dropped = false;
9077
9078         btrfs_debug(fs_info, "Drop subvolume %llu", root->objectid);
9079
9080         path = btrfs_alloc_path();
9081         if (!path) {
9082                 err = -ENOMEM;
9083                 goto out;
9084         }
9085
9086         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9087         if (!wc) {
9088                 btrfs_free_path(path);
9089                 err = -ENOMEM;
9090                 goto out;
9091         }
9092
9093         trans = btrfs_start_transaction(tree_root, 0);
9094         if (IS_ERR(trans)) {
9095                 err = PTR_ERR(trans);
9096                 goto out_free;
9097         }
9098
9099         if (block_rsv)
9100                 trans->block_rsv = block_rsv;
9101
9102         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
9103                 level = btrfs_header_level(root->node);
9104                 path->nodes[level] = btrfs_lock_root_node(root);
9105                 btrfs_set_lock_blocking(path->nodes[level]);
9106                 path->slots[level] = 0;
9107                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9108                 memset(&wc->update_progress, 0,
9109                        sizeof(wc->update_progress));
9110         } else {
9111                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
9112                 memcpy(&wc->update_progress, &key,
9113                        sizeof(wc->update_progress));
9114
9115                 level = root_item->drop_level;
9116                 BUG_ON(level == 0);
9117                 path->lowest_level = level;
9118                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
9119                 path->lowest_level = 0;
9120                 if (ret < 0) {
9121                         err = ret;
9122                         goto out_end_trans;
9123                 }
9124                 WARN_ON(ret > 0);
9125
9126                 /*
9127                  * unlock our path, this is safe because only this
9128                  * function is allowed to delete this snapshot
9129                  */
9130                 btrfs_unlock_up_safe(path, 0);
9131
9132                 level = btrfs_header_level(root->node);
9133                 while (1) {
9134                         btrfs_tree_lock(path->nodes[level]);
9135                         btrfs_set_lock_blocking(path->nodes[level]);
9136                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9137
9138                         ret = btrfs_lookup_extent_info(trans, fs_info,
9139                                                 path->nodes[level]->start,
9140                                                 level, 1, &wc->refs[level],
9141                                                 &wc->flags[level]);
9142                         if (ret < 0) {
9143                                 err = ret;
9144                                 goto out_end_trans;
9145                         }
9146                         BUG_ON(wc->refs[level] == 0);
9147
9148                         if (level == root_item->drop_level)
9149                                 break;
9150
9151                         btrfs_tree_unlock(path->nodes[level]);
9152                         path->locks[level] = 0;
9153                         WARN_ON(wc->refs[level] != 1);
9154                         level--;
9155                 }
9156         }
9157
9158         wc->level = level;
9159         wc->shared_level = -1;
9160         wc->stage = DROP_REFERENCE;
9161         wc->update_ref = update_ref;
9162         wc->keep_locks = 0;
9163         wc->for_reloc = for_reloc;
9164         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
9165
9166         while (1) {
9167
9168                 ret = walk_down_tree(trans, root, path, wc);
9169                 if (ret < 0) {
9170                         err = ret;
9171                         break;
9172                 }
9173
9174                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
9175                 if (ret < 0) {
9176                         err = ret;
9177                         break;
9178                 }
9179
9180                 if (ret > 0) {
9181                         BUG_ON(wc->stage != DROP_REFERENCE);
9182                         break;
9183                 }
9184
9185                 if (wc->stage == DROP_REFERENCE) {
9186                         level = wc->level;
9187                         btrfs_node_key(path->nodes[level],
9188                                        &root_item->drop_progress,
9189                                        path->slots[level]);
9190                         root_item->drop_level = level;
9191                 }
9192
9193                 BUG_ON(wc->level == 0);
9194                 if (btrfs_should_end_transaction(trans) ||
9195                     (!for_reloc && btrfs_need_cleaner_sleep(fs_info))) {
9196                         ret = btrfs_update_root(trans, tree_root,
9197                                                 &root->root_key,
9198                                                 root_item);
9199                         if (ret) {
9200                                 btrfs_abort_transaction(trans, ret);
9201                                 err = ret;
9202                                 goto out_end_trans;
9203                         }
9204
9205                         btrfs_end_transaction_throttle(trans);
9206                         if (!for_reloc && btrfs_need_cleaner_sleep(fs_info)) {
9207                                 btrfs_debug(fs_info,
9208                                             "drop snapshot early exit");
9209                                 err = -EAGAIN;
9210                                 goto out_free;
9211                         }
9212
9213                         trans = btrfs_start_transaction(tree_root, 0);
9214                         if (IS_ERR(trans)) {
9215                                 err = PTR_ERR(trans);
9216                                 goto out_free;
9217                         }
9218                         if (block_rsv)
9219                                 trans->block_rsv = block_rsv;
9220                 }
9221         }
9222         btrfs_release_path(path);
9223         if (err)
9224                 goto out_end_trans;
9225
9226         ret = btrfs_del_root(trans, fs_info, &root->root_key);
9227         if (ret) {
9228                 btrfs_abort_transaction(trans, ret);
9229                 err = ret;
9230                 goto out_end_trans;
9231         }
9232
9233         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
9234                 ret = btrfs_find_root(tree_root, &root->root_key, path,
9235                                       NULL, NULL);
9236                 if (ret < 0) {
9237                         btrfs_abort_transaction(trans, ret);
9238                         err = ret;
9239                         goto out_end_trans;
9240                 } else if (ret > 0) {
9241                         /* if we fail to delete the orphan item this time
9242                          * around, it'll get picked up the next time.
9243                          *
9244                          * The most common failure here is just -ENOENT.
9245                          */
9246                         btrfs_del_orphan_item(trans, tree_root,
9247                                               root->root_key.objectid);
9248                 }
9249         }
9250
9251         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
9252                 btrfs_add_dropped_root(trans, root);
9253         } else {
9254                 free_extent_buffer(root->node);
9255                 free_extent_buffer(root->commit_root);
9256                 btrfs_put_fs_root(root);
9257         }
9258         root_dropped = true;
9259 out_end_trans:
9260         btrfs_end_transaction_throttle(trans);
9261 out_free:
9262         kfree(wc);
9263         btrfs_free_path(path);
9264 out:
9265         /*
9266          * So if we need to stop dropping the snapshot for whatever reason we
9267          * need to make sure to add it back to the dead root list so that we
9268          * keep trying to do the work later.  This also cleans up roots if we
9269          * don't have it in the radix (like when we recover after a power fail
9270          * or unmount) so we don't leak memory.
9271          */
9272         if (!for_reloc && !root_dropped)
9273                 btrfs_add_dead_root(root);
9274         if (err && err != -EAGAIN)
9275                 btrfs_handle_fs_error(fs_info, err, NULL);
9276         return err;
9277 }
9278
9279 /*
9280  * drop subtree rooted at tree block 'node'.
9281  *
9282  * NOTE: this function will unlock and release tree block 'node'
9283  * only used by relocation code
9284  */
9285 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
9286                         struct btrfs_root *root,
9287                         struct extent_buffer *node,
9288                         struct extent_buffer *parent)
9289 {
9290         struct btrfs_fs_info *fs_info = root->fs_info;
9291         struct btrfs_path *path;
9292         struct walk_control *wc;
9293         int level;
9294         int parent_level;
9295         int ret = 0;
9296         int wret;
9297
9298         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
9299
9300         path = btrfs_alloc_path();
9301         if (!path)
9302                 return -ENOMEM;
9303
9304         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9305         if (!wc) {
9306                 btrfs_free_path(path);
9307                 return -ENOMEM;
9308         }
9309
9310         btrfs_assert_tree_locked(parent);
9311         parent_level = btrfs_header_level(parent);
9312         extent_buffer_get(parent);
9313         path->nodes[parent_level] = parent;
9314         path->slots[parent_level] = btrfs_header_nritems(parent);
9315
9316         btrfs_assert_tree_locked(node);
9317         level = btrfs_header_level(node);
9318         path->nodes[level] = node;
9319         path->slots[level] = 0;
9320         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9321
9322         wc->refs[parent_level] = 1;
9323         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
9324         wc->level = level;
9325         wc->shared_level = -1;
9326         wc->stage = DROP_REFERENCE;
9327         wc->update_ref = 0;
9328         wc->keep_locks = 1;
9329         wc->for_reloc = 1;
9330         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
9331
9332         while (1) {
9333                 wret = walk_down_tree(trans, root, path, wc);
9334                 if (wret < 0) {
9335                         ret = wret;
9336                         break;
9337                 }
9338
9339                 wret = walk_up_tree(trans, root, path, wc, parent_level);
9340                 if (wret < 0)
9341                         ret = wret;
9342                 if (wret != 0)
9343                         break;
9344         }
9345
9346         kfree(wc);
9347         btrfs_free_path(path);
9348         return ret;
9349 }
9350
9351 static u64 update_block_group_flags(struct btrfs_fs_info *fs_info, u64 flags)
9352 {
9353         u64 num_devices;
9354         u64 stripped;
9355
9356         /*
9357          * if restripe for this chunk_type is on pick target profile and
9358          * return, otherwise do the usual balance
9359          */
9360         stripped = get_restripe_target(fs_info, flags);
9361         if (stripped)
9362                 return extended_to_chunk(stripped);
9363
9364         num_devices = fs_info->fs_devices->rw_devices;
9365
9366         stripped = BTRFS_BLOCK_GROUP_RAID0 |
9367                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
9368                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
9369
9370         if (num_devices == 1) {
9371                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9372                 stripped = flags & ~stripped;
9373
9374                 /* turn raid0 into single device chunks */
9375                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
9376                         return stripped;
9377
9378                 /* turn mirroring into duplication */
9379                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
9380                              BTRFS_BLOCK_GROUP_RAID10))
9381                         return stripped | BTRFS_BLOCK_GROUP_DUP;
9382         } else {
9383                 /* they already had raid on here, just return */
9384                 if (flags & stripped)
9385                         return flags;
9386
9387                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9388                 stripped = flags & ~stripped;
9389
9390                 /* switch duplicated blocks with raid1 */
9391                 if (flags & BTRFS_BLOCK_GROUP_DUP)
9392                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
9393
9394                 /* this is drive concat, leave it alone */
9395         }
9396
9397         return flags;
9398 }
9399
9400 static int inc_block_group_ro(struct btrfs_block_group_cache *cache, int force)
9401 {
9402         struct btrfs_space_info *sinfo = cache->space_info;
9403         u64 num_bytes;
9404         u64 min_allocable_bytes;
9405         int ret = -ENOSPC;
9406
9407         /*
9408          * We need some metadata space and system metadata space for
9409          * allocating chunks in some corner cases until we force to set
9410          * it to be readonly.
9411          */
9412         if ((sinfo->flags &
9413              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
9414             !force)
9415                 min_allocable_bytes = SZ_1M;
9416         else
9417                 min_allocable_bytes = 0;
9418
9419         spin_lock(&sinfo->lock);
9420         spin_lock(&cache->lock);
9421
9422         if (cache->ro) {
9423                 cache->ro++;
9424                 ret = 0;
9425                 goto out;
9426         }
9427
9428         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
9429                     cache->bytes_super - btrfs_block_group_used(&cache->item);
9430
9431         if (btrfs_space_info_used(sinfo, true) + num_bytes +
9432             min_allocable_bytes <= sinfo->total_bytes) {
9433                 sinfo->bytes_readonly += num_bytes;
9434                 cache->ro++;
9435                 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
9436                 ret = 0;
9437         }
9438 out:
9439         spin_unlock(&cache->lock);
9440         spin_unlock(&sinfo->lock);
9441         return ret;
9442 }
9443
9444 int btrfs_inc_block_group_ro(struct btrfs_fs_info *fs_info,
9445                              struct btrfs_block_group_cache *cache)
9446
9447 {
9448         struct btrfs_trans_handle *trans;
9449         u64 alloc_flags;
9450         int ret;
9451
9452 again:
9453         trans = btrfs_join_transaction(fs_info->extent_root);
9454         if (IS_ERR(trans))
9455                 return PTR_ERR(trans);
9456
9457         /*
9458          * we're not allowed to set block groups readonly after the dirty
9459          * block groups cache has started writing.  If it already started,
9460          * back off and let this transaction commit
9461          */
9462         mutex_lock(&fs_info->ro_block_group_mutex);
9463         if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
9464                 u64 transid = trans->transid;
9465
9466                 mutex_unlock(&fs_info->ro_block_group_mutex);
9467                 btrfs_end_transaction(trans);
9468
9469                 ret = btrfs_wait_for_commit(fs_info, transid);
9470                 if (ret)
9471                         return ret;
9472                 goto again;
9473         }
9474
9475         /*
9476          * if we are changing raid levels, try to allocate a corresponding
9477          * block group with the new raid level.
9478          */
9479         alloc_flags = update_block_group_flags(fs_info, cache->flags);
9480         if (alloc_flags != cache->flags) {
9481                 ret = do_chunk_alloc(trans, fs_info, alloc_flags,
9482                                      CHUNK_ALLOC_FORCE);
9483                 /*
9484                  * ENOSPC is allowed here, we may have enough space
9485                  * already allocated at the new raid level to
9486                  * carry on
9487                  */
9488                 if (ret == -ENOSPC)
9489                         ret = 0;
9490                 if (ret < 0)
9491                         goto out;
9492         }
9493
9494         ret = inc_block_group_ro(cache, 0);
9495         if (!ret)
9496                 goto out;
9497         alloc_flags = get_alloc_profile(fs_info, cache->space_info->flags);
9498         ret = do_chunk_alloc(trans, fs_info, alloc_flags,
9499                              CHUNK_ALLOC_FORCE);
9500         if (ret < 0)
9501                 goto out;
9502         ret = inc_block_group_ro(cache, 0);
9503 out:
9504         if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
9505                 alloc_flags = update_block_group_flags(fs_info, cache->flags);
9506                 mutex_lock(&fs_info->chunk_mutex);
9507                 check_system_chunk(trans, fs_info, alloc_flags);
9508                 mutex_unlock(&fs_info->chunk_mutex);
9509         }
9510         mutex_unlock(&fs_info->ro_block_group_mutex);
9511
9512         btrfs_end_transaction(trans);
9513         return ret;
9514 }
9515
9516 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
9517                             struct btrfs_fs_info *fs_info, u64 type)
9518 {
9519         u64 alloc_flags = get_alloc_profile(fs_info, type);
9520
9521         return do_chunk_alloc(trans, fs_info, alloc_flags, CHUNK_ALLOC_FORCE);
9522 }
9523
9524 /*
9525  * helper to account the unused space of all the readonly block group in the
9526  * space_info. takes mirrors into account.
9527  */
9528 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
9529 {
9530         struct btrfs_block_group_cache *block_group;
9531         u64 free_bytes = 0;
9532         int factor;
9533
9534         /* It's df, we don't care if it's racy */
9535         if (list_empty(&sinfo->ro_bgs))
9536                 return 0;
9537
9538         spin_lock(&sinfo->lock);
9539         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
9540                 spin_lock(&block_group->lock);
9541
9542                 if (!block_group->ro) {
9543                         spin_unlock(&block_group->lock);
9544                         continue;
9545                 }
9546
9547                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
9548                                           BTRFS_BLOCK_GROUP_RAID10 |
9549                                           BTRFS_BLOCK_GROUP_DUP))
9550                         factor = 2;
9551                 else
9552                         factor = 1;
9553
9554                 free_bytes += (block_group->key.offset -
9555                                btrfs_block_group_used(&block_group->item)) *
9556                                factor;
9557
9558                 spin_unlock(&block_group->lock);
9559         }
9560         spin_unlock(&sinfo->lock);
9561
9562         return free_bytes;
9563 }
9564
9565 void btrfs_dec_block_group_ro(struct btrfs_block_group_cache *cache)
9566 {
9567         struct btrfs_space_info *sinfo = cache->space_info;
9568         u64 num_bytes;
9569
9570         BUG_ON(!cache->ro);
9571
9572         spin_lock(&sinfo->lock);
9573         spin_lock(&cache->lock);
9574         if (!--cache->ro) {
9575                 num_bytes = cache->key.offset - cache->reserved -
9576                             cache->pinned - cache->bytes_super -
9577                             btrfs_block_group_used(&cache->item);
9578                 sinfo->bytes_readonly -= num_bytes;
9579                 list_del_init(&cache->ro_list);
9580         }
9581         spin_unlock(&cache->lock);
9582         spin_unlock(&sinfo->lock);
9583 }
9584
9585 /*
9586  * checks to see if its even possible to relocate this block group.
9587  *
9588  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
9589  * ok to go ahead and try.
9590  */
9591 int btrfs_can_relocate(struct btrfs_fs_info *fs_info, u64 bytenr)
9592 {
9593         struct btrfs_root *root = fs_info->extent_root;
9594         struct btrfs_block_group_cache *block_group;
9595         struct btrfs_space_info *space_info;
9596         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
9597         struct btrfs_device *device;
9598         struct btrfs_trans_handle *trans;
9599         u64 min_free;
9600         u64 dev_min = 1;
9601         u64 dev_nr = 0;
9602         u64 target;
9603         int debug;
9604         int index;
9605         int full = 0;
9606         int ret = 0;
9607
9608         debug = btrfs_test_opt(fs_info, ENOSPC_DEBUG);
9609
9610         block_group = btrfs_lookup_block_group(fs_info, bytenr);
9611
9612         /* odd, couldn't find the block group, leave it alone */
9613         if (!block_group) {
9614                 if (debug)
9615                         btrfs_warn(fs_info,
9616                                    "can't find block group for bytenr %llu",
9617                                    bytenr);
9618                 return -1;
9619         }
9620
9621         min_free = btrfs_block_group_used(&block_group->item);
9622
9623         /* no bytes used, we're good */
9624         if (!min_free)
9625                 goto out;
9626
9627         space_info = block_group->space_info;
9628         spin_lock(&space_info->lock);
9629
9630         full = space_info->full;
9631
9632         /*
9633          * if this is the last block group we have in this space, we can't
9634          * relocate it unless we're able to allocate a new chunk below.
9635          *
9636          * Otherwise, we need to make sure we have room in the space to handle
9637          * all of the extents from this block group.  If we can, we're good
9638          */
9639         if ((space_info->total_bytes != block_group->key.offset) &&
9640             (btrfs_space_info_used(space_info, false) + min_free <
9641              space_info->total_bytes)) {
9642                 spin_unlock(&space_info->lock);
9643                 goto out;
9644         }
9645         spin_unlock(&space_info->lock);
9646
9647         /*
9648          * ok we don't have enough space, but maybe we have free space on our
9649          * devices to allocate new chunks for relocation, so loop through our
9650          * alloc devices and guess if we have enough space.  if this block
9651          * group is going to be restriped, run checks against the target
9652          * profile instead of the current one.
9653          */
9654         ret = -1;
9655
9656         /*
9657          * index:
9658          *      0: raid10
9659          *      1: raid1
9660          *      2: dup
9661          *      3: raid0
9662          *      4: single
9663          */
9664         target = get_restripe_target(fs_info, block_group->flags);
9665         if (target) {
9666                 index = btrfs_bg_flags_to_raid_index(extended_to_chunk(target));
9667         } else {
9668                 /*
9669                  * this is just a balance, so if we were marked as full
9670                  * we know there is no space for a new chunk
9671                  */
9672                 if (full) {
9673                         if (debug)
9674                                 btrfs_warn(fs_info,
9675                                            "no space to alloc new chunk for block group %llu",
9676                                            block_group->key.objectid);
9677                         goto out;
9678                 }
9679
9680                 index = btrfs_bg_flags_to_raid_index(block_group->flags);
9681         }
9682
9683         if (index == BTRFS_RAID_RAID10) {
9684                 dev_min = 4;
9685                 /* Divide by 2 */
9686                 min_free >>= 1;
9687         } else if (index == BTRFS_RAID_RAID1) {
9688                 dev_min = 2;
9689         } else if (index == BTRFS_RAID_DUP) {
9690                 /* Multiply by 2 */
9691                 min_free <<= 1;
9692         } else if (index == BTRFS_RAID_RAID0) {
9693                 dev_min = fs_devices->rw_devices;
9694                 min_free = div64_u64(min_free, dev_min);
9695         }
9696
9697         /* We need to do this so that we can look at pending chunks */
9698         trans = btrfs_join_transaction(root);
9699         if (IS_ERR(trans)) {
9700                 ret = PTR_ERR(trans);
9701                 goto out;
9702         }
9703
9704         mutex_lock(&fs_info->chunk_mutex);
9705         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
9706                 u64 dev_offset;
9707
9708                 /*
9709                  * check to make sure we can actually find a chunk with enough
9710                  * space to fit our block group in.
9711                  */
9712                 if (device->total_bytes > device->bytes_used + min_free &&
9713                     !test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state)) {
9714                         ret = find_free_dev_extent(trans, device, min_free,
9715                                                    &dev_offset, NULL);
9716                         if (!ret)
9717                                 dev_nr++;
9718
9719                         if (dev_nr >= dev_min)
9720                                 break;
9721
9722                         ret = -1;
9723                 }
9724         }
9725         if (debug && ret == -1)
9726                 btrfs_warn(fs_info,
9727                            "no space to allocate a new chunk for block group %llu",
9728                            block_group->key.objectid);
9729         mutex_unlock(&fs_info->chunk_mutex);
9730         btrfs_end_transaction(trans);
9731 out:
9732         btrfs_put_block_group(block_group);
9733         return ret;
9734 }
9735
9736 static int find_first_block_group(struct btrfs_fs_info *fs_info,
9737                                   struct btrfs_path *path,
9738                                   struct btrfs_key *key)
9739 {
9740         struct btrfs_root *root = fs_info->extent_root;
9741         int ret = 0;
9742         struct btrfs_key found_key;
9743         struct extent_buffer *leaf;
9744         int slot;
9745
9746         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
9747         if (ret < 0)
9748                 goto out;
9749
9750         while (1) {
9751                 slot = path->slots[0];
9752                 leaf = path->nodes[0];
9753                 if (slot >= btrfs_header_nritems(leaf)) {
9754                         ret = btrfs_next_leaf(root, path);
9755                         if (ret == 0)
9756                                 continue;
9757                         if (ret < 0)
9758                                 goto out;
9759                         break;
9760                 }
9761                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
9762
9763                 if (found_key.objectid >= key->objectid &&
9764                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
9765                         struct extent_map_tree *em_tree;
9766                         struct extent_map *em;
9767
9768                         em_tree = &root->fs_info->mapping_tree.map_tree;
9769                         read_lock(&em_tree->lock);
9770                         em = lookup_extent_mapping(em_tree, found_key.objectid,
9771                                                    found_key.offset);
9772                         read_unlock(&em_tree->lock);
9773                         if (!em) {
9774                                 btrfs_err(fs_info,
9775                         "logical %llu len %llu found bg but no related chunk",
9776                                           found_key.objectid, found_key.offset);
9777                                 ret = -ENOENT;
9778                         } else {
9779                                 ret = 0;
9780                         }
9781                         free_extent_map(em);
9782                         goto out;
9783                 }
9784                 path->slots[0]++;
9785         }
9786 out:
9787         return ret;
9788 }
9789
9790 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
9791 {
9792         struct btrfs_block_group_cache *block_group;
9793         u64 last = 0;
9794
9795         while (1) {
9796                 struct inode *inode;
9797
9798                 block_group = btrfs_lookup_first_block_group(info, last);
9799                 while (block_group) {
9800                         spin_lock(&block_group->lock);
9801                         if (block_group->iref)
9802                                 break;
9803                         spin_unlock(&block_group->lock);
9804                         block_group = next_block_group(info, block_group);
9805                 }
9806                 if (!block_group) {
9807                         if (last == 0)
9808                                 break;
9809                         last = 0;
9810                         continue;
9811                 }
9812
9813                 inode = block_group->inode;
9814                 block_group->iref = 0;
9815                 block_group->inode = NULL;
9816                 spin_unlock(&block_group->lock);
9817                 ASSERT(block_group->io_ctl.inode == NULL);
9818                 iput(inode);
9819                 last = block_group->key.objectid + block_group->key.offset;
9820                 btrfs_put_block_group(block_group);
9821         }
9822 }
9823
9824 /*
9825  * Must be called only after stopping all workers, since we could have block
9826  * group caching kthreads running, and therefore they could race with us if we
9827  * freed the block groups before stopping them.
9828  */
9829 int btrfs_free_block_groups(struct btrfs_fs_info *info)
9830 {
9831         struct btrfs_block_group_cache *block_group;
9832         struct btrfs_space_info *space_info;
9833         struct btrfs_caching_control *caching_ctl;
9834         struct rb_node *n;
9835
9836         down_write(&info->commit_root_sem);
9837         while (!list_empty(&info->caching_block_groups)) {
9838                 caching_ctl = list_entry(info->caching_block_groups.next,
9839                                          struct btrfs_caching_control, list);
9840                 list_del(&caching_ctl->list);
9841                 put_caching_control(caching_ctl);
9842         }
9843         up_write(&info->commit_root_sem);
9844
9845         spin_lock(&info->unused_bgs_lock);
9846         while (!list_empty(&info->unused_bgs)) {
9847                 block_group = list_first_entry(&info->unused_bgs,
9848                                                struct btrfs_block_group_cache,
9849                                                bg_list);
9850                 list_del_init(&block_group->bg_list);
9851                 btrfs_put_block_group(block_group);
9852         }
9853         spin_unlock(&info->unused_bgs_lock);
9854
9855         spin_lock(&info->block_group_cache_lock);
9856         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
9857                 block_group = rb_entry(n, struct btrfs_block_group_cache,
9858                                        cache_node);
9859                 rb_erase(&block_group->cache_node,
9860                          &info->block_group_cache_tree);
9861                 RB_CLEAR_NODE(&block_group->cache_node);
9862                 spin_unlock(&info->block_group_cache_lock);
9863
9864                 down_write(&block_group->space_info->groups_sem);
9865                 list_del(&block_group->list);
9866                 up_write(&block_group->space_info->groups_sem);
9867
9868                 /*
9869                  * We haven't cached this block group, which means we could
9870                  * possibly have excluded extents on this block group.
9871                  */
9872                 if (block_group->cached == BTRFS_CACHE_NO ||
9873                     block_group->cached == BTRFS_CACHE_ERROR)
9874                         free_excluded_extents(info, block_group);
9875
9876                 btrfs_remove_free_space_cache(block_group);
9877                 ASSERT(block_group->cached != BTRFS_CACHE_STARTED);
9878                 ASSERT(list_empty(&block_group->dirty_list));
9879                 ASSERT(list_empty(&block_group->io_list));
9880                 ASSERT(list_empty(&block_group->bg_list));
9881                 ASSERT(atomic_read(&block_group->count) == 1);
9882                 btrfs_put_block_group(block_group);
9883
9884                 spin_lock(&info->block_group_cache_lock);
9885         }
9886         spin_unlock(&info->block_group_cache_lock);
9887
9888         /* now that all the block groups are freed, go through and
9889          * free all the space_info structs.  This is only called during
9890          * the final stages of unmount, and so we know nobody is
9891          * using them.  We call synchronize_rcu() once before we start,
9892          * just to be on the safe side.
9893          */
9894         synchronize_rcu();
9895
9896         release_global_block_rsv(info);
9897
9898         while (!list_empty(&info->space_info)) {
9899                 int i;
9900
9901                 space_info = list_entry(info->space_info.next,
9902                                         struct btrfs_space_info,
9903                                         list);
9904
9905                 /*
9906                  * Do not hide this behind enospc_debug, this is actually
9907                  * important and indicates a real bug if this happens.
9908                  */
9909                 if (WARN_ON(space_info->bytes_pinned > 0 ||
9910                             space_info->bytes_reserved > 0 ||
9911                             space_info->bytes_may_use > 0))
9912                         dump_space_info(info, space_info, 0, 0);
9913                 list_del(&space_info->list);
9914                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
9915                         struct kobject *kobj;
9916                         kobj = space_info->block_group_kobjs[i];
9917                         space_info->block_group_kobjs[i] = NULL;
9918                         if (kobj) {
9919                                 kobject_del(kobj);
9920                                 kobject_put(kobj);
9921                         }
9922                 }
9923                 kobject_del(&space_info->kobj);
9924                 kobject_put(&space_info->kobj);
9925         }
9926         return 0;
9927 }
9928
9929 /* link_block_group will queue up kobjects to add when we're reclaim-safe */
9930 void btrfs_add_raid_kobjects(struct btrfs_fs_info *fs_info)
9931 {
9932         struct btrfs_space_info *space_info;
9933         struct raid_kobject *rkobj;
9934         LIST_HEAD(list);
9935         int index;
9936         int ret = 0;
9937
9938         spin_lock(&fs_info->pending_raid_kobjs_lock);
9939         list_splice_init(&fs_info->pending_raid_kobjs, &list);
9940         spin_unlock(&fs_info->pending_raid_kobjs_lock);
9941
9942         list_for_each_entry(rkobj, &list, list) {
9943                 space_info = __find_space_info(fs_info, rkobj->flags);
9944                 index = btrfs_bg_flags_to_raid_index(rkobj->flags);
9945
9946                 ret = kobject_add(&rkobj->kobj, &space_info->kobj,
9947                                   "%s", get_raid_name(index));
9948                 if (ret) {
9949                         kobject_put(&rkobj->kobj);
9950                         break;
9951                 }
9952         }
9953         if (ret)
9954                 btrfs_warn(fs_info,
9955                            "failed to add kobject for block cache, ignoring");
9956 }
9957
9958 static void link_block_group(struct btrfs_block_group_cache *cache)
9959 {
9960         struct btrfs_space_info *space_info = cache->space_info;
9961         struct btrfs_fs_info *fs_info = cache->fs_info;
9962         int index = btrfs_bg_flags_to_raid_index(cache->flags);
9963         bool first = false;
9964
9965         down_write(&space_info->groups_sem);
9966         if (list_empty(&space_info->block_groups[index]))
9967                 first = true;
9968         list_add_tail(&cache->list, &space_info->block_groups[index]);
9969         up_write(&space_info->groups_sem);
9970
9971         if (first) {
9972                 struct raid_kobject *rkobj = kzalloc(sizeof(*rkobj), GFP_NOFS);
9973                 if (!rkobj) {
9974                         btrfs_warn(cache->fs_info,
9975                                 "couldn't alloc memory for raid level kobject");
9976                         return;
9977                 }
9978                 rkobj->flags = cache->flags;
9979                 kobject_init(&rkobj->kobj, &btrfs_raid_ktype);
9980
9981                 spin_lock(&fs_info->pending_raid_kobjs_lock);
9982                 list_add_tail(&rkobj->list, &fs_info->pending_raid_kobjs);
9983                 spin_unlock(&fs_info->pending_raid_kobjs_lock);
9984                 space_info->block_group_kobjs[index] = &rkobj->kobj;
9985         }
9986 }
9987
9988 static struct btrfs_block_group_cache *
9989 btrfs_create_block_group_cache(struct btrfs_fs_info *fs_info,
9990                                u64 start, u64 size)
9991 {
9992         struct btrfs_block_group_cache *cache;
9993
9994         cache = kzalloc(sizeof(*cache), GFP_NOFS);
9995         if (!cache)
9996                 return NULL;
9997
9998         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
9999                                         GFP_NOFS);
10000         if (!cache->free_space_ctl) {
10001                 kfree(cache);
10002                 return NULL;
10003         }
10004
10005         cache->key.objectid = start;
10006         cache->key.offset = size;
10007         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
10008
10009         cache->fs_info = fs_info;
10010         cache->full_stripe_len = btrfs_full_stripe_len(fs_info, start);
10011         set_free_space_tree_thresholds(cache);
10012
10013         atomic_set(&cache->count, 1);
10014         spin_lock_init(&cache->lock);
10015         init_rwsem(&cache->data_rwsem);
10016         INIT_LIST_HEAD(&cache->list);
10017         INIT_LIST_HEAD(&cache->cluster_list);
10018         INIT_LIST_HEAD(&cache->bg_list);
10019         INIT_LIST_HEAD(&cache->ro_list);
10020         INIT_LIST_HEAD(&cache->dirty_list);
10021         INIT_LIST_HEAD(&cache->io_list);
10022         btrfs_init_free_space_ctl(cache);
10023         atomic_set(&cache->trimming, 0);
10024         mutex_init(&cache->free_space_lock);
10025         btrfs_init_full_stripe_locks_tree(&cache->full_stripe_locks_root);
10026
10027         return cache;
10028 }
10029
10030 int btrfs_read_block_groups(struct btrfs_fs_info *info)
10031 {
10032         struct btrfs_path *path;
10033         int ret;
10034         struct btrfs_block_group_cache *cache;
10035         struct btrfs_space_info *space_info;
10036         struct btrfs_key key;
10037         struct btrfs_key found_key;
10038         struct extent_buffer *leaf;
10039         int need_clear = 0;
10040         u64 cache_gen;
10041         u64 feature;
10042         int mixed;
10043
10044         feature = btrfs_super_incompat_flags(info->super_copy);
10045         mixed = !!(feature & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS);
10046
10047         key.objectid = 0;
10048         key.offset = 0;
10049         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
10050         path = btrfs_alloc_path();
10051         if (!path)
10052                 return -ENOMEM;
10053         path->reada = READA_FORWARD;
10054
10055         cache_gen = btrfs_super_cache_generation(info->super_copy);
10056         if (btrfs_test_opt(info, SPACE_CACHE) &&
10057             btrfs_super_generation(info->super_copy) != cache_gen)
10058                 need_clear = 1;
10059         if (btrfs_test_opt(info, CLEAR_CACHE))
10060                 need_clear = 1;
10061
10062         while (1) {
10063                 ret = find_first_block_group(info, path, &key);
10064                 if (ret > 0)
10065                         break;
10066                 if (ret != 0)
10067                         goto error;
10068
10069                 leaf = path->nodes[0];
10070                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
10071
10072                 cache = btrfs_create_block_group_cache(info, found_key.objectid,
10073                                                        found_key.offset);
10074                 if (!cache) {
10075                         ret = -ENOMEM;
10076                         goto error;
10077                 }
10078
10079                 if (need_clear) {
10080                         /*
10081                          * When we mount with old space cache, we need to
10082                          * set BTRFS_DC_CLEAR and set dirty flag.
10083                          *
10084                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
10085                          *    truncate the old free space cache inode and
10086                          *    setup a new one.
10087                          * b) Setting 'dirty flag' makes sure that we flush
10088                          *    the new space cache info onto disk.
10089                          */
10090                         if (btrfs_test_opt(info, SPACE_CACHE))
10091                                 cache->disk_cache_state = BTRFS_DC_CLEAR;
10092                 }
10093
10094                 read_extent_buffer(leaf, &cache->item,
10095                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
10096                                    sizeof(cache->item));
10097                 cache->flags = btrfs_block_group_flags(&cache->item);
10098                 if (!mixed &&
10099                     ((cache->flags & BTRFS_BLOCK_GROUP_METADATA) &&
10100                     (cache->flags & BTRFS_BLOCK_GROUP_DATA))) {
10101                         btrfs_err(info,
10102 "bg %llu is a mixed block group but filesystem hasn't enabled mixed block groups",
10103                                   cache->key.objectid);
10104                         ret = -EINVAL;
10105                         goto error;
10106                 }
10107
10108                 key.objectid = found_key.objectid + found_key.offset;
10109                 btrfs_release_path(path);
10110
10111                 /*
10112                  * We need to exclude the super stripes now so that the space
10113                  * info has super bytes accounted for, otherwise we'll think
10114                  * we have more space than we actually do.
10115                  */
10116                 ret = exclude_super_stripes(info, cache);
10117                 if (ret) {
10118                         /*
10119                          * We may have excluded something, so call this just in
10120                          * case.
10121                          */
10122                         free_excluded_extents(info, cache);
10123                         btrfs_put_block_group(cache);
10124                         goto error;
10125                 }
10126
10127                 /*
10128                  * check for two cases, either we are full, and therefore
10129                  * don't need to bother with the caching work since we won't
10130                  * find any space, or we are empty, and we can just add all
10131                  * the space in and be done with it.  This saves us _alot_ of
10132                  * time, particularly in the full case.
10133                  */
10134                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
10135                         cache->last_byte_to_unpin = (u64)-1;
10136                         cache->cached = BTRFS_CACHE_FINISHED;
10137                         free_excluded_extents(info, cache);
10138                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10139                         cache->last_byte_to_unpin = (u64)-1;
10140                         cache->cached = BTRFS_CACHE_FINISHED;
10141                         add_new_free_space(cache, info,
10142                                            found_key.objectid,
10143                                            found_key.objectid +
10144                                            found_key.offset);
10145                         free_excluded_extents(info, cache);
10146                 }
10147
10148                 ret = btrfs_add_block_group_cache(info, cache);
10149                 if (ret) {
10150                         btrfs_remove_free_space_cache(cache);
10151                         btrfs_put_block_group(cache);
10152                         goto error;
10153                 }
10154
10155                 trace_btrfs_add_block_group(info, cache, 0);
10156                 update_space_info(info, cache->flags, found_key.offset,
10157                                   btrfs_block_group_used(&cache->item),
10158                                   cache->bytes_super, &space_info);
10159
10160                 cache->space_info = space_info;
10161
10162                 link_block_group(cache);
10163
10164                 set_avail_alloc_bits(info, cache->flags);
10165                 if (btrfs_chunk_readonly(info, cache->key.objectid)) {
10166                         inc_block_group_ro(cache, 1);
10167                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10168                         spin_lock(&info->unused_bgs_lock);
10169                         /* Should always be true but just in case. */
10170                         if (list_empty(&cache->bg_list)) {
10171                                 btrfs_get_block_group(cache);
10172                                 list_add_tail(&cache->bg_list,
10173                                               &info->unused_bgs);
10174                         }
10175                         spin_unlock(&info->unused_bgs_lock);
10176                 }
10177         }
10178
10179         list_for_each_entry_rcu(space_info, &info->space_info, list) {
10180                 if (!(get_alloc_profile(info, space_info->flags) &
10181                       (BTRFS_BLOCK_GROUP_RAID10 |
10182                        BTRFS_BLOCK_GROUP_RAID1 |
10183                        BTRFS_BLOCK_GROUP_RAID5 |
10184                        BTRFS_BLOCK_GROUP_RAID6 |
10185                        BTRFS_BLOCK_GROUP_DUP)))
10186                         continue;
10187                 /*
10188                  * avoid allocating from un-mirrored block group if there are
10189                  * mirrored block groups.
10190                  */
10191                 list_for_each_entry(cache,
10192                                 &space_info->block_groups[BTRFS_RAID_RAID0],
10193                                 list)
10194                         inc_block_group_ro(cache, 1);
10195                 list_for_each_entry(cache,
10196                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
10197                                 list)
10198                         inc_block_group_ro(cache, 1);
10199         }
10200
10201         btrfs_add_raid_kobjects(info);
10202         init_global_block_rsv(info);
10203         ret = 0;
10204 error:
10205         btrfs_free_path(path);
10206         return ret;
10207 }
10208
10209 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans)
10210 {
10211         struct btrfs_fs_info *fs_info = trans->fs_info;
10212         struct btrfs_block_group_cache *block_group, *tmp;
10213         struct btrfs_root *extent_root = fs_info->extent_root;
10214         struct btrfs_block_group_item item;
10215         struct btrfs_key key;
10216         int ret = 0;
10217         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
10218
10219         trans->can_flush_pending_bgs = false;
10220         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) {
10221                 if (ret)
10222                         goto next;
10223
10224                 spin_lock(&block_group->lock);
10225                 memcpy(&item, &block_group->item, sizeof(item));
10226                 memcpy(&key, &block_group->key, sizeof(key));
10227                 spin_unlock(&block_group->lock);
10228
10229                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
10230                                         sizeof(item));
10231                 if (ret)
10232                         btrfs_abort_transaction(trans, ret);
10233                 ret = btrfs_finish_chunk_alloc(trans, fs_info, key.objectid,
10234                                                key.offset);
10235                 if (ret)
10236                         btrfs_abort_transaction(trans, ret);
10237                 add_block_group_free_space(trans, fs_info, block_group);
10238                 /* already aborted the transaction if it failed. */
10239 next:
10240                 list_del_init(&block_group->bg_list);
10241         }
10242         trans->can_flush_pending_bgs = can_flush_pending_bgs;
10243 }
10244
10245 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
10246                            struct btrfs_fs_info *fs_info, u64 bytes_used,
10247                            u64 type, u64 chunk_offset, u64 size)
10248 {
10249         struct btrfs_block_group_cache *cache;
10250         int ret;
10251
10252         btrfs_set_log_full_commit(fs_info, trans);
10253
10254         cache = btrfs_create_block_group_cache(fs_info, chunk_offset, size);
10255         if (!cache)
10256                 return -ENOMEM;
10257
10258         btrfs_set_block_group_used(&cache->item, bytes_used);
10259         btrfs_set_block_group_chunk_objectid(&cache->item,
10260                                              BTRFS_FIRST_CHUNK_TREE_OBJECTID);
10261         btrfs_set_block_group_flags(&cache->item, type);
10262
10263         cache->flags = type;
10264         cache->last_byte_to_unpin = (u64)-1;
10265         cache->cached = BTRFS_CACHE_FINISHED;
10266         cache->needs_free_space = 1;
10267         ret = exclude_super_stripes(fs_info, cache);
10268         if (ret) {
10269                 /*
10270                  * We may have excluded something, so call this just in
10271                  * case.
10272                  */
10273                 free_excluded_extents(fs_info, cache);
10274                 btrfs_put_block_group(cache);
10275                 return ret;
10276         }
10277
10278         add_new_free_space(cache, fs_info, chunk_offset, chunk_offset + size);
10279
10280         free_excluded_extents(fs_info, cache);
10281
10282 #ifdef CONFIG_BTRFS_DEBUG
10283         if (btrfs_should_fragment_free_space(cache)) {
10284                 u64 new_bytes_used = size - bytes_used;
10285
10286                 bytes_used += new_bytes_used >> 1;
10287                 fragment_free_space(cache);
10288         }
10289 #endif
10290         /*
10291          * Ensure the corresponding space_info object is created and
10292          * assigned to our block group. We want our bg to be added to the rbtree
10293          * with its ->space_info set.
10294          */
10295         cache->space_info = __find_space_info(fs_info, cache->flags);
10296         ASSERT(cache->space_info);
10297
10298         ret = btrfs_add_block_group_cache(fs_info, cache);
10299         if (ret) {
10300                 btrfs_remove_free_space_cache(cache);
10301                 btrfs_put_block_group(cache);
10302                 return ret;
10303         }
10304
10305         /*
10306          * Now that our block group has its ->space_info set and is inserted in
10307          * the rbtree, update the space info's counters.
10308          */
10309         trace_btrfs_add_block_group(fs_info, cache, 1);
10310         update_space_info(fs_info, cache->flags, size, bytes_used,
10311                                 cache->bytes_super, &cache->space_info);
10312         update_global_block_rsv(fs_info);
10313
10314         link_block_group(cache);
10315
10316         list_add_tail(&cache->bg_list, &trans->new_bgs);
10317
10318         set_avail_alloc_bits(fs_info, type);
10319         return 0;
10320 }
10321
10322 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
10323 {
10324         u64 extra_flags = chunk_to_extended(flags) &
10325                                 BTRFS_EXTENDED_PROFILE_MASK;
10326
10327         write_seqlock(&fs_info->profiles_lock);
10328         if (flags & BTRFS_BLOCK_GROUP_DATA)
10329                 fs_info->avail_data_alloc_bits &= ~extra_flags;
10330         if (flags & BTRFS_BLOCK_GROUP_METADATA)
10331                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
10332         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
10333                 fs_info->avail_system_alloc_bits &= ~extra_flags;
10334         write_sequnlock(&fs_info->profiles_lock);
10335 }
10336
10337 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
10338                              struct btrfs_fs_info *fs_info, u64 group_start,
10339                              struct extent_map *em)
10340 {
10341         struct btrfs_root *root = fs_info->extent_root;
10342         struct btrfs_path *path;
10343         struct btrfs_block_group_cache *block_group;
10344         struct btrfs_free_cluster *cluster;
10345         struct btrfs_root *tree_root = fs_info->tree_root;
10346         struct btrfs_key key;
10347         struct inode *inode;
10348         struct kobject *kobj = NULL;
10349         int ret;
10350         int index;
10351         int factor;
10352         struct btrfs_caching_control *caching_ctl = NULL;
10353         bool remove_em;
10354
10355         block_group = btrfs_lookup_block_group(fs_info, group_start);
10356         BUG_ON(!block_group);
10357         BUG_ON(!block_group->ro);
10358
10359         /*
10360          * Free the reserved super bytes from this block group before
10361          * remove it.
10362          */
10363         free_excluded_extents(fs_info, block_group);
10364         btrfs_free_ref_tree_range(fs_info, block_group->key.objectid,
10365                                   block_group->key.offset);
10366
10367         memcpy(&key, &block_group->key, sizeof(key));
10368         index = btrfs_bg_flags_to_raid_index(block_group->flags);
10369         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
10370                                   BTRFS_BLOCK_GROUP_RAID1 |
10371                                   BTRFS_BLOCK_GROUP_RAID10))
10372                 factor = 2;
10373         else
10374                 factor = 1;
10375
10376         /* make sure this block group isn't part of an allocation cluster */
10377         cluster = &fs_info->data_alloc_cluster;
10378         spin_lock(&cluster->refill_lock);
10379         btrfs_return_cluster_to_free_space(block_group, cluster);
10380         spin_unlock(&cluster->refill_lock);
10381
10382         /*
10383          * make sure this block group isn't part of a metadata
10384          * allocation cluster
10385          */
10386         cluster = &fs_info->meta_alloc_cluster;
10387         spin_lock(&cluster->refill_lock);
10388         btrfs_return_cluster_to_free_space(block_group, cluster);
10389         spin_unlock(&cluster->refill_lock);
10390
10391         path = btrfs_alloc_path();
10392         if (!path) {
10393                 ret = -ENOMEM;
10394                 goto out;
10395         }
10396
10397         /*
10398          * get the inode first so any iput calls done for the io_list
10399          * aren't the final iput (no unlinks allowed now)
10400          */
10401         inode = lookup_free_space_inode(fs_info, block_group, path);
10402
10403         mutex_lock(&trans->transaction->cache_write_mutex);
10404         /*
10405          * make sure our free spache cache IO is done before remove the
10406          * free space inode
10407          */
10408         spin_lock(&trans->transaction->dirty_bgs_lock);
10409         if (!list_empty(&block_group->io_list)) {
10410                 list_del_init(&block_group->io_list);
10411
10412                 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
10413
10414                 spin_unlock(&trans->transaction->dirty_bgs_lock);
10415                 btrfs_wait_cache_io(trans, block_group, path);
10416                 btrfs_put_block_group(block_group);
10417                 spin_lock(&trans->transaction->dirty_bgs_lock);
10418         }
10419
10420         if (!list_empty(&block_group->dirty_list)) {
10421                 list_del_init(&block_group->dirty_list);
10422                 btrfs_put_block_group(block_group);
10423         }
10424         spin_unlock(&trans->transaction->dirty_bgs_lock);
10425         mutex_unlock(&trans->transaction->cache_write_mutex);
10426
10427         if (!IS_ERR(inode)) {
10428                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
10429                 if (ret) {
10430                         btrfs_add_delayed_iput(inode);
10431                         goto out;
10432                 }
10433                 clear_nlink(inode);
10434                 /* One for the block groups ref */
10435                 spin_lock(&block_group->lock);
10436                 if (block_group->iref) {
10437                         block_group->iref = 0;
10438                         block_group->inode = NULL;
10439                         spin_unlock(&block_group->lock);
10440                         iput(inode);
10441                 } else {
10442                         spin_unlock(&block_group->lock);
10443                 }
10444                 /* One for our lookup ref */
10445                 btrfs_add_delayed_iput(inode);
10446         }
10447
10448         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
10449         key.offset = block_group->key.objectid;
10450         key.type = 0;
10451
10452         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
10453         if (ret < 0)
10454                 goto out;
10455         if (ret > 0)
10456                 btrfs_release_path(path);
10457         if (ret == 0) {
10458                 ret = btrfs_del_item(trans, tree_root, path);
10459                 if (ret)
10460                         goto out;
10461                 btrfs_release_path(path);
10462         }
10463
10464         spin_lock(&fs_info->block_group_cache_lock);
10465         rb_erase(&block_group->cache_node,
10466                  &fs_info->block_group_cache_tree);
10467         RB_CLEAR_NODE(&block_group->cache_node);
10468
10469         if (fs_info->first_logical_byte == block_group->key.objectid)
10470                 fs_info->first_logical_byte = (u64)-1;
10471         spin_unlock(&fs_info->block_group_cache_lock);
10472
10473         down_write(&block_group->space_info->groups_sem);
10474         /*
10475          * we must use list_del_init so people can check to see if they
10476          * are still on the list after taking the semaphore
10477          */
10478         list_del_init(&block_group->list);
10479         if (list_empty(&block_group->space_info->block_groups[index])) {
10480                 kobj = block_group->space_info->block_group_kobjs[index];
10481                 block_group->space_info->block_group_kobjs[index] = NULL;
10482                 clear_avail_alloc_bits(fs_info, block_group->flags);
10483         }
10484         up_write(&block_group->space_info->groups_sem);
10485         if (kobj) {
10486                 kobject_del(kobj);
10487                 kobject_put(kobj);
10488         }
10489
10490         if (block_group->has_caching_ctl)
10491                 caching_ctl = get_caching_control(block_group);
10492         if (block_group->cached == BTRFS_CACHE_STARTED)
10493                 wait_block_group_cache_done(block_group);
10494         if (block_group->has_caching_ctl) {
10495                 down_write(&fs_info->commit_root_sem);
10496                 if (!caching_ctl) {
10497                         struct btrfs_caching_control *ctl;
10498
10499                         list_for_each_entry(ctl,
10500                                     &fs_info->caching_block_groups, list)
10501                                 if (ctl->block_group == block_group) {
10502                                         caching_ctl = ctl;
10503                                         refcount_inc(&caching_ctl->count);
10504                                         break;
10505                                 }
10506                 }
10507                 if (caching_ctl)
10508                         list_del_init(&caching_ctl->list);
10509                 up_write(&fs_info->commit_root_sem);
10510                 if (caching_ctl) {
10511                         /* Once for the caching bgs list and once for us. */
10512                         put_caching_control(caching_ctl);
10513                         put_caching_control(caching_ctl);
10514                 }
10515         }
10516
10517         spin_lock(&trans->transaction->dirty_bgs_lock);
10518         if (!list_empty(&block_group->dirty_list)) {
10519                 WARN_ON(1);
10520         }
10521         if (!list_empty(&block_group->io_list)) {
10522                 WARN_ON(1);
10523         }
10524         spin_unlock(&trans->transaction->dirty_bgs_lock);
10525         btrfs_remove_free_space_cache(block_group);
10526
10527         spin_lock(&block_group->space_info->lock);
10528         list_del_init(&block_group->ro_list);
10529
10530         if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
10531                 WARN_ON(block_group->space_info->total_bytes
10532                         < block_group->key.offset);
10533                 WARN_ON(block_group->space_info->bytes_readonly
10534                         < block_group->key.offset);
10535                 WARN_ON(block_group->space_info->disk_total
10536                         < block_group->key.offset * factor);
10537         }
10538         block_group->space_info->total_bytes -= block_group->key.offset;
10539         block_group->space_info->bytes_readonly -= block_group->key.offset;
10540         block_group->space_info->disk_total -= block_group->key.offset * factor;
10541
10542         spin_unlock(&block_group->space_info->lock);
10543
10544         memcpy(&key, &block_group->key, sizeof(key));
10545
10546         mutex_lock(&fs_info->chunk_mutex);
10547         if (!list_empty(&em->list)) {
10548                 /* We're in the transaction->pending_chunks list. */
10549                 free_extent_map(em);
10550         }
10551         spin_lock(&block_group->lock);
10552         block_group->removed = 1;
10553         /*
10554          * At this point trimming can't start on this block group, because we
10555          * removed the block group from the tree fs_info->block_group_cache_tree
10556          * so no one can't find it anymore and even if someone already got this
10557          * block group before we removed it from the rbtree, they have already
10558          * incremented block_group->trimming - if they didn't, they won't find
10559          * any free space entries because we already removed them all when we
10560          * called btrfs_remove_free_space_cache().
10561          *
10562          * And we must not remove the extent map from the fs_info->mapping_tree
10563          * to prevent the same logical address range and physical device space
10564          * ranges from being reused for a new block group. This is because our
10565          * fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
10566          * completely transactionless, so while it is trimming a range the
10567          * currently running transaction might finish and a new one start,
10568          * allowing for new block groups to be created that can reuse the same
10569          * physical device locations unless we take this special care.
10570          *
10571          * There may also be an implicit trim operation if the file system
10572          * is mounted with -odiscard. The same protections must remain
10573          * in place until the extents have been discarded completely when
10574          * the transaction commit has completed.
10575          */
10576         remove_em = (atomic_read(&block_group->trimming) == 0);
10577         /*
10578          * Make sure a trimmer task always sees the em in the pinned_chunks list
10579          * if it sees block_group->removed == 1 (needs to lock block_group->lock
10580          * before checking block_group->removed).
10581          */
10582         if (!remove_em) {
10583                 /*
10584                  * Our em might be in trans->transaction->pending_chunks which
10585                  * is protected by fs_info->chunk_mutex ([lock|unlock]_chunks),
10586                  * and so is the fs_info->pinned_chunks list.
10587                  *
10588                  * So at this point we must be holding the chunk_mutex to avoid
10589                  * any races with chunk allocation (more specifically at
10590                  * volumes.c:contains_pending_extent()), to ensure it always
10591                  * sees the em, either in the pending_chunks list or in the
10592                  * pinned_chunks list.
10593                  */
10594                 list_move_tail(&em->list, &fs_info->pinned_chunks);
10595         }
10596         spin_unlock(&block_group->lock);
10597
10598         if (remove_em) {
10599                 struct extent_map_tree *em_tree;
10600
10601                 em_tree = &fs_info->mapping_tree.map_tree;
10602                 write_lock(&em_tree->lock);
10603                 /*
10604                  * The em might be in the pending_chunks list, so make sure the
10605                  * chunk mutex is locked, since remove_extent_mapping() will
10606                  * delete us from that list.
10607                  */
10608                 remove_extent_mapping(em_tree, em);
10609                 write_unlock(&em_tree->lock);
10610                 /* once for the tree */
10611                 free_extent_map(em);
10612         }
10613
10614         mutex_unlock(&fs_info->chunk_mutex);
10615
10616         ret = remove_block_group_free_space(trans, fs_info, block_group);
10617         if (ret)
10618                 goto out;
10619
10620         btrfs_put_block_group(block_group);
10621         btrfs_put_block_group(block_group);
10622
10623         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
10624         if (ret > 0)
10625                 ret = -EIO;
10626         if (ret < 0)
10627                 goto out;
10628
10629         ret = btrfs_del_item(trans, root, path);
10630 out:
10631         btrfs_free_path(path);
10632         return ret;
10633 }
10634
10635 struct btrfs_trans_handle *
10636 btrfs_start_trans_remove_block_group(struct btrfs_fs_info *fs_info,
10637                                      const u64 chunk_offset)
10638 {
10639         struct extent_map_tree *em_tree = &fs_info->mapping_tree.map_tree;
10640         struct extent_map *em;
10641         struct map_lookup *map;
10642         unsigned int num_items;
10643
10644         read_lock(&em_tree->lock);
10645         em = lookup_extent_mapping(em_tree, chunk_offset, 1);
10646         read_unlock(&em_tree->lock);
10647         ASSERT(em && em->start == chunk_offset);
10648
10649         /*
10650          * We need to reserve 3 + N units from the metadata space info in order
10651          * to remove a block group (done at btrfs_remove_chunk() and at
10652          * btrfs_remove_block_group()), which are used for:
10653          *
10654          * 1 unit for adding the free space inode's orphan (located in the tree
10655          * of tree roots).
10656          * 1 unit for deleting the block group item (located in the extent
10657          * tree).
10658          * 1 unit for deleting the free space item (located in tree of tree
10659          * roots).
10660          * N units for deleting N device extent items corresponding to each
10661          * stripe (located in the device tree).
10662          *
10663          * In order to remove a block group we also need to reserve units in the
10664          * system space info in order to update the chunk tree (update one or
10665          * more device items and remove one chunk item), but this is done at
10666          * btrfs_remove_chunk() through a call to check_system_chunk().
10667          */
10668         map = em->map_lookup;
10669         num_items = 3 + map->num_stripes;
10670         free_extent_map(em);
10671
10672         return btrfs_start_transaction_fallback_global_rsv(fs_info->extent_root,
10673                                                            num_items, 1);
10674 }
10675
10676 /*
10677  * Process the unused_bgs list and remove any that don't have any allocated
10678  * space inside of them.
10679  */
10680 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
10681 {
10682         struct btrfs_block_group_cache *block_group;
10683         struct btrfs_space_info *space_info;
10684         struct btrfs_trans_handle *trans;
10685         int ret = 0;
10686
10687         if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
10688                 return;
10689
10690         spin_lock(&fs_info->unused_bgs_lock);
10691         while (!list_empty(&fs_info->unused_bgs)) {
10692                 u64 start, end;
10693                 int trimming;
10694
10695                 block_group = list_first_entry(&fs_info->unused_bgs,
10696                                                struct btrfs_block_group_cache,
10697                                                bg_list);
10698                 list_del_init(&block_group->bg_list);
10699
10700                 space_info = block_group->space_info;
10701
10702                 if (ret || btrfs_mixed_space_info(space_info)) {
10703                         btrfs_put_block_group(block_group);
10704                         continue;
10705                 }
10706                 spin_unlock(&fs_info->unused_bgs_lock);
10707
10708                 mutex_lock(&fs_info->delete_unused_bgs_mutex);
10709
10710                 /* Don't want to race with allocators so take the groups_sem */
10711                 down_write(&space_info->groups_sem);
10712                 spin_lock(&block_group->lock);
10713                 if (block_group->reserved ||
10714                     btrfs_block_group_used(&block_group->item) ||
10715                     block_group->ro ||
10716                     list_is_singular(&block_group->list)) {
10717                         /*
10718                          * We want to bail if we made new allocations or have
10719                          * outstanding allocations in this block group.  We do
10720                          * the ro check in case balance is currently acting on
10721                          * this block group.
10722                          */
10723                         spin_unlock(&block_group->lock);
10724                         up_write(&space_info->groups_sem);
10725                         goto next;
10726                 }
10727                 spin_unlock(&block_group->lock);
10728
10729                 /* We don't want to force the issue, only flip if it's ok. */
10730                 ret = inc_block_group_ro(block_group, 0);
10731                 up_write(&space_info->groups_sem);
10732                 if (ret < 0) {
10733                         ret = 0;
10734                         goto next;
10735                 }
10736
10737                 /*
10738                  * Want to do this before we do anything else so we can recover
10739                  * properly if we fail to join the transaction.
10740                  */
10741                 trans = btrfs_start_trans_remove_block_group(fs_info,
10742                                                      block_group->key.objectid);
10743                 if (IS_ERR(trans)) {
10744                         btrfs_dec_block_group_ro(block_group);
10745                         ret = PTR_ERR(trans);
10746                         goto next;
10747                 }
10748
10749                 /*
10750                  * We could have pending pinned extents for this block group,
10751                  * just delete them, we don't care about them anymore.
10752                  */
10753                 start = block_group->key.objectid;
10754                 end = start + block_group->key.offset - 1;
10755                 /*
10756                  * Hold the unused_bg_unpin_mutex lock to avoid racing with
10757                  * btrfs_finish_extent_commit(). If we are at transaction N,
10758                  * another task might be running finish_extent_commit() for the
10759                  * previous transaction N - 1, and have seen a range belonging
10760                  * to the block group in freed_extents[] before we were able to
10761                  * clear the whole block group range from freed_extents[]. This
10762                  * means that task can lookup for the block group after we
10763                  * unpinned it from freed_extents[] and removed it, leading to
10764                  * a BUG_ON() at btrfs_unpin_extent_range().
10765                  */
10766                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
10767                 ret = clear_extent_bits(&fs_info->freed_extents[0], start, end,
10768                                   EXTENT_DIRTY);
10769                 if (ret) {
10770                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10771                         btrfs_dec_block_group_ro(block_group);
10772                         goto end_trans;
10773                 }
10774                 ret = clear_extent_bits(&fs_info->freed_extents[1], start, end,
10775                                   EXTENT_DIRTY);
10776                 if (ret) {
10777                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10778                         btrfs_dec_block_group_ro(block_group);
10779                         goto end_trans;
10780                 }
10781                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10782
10783                 /* Reset pinned so btrfs_put_block_group doesn't complain */
10784                 spin_lock(&space_info->lock);
10785                 spin_lock(&block_group->lock);
10786
10787                 space_info->bytes_pinned -= block_group->pinned;
10788                 space_info->bytes_readonly += block_group->pinned;
10789                 percpu_counter_add(&space_info->total_bytes_pinned,
10790                                    -block_group->pinned);
10791                 block_group->pinned = 0;
10792
10793                 spin_unlock(&block_group->lock);
10794                 spin_unlock(&space_info->lock);
10795
10796                 /* DISCARD can flip during remount */
10797                 trimming = btrfs_test_opt(fs_info, DISCARD);
10798
10799                 /* Implicit trim during transaction commit. */
10800                 if (trimming)
10801                         btrfs_get_block_group_trimming(block_group);
10802
10803                 /*
10804                  * Btrfs_remove_chunk will abort the transaction if things go
10805                  * horribly wrong.
10806                  */
10807                 ret = btrfs_remove_chunk(trans, fs_info,
10808                                          block_group->key.objectid);
10809
10810                 if (ret) {
10811                         if (trimming)
10812                                 btrfs_put_block_group_trimming(block_group);
10813                         goto end_trans;
10814                 }
10815
10816                 /*
10817                  * If we're not mounted with -odiscard, we can just forget
10818                  * about this block group. Otherwise we'll need to wait
10819                  * until transaction commit to do the actual discard.
10820                  */
10821                 if (trimming) {
10822                         spin_lock(&fs_info->unused_bgs_lock);
10823                         /*
10824                          * A concurrent scrub might have added us to the list
10825                          * fs_info->unused_bgs, so use a list_move operation
10826                          * to add the block group to the deleted_bgs list.
10827                          */
10828                         list_move(&block_group->bg_list,
10829                                   &trans->transaction->deleted_bgs);
10830                         spin_unlock(&fs_info->unused_bgs_lock);
10831                         btrfs_get_block_group(block_group);
10832                 }
10833 end_trans:
10834                 btrfs_end_transaction(trans);
10835 next:
10836                 mutex_unlock(&fs_info->delete_unused_bgs_mutex);
10837                 btrfs_put_block_group(block_group);
10838                 spin_lock(&fs_info->unused_bgs_lock);
10839         }
10840         spin_unlock(&fs_info->unused_bgs_lock);
10841 }
10842
10843 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
10844 {
10845         struct btrfs_space_info *space_info;
10846         struct btrfs_super_block *disk_super;
10847         u64 features;
10848         u64 flags;
10849         int mixed = 0;
10850         int ret;
10851
10852         disk_super = fs_info->super_copy;
10853         if (!btrfs_super_root(disk_super))
10854                 return -EINVAL;
10855
10856         features = btrfs_super_incompat_flags(disk_super);
10857         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
10858                 mixed = 1;
10859
10860         flags = BTRFS_BLOCK_GROUP_SYSTEM;
10861         ret = create_space_info(fs_info, flags, &space_info);
10862         if (ret)
10863                 goto out;
10864
10865         if (mixed) {
10866                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
10867                 ret = create_space_info(fs_info, flags, &space_info);
10868         } else {
10869                 flags = BTRFS_BLOCK_GROUP_METADATA;
10870                 ret = create_space_info(fs_info, flags, &space_info);
10871                 if (ret)
10872                         goto out;
10873
10874                 flags = BTRFS_BLOCK_GROUP_DATA;
10875                 ret = create_space_info(fs_info, flags, &space_info);
10876         }
10877 out:
10878         return ret;
10879 }
10880
10881 int btrfs_error_unpin_extent_range(struct btrfs_fs_info *fs_info,
10882                                    u64 start, u64 end)
10883 {
10884         return unpin_extent_range(fs_info, start, end, false);
10885 }
10886
10887 /*
10888  * It used to be that old block groups would be left around forever.
10889  * Iterating over them would be enough to trim unused space.  Since we
10890  * now automatically remove them, we also need to iterate over unallocated
10891  * space.
10892  *
10893  * We don't want a transaction for this since the discard may take a
10894  * substantial amount of time.  We don't require that a transaction be
10895  * running, but we do need to take a running transaction into account
10896  * to ensure that we're not discarding chunks that were released in
10897  * the current transaction.
10898  *
10899  * Holding the chunks lock will prevent other threads from allocating
10900  * or releasing chunks, but it won't prevent a running transaction
10901  * from committing and releasing the memory that the pending chunks
10902  * list head uses.  For that, we need to take a reference to the
10903  * transaction.
10904  */
10905 static int btrfs_trim_free_extents(struct btrfs_device *device,
10906                                    u64 minlen, u64 *trimmed)
10907 {
10908         u64 start = 0, len = 0;
10909         int ret;
10910
10911         *trimmed = 0;
10912
10913         /* Not writeable = nothing to do. */
10914         if (!test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state))
10915                 return 0;
10916
10917         /* No free space = nothing to do. */
10918         if (device->total_bytes <= device->bytes_used)
10919                 return 0;
10920
10921         ret = 0;
10922
10923         while (1) {
10924                 struct btrfs_fs_info *fs_info = device->fs_info;
10925                 struct btrfs_transaction *trans;
10926                 u64 bytes;
10927
10928                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
10929                 if (ret)
10930                         return ret;
10931
10932                 down_read(&fs_info->commit_root_sem);
10933
10934                 spin_lock(&fs_info->trans_lock);
10935                 trans = fs_info->running_transaction;
10936                 if (trans)
10937                         refcount_inc(&trans->use_count);
10938                 spin_unlock(&fs_info->trans_lock);
10939
10940                 ret = find_free_dev_extent_start(trans, device, minlen, start,
10941                                                  &start, &len);
10942                 if (trans)
10943                         btrfs_put_transaction(trans);
10944
10945                 if (ret) {
10946                         up_read(&fs_info->commit_root_sem);
10947                         mutex_unlock(&fs_info->chunk_mutex);
10948                         if (ret == -ENOSPC)
10949                                 ret = 0;
10950                         break;
10951                 }
10952
10953                 ret = btrfs_issue_discard(device->bdev, start, len, &bytes);
10954                 up_read(&fs_info->commit_root_sem);
10955                 mutex_unlock(&fs_info->chunk_mutex);
10956
10957                 if (ret)
10958                         break;
10959
10960                 start += len;
10961                 *trimmed += bytes;
10962
10963                 if (fatal_signal_pending(current)) {
10964                         ret = -ERESTARTSYS;
10965                         break;
10966                 }
10967
10968                 cond_resched();
10969         }
10970
10971         return ret;
10972 }
10973
10974 int btrfs_trim_fs(struct btrfs_fs_info *fs_info, struct fstrim_range *range)
10975 {
10976         struct btrfs_block_group_cache *cache = NULL;
10977         struct btrfs_device *device;
10978         struct list_head *devices;
10979         u64 group_trimmed;
10980         u64 start;
10981         u64 end;
10982         u64 trimmed = 0;
10983         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
10984         int ret = 0;
10985
10986         /*
10987          * try to trim all FS space, our block group may start from non-zero.
10988          */
10989         if (range->len == total_bytes)
10990                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
10991         else
10992                 cache = btrfs_lookup_block_group(fs_info, range->start);
10993
10994         while (cache) {
10995                 if (cache->key.objectid >= (range->start + range->len)) {
10996                         btrfs_put_block_group(cache);
10997                         break;
10998                 }
10999
11000                 start = max(range->start, cache->key.objectid);
11001                 end = min(range->start + range->len,
11002                                 cache->key.objectid + cache->key.offset);
11003
11004                 if (end - start >= range->minlen) {
11005                         if (!block_group_cache_done(cache)) {
11006                                 ret = cache_block_group(cache, 0);
11007                                 if (ret) {
11008                                         btrfs_put_block_group(cache);
11009                                         break;
11010                                 }
11011                                 ret = wait_block_group_cache_done(cache);
11012                                 if (ret) {
11013                                         btrfs_put_block_group(cache);
11014                                         break;
11015                                 }
11016                         }
11017                         ret = btrfs_trim_block_group(cache,
11018                                                      &group_trimmed,
11019                                                      start,
11020                                                      end,
11021                                                      range->minlen);
11022
11023                         trimmed += group_trimmed;
11024                         if (ret) {
11025                                 btrfs_put_block_group(cache);
11026                                 break;
11027                         }
11028                 }
11029
11030                 cache = next_block_group(fs_info, cache);
11031         }
11032
11033         mutex_lock(&fs_info->fs_devices->device_list_mutex);
11034         devices = &fs_info->fs_devices->alloc_list;
11035         list_for_each_entry(device, devices, dev_alloc_list) {
11036                 ret = btrfs_trim_free_extents(device, range->minlen,
11037                                               &group_trimmed);
11038                 if (ret)
11039                         break;
11040
11041                 trimmed += group_trimmed;
11042         }
11043         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
11044
11045         range->len = trimmed;
11046         return ret;
11047 }
11048
11049 /*
11050  * btrfs_{start,end}_write_no_snapshotting() are similar to
11051  * mnt_{want,drop}_write(), they are used to prevent some tasks from writing
11052  * data into the page cache through nocow before the subvolume is snapshoted,
11053  * but flush the data into disk after the snapshot creation, or to prevent
11054  * operations while snapshotting is ongoing and that cause the snapshot to be
11055  * inconsistent (writes followed by expanding truncates for example).
11056  */
11057 void btrfs_end_write_no_snapshotting(struct btrfs_root *root)
11058 {
11059         percpu_counter_dec(&root->subv_writers->counter);
11060         /*
11061          * Make sure counter is updated before we wake up waiters.
11062          */
11063         smp_mb();
11064         if (waitqueue_active(&root->subv_writers->wait))
11065                 wake_up(&root->subv_writers->wait);
11066 }
11067
11068 int btrfs_start_write_no_snapshotting(struct btrfs_root *root)
11069 {
11070         if (atomic_read(&root->will_be_snapshotted))
11071                 return 0;
11072
11073         percpu_counter_inc(&root->subv_writers->counter);
11074         /*
11075          * Make sure counter is updated before we check for snapshot creation.
11076          */
11077         smp_mb();
11078         if (atomic_read(&root->will_be_snapshotted)) {
11079                 btrfs_end_write_no_snapshotting(root);
11080                 return 0;
11081         }
11082         return 1;
11083 }
11084
11085 void btrfs_wait_for_snapshot_creation(struct btrfs_root *root)
11086 {
11087         while (true) {
11088                 int ret;
11089
11090                 ret = btrfs_start_write_no_snapshotting(root);
11091                 if (ret)
11092                         break;
11093                 wait_on_atomic_t(&root->will_be_snapshotted, atomic_t_wait,
11094                                  TASK_UNINTERRUPTIBLE);
11095         }
11096 }