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 "hash.h"
  30 #include "tree-log.h"
  31 #include "disk-io.h"
  32 #include "print-tree.h"
  33 #include "volumes.h"
  34 #include "raid56.h"
  35 #include "locking.h"
  36 #include "free-space-cache.h"
  37 #include "free-space-tree.h"
  38 #include "math.h"
  39 #include "sysfs.h"
  40 #include "qgroup.h"
  41
  42 #undef SCRAMBLE_DELAYED_REFS
  43
  44 /*
  45  * control flags for do_chunk_alloc's force field
  46  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
  47  * if we really need one.
  48  *
  49  * CHUNK_ALLOC_LIMITED means to only try and allocate one
  50  * if we have very few chunks already allocated.  This is
  51  * used as part of the clustering code to help make sure
  52  * we have a good pool of storage to cluster in, without
  53  * filling the FS with empty chunks
  54  *
  55  * CHUNK_ALLOC_FORCE means it must try to allocate one
  56  *
  57  */
  58 enum {
  59         CHUNK_ALLOC_NO_FORCE = 0,
  60         CHUNK_ALLOC_LIMITED = 1,
  61         CHUNK_ALLOC_FORCE = 2,
  62 };
  63
  64 static int update_block_group(struct btrfs_trans_handle *trans,
  65                               struct btrfs_fs_info *fs_info, u64 bytenr,
  66                               u64 num_bytes, int alloc);
  67 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  68                                struct btrfs_fs_info *fs_info,
  69                                 struct btrfs_delayed_ref_node *node, u64 parent,
  70                                 u64 root_objectid, u64 owner_objectid,
  71                                 u64 owner_offset, int refs_to_drop,
  72                                 struct btrfs_delayed_extent_op *extra_op);
  73 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  74                                     struct extent_buffer *leaf,
  75                                     struct btrfs_extent_item *ei);
  76 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  77                                       struct btrfs_fs_info *fs_info,
  78                                       u64 parent, u64 root_objectid,
  79                                       u64 flags, u64 owner, u64 offset,
  80                                       struct btrfs_key *ins, int ref_mod);
  81 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  82                                      struct btrfs_fs_info *fs_info,
  83                                      u64 parent, u64 root_objectid,
  84                                      u64 flags, struct btrfs_disk_key *key,
  85                                      int level, struct btrfs_key *ins);
  86 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
  87                           struct btrfs_fs_info *fs_info, u64 flags,
  88                           int force);
  89 static int find_next_key(struct btrfs_path *path, int level,
  90                          struct btrfs_key *key);
  91 static void dump_space_info(struct btrfs_fs_info *fs_info,
  92                             struct btrfs_space_info *info, u64 bytes,
  93                             int dump_block_groups);
  94 static int btrfs_add_reserved_bytes(struct btrfs_block_group_cache *cache,
  95                                     u64 ram_bytes, u64 num_bytes, int delalloc);
  96 static int btrfs_free_reserved_bytes(struct btrfs_block_group_cache *cache,
  97                                      u64 num_bytes, int delalloc);
  98 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
  99                                u64 num_bytes);
 100 static int __reserve_metadata_bytes(struct btrfs_fs_info *fs_info,
 101                                     struct btrfs_space_info *space_info,
 102                                     u64 orig_bytes,
 103                                     enum btrfs_reserve_flush_enum flush,
 104                                     bool system_chunk);
 105 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
 106                                      struct btrfs_space_info *space_info,
 107                                      u64 num_bytes);
 108 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
 109                                      struct btrfs_space_info *space_info,
 110                                      u64 num_bytes);
 111
 112 static noinline int
 113 block_group_cache_done(struct btrfs_block_group_cache *cache)
 114 {
 115         smp_mb();
 116         return cache->cached == BTRFS_CACHE_FINISHED ||
 117                 cache->cached == BTRFS_CACHE_ERROR;
 118 }
 119
 120 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
 121 {
 122         return (cache->flags & bits) == bits;
 123 }
 124
 125 void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 126 {
 127         atomic_inc(&cache->count);
 128 }
 129
 130 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 131 {
 132         if (atomic_dec_and_test(&cache->count)) {
 133                 WARN_ON(cache->pinned > 0);
 134                 WARN_ON(cache->reserved > 0);
 135
 136                 /*
 137                  * If not empty, someone is still holding mutex of
 138                  * full_stripe_lock, which can only be released by caller.
 139                  * And it will definitely cause use-after-free when caller
 140                  * tries to release full stripe lock.
 141                  *
 142                  * No better way to resolve, but only to warn.
 143                  */
 144                 WARN_ON(!RB_EMPTY_ROOT(&cache->full_stripe_locks_root.root));
 145                 kfree(cache->free_space_ctl);
 146                 kfree(cache);
 147         }
 148 }
 149
 150 /*
 151  * this adds the block group to the fs_info rb tree for the block group
 152  * cache
 153  */
 154 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 155                                 struct btrfs_block_group_cache *block_group)
 156 {
 157         struct rb_node **p;
 158         struct rb_node *parent = NULL;
 159         struct btrfs_block_group_cache *cache;
 160
 161         spin_lock(&info->block_group_cache_lock);
 162         p = &info->block_group_cache_tree.rb_node;
 163
 164         while (*p) {
 165                 parent = *p;
 166                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 167                                  cache_node);
 168                 if (block_group->key.objectid < cache->key.objectid) {
 169                         p = &(*p)->rb_left;
 170                 } else if (block_group->key.objectid > cache->key.objectid) {
 171                         p = &(*p)->rb_right;
 172                 } else {
 173                         spin_unlock(&info->block_group_cache_lock);
 174                         return -EEXIST;
 175                 }
 176         }
 177
 178         rb_link_node(&block_group->cache_node, parent, p);
 179         rb_insert_color(&block_group->cache_node,
 180                         &info->block_group_cache_tree);
 181
 182         if (info->first_logical_byte > block_group->key.objectid)
 183                 info->first_logical_byte = block_group->key.objectid;
 184
 185         spin_unlock(&info->block_group_cache_lock);
 186
 187         return 0;
 188 }
 189
 190 /*
 191  * This will return the block group at or after bytenr if contains is 0, else
 192  * it will return the block group that contains the bytenr
 193  */
 194 static struct btrfs_block_group_cache *
 195 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 196                               int contains)
 197 {
 198         struct btrfs_block_group_cache *cache, *ret = NULL;
 199         struct rb_node *n;
 200         u64 end, start;
 201
 202         spin_lock(&info->block_group_cache_lock);
 203         n = info->block_group_cache_tree.rb_node;
 204
 205         while (n) {
 206                 cache = rb_entry(n, struct btrfs_block_group_cache,
 207                                  cache_node);
 208                 end = cache->key.objectid + cache->key.offset - 1;
 209                 start = cache->key.objectid;
 210
 211                 if (bytenr < start) {
 212                         if (!contains && (!ret || start < ret->key.objectid))
 213                                 ret = cache;
 214                         n = n->rb_left;
 215                 } else if (bytenr > start) {
 216                         if (contains && bytenr <= end) {
 217                                 ret = cache;
 218                                 break;
 219                         }
 220                         n = n->rb_right;
 221                 } else {
 222                         ret = cache;
 223                         break;
 224                 }
 225         }
 226         if (ret) {
 227                 btrfs_get_block_group(ret);
 228                 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
 229                         info->first_logical_byte = ret->key.objectid;
 230         }
 231         spin_unlock(&info->block_group_cache_lock);
 232
 233         return ret;
 234 }
 235
 236 static int add_excluded_extent(struct btrfs_fs_info *fs_info,
 237                                u64 start, u64 num_bytes)
 238 {
 239         u64 end = start + num_bytes - 1;
 240         set_extent_bits(&fs_info->freed_extents[0],
 241                         start, end, EXTENT_UPTODATE);
 242         set_extent_bits(&fs_info->freed_extents[1],
 243                         start, end, EXTENT_UPTODATE);
 244         return 0;
 245 }
 246
 247 static void free_excluded_extents(struct btrfs_fs_info *fs_info,
 248                                   struct btrfs_block_group_cache *cache)
 249 {
 250         u64 start, end;
 251
 252         start = cache->key.objectid;
 253         end = start + cache->key.offset - 1;
 254
 255         clear_extent_bits(&fs_info->freed_extents[0],
 256                           start, end, EXTENT_UPTODATE);
 257         clear_extent_bits(&fs_info->freed_extents[1],
 258                           start, end, EXTENT_UPTODATE);
 259 }
 260
 261 static int exclude_super_stripes(struct btrfs_fs_info *fs_info,
 262                                  struct btrfs_block_group_cache *cache)
 263 {
 264         u64 bytenr;
 265         u64 *logical;
 266         int stripe_len;
 267         int i, nr, ret;
 268
 269         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 270                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 271                 cache->bytes_super += stripe_len;
 272                 ret = add_excluded_extent(fs_info, cache->key.objectid,
 273                                           stripe_len);
 274                 if (ret)
 275                         return ret;
 276         }
 277
 278         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 279                 bytenr = btrfs_sb_offset(i);
 280                 ret = btrfs_rmap_block(fs_info, cache->key.objectid,
 281                                        bytenr, 0, &logical, &nr, &stripe_len);
 282                 if (ret)
 283                         return ret;
 284
 285                 while (nr--) {
 286                         u64 start, len;
 287
 288                         if (logical[nr] > cache->key.objectid +
 289                             cache->key.offset)
 290                                 continue;
 291
 292                         if (logical[nr] + stripe_len <= cache->key.objectid)
 293                                 continue;
 294
 295                         start = logical[nr];
 296                         if (start < cache->key.objectid) {
 297                                 start = cache->key.objectid;
 298                                 len = (logical[nr] + stripe_len) - start;
 299                         } else {
 300                                 len = min_t(u64, stripe_len,
 301                                             cache->key.objectid +
 302                                             cache->key.offset - start);
 303                         }
 304
 305                         cache->bytes_super += len;
 306                         ret = add_excluded_extent(fs_info, start, len);
 307                         if (ret) {
 308                                 kfree(logical);
 309                                 return ret;
 310                         }
 311                 }
 312
 313                 kfree(logical);
 314         }
 315         return 0;
 316 }
 317
 318 static struct btrfs_caching_control *
 319 get_caching_control(struct btrfs_block_group_cache *cache)
 320 {
 321         struct btrfs_caching_control *ctl;
 322
 323         spin_lock(&cache->lock);
 324         if (!cache->caching_ctl) {
 325                 spin_unlock(&cache->lock);
 326                 return NULL;
 327         }
 328
 329         ctl = cache->caching_ctl;
 330         refcount_inc(&ctl->count);
 331         spin_unlock(&cache->lock);
 332         return ctl;
 333 }
 334
 335 static void put_caching_control(struct btrfs_caching_control *ctl)
 336 {
 337         if (refcount_dec_and_test(&ctl->count))
 338                 kfree(ctl);
 339 }
 340
 341 #ifdef CONFIG_BTRFS_DEBUG
 342 static void fragment_free_space(struct btrfs_block_group_cache *block_group)
 343 {
 344         struct btrfs_fs_info *fs_info = block_group->fs_info;
 345         u64 start = block_group->key.objectid;
 346         u64 len = block_group->key.offset;
 347         u64 chunk = block_group->flags & BTRFS_BLOCK_GROUP_METADATA ?
 348                 fs_info->nodesize : fs_info->sectorsize;
 349         u64 step = chunk << 1;
 350
 351         while (len > chunk) {
 352                 btrfs_remove_free_space(block_group, start, chunk);
 353                 start += step;
 354                 if (len < step)
 355                         len = 0;
 356                 else
 357                         len -= step;
 358         }
 359 }
 360 #endif
 361
 362 /*
 363  * this is only called by cache_block_group, since we could have freed extents
 364  * we need to check the pinned_extents for any extents that can't be used yet
 365  * since their free space will be released as soon as the transaction commits.
 366  */
 367 u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 368                        struct btrfs_fs_info *info, u64 start, u64 end)
 369 {
 370         u64 extent_start, extent_end, size, total_added = 0;
 371         int ret;
 372
 373         while (start < end) {
 374                 ret = find_first_extent_bit(info->pinned_extents, start,
 375                                             &extent_start, &extent_end,
 376                                             EXTENT_DIRTY | EXTENT_UPTODATE,
 377                                             NULL);
 378                 if (ret)
 379                         break;
 380
 381                 if (extent_start <= start) {
 382                         start = extent_end + 1;
 383                 } else if (extent_start > start && extent_start < end) {
 384                         size = extent_start - start;
 385                         total_added += size;
 386                         ret = btrfs_add_free_space(block_group, start,
 387                                                    size);
 388                         BUG_ON(ret); /* -ENOMEM or logic error */
 389                         start = extent_end + 1;
 390                 } else {
 391                         break;
 392                 }
 393         }
 394
 395         if (start < end) {
 396                 size = end - start;
 397                 total_added += size;
 398                 ret = btrfs_add_free_space(block_group, start, size);
 399                 BUG_ON(ret); /* -ENOMEM or logic error */
 400         }
 401
 402         return total_added;
 403 }
 404
 405 static int load_extent_tree_free(struct btrfs_caching_control *caching_ctl)
 406 {
 407         struct btrfs_block_group_cache *block_group = caching_ctl->block_group;
 408         struct btrfs_fs_info *fs_info = block_group->fs_info;
 409         struct btrfs_root *extent_root = fs_info->extent_root;
 410         struct btrfs_path *path;
 411         struct extent_buffer *leaf;
 412         struct btrfs_key key;
 413         u64 total_found = 0;
 414         u64 last = 0;
 415         u32 nritems;
 416         int ret;
 417         bool wakeup = true;
 418
 419         path = btrfs_alloc_path();
 420         if (!path)
 421                 return -ENOMEM;
 422
 423         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 424
 425 #ifdef CONFIG_BTRFS_DEBUG
 426         /*
 427          * If we're fragmenting we don't want to make anybody think we can
 428          * allocate from this block group until we've had a chance to fragment
 429          * the free space.
 430          */
 431         if (btrfs_should_fragment_free_space(block_group))
 432                 wakeup = false;
 433 #endif
 434         /*
 435          * We don't want to deadlock with somebody trying to allocate a new
 436          * extent for the extent root while also trying to search the extent
 437          * root to add free space.  So we skip locking and search the commit
 438          * root, since its read-only
 439          */
 440         path->skip_locking = 1;
 441         path->search_commit_root = 1;
 442         path->reada = READA_FORWARD;
 443
 444         key.objectid = last;
 445         key.offset = 0;
 446         key.type = BTRFS_EXTENT_ITEM_KEY;
 447
 448 next:
 449         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 450         if (ret < 0)
 451                 goto out;
 452
 453         leaf = path->nodes[0];
 454         nritems = btrfs_header_nritems(leaf);
 455
 456         while (1) {
 457                 if (btrfs_fs_closing(fs_info) > 1) {
 458                         last = (u64)-1;
 459                         break;
 460                 }
 461
 462                 if (path->slots[0] < nritems) {
 463                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 464                 } else {
 465                         ret = find_next_key(path, 0, &key);
 466                         if (ret)
 467                                 break;
 468
 469                         if (need_resched() ||
 470                             rwsem_is_contended(&fs_info->commit_root_sem)) {
 471                                 if (wakeup)
 472                                         caching_ctl->progress = last;
 473                                 btrfs_release_path(path);
 474                                 up_read(&fs_info->commit_root_sem);
 475                                 mutex_unlock(&caching_ctl->mutex);
 476                                 cond_resched();
 477                                 mutex_lock(&caching_ctl->mutex);
 478                                 down_read(&fs_info->commit_root_sem);
 479                                 goto next;
 480                         }
 481
 482                         ret = btrfs_next_leaf(extent_root, path);
 483                         if (ret < 0)
 484                                 goto out;
 485                         if (ret)
 486                                 break;
 487                         leaf = path->nodes[0];
 488                         nritems = btrfs_header_nritems(leaf);
 489                         continue;
 490                 }
 491
 492                 if (key.objectid < last) {
 493                         key.objectid = last;
 494                         key.offset = 0;
 495                         key.type = BTRFS_EXTENT_ITEM_KEY;
 496
 497                         if (wakeup)
 498                                 caching_ctl->progress = last;
 499                         btrfs_release_path(path);
 500                         goto next;
 501                 }
 502
 503                 if (key.objectid < block_group->key.objectid) {
 504                         path->slots[0]++;
 505                         continue;
 506                 }
 507
 508                 if (key.objectid >= block_group->key.objectid +
 509                     block_group->key.offset)
 510                         break;
 511
 512                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 513                     key.type == BTRFS_METADATA_ITEM_KEY) {
 514                         total_found += add_new_free_space(block_group,
 515                                                           fs_info, last,
 516                                                           key.objectid);
 517                         if (key.type == BTRFS_METADATA_ITEM_KEY)
 518                                 last = key.objectid +
 519                                         fs_info->nodesize;
 520                         else
 521                                 last = key.objectid + key.offset;
 522
 523                         if (total_found > CACHING_CTL_WAKE_UP) {
 524                                 total_found = 0;
 525                                 if (wakeup)
 526                                         wake_up(&caching_ctl->wait);
 527                         }
 528                 }
 529                 path->slots[0]++;
 530         }
 531         ret = 0;
 532
 533         total_found += add_new_free_space(block_group, fs_info, last,
 534                                           block_group->key.objectid +
 535                                           block_group->key.offset);
 536         caching_ctl->progress = (u64)-1;
 537
 538 out:
 539         btrfs_free_path(path);
 540         return ret;
 541 }
 542
 543 static noinline void caching_thread(struct btrfs_work *work)
 544 {
 545         struct btrfs_block_group_cache *block_group;
 546         struct btrfs_fs_info *fs_info;
 547         struct btrfs_caching_control *caching_ctl;
 548         struct btrfs_root *extent_root;
 549         int ret;
 550
 551         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 552         block_group = caching_ctl->block_group;
 553         fs_info = block_group->fs_info;
 554         extent_root = fs_info->extent_root;
 555
 556         mutex_lock(&caching_ctl->mutex);
 557         down_read(&fs_info->commit_root_sem);
 558
 559         if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
 560                 ret = load_free_space_tree(caching_ctl);
 561         else
 562                 ret = load_extent_tree_free(caching_ctl);
 563
 564         spin_lock(&block_group->lock);
 565         block_group->caching_ctl = NULL;
 566         block_group->cached = ret ? BTRFS_CACHE_ERROR : BTRFS_CACHE_FINISHED;
 567         spin_unlock(&block_group->lock);
 568
 569 #ifdef CONFIG_BTRFS_DEBUG
 570         if (btrfs_should_fragment_free_space(block_group)) {
 571                 u64 bytes_used;
 572
 573                 spin_lock(&block_group->space_info->lock);
 574                 spin_lock(&block_group->lock);
 575                 bytes_used = block_group->key.offset -
 576                         btrfs_block_group_used(&block_group->item);
 577                 block_group->space_info->bytes_used += bytes_used >> 1;
 578                 spin_unlock(&block_group->lock);
 579                 spin_unlock(&block_group->space_info->lock);
 580                 fragment_free_space(block_group);
 581         }
 582 #endif
 583
 584         caching_ctl->progress = (u64)-1;
 585
 586         up_read(&fs_info->commit_root_sem);
 587         free_excluded_extents(fs_info, block_group);
 588         mutex_unlock(&caching_ctl->mutex);
 589
 590         wake_up(&caching_ctl->wait);
 591
 592         put_caching_control(caching_ctl);
 593         btrfs_put_block_group(block_group);
 594 }
 595
 596 static int cache_block_group(struct btrfs_block_group_cache *cache,
 597                              int load_cache_only)
 598 {
 599         DEFINE_WAIT(wait);
 600         struct btrfs_fs_info *fs_info = cache->fs_info;
 601         struct btrfs_caching_control *caching_ctl;
 602         int ret = 0;
 603
 604         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 605         if (!caching_ctl)
 606                 return -ENOMEM;
 607
 608         INIT_LIST_HEAD(&caching_ctl->list);
 609         mutex_init(&caching_ctl->mutex);
 610         init_waitqueue_head(&caching_ctl->wait);
 611         caching_ctl->block_group = cache;
 612         caching_ctl->progress = cache->key.objectid;
 613         refcount_set(&caching_ctl->count, 1);
 614         btrfs_init_work(&caching_ctl->work, btrfs_cache_helper,
 615                         caching_thread, NULL, NULL);
 616
 617         spin_lock(&cache->lock);
 618         /*
 619          * This should be a rare occasion, but this could happen I think in the
 620          * case where one thread starts to load the space cache info, and then
 621          * some other thread starts a transaction commit which tries to do an
 622          * allocation while the other thread is still loading the space cache
 623          * info.  The previous loop should have kept us from choosing this block
 624          * group, but if we've moved to the state where we will wait on caching
 625          * block groups we need to first check if we're doing a fast load here,
 626          * so we can wait for it to finish, otherwise we could end up allocating
 627          * from a block group who's cache gets evicted for one reason or
 628          * another.
 629          */
 630         while (cache->cached == BTRFS_CACHE_FAST) {
 631                 struct btrfs_caching_control *ctl;
 632
 633                 ctl = cache->caching_ctl;
 634                 refcount_inc(&ctl->count);
 635                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 636                 spin_unlock(&cache->lock);
 637
 638                 schedule();
 639
 640                 finish_wait(&ctl->wait, &wait);
 641                 put_caching_control(ctl);
 642                 spin_lock(&cache->lock);
 643         }
 644
 645         if (cache->cached != BTRFS_CACHE_NO) {
 646                 spin_unlock(&cache->lock);
 647                 kfree(caching_ctl);
 648                 return 0;
 649         }
 650         WARN_ON(cache->caching_ctl);
 651         cache->caching_ctl = caching_ctl;
 652         cache->cached = BTRFS_CACHE_FAST;
 653         spin_unlock(&cache->lock);
 654
 655         if (fs_info->mount_opt & BTRFS_MOUNT_SPACE_CACHE) {
 656                 mutex_lock(&caching_ctl->mutex);
 657                 ret = load_free_space_cache(fs_info, cache);
 658
 659                 spin_lock(&cache->lock);
 660                 if (ret == 1) {
 661                         cache->caching_ctl = NULL;
 662                         cache->cached = BTRFS_CACHE_FINISHED;
 663                         cache->last_byte_to_unpin = (u64)-1;
 664                         caching_ctl->progress = (u64)-1;
 665                 } else {
 666                         if (load_cache_only) {
 667                                 cache->caching_ctl = NULL;
 668                                 cache->cached = BTRFS_CACHE_NO;
 669                         } else {
 670                                 cache->cached = BTRFS_CACHE_STARTED;
 671                                 cache->has_caching_ctl = 1;
 672                         }
 673                 }
 674                 spin_unlock(&cache->lock);
 675 #ifdef CONFIG_BTRFS_DEBUG
 676                 if (ret == 1 &&
 677                     btrfs_should_fragment_free_space(cache)) {
 678                         u64 bytes_used;
 679
 680                         spin_lock(&cache->space_info->lock);
 681                         spin_lock(&cache->lock);
 682                         bytes_used = cache->key.offset -
 683                                 btrfs_block_group_used(&cache->item);
 684                         cache->space_info->bytes_used += bytes_used >> 1;
 685                         spin_unlock(&cache->lock);
 686                         spin_unlock(&cache->space_info->lock);
 687                         fragment_free_space(cache);
 688                 }
 689 #endif
 690                 mutex_unlock(&caching_ctl->mutex);
 691
 692                 wake_up(&caching_ctl->wait);
 693                 if (ret == 1) {
 694                         put_caching_control(caching_ctl);
 695                         free_excluded_extents(fs_info, cache);
 696                         return 0;
 697                 }
 698         } else {
 699                 /*
 700                  * We're either using the free space tree or no caching at all.
 701                  * Set cached to the appropriate value and wakeup any waiters.
 702                  */
 703                 spin_lock(&cache->lock);
 704                 if (load_cache_only) {
 705                         cache->caching_ctl = NULL;
 706                         cache->cached = BTRFS_CACHE_NO;
 707                 } else {
 708                         cache->cached = BTRFS_CACHE_STARTED;
 709                         cache->has_caching_ctl = 1;
 710                 }
 711                 spin_unlock(&cache->lock);
 712                 wake_up(&caching_ctl->wait);
 713         }
 714
 715         if (load_cache_only) {
 716                 put_caching_control(caching_ctl);
 717                 return 0;
 718         }
 719
 720         down_write(&fs_info->commit_root_sem);
 721         refcount_inc(&caching_ctl->count);
 722         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 723         up_write(&fs_info->commit_root_sem);
 724
 725         btrfs_get_block_group(cache);
 726
 727         btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
 728
 729         return ret;
 730 }
 731
 732 /*
 733  * return the block group that starts at or after bytenr
 734  */
 735 static struct btrfs_block_group_cache *
 736 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 737 {
 738         return block_group_cache_tree_search(info, bytenr, 0);
 739 }
 740
 741 /*
 742  * return the block group that contains the given bytenr
 743  */
 744 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 745                                                  struct btrfs_fs_info *info,
 746                                                  u64 bytenr)
 747 {
 748         return block_group_cache_tree_search(info, bytenr, 1);
 749 }
 750
 751 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 752                                                   u64 flags)
 753 {
 754         struct list_head *head = &info->space_info;
 755         struct btrfs_space_info *found;
 756
 757         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 758
 759         rcu_read_lock();
 760         list_for_each_entry_rcu(found, head, list) {
 761                 if (found->flags & flags) {
 762                         rcu_read_unlock();
 763                         return found;
 764                 }
 765         }
 766         rcu_read_unlock();
 767         return NULL;
 768 }
 769
 770 static void add_pinned_bytes(struct btrfs_fs_info *fs_info, s64 num_bytes,
 771                              u64 owner, u64 root_objectid)
 772 {
 773         struct btrfs_space_info *space_info;
 774         u64 flags;
 775
 776         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
 777                 if (root_objectid == BTRFS_CHUNK_TREE_OBJECTID)
 778                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
 779                 else
 780                         flags = BTRFS_BLOCK_GROUP_METADATA;
 781         } else {
 782                 flags = BTRFS_BLOCK_GROUP_DATA;
 783         }
 784
 785         space_info = __find_space_info(fs_info, flags);
 786         ASSERT(space_info);
 787         percpu_counter_add(&space_info->total_bytes_pinned, num_bytes);
 788 }
 789
 790 /*
 791  * after adding space to the filesystem, we need to clear the full flags
 792  * on all the space infos.
 793  */
 794 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 795 {
 796         struct list_head *head = &info->space_info;
 797         struct btrfs_space_info *found;
 798
 799         rcu_read_lock();
 800         list_for_each_entry_rcu(found, head, list)
 801                 found->full = 0;
 802         rcu_read_unlock();
 803 }
 804
 805 /* simple helper to search for an existing data extent at a given offset */
 806 int btrfs_lookup_data_extent(struct btrfs_fs_info *fs_info, u64 start, u64 len)
 807 {
 808         int ret;
 809         struct btrfs_key key;
 810         struct btrfs_path *path;
 811
 812         path = btrfs_alloc_path();
 813         if (!path)
 814                 return -ENOMEM;
 815
 816         key.objectid = start;
 817         key.offset = len;
 818         key.type = BTRFS_EXTENT_ITEM_KEY;
 819         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
 820         btrfs_free_path(path);
 821         return ret;
 822 }
 823
 824 /*
 825  * helper function to lookup reference count and flags of a tree block.
 826  *
 827  * the head node for delayed ref is used to store the sum of all the
 828  * reference count modifications queued up in the rbtree. the head
 829  * node may also store the extent flags to set. This way you can check
 830  * to see what the reference count and extent flags would be if all of
 831  * the delayed refs are not processed.
 832  */
 833 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 834                              struct btrfs_fs_info *fs_info, u64 bytenr,
 835                              u64 offset, int metadata, u64 *refs, u64 *flags)
 836 {
 837         struct btrfs_delayed_ref_head *head;
 838         struct btrfs_delayed_ref_root *delayed_refs;
 839         struct btrfs_path *path;
 840         struct btrfs_extent_item *ei;
 841         struct extent_buffer *leaf;
 842         struct btrfs_key key;
 843         u32 item_size;
 844         u64 num_refs;
 845         u64 extent_flags;
 846         int ret;
 847
 848         /*
 849          * If we don't have skinny metadata, don't bother doing anything
 850          * different
 851          */
 852         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA)) {
 853                 offset = fs_info->nodesize;
 854                 metadata = 0;
 855         }
 856
 857         path = btrfs_alloc_path();
 858         if (!path)
 859                 return -ENOMEM;
 860
 861         if (!trans) {
 862                 path->skip_locking = 1;
 863                 path->search_commit_root = 1;
 864         }
 865
 866 search_again:
 867         key.objectid = bytenr;
 868         key.offset = offset;
 869         if (metadata)
 870                 key.type = BTRFS_METADATA_ITEM_KEY;
 871         else
 872                 key.type = BTRFS_EXTENT_ITEM_KEY;
 873
 874         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 0);
 875         if (ret < 0)
 876                 goto out_free;
 877
 878         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 879                 if (path->slots[0]) {
 880                         path->slots[0]--;
 881                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 882                                               path->slots[0]);
 883                         if (key.objectid == bytenr &&
 884                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 885                             key.offset == fs_info->nodesize)
 886                                 ret = 0;
 887                 }
 888         }
 889
 890         if (ret == 0) {
 891                 leaf = path->nodes[0];
 892                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 893                 if (item_size >= sizeof(*ei)) {
 894                         ei = btrfs_item_ptr(leaf, path->slots[0],
 895                                             struct btrfs_extent_item);
 896                         num_refs = btrfs_extent_refs(leaf, ei);
 897                         extent_flags = btrfs_extent_flags(leaf, ei);
 898                 } else {
 899 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 900                         struct btrfs_extent_item_v0 *ei0;
 901                         BUG_ON(item_size != sizeof(*ei0));
 902                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 903                                              struct btrfs_extent_item_v0);
 904                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
 905                         /* FIXME: this isn't correct for data */
 906                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 907 #else
 908                         BUG();
 909 #endif
 910                 }
 911                 BUG_ON(num_refs == 0);
 912         } else {
 913                 num_refs = 0;
 914                 extent_flags = 0;
 915                 ret = 0;
 916         }
 917
 918         if (!trans)
 919                 goto out;
 920
 921         delayed_refs = &trans->transaction->delayed_refs;
 922         spin_lock(&delayed_refs->lock);
 923         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
 924         if (head) {
 925                 if (!mutex_trylock(&head->mutex)) {
 926                         refcount_inc(&head->node.refs);
 927                         spin_unlock(&delayed_refs->lock);
 928
 929                         btrfs_release_path(path);
 930
 931                         /*
 932                          * Mutex was contended, block until it's released and try
 933                          * again
 934                          */
 935                         mutex_lock(&head->mutex);
 936                         mutex_unlock(&head->mutex);
 937                         btrfs_put_delayed_ref(&head->node);
 938                         goto search_again;
 939                 }
 940                 spin_lock(&head->lock);
 941                 if (head->extent_op && head->extent_op->update_flags)
 942                         extent_flags |= head->extent_op->flags_to_set;
 943                 else
 944                         BUG_ON(num_refs == 0);
 945
 946                 num_refs += head->node.ref_mod;
 947                 spin_unlock(&head->lock);
 948                 mutex_unlock(&head->mutex);
 949         }
 950         spin_unlock(&delayed_refs->lock);
 951 out:
 952         WARN_ON(num_refs == 0);
 953         if (refs)
 954                 *refs = num_refs;
 955         if (flags)
 956                 *flags = extent_flags;
 957 out_free:
 958         btrfs_free_path(path);
 959         return ret;
 960 }
 961
 962 /*
 963  * Back reference rules.  Back refs have three main goals:
 964  *
 965  * 1) differentiate between all holders of references to an extent so that
 966  *    when a reference is dropped we can make sure it was a valid reference
 967  *    before freeing the extent.
 968  *
 969  * 2) Provide enough information to quickly find the holders of an extent
 970  *    if we notice a given block is corrupted or bad.
 971  *
 972  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 973  *    maintenance.  This is actually the same as #2, but with a slightly
 974  *    different use case.
 975  *
 976  * There are two kinds of back refs. The implicit back refs is optimized
 977  * for pointers in non-shared tree blocks. For a given pointer in a block,
 978  * back refs of this kind provide information about the block's owner tree
 979  * and the pointer's key. These information allow us to find the block by
 980  * b-tree searching. The full back refs is for pointers in tree blocks not
 981  * referenced by their owner trees. The location of tree block is recorded
 982  * in the back refs. Actually the full back refs is generic, and can be
 983  * used in all cases the implicit back refs is used. The major shortcoming
 984  * of the full back refs is its overhead. Every time a tree block gets
 985  * COWed, we have to update back refs entry for all pointers in it.
 986  *
 987  * For a newly allocated tree block, we use implicit back refs for
 988  * pointers in it. This means most tree related operations only involve
 989  * implicit back refs. For a tree block created in old transaction, the
 990  * only way to drop a reference to it is COW it. So we can detect the
 991  * event that tree block loses its owner tree's reference and do the
 992  * back refs conversion.
 993  *
 994  * When a tree block is COWed through a tree, there are four cases:
 995  *
 996  * The reference count of the block is one and the tree is the block's
 997  * owner tree. Nothing to do in this case.
 998  *
 999  * The reference count of the block is one and the tree is not the
1000  * block's owner tree. In this case, full back refs is used for pointers
1001  * in the block. Remove these full back refs, add implicit back refs for
1002  * every pointers in the new block.
1003  *
1004  * The reference count of the block is greater than one and the tree is
1005  * the block's owner tree. In this case, implicit back refs is used for
1006  * pointers in the block. Add full back refs for every pointers in the
1007  * block, increase lower level extents' reference counts. The original
1008  * implicit back refs are entailed to the new block.
1009  *
1010  * The reference count of the block is greater than one and the tree is
1011  * not the block's owner tree. Add implicit back refs for every pointer in
1012  * the new block, increase lower level extents' reference count.
1013  *
1014  * Back Reference Key composing:
1015  *
1016  * The key objectid corresponds to the first byte in the extent,
1017  * The key type is used to differentiate between types of back refs.
1018  * There are different meanings of the key offset for different types
1019  * of back refs.
1020  *
1021  * File extents can be referenced by:
1022  *
1023  * - multiple snapshots, subvolumes, or different generations in one subvol
1024  * - different files inside a single subvolume
1025  * - different offsets inside a file (bookend extents in file.c)
1026  *
1027  * The extent ref structure for the implicit back refs has fields for:
1028  *
1029  * - Objectid of the subvolume root
1030  * - objectid of the file holding the reference
1031  * - original offset in the file
1032  * - how many bookend extents
1033  *
1034  * The key offset for the implicit back refs is hash of the first
1035  * three fields.
1036  *
1037  * The extent ref structure for the full back refs has field for:
1038  *
1039  * - number of pointers in the tree leaf
1040  *
1041  * The key offset for the implicit back refs is the first byte of
1042  * the tree leaf
1043  *
1044  * When a file extent is allocated, The implicit back refs is used.
1045  * the fields are filled in:
1046  *
1047  *     (root_key.objectid, inode objectid, offset in file, 1)
1048  *
1049  * When a file extent is removed file truncation, we find the
1050  * corresponding implicit back refs and check the following fields:
1051  *
1052  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
1053  *
1054  * Btree extents can be referenced by:
1055  *
1056  * - Different subvolumes
1057  *
1058  * Both the implicit back refs and the full back refs for tree blocks
1059  * only consist of key. The key offset for the implicit back refs is
1060  * objectid of block's owner tree. The key offset for the full back refs
1061  * is the first byte of parent block.
1062  *
1063  * When implicit back refs is used, information about the lowest key and
1064  * level of the tree block are required. These information are stored in
1065  * tree block info structure.
1066  */
1067
1068 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1069 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
1070                                   struct btrfs_fs_info *fs_info,
1071                                   struct btrfs_path *path,
1072                                   u64 owner, u32 extra_size)
1073 {
1074         struct btrfs_root *root = fs_info->extent_root;
1075         struct btrfs_extent_item *item;
1076         struct btrfs_extent_item_v0 *ei0;
1077         struct btrfs_extent_ref_v0 *ref0;
1078         struct btrfs_tree_block_info *bi;
1079         struct extent_buffer *leaf;
1080         struct btrfs_key key;
1081         struct btrfs_key found_key;
1082         u32 new_size = sizeof(*item);
1083         u64 refs;
1084         int ret;
1085
1086         leaf = path->nodes[0];
1087         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
1088
1089         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1090         ei0 = btrfs_item_ptr(leaf, path->slots[0],
1091                              struct btrfs_extent_item_v0);
1092         refs = btrfs_extent_refs_v0(leaf, ei0);
1093
1094         if (owner == (u64)-1) {
1095                 while (1) {
1096                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1097                                 ret = btrfs_next_leaf(root, path);
1098                                 if (ret < 0)
1099                                         return ret;
1100                                 BUG_ON(ret > 0); /* Corruption */
1101                                 leaf = path->nodes[0];
1102                         }
1103                         btrfs_item_key_to_cpu(leaf, &found_key,
1104                                               path->slots[0]);
1105                         BUG_ON(key.objectid != found_key.objectid);
1106                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
1107                                 path->slots[0]++;
1108                                 continue;
1109                         }
1110                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1111                                               struct btrfs_extent_ref_v0);
1112                         owner = btrfs_ref_objectid_v0(leaf, ref0);
1113                         break;
1114                 }
1115         }
1116         btrfs_release_path(path);
1117
1118         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1119                 new_size += sizeof(*bi);
1120
1121         new_size -= sizeof(*ei0);
1122         ret = btrfs_search_slot(trans, root, &key, path,
1123                                 new_size + extra_size, 1);
1124         if (ret < 0)
1125                 return ret;
1126         BUG_ON(ret); /* Corruption */
1127
1128         btrfs_extend_item(fs_info, path, new_size);
1129
1130         leaf = path->nodes[0];
1131         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1132         btrfs_set_extent_refs(leaf, item, refs);
1133         /* FIXME: get real generation */
1134         btrfs_set_extent_generation(leaf, item, 0);
1135         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1136                 btrfs_set_extent_flags(leaf, item,
1137                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1138                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1139                 bi = (struct btrfs_tree_block_info *)(item + 1);
1140                 /* FIXME: get first key of the block */
1141                 memzero_extent_buffer(leaf, (unsigned long)bi, sizeof(*bi));
1142                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1143         } else {
1144                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1145         }
1146         btrfs_mark_buffer_dirty(leaf);
1147         return 0;
1148 }
1149 #endif
1150
1151 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1152 {
1153         u32 high_crc = ~(u32)0;
1154         u32 low_crc = ~(u32)0;
1155         __le64 lenum;
1156
1157         lenum = cpu_to_le64(root_objectid);
1158         high_crc = btrfs_crc32c(high_crc, &lenum, sizeof(lenum));
1159         lenum = cpu_to_le64(owner);
1160         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
1161         lenum = cpu_to_le64(offset);
1162         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
1163
1164         return ((u64)high_crc << 31) ^ (u64)low_crc;
1165 }
1166
1167 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1168                                      struct btrfs_extent_data_ref *ref)
1169 {
1170         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1171                                     btrfs_extent_data_ref_objectid(leaf, ref),
1172                                     btrfs_extent_data_ref_offset(leaf, ref));
1173 }
1174
1175 static int match_extent_data_ref(struct extent_buffer *leaf,
1176                                  struct btrfs_extent_data_ref *ref,
1177                                  u64 root_objectid, u64 owner, u64 offset)
1178 {
1179         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1180             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1181             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1182                 return 0;
1183         return 1;
1184 }
1185
1186 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1187                                            struct btrfs_fs_info *fs_info,
1188                                            struct btrfs_path *path,
1189                                            u64 bytenr, u64 parent,
1190                                            u64 root_objectid,
1191                                            u64 owner, u64 offset)
1192 {
1193         struct btrfs_root *root = fs_info->extent_root;
1194         struct btrfs_key key;
1195         struct btrfs_extent_data_ref *ref;
1196         struct extent_buffer *leaf;
1197         u32 nritems;
1198         int ret;
1199         int recow;
1200         int err = -ENOENT;
1201
1202         key.objectid = bytenr;
1203         if (parent) {
1204                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1205                 key.offset = parent;
1206         } else {
1207                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1208                 key.offset = hash_extent_data_ref(root_objectid,
1209                                                   owner, offset);
1210         }
1211 again:
1212         recow = 0;
1213         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1214         if (ret < 0) {
1215                 err = ret;
1216                 goto fail;
1217         }
1218
1219         if (parent) {
1220                 if (!ret)
1221                         return 0;
1222 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1223                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1224                 btrfs_release_path(path);
1225                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1226                 if (ret < 0) {
1227                         err = ret;
1228                         goto fail;
1229                 }
1230                 if (!ret)
1231                         return 0;
1232 #endif
1233                 goto fail;
1234         }
1235
1236         leaf = path->nodes[0];
1237         nritems = btrfs_header_nritems(leaf);
1238         while (1) {
1239                 if (path->slots[0] >= nritems) {
1240                         ret = btrfs_next_leaf(root, path);
1241                         if (ret < 0)
1242                                 err = ret;
1243                         if (ret)
1244                                 goto fail;
1245
1246                         leaf = path->nodes[0];
1247                         nritems = btrfs_header_nritems(leaf);
1248                         recow = 1;
1249                 }
1250
1251                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1252                 if (key.objectid != bytenr ||
1253                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1254                         goto fail;
1255
1256                 ref = btrfs_item_ptr(leaf, path->slots[0],
1257                                      struct btrfs_extent_data_ref);
1258
1259                 if (match_extent_data_ref(leaf, ref, root_objectid,
1260                                           owner, offset)) {
1261                         if (recow) {
1262                                 btrfs_release_path(path);
1263                                 goto again;
1264                         }
1265                         err = 0;
1266                         break;
1267                 }
1268                 path->slots[0]++;
1269         }
1270 fail:
1271         return err;
1272 }
1273
1274 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1275                                            struct btrfs_fs_info *fs_info,
1276                                            struct btrfs_path *path,
1277                                            u64 bytenr, u64 parent,
1278                                            u64 root_objectid, u64 owner,
1279                                            u64 offset, int refs_to_add)
1280 {
1281         struct btrfs_root *root = fs_info->extent_root;
1282         struct btrfs_key key;
1283         struct extent_buffer *leaf;
1284         u32 size;
1285         u32 num_refs;
1286         int ret;
1287
1288         key.objectid = bytenr;
1289         if (parent) {
1290                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1291                 key.offset = parent;
1292                 size = sizeof(struct btrfs_shared_data_ref);
1293         } else {
1294                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1295                 key.offset = hash_extent_data_ref(root_objectid,
1296                                                   owner, offset);
1297                 size = sizeof(struct btrfs_extent_data_ref);
1298         }
1299
1300         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1301         if (ret && ret != -EEXIST)
1302                 goto fail;
1303
1304         leaf = path->nodes[0];
1305         if (parent) {
1306                 struct btrfs_shared_data_ref *ref;
1307                 ref = btrfs_item_ptr(leaf, path->slots[0],
1308                                      struct btrfs_shared_data_ref);
1309                 if (ret == 0) {
1310                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1311                 } else {
1312                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1313                         num_refs += refs_to_add;
1314                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1315                 }
1316         } else {
1317                 struct btrfs_extent_data_ref *ref;
1318                 while (ret == -EEXIST) {
1319                         ref = btrfs_item_ptr(leaf, path->slots[0],
1320                                              struct btrfs_extent_data_ref);
1321                         if (match_extent_data_ref(leaf, ref, root_objectid,
1322                                                   owner, offset))
1323                                 break;
1324                         btrfs_release_path(path);
1325                         key.offset++;
1326                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1327                                                       size);
1328                         if (ret && ret != -EEXIST)
1329                                 goto fail;
1330
1331                         leaf = path->nodes[0];
1332                 }
1333                 ref = btrfs_item_ptr(leaf, path->slots[0],
1334                                      struct btrfs_extent_data_ref);
1335                 if (ret == 0) {
1336                         btrfs_set_extent_data_ref_root(leaf, ref,
1337                                                        root_objectid);
1338                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1339                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1340                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1341                 } else {
1342                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1343                         num_refs += refs_to_add;
1344                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1345                 }
1346         }
1347         btrfs_mark_buffer_dirty(leaf);
1348         ret = 0;
1349 fail:
1350         btrfs_release_path(path);
1351         return ret;
1352 }
1353
1354 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1355                                            struct btrfs_fs_info *fs_info,
1356                                            struct btrfs_path *path,
1357                                            int refs_to_drop, int *last_ref)
1358 {
1359         struct btrfs_key key;
1360         struct btrfs_extent_data_ref *ref1 = NULL;
1361         struct btrfs_shared_data_ref *ref2 = NULL;
1362         struct extent_buffer *leaf;
1363         u32 num_refs = 0;
1364         int ret = 0;
1365
1366         leaf = path->nodes[0];
1367         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1368
1369         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1370                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1371                                       struct btrfs_extent_data_ref);
1372                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1373         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1374                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1375                                       struct btrfs_shared_data_ref);
1376                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1377 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1378         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1379                 struct btrfs_extent_ref_v0 *ref0;
1380                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1381                                       struct btrfs_extent_ref_v0);
1382                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1383 #endif
1384         } else {
1385                 BUG();
1386         }
1387
1388         BUG_ON(num_refs < refs_to_drop);
1389         num_refs -= refs_to_drop;
1390
1391         if (num_refs == 0) {
1392                 ret = btrfs_del_item(trans, fs_info->extent_root, path);
1393                 *last_ref = 1;
1394         } else {
1395                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1396                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1397                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1398                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1399 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1400                 else {
1401                         struct btrfs_extent_ref_v0 *ref0;
1402                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1403                                         struct btrfs_extent_ref_v0);
1404                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1405                 }
1406 #endif
1407                 btrfs_mark_buffer_dirty(leaf);
1408         }
1409         return ret;
1410 }
1411
1412 static noinline u32 extent_data_ref_count(struct btrfs_path *path,
1413                                           struct btrfs_extent_inline_ref *iref)
1414 {
1415         struct btrfs_key key;
1416         struct extent_buffer *leaf;
1417         struct btrfs_extent_data_ref *ref1;
1418         struct btrfs_shared_data_ref *ref2;
1419         u32 num_refs = 0;
1420
1421         leaf = path->nodes[0];
1422         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1423         if (iref) {
1424                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1425                     BTRFS_EXTENT_DATA_REF_KEY) {
1426                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1427                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1428                 } else {
1429                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1430                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1431                 }
1432         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1433                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1434                                       struct btrfs_extent_data_ref);
1435                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1436         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1437                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1438                                       struct btrfs_shared_data_ref);
1439                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1440 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1441         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1442                 struct btrfs_extent_ref_v0 *ref0;
1443                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1444                                       struct btrfs_extent_ref_v0);
1445                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1446 #endif
1447         } else {
1448                 WARN_ON(1);
1449         }
1450         return num_refs;
1451 }
1452
1453 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1454                                           struct btrfs_fs_info *fs_info,
1455                                           struct btrfs_path *path,
1456                                           u64 bytenr, u64 parent,
1457                                           u64 root_objectid)
1458 {
1459         struct btrfs_root *root = fs_info->extent_root;
1460         struct btrfs_key key;
1461         int ret;
1462
1463         key.objectid = bytenr;
1464         if (parent) {
1465                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1466                 key.offset = parent;
1467         } else {
1468                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1469                 key.offset = root_objectid;
1470         }
1471
1472         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1473         if (ret > 0)
1474                 ret = -ENOENT;
1475 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1476         if (ret == -ENOENT && parent) {
1477                 btrfs_release_path(path);
1478                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1479                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1480                 if (ret > 0)
1481                         ret = -ENOENT;
1482         }
1483 #endif
1484         return ret;
1485 }
1486
1487 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1488                                           struct btrfs_fs_info *fs_info,
1489                                           struct btrfs_path *path,
1490                                           u64 bytenr, u64 parent,
1491                                           u64 root_objectid)
1492 {
1493         struct btrfs_key key;
1494         int ret;
1495
1496         key.objectid = bytenr;
1497         if (parent) {
1498                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1499                 key.offset = parent;
1500         } else {
1501                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1502                 key.offset = root_objectid;
1503         }
1504
1505         ret = btrfs_insert_empty_item(trans, fs_info->extent_root,
1506                                       path, &key, 0);
1507         btrfs_release_path(path);
1508         return ret;
1509 }
1510
1511 static inline int extent_ref_type(u64 parent, u64 owner)
1512 {
1513         int type;
1514         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1515                 if (parent > 0)
1516                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1517                 else
1518                         type = BTRFS_TREE_BLOCK_REF_KEY;
1519         } else {
1520                 if (parent > 0)
1521                         type = BTRFS_SHARED_DATA_REF_KEY;
1522                 else
1523                         type = BTRFS_EXTENT_DATA_REF_KEY;
1524         }
1525         return type;
1526 }
1527
1528 static int find_next_key(struct btrfs_path *path, int level,
1529                          struct btrfs_key *key)
1530
1531 {
1532         for (; level < BTRFS_MAX_LEVEL; level++) {
1533                 if (!path->nodes[level])
1534                         break;
1535                 if (path->slots[level] + 1 >=
1536                     btrfs_header_nritems(path->nodes[level]))
1537                         continue;
1538                 if (level == 0)
1539                         btrfs_item_key_to_cpu(path->nodes[level], key,
1540                                               path->slots[level] + 1);
1541                 else
1542                         btrfs_node_key_to_cpu(path->nodes[level], key,
1543                                               path->slots[level] + 1);
1544                 return 0;
1545         }
1546         return 1;
1547 }
1548
1549 /*
1550  * look for inline back ref. if back ref is found, *ref_ret is set
1551  * to the address of inline back ref, and 0 is returned.
1552  *
1553  * if back ref isn't found, *ref_ret is set to the address where it
1554  * should be inserted, and -ENOENT is returned.
1555  *
1556  * if insert is true and there are too many inline back refs, the path
1557  * points to the extent item, and -EAGAIN is returned.
1558  *
1559  * NOTE: inline back refs are ordered in the same way that back ref
1560  *       items in the tree are ordered.
1561  */
1562 static noinline_for_stack
1563 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1564                                  struct btrfs_fs_info *fs_info,
1565                                  struct btrfs_path *path,
1566                                  struct btrfs_extent_inline_ref **ref_ret,
1567                                  u64 bytenr, u64 num_bytes,
1568                                  u64 parent, u64 root_objectid,
1569                                  u64 owner, u64 offset, int insert)
1570 {
1571         struct btrfs_root *root = fs_info->extent_root;
1572         struct btrfs_key key;
1573         struct extent_buffer *leaf;
1574         struct btrfs_extent_item *ei;
1575         struct btrfs_extent_inline_ref *iref;
1576         u64 flags;
1577         u64 item_size;
1578         unsigned long ptr;
1579         unsigned long end;
1580         int extra_size;
1581         int type;
1582         int want;
1583         int ret;
1584         int err = 0;
1585         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
1586
1587         key.objectid = bytenr;
1588         key.type = BTRFS_EXTENT_ITEM_KEY;
1589         key.offset = num_bytes;
1590
1591         want = extent_ref_type(parent, owner);
1592         if (insert) {
1593                 extra_size = btrfs_extent_inline_ref_size(want);
1594                 path->keep_locks = 1;
1595         } else
1596                 extra_size = -1;
1597
1598         /*
1599          * Owner is our parent level, so we can just add one to get the level
1600          * for the block we are interested in.
1601          */
1602         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
1603                 key.type = BTRFS_METADATA_ITEM_KEY;
1604                 key.offset = owner;
1605         }
1606
1607 again:
1608         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1609         if (ret < 0) {
1610                 err = ret;
1611                 goto out;
1612         }
1613
1614         /*
1615          * We may be a newly converted file system which still has the old fat
1616          * extent entries for metadata, so try and see if we have one of those.
1617          */
1618         if (ret > 0 && skinny_metadata) {
1619                 skinny_metadata = false;
1620                 if (path->slots[0]) {
1621                         path->slots[0]--;
1622                         btrfs_item_key_to_cpu(path->nodes[0], &key,
1623                                               path->slots[0]);
1624                         if (key.objectid == bytenr &&
1625                             key.type == BTRFS_EXTENT_ITEM_KEY &&
1626                             key.offset == num_bytes)
1627                                 ret = 0;
1628                 }
1629                 if (ret) {
1630                         key.objectid = bytenr;
1631                         key.type = BTRFS_EXTENT_ITEM_KEY;
1632                         key.offset = num_bytes;
1633                         btrfs_release_path(path);
1634                         goto again;
1635                 }
1636         }
1637
1638         if (ret && !insert) {
1639                 err = -ENOENT;
1640                 goto out;
1641         } else if (WARN_ON(ret)) {
1642                 err = -EIO;
1643                 goto out;
1644         }
1645
1646         leaf = path->nodes[0];
1647         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1648 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1649         if (item_size < sizeof(*ei)) {
1650                 if (!insert) {
1651                         err = -ENOENT;
1652                         goto out;
1653                 }
1654                 ret = convert_extent_item_v0(trans, fs_info, path, owner,
1655                                              extra_size);
1656                 if (ret < 0) {
1657                         err = ret;
1658                         goto out;
1659                 }
1660                 leaf = path->nodes[0];
1661                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1662         }
1663 #endif
1664         BUG_ON(item_size < sizeof(*ei));
1665
1666         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1667         flags = btrfs_extent_flags(leaf, ei);
1668
1669         ptr = (unsigned long)(ei + 1);
1670         end = (unsigned long)ei + item_size;
1671
1672         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
1673                 ptr += sizeof(struct btrfs_tree_block_info);
1674                 BUG_ON(ptr > end);
1675         }
1676
1677         err = -ENOENT;
1678         while (1) {
1679                 if (ptr >= end) {
1680                         WARN_ON(ptr > end);
1681                         break;
1682                 }
1683                 iref = (struct btrfs_extent_inline_ref *)ptr;
1684                 type = btrfs_extent_inline_ref_type(leaf, iref);
1685                 if (want < type)
1686                         break;
1687                 if (want > type) {
1688                         ptr += btrfs_extent_inline_ref_size(type);
1689                         continue;
1690                 }
1691
1692                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1693                         struct btrfs_extent_data_ref *dref;
1694                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1695                         if (match_extent_data_ref(leaf, dref, root_objectid,
1696                                                   owner, offset)) {
1697                                 err = 0;
1698                                 break;
1699                         }
1700                         if (hash_extent_data_ref_item(leaf, dref) <
1701                             hash_extent_data_ref(root_objectid, owner, offset))
1702                                 break;
1703                 } else {
1704                         u64 ref_offset;
1705                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1706                         if (parent > 0) {
1707                                 if (parent == ref_offset) {
1708                                         err = 0;
1709                                         break;
1710                                 }
1711                                 if (ref_offset < parent)
1712                                         break;
1713                         } else {
1714                                 if (root_objectid == ref_offset) {
1715                                         err = 0;
1716                                         break;
1717                                 }
1718                                 if (ref_offset < root_objectid)
1719                                         break;
1720                         }
1721                 }
1722                 ptr += btrfs_extent_inline_ref_size(type);
1723         }
1724         if (err == -ENOENT && insert) {
1725                 if (item_size + extra_size >=
1726                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1727                         err = -EAGAIN;
1728                         goto out;
1729                 }
1730                 /*
1731                  * To add new inline back ref, we have to make sure
1732                  * there is no corresponding back ref item.
1733                  * For simplicity, we just do not add new inline back
1734                  * ref if there is any kind of item for this block
1735                  */
1736                 if (find_next_key(path, 0, &key) == 0 &&
1737                     key.objectid == bytenr &&
1738                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1739                         err = -EAGAIN;
1740                         goto out;
1741                 }
1742         }
1743         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1744 out:
1745         if (insert) {
1746                 path->keep_locks = 0;
1747                 btrfs_unlock_up_safe(path, 1);
1748         }
1749         return err;
1750 }
1751
1752 /*
1753  * helper to add new inline back ref
1754  */
1755 static noinline_for_stack
1756 void setup_inline_extent_backref(struct btrfs_fs_info *fs_info,
1757                                  struct btrfs_path *path,
1758                                  struct btrfs_extent_inline_ref *iref,
1759                                  u64 parent, u64 root_objectid,
1760                                  u64 owner, u64 offset, int refs_to_add,
1761                                  struct btrfs_delayed_extent_op *extent_op)
1762 {
1763         struct extent_buffer *leaf;
1764         struct btrfs_extent_item *ei;
1765         unsigned long ptr;
1766         unsigned long end;
1767         unsigned long item_offset;
1768         u64 refs;
1769         int size;
1770         int type;
1771
1772         leaf = path->nodes[0];
1773         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1774         item_offset = (unsigned long)iref - (unsigned long)ei;
1775
1776         type = extent_ref_type(parent, owner);
1777         size = btrfs_extent_inline_ref_size(type);
1778
1779         btrfs_extend_item(fs_info, path, size);
1780
1781         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1782         refs = btrfs_extent_refs(leaf, ei);
1783         refs += refs_to_add;
1784         btrfs_set_extent_refs(leaf, ei, refs);
1785         if (extent_op)
1786                 __run_delayed_extent_op(extent_op, leaf, ei);
1787
1788         ptr = (unsigned long)ei + item_offset;
1789         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1790         if (ptr < end - size)
1791                 memmove_extent_buffer(leaf, ptr + size, ptr,
1792                                       end - size - ptr);
1793
1794         iref = (struct btrfs_extent_inline_ref *)ptr;
1795         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1796         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1797                 struct btrfs_extent_data_ref *dref;
1798                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1799                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1800                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1801                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1802                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1803         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1804                 struct btrfs_shared_data_ref *sref;
1805                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1806                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1807                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1808         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1809                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1810         } else {
1811                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1812         }
1813         btrfs_mark_buffer_dirty(leaf);
1814 }
1815
1816 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1817                                  struct btrfs_fs_info *fs_info,
1818                                  struct btrfs_path *path,
1819                                  struct btrfs_extent_inline_ref **ref_ret,
1820                                  u64 bytenr, u64 num_bytes, u64 parent,
1821                                  u64 root_objectid, u64 owner, u64 offset)
1822 {
1823         int ret;
1824
1825         ret = lookup_inline_extent_backref(trans, fs_info, path, ref_ret,
1826                                            bytenr, num_bytes, parent,
1827                                            root_objectid, owner, offset, 0);
1828         if (ret != -ENOENT)
1829                 return ret;
1830
1831         btrfs_release_path(path);
1832         *ref_ret = NULL;
1833
1834         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1835                 ret = lookup_tree_block_ref(trans, fs_info, path, bytenr,
1836                                             parent, root_objectid);
1837         } else {
1838                 ret = lookup_extent_data_ref(trans, fs_info, path, bytenr,
1839                                              parent, root_objectid, owner,
1840                                              offset);
1841         }
1842         return ret;
1843 }
1844
1845 /*
1846  * helper to update/remove inline back ref
1847  */
1848 static noinline_for_stack
1849 void update_inline_extent_backref(struct btrfs_fs_info *fs_info,
1850                                   struct btrfs_path *path,
1851                                   struct btrfs_extent_inline_ref *iref,
1852                                   int refs_to_mod,
1853                                   struct btrfs_delayed_extent_op *extent_op,
1854                                   int *last_ref)
1855 {
1856         struct extent_buffer *leaf;
1857         struct btrfs_extent_item *ei;
1858         struct btrfs_extent_data_ref *dref = NULL;
1859         struct btrfs_shared_data_ref *sref = NULL;
1860         unsigned long ptr;
1861         unsigned long end;
1862         u32 item_size;
1863         int size;
1864         int type;
1865         u64 refs;
1866
1867         leaf = path->nodes[0];
1868         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1869         refs = btrfs_extent_refs(leaf, ei);
1870         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1871         refs += refs_to_mod;
1872         btrfs_set_extent_refs(leaf, ei, refs);
1873         if (extent_op)
1874                 __run_delayed_extent_op(extent_op, leaf, ei);
1875
1876         type = btrfs_extent_inline_ref_type(leaf, iref);
1877
1878         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1879                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1880                 refs = btrfs_extent_data_ref_count(leaf, dref);
1881         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1882                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1883                 refs = btrfs_shared_data_ref_count(leaf, sref);
1884         } else {
1885                 refs = 1;
1886                 BUG_ON(refs_to_mod != -1);
1887         }
1888
1889         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1890         refs += refs_to_mod;
1891
1892         if (refs > 0) {
1893                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1894                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1895                 else
1896                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1897         } else {
1898                 *last_ref = 1;
1899                 size =  btrfs_extent_inline_ref_size(type);
1900                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1901                 ptr = (unsigned long)iref;
1902                 end = (unsigned long)ei + item_size;
1903                 if (ptr + size < end)
1904                         memmove_extent_buffer(leaf, ptr, ptr + size,
1905                                               end - ptr - size);
1906                 item_size -= size;
1907                 btrfs_truncate_item(fs_info, path, item_size, 1);
1908         }
1909         btrfs_mark_buffer_dirty(leaf);
1910 }
1911
1912 static noinline_for_stack
1913 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1914                                  struct btrfs_fs_info *fs_info,
1915                                  struct btrfs_path *path,
1916                                  u64 bytenr, u64 num_bytes, u64 parent,
1917                                  u64 root_objectid, u64 owner,
1918                                  u64 offset, int refs_to_add,
1919                                  struct btrfs_delayed_extent_op *extent_op)
1920 {
1921         struct btrfs_extent_inline_ref *iref;
1922         int ret;
1923
1924         ret = lookup_inline_extent_backref(trans, fs_info, path, &iref,
1925                                            bytenr, num_bytes, parent,
1926                                            root_objectid, owner, offset, 1);
1927         if (ret == 0) {
1928                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1929                 update_inline_extent_backref(fs_info, path, iref,
1930                                              refs_to_add, extent_op, NULL);
1931         } else if (ret == -ENOENT) {
1932                 setup_inline_extent_backref(fs_info, path, iref, parent,
1933                                             root_objectid, owner, offset,
1934                                             refs_to_add, extent_op);
1935                 ret = 0;
1936         }
1937         return ret;
1938 }
1939
1940 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1941                                  struct btrfs_fs_info *fs_info,
1942                                  struct btrfs_path *path,
1943                                  u64 bytenr, u64 parent, u64 root_objectid,
1944                                  u64 owner, u64 offset, int refs_to_add)
1945 {
1946         int ret;
1947         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1948                 BUG_ON(refs_to_add != 1);
1949                 ret = insert_tree_block_ref(trans, fs_info, path, bytenr,
1950                                             parent, root_objectid);
1951         } else {
1952                 ret = insert_extent_data_ref(trans, fs_info, path, bytenr,
1953                                              parent, root_objectid,
1954                                              owner, offset, refs_to_add);
1955         }
1956         return ret;
1957 }
1958
1959 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1960                                  struct btrfs_fs_info *fs_info,
1961                                  struct btrfs_path *path,
1962                                  struct btrfs_extent_inline_ref *iref,
1963                                  int refs_to_drop, int is_data, int *last_ref)
1964 {
1965         int ret = 0;
1966
1967         BUG_ON(!is_data && refs_to_drop != 1);
1968         if (iref) {
1969                 update_inline_extent_backref(fs_info, path, iref,
1970                                              -refs_to_drop, NULL, last_ref);
1971         } else if (is_data) {
1972                 ret = remove_extent_data_ref(trans, fs_info, path, refs_to_drop,
1973                                              last_ref);
1974         } else {
1975                 *last_ref = 1;
1976                 ret = btrfs_del_item(trans, fs_info->extent_root, path);
1977         }
1978         return ret;
1979 }
1980
1981 #define in_range(b, first, len)        ((b) >= (first) && (b) < (first) + (len))
1982 static int btrfs_issue_discard(struct block_device *bdev, u64 start, u64 len,
1983                                u64 *discarded_bytes)
1984 {
1985         int j, ret = 0;
1986         u64 bytes_left, end;
1987         u64 aligned_start = ALIGN(start, 1 << 9);
1988
1989         if (WARN_ON(start != aligned_start)) {
1990                 len -= aligned_start - start;
1991                 len = round_down(len, 1 << 9);
1992                 start = aligned_start;
1993         }
1994
1995         *discarded_bytes = 0;
1996
1997         if (!len)
1998                 return 0;
1999
2000         end = start + len;
2001         bytes_left = len;
2002
2003         /* Skip any superblocks on this device. */
2004         for (j = 0; j < BTRFS_SUPER_MIRROR_MAX; j++) {
2005                 u64 sb_start = btrfs_sb_offset(j);
2006                 u64 sb_end = sb_start + BTRFS_SUPER_INFO_SIZE;
2007                 u64 size = sb_start - start;
2008
2009                 if (!in_range(sb_start, start, bytes_left) &&
2010                     !in_range(sb_end, start, bytes_left) &&
2011                     !in_range(start, sb_start, BTRFS_SUPER_INFO_SIZE))
2012                         continue;
2013
2014                 /*
2015                  * Superblock spans beginning of range.  Adjust start and
2016                  * try again.
2017                  */
2018                 if (sb_start <= start) {
2019                         start += sb_end - start;
2020                         if (start > end) {
2021                                 bytes_left = 0;
2022                                 break;
2023                         }
2024                         bytes_left = end - start;
2025                         continue;
2026                 }
2027
2028                 if (size) {
2029                         ret = blkdev_issue_discard(bdev, start >> 9, size >> 9,
2030                                                    GFP_NOFS, 0);
2031                         if (!ret)
2032                                 *discarded_bytes += size;
2033                         else if (ret != -EOPNOTSUPP)
2034                                 return ret;
2035                 }
2036
2037                 start = sb_end;
2038                 if (start > end) {
2039                         bytes_left = 0;
2040                         break;
2041                 }
2042                 bytes_left = end - start;
2043         }
2044
2045         if (bytes_left) {
2046                 ret = blkdev_issue_discard(bdev, start >> 9, bytes_left >> 9,
2047                                            GFP_NOFS, 0);
2048                 if (!ret)
2049                         *discarded_bytes += bytes_left;
2050         }
2051         return ret;
2052 }
2053
2054 int btrfs_discard_extent(struct btrfs_fs_info *fs_info, u64 bytenr,
2055                          u64 num_bytes, u64 *actual_bytes)
2056 {
2057         int ret;
2058         u64 discarded_bytes = 0;
2059         struct btrfs_bio *bbio = NULL;
2060
2061
2062         /*
2063          * Avoid races with device replace and make sure our bbio has devices
2064          * associated to its stripes that don't go away while we are discarding.
2065          */
2066         btrfs_bio_counter_inc_blocked(fs_info);
2067         /* Tell the block device(s) that the sectors can be discarded */
2068         ret = btrfs_map_block(fs_info, BTRFS_MAP_DISCARD, bytenr, &num_bytes,
2069                               &bbio, 0);
2070         /* Error condition is -ENOMEM */
2071         if (!ret) {
2072                 struct btrfs_bio_stripe *stripe = bbio->stripes;
2073                 int i;
2074
2075
2076                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
2077                         u64 bytes;
2078                         if (!stripe->dev->can_discard)
2079                                 continue;
2080
2081                         ret = btrfs_issue_discard(stripe->dev->bdev,
2082                                                   stripe->physical,
2083                                                   stripe->length,
2084                                                   &bytes);
2085                         if (!ret)
2086                                 discarded_bytes += bytes;
2087                         else if (ret != -EOPNOTSUPP)
2088                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
2089
2090                         /*
2091                          * Just in case we get back EOPNOTSUPP for some reason,
2092                          * just ignore the return value so we don't screw up
2093                          * people calling discard_extent.
2094                          */
2095                         ret = 0;
2096                 }
2097                 btrfs_put_bbio(bbio);
2098         }
2099         btrfs_bio_counter_dec(fs_info);
2100
2101         if (actual_bytes)
2102                 *actual_bytes = discarded_bytes;
2103
2104
2105         if (ret == -EOPNOTSUPP)
2106                 ret = 0;
2107         return ret;
2108 }
2109
2110 /* Can return -ENOMEM */
2111 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2112                          struct btrfs_fs_info *fs_info,
2113                          u64 bytenr, u64 num_bytes, u64 parent,
2114                          u64 root_objectid, u64 owner, u64 offset)
2115 {
2116         int old_ref_mod, new_ref_mod;
2117         int ret;
2118
2119         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
2120                root_objectid == BTRFS_TREE_LOG_OBJECTID);
2121
2122         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
2123                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
2124                                                  num_bytes, parent,
2125                                                  root_objectid, (int)owner,
2126                                                  BTRFS_ADD_DELAYED_REF, NULL,
2127                                                  &old_ref_mod, &new_ref_mod);
2128         } else {
2129                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
2130                                                  num_bytes, parent,
2131                                                  root_objectid, owner, offset,
2132                                                  0, BTRFS_ADD_DELAYED_REF,
2133                                                  &old_ref_mod, &new_ref_mod);
2134         }
2135
2136         if (ret == 0 && old_ref_mod < 0 && new_ref_mod >= 0)
2137                 add_pinned_bytes(fs_info, -num_bytes, owner, root_objectid);
2138
2139         return ret;
2140 }
2141
2142 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2143                                   struct btrfs_fs_info *fs_info,
2144                                   struct btrfs_delayed_ref_node *node,
2145                                   u64 parent, u64 root_objectid,
2146                                   u64 owner, u64 offset, int refs_to_add,
2147                                   struct btrfs_delayed_extent_op *extent_op)
2148 {
2149         struct btrfs_path *path;
2150         struct extent_buffer *leaf;
2151         struct btrfs_extent_item *item;
2152         struct btrfs_key key;
2153         u64 bytenr = node->bytenr;
2154         u64 num_bytes = node->num_bytes;
2155         u64 refs;
2156         int ret;
2157
2158         path = btrfs_alloc_path();
2159         if (!path)
2160                 return -ENOMEM;
2161
2162         path->reada = READA_FORWARD;
2163         path->leave_spinning = 1;
2164         /* this will setup the path even if it fails to insert the back ref */
2165         ret = insert_inline_extent_backref(trans, fs_info, path, bytenr,
2166                                            num_bytes, parent, root_objectid,
2167                                            owner, offset,
2168                                            refs_to_add, extent_op);
2169         if ((ret < 0 && ret != -EAGAIN) || !ret)
2170                 goto out;
2171
2172         /*
2173          * Ok we had -EAGAIN which means we didn't have space to insert and
2174          * inline extent ref, so just update the reference count and add a
2175          * normal backref.
2176          */
2177         leaf = path->nodes[0];
2178         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2179         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2180         refs = btrfs_extent_refs(leaf, item);
2181         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
2182         if (extent_op)
2183                 __run_delayed_extent_op(extent_op, leaf, item);
2184
2185         btrfs_mark_buffer_dirty(leaf);
2186         btrfs_release_path(path);
2187
2188         path->reada = READA_FORWARD;
2189         path->leave_spinning = 1;
2190         /* now insert the actual backref */
2191         ret = insert_extent_backref(trans, fs_info, path, bytenr, parent,
2192                                     root_objectid, owner, offset, refs_to_add);
2193         if (ret)
2194                 btrfs_abort_transaction(trans, ret);
2195 out:
2196         btrfs_free_path(path);
2197         return ret;
2198 }
2199
2200 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
2201                                 struct btrfs_fs_info *fs_info,
2202                                 struct btrfs_delayed_ref_node *node,
2203                                 struct btrfs_delayed_extent_op *extent_op,
2204                                 int insert_reserved)
2205 {
2206         int ret = 0;
2207         struct btrfs_delayed_data_ref *ref;
2208         struct btrfs_key ins;
2209         u64 parent = 0;
2210         u64 ref_root = 0;
2211         u64 flags = 0;
2212
2213         ins.objectid = node->bytenr;
2214         ins.offset = node->num_bytes;
2215         ins.type = BTRFS_EXTENT_ITEM_KEY;
2216
2217         ref = btrfs_delayed_node_to_data_ref(node);
2218         trace_run_delayed_data_ref(fs_info, node, ref, node->action);
2219
2220         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2221                 parent = ref->parent;
2222         ref_root = ref->root;
2223
2224         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2225                 if (extent_op)
2226                         flags |= extent_op->flags_to_set;
2227                 ret = alloc_reserved_file_extent(trans, fs_info,
2228                                                  parent, ref_root, flags,
2229                                                  ref->objectid, ref->offset,
2230                                                  &ins, node->ref_mod);
2231         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2232                 ret = __btrfs_inc_extent_ref(trans, fs_info, node, parent,
2233                                              ref_root, ref->objectid,
2234                                              ref->offset, node->ref_mod,
2235                                              extent_op);
2236         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2237                 ret = __btrfs_free_extent(trans, fs_info, node, parent,
2238                                           ref_root, ref->objectid,
2239                                           ref->offset, node->ref_mod,
2240                                           extent_op);
2241         } else {
2242                 BUG();
2243         }
2244         return ret;
2245 }
2246
2247 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2248                                     struct extent_buffer *leaf,
2249                                     struct btrfs_extent_item *ei)
2250 {
2251         u64 flags = btrfs_extent_flags(leaf, ei);
2252         if (extent_op->update_flags) {
2253                 flags |= extent_op->flags_to_set;
2254                 btrfs_set_extent_flags(leaf, ei, flags);
2255         }
2256
2257         if (extent_op->update_key) {
2258                 struct btrfs_tree_block_info *bi;
2259                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2260                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2261                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2262         }
2263 }
2264
2265 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2266                                  struct btrfs_fs_info *fs_info,
2267                                  struct btrfs_delayed_ref_node *node,
2268                                  struct btrfs_delayed_extent_op *extent_op)
2269 {
2270         struct btrfs_key key;
2271         struct btrfs_path *path;
2272         struct btrfs_extent_item *ei;
2273         struct extent_buffer *leaf;
2274         u32 item_size;
2275         int ret;
2276         int err = 0;
2277         int metadata = !extent_op->is_data;
2278
2279         if (trans->aborted)
2280                 return 0;
2281
2282         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2283                 metadata = 0;
2284
2285         path = btrfs_alloc_path();
2286         if (!path)
2287                 return -ENOMEM;
2288
2289         key.objectid = node->bytenr;
2290
2291         if (metadata) {
2292                 key.type = BTRFS_METADATA_ITEM_KEY;
2293                 key.offset = extent_op->level;
2294         } else {
2295                 key.type = BTRFS_EXTENT_ITEM_KEY;
2296                 key.offset = node->num_bytes;
2297         }
2298
2299 again:
2300         path->reada = READA_FORWARD;
2301         path->leave_spinning = 1;
2302         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 1);
2303         if (ret < 0) {
2304                 err = ret;
2305                 goto out;
2306         }
2307         if (ret > 0) {
2308                 if (metadata) {
2309                         if (path->slots[0] > 0) {
2310                                 path->slots[0]--;
2311                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
2312                                                       path->slots[0]);
2313                                 if (key.objectid == node->bytenr &&
2314                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
2315                                     key.offset == node->num_bytes)
2316                                         ret = 0;
2317                         }
2318                         if (ret > 0) {
2319                                 btrfs_release_path(path);
2320                                 metadata = 0;
2321
2322                                 key.objectid = node->bytenr;
2323                                 key.offset = node->num_bytes;
2324                                 key.type = BTRFS_EXTENT_ITEM_KEY;
2325                                 goto again;
2326                         }
2327                 } else {
2328                         err = -EIO;
2329                         goto out;
2330                 }
2331         }
2332
2333         leaf = path->nodes[0];
2334         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2335 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2336         if (item_size < sizeof(*ei)) {
2337                 ret = convert_extent_item_v0(trans, fs_info, path, (u64)-1, 0);
2338                 if (ret < 0) {
2339                         err = ret;
2340                         goto out;
2341                 }
2342                 leaf = path->nodes[0];
2343                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2344         }
2345 #endif
2346         BUG_ON(item_size < sizeof(*ei));
2347         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2348         __run_delayed_extent_op(extent_op, leaf, ei);
2349
2350         btrfs_mark_buffer_dirty(leaf);
2351 out:
2352         btrfs_free_path(path);
2353         return err;
2354 }
2355
2356 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2357                                 struct btrfs_fs_info *fs_info,
2358                                 struct btrfs_delayed_ref_node *node,
2359                                 struct btrfs_delayed_extent_op *extent_op,
2360                                 int insert_reserved)
2361 {
2362         int ret = 0;
2363         struct btrfs_delayed_tree_ref *ref;
2364         struct btrfs_key ins;
2365         u64 parent = 0;
2366         u64 ref_root = 0;
2367         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
2368
2369         ref = btrfs_delayed_node_to_tree_ref(node);
2370         trace_run_delayed_tree_ref(fs_info, node, ref, node->action);
2371
2372         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2373                 parent = ref->parent;
2374         ref_root = ref->root;
2375
2376         ins.objectid = node->bytenr;
2377         if (skinny_metadata) {
2378                 ins.offset = ref->level;
2379                 ins.type = BTRFS_METADATA_ITEM_KEY;
2380         } else {
2381                 ins.offset = node->num_bytes;
2382                 ins.type = BTRFS_EXTENT_ITEM_KEY;
2383         }
2384
2385         if (node->ref_mod != 1) {
2386                 btrfs_err(fs_info,
2387         "btree block(%llu) has %d references rather than 1: action %d ref_root %llu parent %llu",
2388                           node->bytenr, node->ref_mod, node->action, ref_root,
2389                           parent);
2390                 return -EIO;
2391         }
2392         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2393                 BUG_ON(!extent_op || !extent_op->update_flags);
2394                 ret = alloc_reserved_tree_block(trans, fs_info,
2395                                                 parent, ref_root,
2396                                                 extent_op->flags_to_set,
2397                                                 &extent_op->key,
2398                                                 ref->level, &ins);
2399         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2400                 ret = __btrfs_inc_extent_ref(trans, fs_info, node,
2401                                              parent, ref_root,
2402                                              ref->level, 0, 1,
2403                                              extent_op);
2404         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2405                 ret = __btrfs_free_extent(trans, fs_info, node,
2406                                           parent, ref_root,
2407                                           ref->level, 0, 1, extent_op);
2408         } else {
2409                 BUG();
2410         }
2411         return ret;
2412 }
2413
2414 /* helper function to actually process a single delayed ref entry */
2415 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2416                                struct btrfs_fs_info *fs_info,
2417                                struct btrfs_delayed_ref_node *node,
2418                                struct btrfs_delayed_extent_op *extent_op,
2419                                int insert_reserved)
2420 {
2421         int ret = 0;
2422
2423         if (trans->aborted) {
2424                 if (insert_reserved)
2425                         btrfs_pin_extent(fs_info, node->bytenr,
2426                                          node->num_bytes, 1);
2427                 return 0;
2428         }
2429
2430         if (btrfs_delayed_ref_is_head(node)) {
2431                 struct btrfs_delayed_ref_head *head;
2432                 /*
2433                  * we've hit the end of the chain and we were supposed
2434                  * to insert this extent into the tree.  But, it got
2435                  * deleted before we ever needed to insert it, so all
2436                  * we have to do is clean up the accounting
2437                  */
2438                 BUG_ON(extent_op);
2439                 head = btrfs_delayed_node_to_head(node);
2440                 trace_run_delayed_ref_head(fs_info, node, head, node->action);
2441
2442                 if (head->total_ref_mod < 0) {
2443                         struct btrfs_block_group_cache *cache;
2444
2445                         cache = btrfs_lookup_block_group(fs_info, node->bytenr);
2446                         ASSERT(cache);
2447                         percpu_counter_add(&cache->space_info->total_bytes_pinned,
2448                                            -node->num_bytes);
2449                         btrfs_put_block_group(cache);
2450                 }
2451
2452                 if (insert_reserved) {
2453                         btrfs_pin_extent(fs_info, node->bytenr,
2454                                          node->num_bytes, 1);
2455                         if (head->is_data) {
2456                                 ret = btrfs_del_csums(trans, fs_info,
2457                                                       node->bytenr,
2458                                                       node->num_bytes);
2459                         }
2460                 }
2461
2462                 /* Also free its reserved qgroup space */
2463                 btrfs_qgroup_free_delayed_ref(fs_info, head->qgroup_ref_root,
2464                                               head->qgroup_reserved);
2465                 return ret;
2466         }
2467
2468         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2469             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2470                 ret = run_delayed_tree_ref(trans, fs_info, node, extent_op,
2471                                            insert_reserved);
2472         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2473                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2474                 ret = run_delayed_data_ref(trans, fs_info, node, extent_op,
2475                                            insert_reserved);
2476         else
2477                 BUG();
2478         return ret;
2479 }
2480
2481 static inline struct btrfs_delayed_ref_node *
2482 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2483 {
2484         struct btrfs_delayed_ref_node *ref;
2485
2486         if (list_empty(&head->ref_list))
2487                 return NULL;
2488
2489         /*
2490          * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
2491          * This is to prevent a ref count from going down to zero, which deletes
2492          * the extent item from the extent tree, when there still are references
2493          * to add, which would fail because they would not find the extent item.
2494          */
2495         if (!list_empty(&head->ref_add_list))
2496                 return list_first_entry(&head->ref_add_list,
2497                                 struct btrfs_delayed_ref_node, add_list);
2498
2499         ref = list_first_entry(&head->ref_list, struct btrfs_delayed_ref_node,
2500                                list);
2501         ASSERT(list_empty(&ref->add_list));
2502         return ref;
2503 }
2504
2505 /*
2506  * Returns 0 on success or if called with an already aborted transaction.
2507  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2508  */
2509 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2510                                              struct btrfs_fs_info *fs_info,
2511                                              unsigned long nr)
2512 {
2513         struct btrfs_delayed_ref_root *delayed_refs;
2514         struct btrfs_delayed_ref_node *ref;
2515         struct btrfs_delayed_ref_head *locked_ref = NULL;
2516         struct btrfs_delayed_extent_op *extent_op;
2517         ktime_t start = ktime_get();
2518         int ret;
2519         unsigned long count = 0;
2520         unsigned long actual_count = 0;
2521         int must_insert_reserved = 0;
2522
2523         delayed_refs = &trans->transaction->delayed_refs;
2524         while (1) {
2525                 if (!locked_ref) {
2526                         if (count >= nr)
2527                                 break;
2528
2529                         spin_lock(&delayed_refs->lock);
2530                         locked_ref = btrfs_select_ref_head(trans);
2531                         if (!locked_ref) {
2532                                 spin_unlock(&delayed_refs->lock);
2533                                 break;
2534                         }
2535
2536                         /* grab the lock that says we are going to process
2537                          * all the refs for this head */
2538                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2539                         spin_unlock(&delayed_refs->lock);
2540                         /*
2541                          * we may have dropped the spin lock to get the head
2542                          * mutex lock, and that might have given someone else
2543                          * time to free the head.  If that's true, it has been
2544                          * removed from our list and we can move on.
2545                          */
2546                         if (ret == -EAGAIN) {
2547                                 locked_ref = NULL;
2548                                 count++;
2549                                 continue;
2550                         }
2551                 }
2552
2553                 /*
2554                  * We need to try and merge add/drops of the same ref since we
2555                  * can run into issues with relocate dropping the implicit ref
2556                  * and then it being added back again before the drop can
2557                  * finish.  If we merged anything we need to re-loop so we can
2558                  * get a good ref.
2559                  * Or we can get node references of the same type that weren't
2560                  * merged when created due to bumps in the tree mod seq, and
2561                  * we need to merge them to prevent adding an inline extent
2562                  * backref before dropping it (triggering a BUG_ON at
2563                  * insert_inline_extent_backref()).
2564                  */
2565                 spin_lock(&locked_ref->lock);
2566                 btrfs_merge_delayed_refs(trans, fs_info, delayed_refs,
2567                                          locked_ref);
2568
2569                 /*
2570                  * locked_ref is the head node, so we have to go one
2571                  * node back for any delayed ref updates
2572                  */
2573                 ref = select_delayed_ref(locked_ref);
2574
2575                 if (ref && ref->seq &&
2576                     btrfs_check_delayed_seq(fs_info, delayed_refs, ref->seq)) {
2577                         spin_unlock(&locked_ref->lock);
2578                         spin_lock(&delayed_refs->lock);
2579                         locked_ref->processing = 0;
2580                         delayed_refs->num_heads_ready++;
2581                         spin_unlock(&delayed_refs->lock);
2582                         btrfs_delayed_ref_unlock(locked_ref);
2583                         locked_ref = NULL;
2584                         cond_resched();
2585                         count++;
2586                         continue;
2587                 }
2588
2589                 /*
2590                  * record the must insert reserved flag before we
2591                  * drop the spin lock.
2592                  */
2593                 must_insert_reserved = locked_ref->must_insert_reserved;
2594                 locked_ref->must_insert_reserved = 0;
2595
2596                 extent_op = locked_ref->extent_op;
2597                 locked_ref->extent_op = NULL;
2598
2599                 if (!ref) {
2600
2601
2602                         /* All delayed refs have been processed, Go ahead
2603                          * and send the head node to run_one_delayed_ref,
2604                          * so that any accounting fixes can happen
2605                          */
2606                         ref = &locked_ref->node;
2607
2608                         if (extent_op && must_insert_reserved) {
2609                                 btrfs_free_delayed_extent_op(extent_op);
2610                                 extent_op = NULL;
2611                         }
2612
2613                         if (extent_op) {
2614                                 spin_unlock(&locked_ref->lock);
2615                                 ret = run_delayed_extent_op(trans, fs_info,
2616                                                             ref, extent_op);
2617                                 btrfs_free_delayed_extent_op(extent_op);
2618
2619                                 if (ret) {
2620                                         /*
2621                                          * Need to reset must_insert_reserved if
2622                                          * there was an error so the abort stuff
2623                                          * can cleanup the reserved space
2624                                          * properly.
2625                                          */
2626                                         if (must_insert_reserved)
2627                                                 locked_ref->must_insert_reserved = 1;
2628                                         spin_lock(&delayed_refs->lock);
2629                                         locked_ref->processing = 0;
2630                                         delayed_refs->num_heads_ready++;
2631                                         spin_unlock(&delayed_refs->lock);
2632                                         btrfs_debug(fs_info,
2633                                                     "run_delayed_extent_op returned %d",
2634                                                     ret);
2635                                         btrfs_delayed_ref_unlock(locked_ref);
2636                                         return ret;
2637                                 }
2638                                 continue;
2639                         }
2640
2641                         /*
2642                          * Need to drop our head ref lock and re-acquire the
2643                          * delayed ref lock and then re-check to make sure
2644                          * nobody got added.
2645                          */
2646                         spin_unlock(&locked_ref->lock);
2647                         spin_lock(&delayed_refs->lock);
2648                         spin_lock(&locked_ref->lock);
2649                         if (!list_empty(&locked_ref->ref_list) ||
2650                             locked_ref->extent_op) {
2651                                 spin_unlock(&locked_ref->lock);
2652                                 spin_unlock(&delayed_refs->lock);
2653                                 continue;
2654                         }
2655                         ref->in_tree = 0;
2656                         delayed_refs->num_heads--;
2657                         rb_erase(&locked_ref->href_node,
2658                                  &delayed_refs->href_root);
2659                         spin_unlock(&delayed_refs->lock);
2660                 } else {
2661                         actual_count++;
2662                         ref->in_tree = 0;
2663                         list_del(&ref->list);
2664                         if (!list_empty(&ref->add_list))
2665                                 list_del(&ref->add_list);
2666                 }
2667                 atomic_dec(&delayed_refs->num_entries);
2668
2669                 if (!btrfs_delayed_ref_is_head(ref)) {
2670                         /*
2671                          * when we play the delayed ref, also correct the
2672                          * ref_mod on head
2673                          */
2674                         switch (ref->action) {
2675                         case BTRFS_ADD_DELAYED_REF:
2676                         case BTRFS_ADD_DELAYED_EXTENT:
2677                                 locked_ref->node.ref_mod -= ref->ref_mod;
2678                                 break;
2679                         case BTRFS_DROP_DELAYED_REF:
2680                                 locked_ref->node.ref_mod += ref->ref_mod;
2681                                 break;
2682                         default:
2683                                 WARN_ON(1);
2684                         }
2685                 }
2686                 spin_unlock(&locked_ref->lock);
2687
2688                 ret = run_one_delayed_ref(trans, fs_info, ref, extent_op,
2689                                           must_insert_reserved);
2690
2691                 btrfs_free_delayed_extent_op(extent_op);
2692                 if (ret) {
2693                         spin_lock(&delayed_refs->lock);
2694                         locked_ref->processing = 0;
2695                         delayed_refs->num_heads_ready++;
2696                         spin_unlock(&delayed_refs->lock);
2697                         btrfs_delayed_ref_unlock(locked_ref);
2698                         btrfs_put_delayed_ref(ref);
2699                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d",
2700                                     ret);
2701                         return ret;
2702                 }
2703
2704                 /*
2705                  * If this node is a head, that means all the refs in this head
2706                  * have been dealt with, and we will pick the next head to deal
2707                  * with, so we must unlock the head and drop it from the cluster
2708                  * list before we release it.
2709                  */
2710                 if (btrfs_delayed_ref_is_head(ref)) {
2711                         if (locked_ref->is_data &&
2712                             locked_ref->total_ref_mod < 0) {
2713                                 spin_lock(&delayed_refs->lock);
2714                                 delayed_refs->pending_csums -= ref->num_bytes;
2715                                 spin_unlock(&delayed_refs->lock);
2716                         }
2717                         btrfs_delayed_ref_unlock(locked_ref);
2718                         locked_ref = NULL;
2719                 }
2720                 btrfs_put_delayed_ref(ref);
2721                 count++;
2722                 cond_resched();
2723         }
2724
2725         /*
2726          * We don't want to include ref heads since we can have empty ref heads
2727          * and those will drastically skew our runtime down since we just do
2728          * accounting, no actual extent tree updates.
2729          */
2730         if (actual_count > 0) {
2731                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2732                 u64 avg;
2733
2734                 /*
2735                  * We weigh the current average higher than our current runtime
2736                  * to avoid large swings in the average.
2737                  */
2738                 spin_lock(&delayed_refs->lock);
2739                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2740                 fs_info->avg_delayed_ref_runtime = avg >> 2;    /* div by 4 */
2741                 spin_unlock(&delayed_refs->lock);
2742         }
2743         return 0;
2744 }
2745
2746 #ifdef SCRAMBLE_DELAYED_REFS
2747 /*
2748  * Normally delayed refs get processed in ascending bytenr order. This
2749  * correlates in most cases to the order added. To expose dependencies on this
2750  * order, we start to process the tree in the middle instead of the beginning
2751  */
2752 static u64 find_middle(struct rb_root *root)
2753 {
2754         struct rb_node *n = root->rb_node;
2755         struct btrfs_delayed_ref_node *entry;
2756         int alt = 1;
2757         u64 middle;
2758         u64 first = 0, last = 0;
2759
2760         n = rb_first(root);
2761         if (n) {
2762                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2763                 first = entry->bytenr;
2764         }
2765         n = rb_last(root);
2766         if (n) {
2767                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2768                 last = entry->bytenr;
2769         }
2770         n = root->rb_node;
2771
2772         while (n) {
2773                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2774                 WARN_ON(!entry->in_tree);
2775
2776                 middle = entry->bytenr;
2777
2778                 if (alt)
2779                         n = n->rb_left;
2780                 else
2781                         n = n->rb_right;
2782
2783                 alt = 1 - alt;
2784         }
2785         return middle;
2786 }
2787 #endif
2788
2789 static inline u64 heads_to_leaves(struct btrfs_fs_info *fs_info, u64 heads)
2790 {
2791         u64 num_bytes;
2792
2793         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2794                              sizeof(struct btrfs_extent_inline_ref));
2795         if (!btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2796                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2797
2798         /*
2799          * We don't ever fill up leaves all the way so multiply by 2 just to be
2800          * closer to what we're really going to want to use.
2801          */
2802         return div_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(fs_info));
2803 }
2804
2805 /*
2806  * Takes the number of bytes to be csumm'ed and figures out how many leaves it
2807  * would require to store the csums for that many bytes.
2808  */
2809 u64 btrfs_csum_bytes_to_leaves(struct btrfs_fs_info *fs_info, u64 csum_bytes)
2810 {
2811         u64 csum_size;
2812         u64 num_csums_per_leaf;
2813         u64 num_csums;
2814
2815         csum_size = BTRFS_MAX_ITEM_SIZE(fs_info);
2816         num_csums_per_leaf = div64_u64(csum_size,
2817                         (u64)btrfs_super_csum_size(fs_info->super_copy));
2818         num_csums = div64_u64(csum_bytes, fs_info->sectorsize);
2819         num_csums += num_csums_per_leaf - 1;
2820         num_csums = div64_u64(num_csums, num_csums_per_leaf);
2821         return num_csums;
2822 }
2823
2824 int btrfs_check_space_for_delayed_refs(struct btrfs_trans_handle *trans,
2825                                        struct btrfs_fs_info *fs_info)
2826 {
2827         struct btrfs_block_rsv *global_rsv;
2828         u64 num_heads = trans->transaction->delayed_refs.num_heads_ready;
2829         u64 csum_bytes = trans->transaction->delayed_refs.pending_csums;
2830         u64 num_dirty_bgs = trans->transaction->num_dirty_bgs;
2831         u64 num_bytes, num_dirty_bgs_bytes;
2832         int ret = 0;
2833
2834         num_bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
2835         num_heads = heads_to_leaves(fs_info, num_heads);
2836         if (num_heads > 1)
2837                 num_bytes += (num_heads - 1) * fs_info->nodesize;
2838         num_bytes <<= 1;
2839         num_bytes += btrfs_csum_bytes_to_leaves(fs_info, csum_bytes) *
2840                                                         fs_info->nodesize;
2841         num_dirty_bgs_bytes = btrfs_calc_trans_metadata_size(fs_info,
2842                                                              num_dirty_bgs);
2843         global_rsv = &fs_info->global_block_rsv;
2844
2845         /*
2846          * If we can't allocate any more chunks lets make sure we have _lots_ of
2847          * wiggle room since running delayed refs can create more delayed refs.
2848          */
2849         if (global_rsv->space_info->full) {
2850                 num_dirty_bgs_bytes <<= 1;
2851                 num_bytes <<= 1;
2852         }
2853
2854         spin_lock(&global_rsv->lock);
2855         if (global_rsv->reserved <= num_bytes + num_dirty_bgs_bytes)
2856                 ret = 1;
2857         spin_unlock(&global_rsv->lock);
2858         return ret;
2859 }
2860
2861 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans,
2862                                        struct btrfs_fs_info *fs_info)
2863 {
2864         u64 num_entries =
2865                 atomic_read(&trans->transaction->delayed_refs.num_entries);
2866         u64 avg_runtime;
2867         u64 val;
2868
2869         smp_mb();
2870         avg_runtime = fs_info->avg_delayed_ref_runtime;
2871         val = num_entries * avg_runtime;
2872         if (val >= NSEC_PER_SEC)
2873                 return 1;
2874         if (val >= NSEC_PER_SEC / 2)
2875                 return 2;
2876
2877         return btrfs_check_space_for_delayed_refs(trans, fs_info);
2878 }
2879
2880 struct async_delayed_refs {
2881         struct btrfs_root *root;
2882         u64 transid;
2883         int count;
2884         int error;
2885         int sync;
2886         struct completion wait;
2887         struct btrfs_work work;
2888 };
2889
2890 static inline struct async_delayed_refs *
2891 to_async_delayed_refs(struct btrfs_work *work)
2892 {
2893         return container_of(work, struct async_delayed_refs, work);
2894 }
2895
2896 static void delayed_ref_async_start(struct btrfs_work *work)
2897 {
2898         struct async_delayed_refs *async = to_async_delayed_refs(work);
2899         struct btrfs_trans_handle *trans;
2900         struct btrfs_fs_info *fs_info = async->root->fs_info;
2901         int ret;
2902
2903         /* if the commit is already started, we don't need to wait here */
2904         if (btrfs_transaction_blocked(fs_info))
2905                 goto done;
2906
2907         trans = btrfs_join_transaction(async->root);
2908         if (IS_ERR(trans)) {
2909                 async->error = PTR_ERR(trans);
2910                 goto done;
2911         }
2912
2913         /*
2914          * trans->sync means that when we call end_transaction, we won't
2915          * wait on delayed refs
2916          */
2917         trans->sync = true;
2918
2919         /* Don't bother flushing if we got into a different transaction */
2920         if (trans->transid > async->transid)
2921                 goto end;
2922
2923         ret = btrfs_run_delayed_refs(trans, fs_info, async->count);
2924         if (ret)
2925                 async->error = ret;
2926 end:
2927         ret = btrfs_end_transaction(trans);
2928         if (ret && !async->error)
2929                 async->error = ret;
2930 done:
2931         if (async->sync)
2932                 complete(&async->wait);
2933         else
2934                 kfree(async);
2935 }
2936
2937 int btrfs_async_run_delayed_refs(struct btrfs_fs_info *fs_info,
2938                                  unsigned long count, u64 transid, int wait)
2939 {
2940         struct async_delayed_refs *async;
2941         int ret;
2942
2943         async = kmalloc(sizeof(*async), GFP_NOFS);
2944         if (!async)
2945                 return -ENOMEM;
2946
2947         async->root = fs_info->tree_root;
2948         async->count = count;
2949         async->error = 0;
2950         async->transid = transid;
2951         if (wait)
2952                 async->sync = 1;
2953         else
2954                 async->sync = 0;
2955         init_completion(&async->wait);
2956
2957         btrfs_init_work(&async->work, btrfs_extent_refs_helper,
2958                         delayed_ref_async_start, NULL, NULL);
2959
2960         btrfs_queue_work(fs_info->extent_workers, &async->work);
2961
2962         if (wait) {
2963                 wait_for_completion(&async->wait);
2964                 ret = async->error;
2965                 kfree(async);
2966                 return ret;
2967         }
2968         return 0;
2969 }
2970
2971 /*
2972  * this starts processing the delayed reference count updates and
2973  * extent insertions we have queued up so far.  count can be
2974  * 0, which means to process everything in the tree at the start
2975  * of the run (but not newly added entries), or it can be some target
2976  * number you'd like to process.
2977  *
2978  * Returns 0 on success or if called with an aborted transaction
2979  * Returns <0 on error and aborts the transaction
2980  */
2981 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2982                            struct btrfs_fs_info *fs_info, unsigned long count)
2983 {
2984         struct rb_node *node;
2985         struct btrfs_delayed_ref_root *delayed_refs;
2986         struct btrfs_delayed_ref_head *head;
2987         int ret;
2988         int run_all = count == (unsigned long)-1;
2989         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
2990
2991         /* We'll clean this up in btrfs_cleanup_transaction */
2992         if (trans->aborted)
2993                 return 0;
2994
2995         if (test_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags))
2996                 return 0;
2997
2998         delayed_refs = &trans->transaction->delayed_refs;
2999         if (count == 0)
3000                 count = atomic_read(&delayed_refs->num_entries) * 2;
3001
3002 again:
3003 #ifdef SCRAMBLE_DELAYED_REFS
3004         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
3005 #endif
3006         trans->can_flush_pending_bgs = false;
3007         ret = __btrfs_run_delayed_refs(trans, fs_info, count);
3008         if (ret < 0) {
3009                 btrfs_abort_transaction(trans, ret);
3010                 return ret;
3011         }
3012
3013         if (run_all) {
3014                 if (!list_empty(&trans->new_bgs))
3015                         btrfs_create_pending_block_groups(trans, fs_info);
3016
3017                 spin_lock(&delayed_refs->lock);
3018                 node = rb_first(&delayed_refs->href_root);
3019                 if (!node) {
3020                         spin_unlock(&delayed_refs->lock);
3021                         goto out;
3022                 }
3023
3024                 while (node) {
3025                         head = rb_entry(node, struct btrfs_delayed_ref_head,
3026                                         href_node);
3027                         if (btrfs_delayed_ref_is_head(&head->node)) {
3028                                 struct btrfs_delayed_ref_node *ref;
3029
3030                                 ref = &head->node;
3031                                 refcount_inc(&ref->refs);
3032
3033                                 spin_unlock(&delayed_refs->lock);
3034                                 /*
3035                                  * Mutex was contended, block until it's
3036                                  * released and try again
3037                                  */
3038                                 mutex_lock(&head->mutex);
3039                                 mutex_unlock(&head->mutex);
3040
3041                                 btrfs_put_delayed_ref(ref);
3042                                 cond_resched();
3043                                 goto again;
3044                         } else {
3045                                 WARN_ON(1);
3046                         }
3047                         node = rb_next(node);
3048                 }
3049                 spin_unlock(&delayed_refs->lock);
3050                 cond_resched();
3051                 goto again;
3052         }
3053 out:
3054         trans->can_flush_pending_bgs = can_flush_pending_bgs;
3055         return 0;
3056 }
3057
3058 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
3059                                 struct btrfs_fs_info *fs_info,
3060                                 u64 bytenr, u64 num_bytes, u64 flags,
3061                                 int level, int is_data)
3062 {
3063         struct btrfs_delayed_extent_op *extent_op;
3064         int ret;
3065
3066         extent_op = btrfs_alloc_delayed_extent_op();
3067         if (!extent_op)
3068                 return -ENOMEM;
3069
3070         extent_op->flags_to_set = flags;
3071         extent_op->update_flags = true;
3072         extent_op->update_key = false;
3073         extent_op->is_data = is_data ? true : false;
3074         extent_op->level = level;
3075
3076         ret = btrfs_add_delayed_extent_op(fs_info, trans, bytenr,
3077                                           num_bytes, extent_op);
3078         if (ret)
3079                 btrfs_free_delayed_extent_op(extent_op);
3080         return ret;
3081 }
3082
3083 static noinline int check_delayed_ref(struct btrfs_root *root,
3084                                       struct btrfs_path *path,
3085                                       u64 objectid, u64 offset, u64 bytenr)
3086 {
3087         struct btrfs_delayed_ref_head *head;
3088         struct btrfs_delayed_ref_node *ref;
3089         struct btrfs_delayed_data_ref *data_ref;
3090         struct btrfs_delayed_ref_root *delayed_refs;
3091         struct btrfs_transaction *cur_trans;
3092         int ret = 0;
3093
3094         cur_trans = root->fs_info->running_transaction;
3095         if (!cur_trans)
3096                 return 0;
3097
3098         delayed_refs = &cur_trans->delayed_refs;
3099         spin_lock(&delayed_refs->lock);
3100         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
3101         if (!head) {
3102                 spin_unlock(&delayed_refs->lock);
3103                 return 0;
3104         }
3105
3106         if (!mutex_trylock(&head->mutex)) {
3107                 refcount_inc(&head->node.refs);
3108                 spin_unlock(&delayed_refs->lock);
3109
3110                 btrfs_release_path(path);
3111
3112                 /*
3113                  * Mutex was contended, block until it's released and let
3114                  * caller try again
3115                  */
3116                 mutex_lock(&head->mutex);
3117                 mutex_unlock(&head->mutex);
3118                 btrfs_put_delayed_ref(&head->node);
3119                 return -EAGAIN;
3120         }
3121         spin_unlock(&delayed_refs->lock);
3122
3123         spin_lock(&head->lock);
3124         list_for_each_entry(ref, &head->ref_list, list) {
3125                 /* If it's a shared ref we know a cross reference exists */
3126                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
3127                         ret = 1;
3128                         break;
3129                 }
3130
3131                 data_ref = btrfs_delayed_node_to_data_ref(ref);
3132
3133                 /*
3134                  * If our ref doesn't match the one we're currently looking at
3135                  * then we have a cross reference.
3136                  */
3137                 if (data_ref->root != root->root_key.objectid ||
3138                     data_ref->objectid != objectid ||
3139                     data_ref->offset != offset) {
3140                         ret = 1;
3141                         break;
3142                 }
3143         }
3144         spin_unlock(&head->lock);
3145         mutex_unlock(&head->mutex);
3146         return ret;
3147 }
3148
3149 static noinline int check_committed_ref(struct btrfs_root *root,
3150                                         struct btrfs_path *path,
3151                                         u64 objectid, u64 offset, u64 bytenr)
3152 {
3153         struct btrfs_fs_info *fs_info = root->fs_info;
3154         struct btrfs_root *extent_root = fs_info->extent_root;
3155         struct extent_buffer *leaf;
3156         struct btrfs_extent_data_ref *ref;
3157         struct btrfs_extent_inline_ref *iref;
3158         struct btrfs_extent_item *ei;
3159         struct btrfs_key key;
3160         u32 item_size;
3161         int ret;
3162
3163         key.objectid = bytenr;
3164         key.offset = (u64)-1;
3165         key.type = BTRFS_EXTENT_ITEM_KEY;
3166
3167         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
3168         if (ret < 0)
3169                 goto out;
3170         BUG_ON(ret == 0); /* Corruption */
3171
3172         ret = -ENOENT;
3173         if (path->slots[0] == 0)
3174                 goto out;
3175
3176         path->slots[0]--;
3177         leaf = path->nodes[0];
3178         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3179
3180         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
3181                 goto out;
3182
3183         ret = 1;
3184         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
3185 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3186         if (item_size < sizeof(*ei)) {
3187                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3188                 goto out;
3189         }
3190 #endif
3191         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
3192
3193         if (item_size != sizeof(*ei) +
3194             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
3195                 goto out;
3196
3197         if (btrfs_extent_generation(leaf, ei) <=
3198             btrfs_root_last_snapshot(&root->root_item))
3199                 goto out;
3200
3201         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
3202         if (btrfs_extent_inline_ref_type(leaf, iref) !=
3203             BTRFS_EXTENT_DATA_REF_KEY)
3204                 goto out;
3205
3206         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
3207         if (btrfs_extent_refs(leaf, ei) !=
3208             btrfs_extent_data_ref_count(leaf, ref) ||
3209             btrfs_extent_data_ref_root(leaf, ref) !=
3210             root->root_key.objectid ||
3211             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
3212             btrfs_extent_data_ref_offset(leaf, ref) != offset)
3213                 goto out;
3214
3215         ret = 0;
3216 out:
3217         return ret;
3218 }
3219
3220 int btrfs_cross_ref_exist(struct btrfs_root *root, u64 objectid, u64 offset,
3221                           u64 bytenr)
3222 {
3223         struct btrfs_path *path;
3224         int ret;
3225         int ret2;
3226
3227         path = btrfs_alloc_path();
3228         if (!path)
3229                 return -ENOENT;
3230
3231         do {
3232                 ret = check_committed_ref(root, path, objectid,
3233                                           offset, bytenr);
3234                 if (ret && ret != -ENOENT)
3235                         goto out;
3236
3237                 ret2 = check_delayed_ref(root, path, objectid,
3238                                          offset, bytenr);
3239         } while (ret2 == -EAGAIN);
3240
3241         if (ret2 && ret2 != -ENOENT) {
3242                 ret = ret2;
3243                 goto out;
3244         }
3245
3246         if (ret != -ENOENT || ret2 != -ENOENT)
3247                 ret = 0;
3248 out:
3249         btrfs_free_path(path);
3250         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
3251                 WARN_ON(ret > 0);
3252         return ret;
3253 }
3254
3255 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
3256                            struct btrfs_root *root,
3257                            struct extent_buffer *buf,
3258                            int full_backref, int inc)
3259 {
3260         struct btrfs_fs_info *fs_info = root->fs_info;
3261         u64 bytenr;
3262         u64 num_bytes;
3263         u64 parent;
3264         u64 ref_root;
3265         u32 nritems;
3266         struct btrfs_key key;
3267         struct btrfs_file_extent_item *fi;
3268         int i;
3269         int level;
3270         int ret = 0;
3271         int (*process_func)(struct btrfs_trans_handle *,
3272                             struct btrfs_fs_info *,
3273                             u64, u64, u64, u64, u64, u64);
3274
3275
3276         if (btrfs_is_testing(fs_info))
3277                 return 0;
3278
3279         ref_root = btrfs_header_owner(buf);
3280         nritems = btrfs_header_nritems(buf);
3281         level = btrfs_header_level(buf);
3282
3283         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state) && level == 0)
3284                 return 0;
3285
3286         if (inc)
3287                 process_func = btrfs_inc_extent_ref;
3288         else
3289                 process_func = btrfs_free_extent;
3290
3291         if (full_backref)
3292                 parent = buf->start;
3293         else
3294                 parent = 0;
3295
3296         for (i = 0; i < nritems; i++) {
3297                 if (level == 0) {
3298                         btrfs_item_key_to_cpu(buf, &key, i);
3299                         if (key.type != BTRFS_EXTENT_DATA_KEY)
3300                                 continue;
3301                         fi = btrfs_item_ptr(buf, i,
3302                                             struct btrfs_file_extent_item);
3303                         if (btrfs_file_extent_type(buf, fi) ==
3304                             BTRFS_FILE_EXTENT_INLINE)
3305                                 continue;
3306                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3307                         if (bytenr == 0)
3308                                 continue;
3309
3310                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3311                         key.offset -= btrfs_file_extent_offset(buf, fi);
3312                         ret = process_func(trans, fs_info, bytenr, num_bytes,
3313                                            parent, ref_root, key.objectid,
3314                                            key.offset);
3315                         if (ret)
3316                                 goto fail;
3317                 } else {
3318                         bytenr = btrfs_node_blockptr(buf, i);
3319                         num_bytes = fs_info->nodesize;
3320                         ret = process_func(trans, fs_info, bytenr, num_bytes,
3321                                            parent, ref_root, level - 1, 0);
3322                         if (ret)
3323                                 goto fail;
3324                 }
3325         }
3326         return 0;
3327 fail:
3328         return ret;
3329 }
3330
3331 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3332                   struct extent_buffer *buf, int full_backref)
3333 {
3334         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
3335 }
3336
3337 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3338                   struct extent_buffer *buf, int full_backref)
3339 {
3340         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
3341 }
3342
3343 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3344                                  struct btrfs_fs_info *fs_info,
3345                                  struct btrfs_path *path,
3346                                  struct btrfs_block_group_cache *cache)
3347 {
3348         int ret;
3349         struct btrfs_root *extent_root = fs_info->extent_root;
3350         unsigned long bi;
3351         struct extent_buffer *leaf;
3352
3353         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3354         if (ret) {
3355                 if (ret > 0)
3356                         ret = -ENOENT;
3357                 goto fail;
3358         }
3359
3360         leaf = path->nodes[0];
3361         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3362         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3363         btrfs_mark_buffer_dirty(leaf);
3364 fail:
3365         btrfs_release_path(path);
3366         return ret;
3367
3368 }
3369
3370 static struct btrfs_block_group_cache *
3371 next_block_group(struct btrfs_fs_info *fs_info,
3372                  struct btrfs_block_group_cache *cache)
3373 {
3374         struct rb_node *node;
3375
3376         spin_lock(&fs_info->block_group_cache_lock);
3377
3378         /* If our block group was removed, we need a full search. */
3379         if (RB_EMPTY_NODE(&cache->cache_node)) {
3380                 const u64 next_bytenr = cache->key.objectid + cache->key.offset;
3381
3382                 spin_unlock(&fs_info->block_group_cache_lock);
3383                 btrfs_put_block_group(cache);
3384                 cache = btrfs_lookup_first_block_group(fs_info, next_bytenr); return cache;
3385         }
3386         node = rb_next(&cache->cache_node);
3387         btrfs_put_block_group(cache);
3388         if (node) {
3389                 cache = rb_entry(node, struct btrfs_block_group_cache,
3390                                  cache_node);
3391                 btrfs_get_block_group(cache);
3392         } else
3393                 cache = NULL;
3394         spin_unlock(&fs_info->block_group_cache_lock);
3395         return cache;
3396 }
3397
3398 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3399                             struct btrfs_trans_handle *trans,
3400                             struct btrfs_path *path)
3401 {
3402         struct btrfs_fs_info *fs_info = block_group->fs_info;
3403         struct btrfs_root *root = fs_info->tree_root;
3404         struct inode *inode = NULL;
3405         struct extent_changeset *data_reserved = NULL;
3406         u64 alloc_hint = 0;
3407         int dcs = BTRFS_DC_ERROR;
3408         u64 num_pages = 0;
3409         int retries = 0;
3410         int ret = 0;
3411
3412         /*
3413          * If this block group is smaller than 100 megs don't bother caching the
3414          * block group.
3415          */
3416         if (block_group->key.offset < (100 * SZ_1M)) {
3417                 spin_lock(&block_group->lock);
3418                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3419                 spin_unlock(&block_group->lock);
3420                 return 0;
3421         }
3422
3423         if (trans->aborted)
3424                 return 0;
3425 again:
3426         inode = lookup_free_space_inode(fs_info, block_group, path);
3427         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3428                 ret = PTR_ERR(inode);
3429                 btrfs_release_path(path);
3430                 goto out;
3431         }
3432
3433         if (IS_ERR(inode)) {
3434                 BUG_ON(retries);
3435                 retries++;
3436
3437                 if (block_group->ro)
3438                         goto out_free;
3439
3440                 ret = create_free_space_inode(fs_info, trans, block_group,
3441                                               path);
3442                 if (ret)
3443                         goto out_free;
3444                 goto again;
3445         }
3446
3447         /* We've already setup this transaction, go ahead and exit */
3448         if (block_group->cache_generation == trans->transid &&
3449             i_size_read(inode)) {
3450                 dcs = BTRFS_DC_SETUP;
3451                 goto out_put;
3452         }
3453
3454         /*
3455          * We want to set the generation to 0, that way if anything goes wrong
3456          * from here on out we know not to trust this cache when we load up next
3457          * time.
3458          */
3459         BTRFS_I(inode)->generation = 0;
3460         ret = btrfs_update_inode(trans, root, inode);
3461         if (ret) {
3462                 /*
3463                  * So theoretically we could recover from this, simply set the
3464                  * super cache generation to 0 so we know to invalidate the
3465                  * cache, but then we'd have to keep track of the block groups
3466                  * that fail this way so we know we _have_ to reset this cache
3467                  * before the next commit or risk reading stale cache.  So to
3468                  * limit our exposure to horrible edge cases lets just abort the
3469                  * transaction, this only happens in really bad situations
3470                  * anyway.
3471                  */
3472                 btrfs_abort_transaction(trans, ret);
3473                 goto out_put;
3474         }
3475         WARN_ON(ret);
3476
3477         if (i_size_read(inode) > 0) {
3478                 ret = btrfs_check_trunc_cache_free_space(fs_info,
3479                                         &fs_info->global_block_rsv);
3480                 if (ret)
3481                         goto out_put;
3482
3483                 ret = btrfs_truncate_free_space_cache(trans, NULL, inode);
3484                 if (ret)
3485                         goto out_put;
3486         }
3487
3488         spin_lock(&block_group->lock);
3489         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3490             !btrfs_test_opt(fs_info, SPACE_CACHE)) {
3491                 /*
3492                  * don't bother trying to write stuff out _if_
3493                  * a) we're not cached,
3494                  * b) we're with nospace_cache mount option,
3495                  * c) we're with v2 space_cache (FREE_SPACE_TREE).
3496                  */
3497                 dcs = BTRFS_DC_WRITTEN;
3498                 spin_unlock(&block_group->lock);
3499                 goto out_put;
3500         }
3501         spin_unlock(&block_group->lock);
3502
3503         /*
3504          * We hit an ENOSPC when setting up the cache in this transaction, just
3505          * skip doing the setup, we've already cleared the cache so we're safe.
3506          */
3507         if (test_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags)) {
3508                 ret = -ENOSPC;
3509                 goto out_put;
3510         }
3511
3512         /*
3513          * Try to preallocate enough space based on how big the block group is.
3514          * Keep in mind this has to include any pinned space which could end up
3515          * taking up quite a bit since it's not folded into the other space
3516          * cache.
3517          */
3518         num_pages = div_u64(block_group->key.offset, SZ_256M);
3519         if (!num_pages)
3520                 num_pages = 1;
3521
3522         num_pages *= 16;
3523         num_pages *= PAGE_SIZE;
3524
3525         ret = btrfs_check_data_free_space(inode, &data_reserved, 0, num_pages);
3526         if (ret)
3527                 goto out_put;
3528
3529         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3530                                               num_pages, num_pages,
3531                                               &alloc_hint);
3532         /*
3533          * Our cache requires contiguous chunks so that we don't modify a bunch
3534          * of metadata or split extents when writing the cache out, which means
3535          * we can enospc if we are heavily fragmented in addition to just normal
3536          * out of space conditions.  So if we hit this just skip setting up any
3537          * other block groups for this transaction, maybe we'll unpin enough
3538          * space the next time around.
3539          */
3540         if (!ret)
3541                 dcs = BTRFS_DC_SETUP;
3542         else if (ret == -ENOSPC)
3543                 set_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags);
3544
3545 out_put:
3546         iput(inode);
3547 out_free:
3548         btrfs_release_path(path);
3549 out:
3550         spin_lock(&block_group->lock);
3551         if (!ret && dcs == BTRFS_DC_SETUP)
3552                 block_group->cache_generation = trans->transid;
3553         block_group->disk_cache_state = dcs;
3554         spin_unlock(&block_group->lock);
3555
3556         extent_changeset_free(data_reserved);
3557         return ret;
3558 }
3559
3560 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans,
3561                             struct btrfs_fs_info *fs_info)
3562 {
3563         struct btrfs_block_group_cache *cache, *tmp;
3564         struct btrfs_transaction *cur_trans = trans->transaction;
3565         struct btrfs_path *path;
3566
3567         if (list_empty(&cur_trans->dirty_bgs) ||
3568             !btrfs_test_opt(fs_info, SPACE_CACHE))
3569                 return 0;
3570
3571         path = btrfs_alloc_path();
3572         if (!path)
3573                 return -ENOMEM;
3574
3575         /* Could add new block groups, use _safe just in case */
3576         list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3577                                  dirty_list) {
3578                 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3579                         cache_save_setup(cache, trans, path);
3580         }
3581
3582         btrfs_free_path(path);
3583         return 0;
3584 }
3585
3586 /*
3587  * transaction commit does final block group cache writeback during a
3588  * critical section where nothing is allowed to change the FS.  This is
3589  * required in order for the cache to actually match the block group,
3590  * but can introduce a lot of latency into the commit.
3591  *
3592  * So, btrfs_start_dirty_block_groups is here to kick off block group
3593  * cache IO.  There's a chance we'll have to redo some of it if the
3594  * block group changes again during the commit, but it greatly reduces
3595  * the commit latency by getting rid of the easy block groups while
3596  * we're still allowing others to join the commit.
3597  */
3598 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans,
3599                                    struct btrfs_fs_info *fs_info)
3600 {
3601         struct btrfs_block_group_cache *cache;
3602         struct btrfs_transaction *cur_trans = trans->transaction;
3603         int ret = 0;
3604         int should_put;
3605         struct btrfs_path *path = NULL;
3606         LIST_HEAD(dirty);
3607         struct list_head *io = &cur_trans->io_bgs;
3608         int num_started = 0;
3609         int loops = 0;
3610
3611         spin_lock(&cur_trans->dirty_bgs_lock);
3612         if (list_empty(&cur_trans->dirty_bgs)) {
3613                 spin_unlock(&cur_trans->dirty_bgs_lock);
3614                 return 0;
3615         }
3616         list_splice_init(&cur_trans->dirty_bgs, &dirty);
3617         spin_unlock(&cur_trans->dirty_bgs_lock);
3618
3619 again:
3620         /*
3621          * make sure all the block groups on our dirty list actually
3622          * exist
3623          */
3624         btrfs_create_pending_block_groups(trans, fs_info);
3625
3626         if (!path) {
3627                 path = btrfs_alloc_path();
3628                 if (!path)
3629                         return -ENOMEM;
3630         }
3631
3632         /*
3633          * cache_write_mutex is here only to save us from balance or automatic
3634          * removal of empty block groups deleting this block group while we are
3635          * writing out the cache
3636          */
3637         mutex_lock(&trans->transaction->cache_write_mutex);
3638         while (!list_empty(&dirty)) {
3639                 cache = list_first_entry(&dirty,
3640                                          struct btrfs_block_group_cache,
3641                                          dirty_list);
3642                 /*
3643                  * this can happen if something re-dirties a block
3644                  * group that is already under IO.  Just wait for it to
3645                  * finish and then do it all again
3646                  */
3647                 if (!list_empty(&cache->io_list)) {
3648                         list_del_init(&cache->io_list);
3649                         btrfs_wait_cache_io(trans, cache, path);
3650                         btrfs_put_block_group(cache);
3651                 }
3652
3653
3654                 /*
3655                  * btrfs_wait_cache_io uses the cache->dirty_list to decide
3656                  * if it should update the cache_state.  Don't delete
3657                  * until after we wait.
3658                  *
3659                  * Since we're not running in the commit critical section
3660                  * we need the dirty_bgs_lock to protect from update_block_group
3661                  */
3662                 spin_lock(&cur_trans->dirty_bgs_lock);
3663                 list_del_init(&cache->dirty_list);
3664                 spin_unlock(&cur_trans->dirty_bgs_lock);
3665
3666                 should_put = 1;
3667
3668                 cache_save_setup(cache, trans, path);
3669
3670                 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
3671                         cache->io_ctl.inode = NULL;
3672                         ret = btrfs_write_out_cache(fs_info, trans,
3673                                                     cache, path);
3674                         if (ret == 0 && cache->io_ctl.inode) {
3675                                 num_started++;
3676                                 should_put = 0;
3677
3678                                 /*
3679                                  * the cache_write_mutex is protecting
3680                                  * the io_list
3681                                  */
3682                                 list_add_tail(&cache->io_list, io);
3683                         } else {
3684                                 /*
3685                                  * if we failed to write the cache, the
3686                                  * generation will be bad and life goes on
3687                                  */
3688                                 ret = 0;
3689                         }
3690                 }
3691                 if (!ret) {
3692                         ret = write_one_cache_group(trans, fs_info,
3693                                                     path, cache);
3694                         /*
3695                          * Our block group might still be attached to the list
3696                          * of new block groups in the transaction handle of some
3697                          * other task (struct btrfs_trans_handle->new_bgs). This
3698                          * means its block group item isn't yet in the extent
3699                          * tree. If this happens ignore the error, as we will
3700                          * try again later in the critical section of the
3701                          * transaction commit.
3702                          */
3703                         if (ret == -ENOENT) {
3704                                 ret = 0;
3705                                 spin_lock(&cur_trans->dirty_bgs_lock);
3706                                 if (list_empty(&cache->dirty_list)) {
3707                                         list_add_tail(&cache->dirty_list,
3708                                                       &cur_trans->dirty_bgs);
3709                                         btrfs_get_block_group(cache);
3710                                 }
3711                                 spin_unlock(&cur_trans->dirty_bgs_lock);
3712                         } else if (ret) {
3713                                 btrfs_abort_transaction(trans, ret);
3714                         }
3715                 }
3716
3717                 /* if its not on the io list, we need to put the block group */
3718                 if (should_put)
3719                         btrfs_put_block_group(cache);
3720
3721                 if (ret)
3722                         break;
3723
3724                 /*
3725                  * Avoid blocking other tasks for too long. It might even save
3726                  * us from writing caches for block groups that are going to be
3727                  * removed.
3728                  */
3729                 mutex_unlock(&trans->transaction->cache_write_mutex);
3730                 mutex_lock(&trans->transaction->cache_write_mutex);
3731         }
3732         mutex_unlock(&trans->transaction->cache_write_mutex);
3733
3734         /*
3735          * go through delayed refs for all the stuff we've just kicked off
3736          * and then loop back (just once)
3737          */
3738         ret = btrfs_run_delayed_refs(trans, fs_info, 0);
3739         if (!ret && loops == 0) {
3740                 loops++;
3741                 spin_lock(&cur_trans->dirty_bgs_lock);
3742                 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3743                 /*
3744                  * dirty_bgs_lock protects us from concurrent block group
3745                  * deletes too (not just cache_write_mutex).
3746                  */
3747                 if (!list_empty(&dirty)) {
3748                         spin_unlock(&cur_trans->dirty_bgs_lock);
3749                         goto again;
3750                 }
3751                 spin_unlock(&cur_trans->dirty_bgs_lock);
3752         } else if (ret < 0) {
3753                 btrfs_cleanup_dirty_bgs(cur_trans, fs_info);
3754         }
3755
3756         btrfs_free_path(path);
3757         return ret;
3758 }
3759
3760 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3761                                    struct btrfs_fs_info *fs_info)
3762 {
3763         struct btrfs_block_group_cache *cache;
3764         struct btrfs_transaction *cur_trans = trans->transaction;
3765         int ret = 0;
3766         int should_put;
3767         struct btrfs_path *path;
3768         struct list_head *io = &cur_trans->io_bgs;
3769         int num_started = 0;
3770
3771         path = btrfs_alloc_path();
3772         if (!path)
3773                 return -ENOMEM;
3774
3775         /*
3776          * Even though we are in the critical section of the transaction commit,
3777          * we can still have concurrent tasks adding elements to this
3778          * transaction's list of dirty block groups. These tasks correspond to
3779          * endio free space workers started when writeback finishes for a
3780          * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
3781          * allocate new block groups as a result of COWing nodes of the root
3782          * tree when updating the free space inode. The writeback for the space
3783          * caches is triggered by an earlier call to
3784          * btrfs_start_dirty_block_groups() and iterations of the following
3785          * loop.
3786          * Also we want to do the cache_save_setup first and then run the
3787          * delayed refs to make sure we have the best chance at doing this all
3788          * in one shot.
3789          */
3790         spin_lock(&cur_trans->dirty_bgs_lock);
3791         while (!list_empty(&cur_trans->dirty_bgs)) {
3792                 cache = list_first_entry(&cur_trans->dirty_bgs,
3793                                          struct btrfs_block_group_cache,
3794                                          dirty_list);
3795
3796                 /*
3797                  * this can happen if cache_save_setup re-dirties a block
3798                  * group that is already under IO.  Just wait for it to
3799                  * finish and then do it all again
3800                  */
3801                 if (!list_empty(&cache->io_list)) {
3802                         spin_unlock(&cur_trans->dirty_bgs_lock);
3803                         list_del_init(&cache->io_list);
3804                         btrfs_wait_cache_io(trans, cache, path);
3805                         btrfs_put_block_group(cache);
3806                         spin_lock(&cur_trans->dirty_bgs_lock);
3807                 }
3808
3809                 /*
3810                  * don't remove from the dirty list until after we've waited
3811                  * on any pending IO
3812                  */
3813                 list_del_init(&cache->dirty_list);
3814                 spin_unlock(&cur_trans->dirty_bgs_lock);
3815                 should_put = 1;
3816
3817                 cache_save_setup(cache, trans, path);
3818
3819                 if (!ret)
3820                         ret = btrfs_run_delayed_refs(trans, fs_info,
3821                                                      (unsigned long) -1);
3822
3823                 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3824                         cache->io_ctl.inode = NULL;
3825                         ret = btrfs_write_out_cache(fs_info, trans,
3826                                                     cache, path);
3827                         if (ret == 0 && cache->io_ctl.inode) {
3828                                 num_started++;
3829                                 should_put = 0;
3830                                 list_add_tail(&cache->io_list, io);
3831                         } else {
3832                                 /*
3833                                  * if we failed to write the cache, the
3834                                  * generation will be bad and life goes on
3835                                  */
3836                                 ret = 0;
3837                         }
3838                 }
3839                 if (!ret) {
3840                         ret = write_one_cache_group(trans, fs_info,
3841                                                     path, cache);
3842                         /*
3843                          * One of the free space endio workers might have
3844                          * created a new block group while updating a free space
3845                          * cache's inode (at inode.c:btrfs_finish_ordered_io())
3846                          * and hasn't released its transaction handle yet, in
3847                          * which case the new block group is still attached to
3848                          * its transaction handle and its creation has not
3849                          * finished yet (no block group item in the extent tree
3850                          * yet, etc). If this is the case, wait for all free
3851                          * space endio workers to finish and retry. This is a
3852                          * a very rare case so no need for a more efficient and
3853                          * complex approach.
3854                          */
3855                         if (ret == -ENOENT) {
3856                                 wait_event(cur_trans->writer_wait,
3857                                    atomic_read(&cur_trans->num_writers) == 1);
3858                                 ret = write_one_cache_group(trans, fs_info,
3859                                                             path, cache);
3860                         }
3861                         if (ret)
3862                                 btrfs_abort_transaction(trans, ret);
3863                 }
3864
3865                 /* if its not on the io list, we need to put the block group */
3866                 if (should_put)
3867                         btrfs_put_block_group(cache);
3868                 spin_lock(&cur_trans->dirty_bgs_lock);
3869         }
3870         spin_unlock(&cur_trans->dirty_bgs_lock);
3871
3872         while (!list_empty(io)) {
3873                 cache = list_first_entry(io, struct btrfs_block_group_cache,
3874                                          io_list);
3875                 list_del_init(&cache->io_list);
3876                 btrfs_wait_cache_io(trans, cache, path);
3877                 btrfs_put_block_group(cache);
3878         }
3879
3880         btrfs_free_path(path);
3881         return ret;
3882 }
3883
3884 int btrfs_extent_readonly(struct btrfs_fs_info *fs_info, u64 bytenr)
3885 {
3886         struct btrfs_block_group_cache *block_group;
3887         int readonly = 0;
3888
3889         block_group = btrfs_lookup_block_group(fs_info, bytenr);
3890         if (!block_group || block_group->ro)
3891                 readonly = 1;
3892         if (block_group)
3893                 btrfs_put_block_group(block_group);
3894         return readonly;
3895 }
3896
3897 bool btrfs_inc_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3898 {
3899         struct btrfs_block_group_cache *bg;
3900         bool ret = true;
3901
3902         bg = btrfs_lookup_block_group(fs_info, bytenr);
3903         if (!bg)
3904                 return false;
3905
3906         spin_lock(&bg->lock);
3907         if (bg->ro)
3908                 ret = false;
3909         else
3910                 atomic_inc(&bg->nocow_writers);
3911         spin_unlock(&bg->lock);
3912
3913         /* no put on block group, done by btrfs_dec_nocow_writers */
3914         if (!ret)
3915                 btrfs_put_block_group(bg);
3916
3917         return ret;
3918
3919 }
3920
3921 void btrfs_dec_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3922 {
3923         struct btrfs_block_group_cache *bg;
3924
3925         bg = btrfs_lookup_block_group(fs_info, bytenr);
3926         ASSERT(bg);
3927         if (atomic_dec_and_test(&bg->nocow_writers))
3928                 wake_up_atomic_t(&bg->nocow_writers);
3929         /*
3930          * Once for our lookup and once for the lookup done by a previous call
3931          * to btrfs_inc_nocow_writers()
3932          */
3933         btrfs_put_block_group(bg);
3934         btrfs_put_block_group(bg);
3935 }
3936
3937 static int btrfs_wait_nocow_writers_atomic_t(atomic_t *a)
3938 {
3939         schedule();
3940         return 0;
3941 }
3942
3943 void btrfs_wait_nocow_writers(struct btrfs_block_group_cache *bg)
3944 {
3945         wait_on_atomic_t(&bg->nocow_writers,
3946                          btrfs_wait_nocow_writers_atomic_t,
3947                          TASK_UNINTERRUPTIBLE);
3948 }
3949
3950 static const char *alloc_name(u64 flags)
3951 {
3952         switch (flags) {
3953         case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA:
3954                 return "mixed";
3955         case BTRFS_BLOCK_GROUP_METADATA:
3956                 return "metadata";
3957         case BTRFS_BLOCK_GROUP_DATA:
3958                 return "data";
3959         case BTRFS_BLOCK_GROUP_SYSTEM:
3960                 return "system";
3961         default:
3962                 WARN_ON(1);
3963                 return "invalid-combination";
3964         };
3965 }
3966
3967 static int create_space_info(struct btrfs_fs_info *info, u64 flags,
3968                              struct btrfs_space_info **new)
3969 {
3970
3971         struct btrfs_space_info *space_info;
3972         int i;
3973         int ret;
3974
3975         space_info = kzalloc(sizeof(*space_info), GFP_NOFS);
3976         if (!space_info)
3977                 return -ENOMEM;
3978
3979         ret = percpu_counter_init(&space_info->total_bytes_pinned, 0,
3980                                  GFP_KERNEL);
3981         if (ret) {
3982                 kfree(space_info);
3983                 return ret;
3984         }
3985
3986         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3987                 INIT_LIST_HEAD(&space_info->block_groups[i]);
3988         init_rwsem(&space_info->groups_sem);
3989         spin_lock_init(&space_info->lock);
3990         space_info->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3991         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3992         init_waitqueue_head(&space_info->wait);
3993         INIT_LIST_HEAD(&space_info->ro_bgs);
3994         INIT_LIST_HEAD(&space_info->tickets);
3995         INIT_LIST_HEAD(&space_info->priority_tickets);
3996
3997         ret = kobject_init_and_add(&space_info->kobj, &space_info_ktype,
3998                                     info->space_info_kobj, "%s",
3999                                     alloc_name(space_info->flags));
4000         if (ret) {
4001                 percpu_counter_destroy(&space_info->total_bytes_pinned);
4002                 kfree(space_info);
4003                 return ret;
4004         }
4005
4006         *new = space_info;
4007         list_add_rcu(&space_info->list, &info->space_info);
4008         if (flags & BTRFS_BLOCK_GROUP_DATA)
4009                 info->data_sinfo = space_info;
4010
4011         return ret;
4012 }
4013
4014 static void update_space_info(struct btrfs_fs_info *info, u64 flags,
4015                              u64 total_bytes, u64 bytes_used,
4016                              u64 bytes_readonly,
4017                              struct btrfs_space_info **space_info)
4018 {
4019         struct btrfs_space_info *found;
4020         int factor;
4021
4022         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
4023                      BTRFS_BLOCK_GROUP_RAID10))
4024                 factor = 2;
4025         else
4026                 factor = 1;
4027
4028         found = __find_space_info(info, flags);
4029         ASSERT(found);
4030         spin_lock(&found->lock);
4031         found->total_bytes += total_bytes;
4032         found->disk_total += total_bytes * factor;
4033         found->bytes_used += bytes_used;
4034         found->disk_used += bytes_used * factor;
4035         found->bytes_readonly += bytes_readonly;
4036         if (total_bytes > 0)
4037                 found->full = 0;
4038         space_info_add_new_bytes(info, found, total_bytes -
4039                                  bytes_used - bytes_readonly);
4040         spin_unlock(&found->lock);
4041         *space_info = found;
4042 }
4043
4044 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
4045 {
4046         u64 extra_flags = chunk_to_extended(flags) &
4047                                 BTRFS_EXTENDED_PROFILE_MASK;
4048
4049         write_seqlock(&fs_info->profiles_lock);
4050         if (flags & BTRFS_BLOCK_GROUP_DATA)
4051                 fs_info->avail_data_alloc_bits |= extra_flags;
4052         if (flags & BTRFS_BLOCK_GROUP_METADATA)
4053                 fs_info->avail_metadata_alloc_bits |= extra_flags;
4054         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4055                 fs_info->avail_system_alloc_bits |= extra_flags;
4056         write_sequnlock(&fs_info->profiles_lock);
4057 }
4058
4059 /*
4060  * returns target flags in extended format or 0 if restripe for this
4061  * chunk_type is not in progress
4062  *
4063  * should be called with either volume_mutex or balance_lock held
4064  */
4065 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
4066 {
4067         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
4068         u64 target = 0;
4069
4070         if (!bctl)
4071                 return 0;
4072
4073         if (flags & BTRFS_BLOCK_GROUP_DATA &&
4074             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4075                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
4076         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
4077                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4078                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
4079         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
4080                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4081                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
4082         }
4083
4084         return target;
4085 }
4086
4087 /*
4088  * @flags: available profiles in extended format (see ctree.h)
4089  *
4090  * Returns reduced profile in chunk format.  If profile changing is in
4091  * progress (either running or paused) picks the target profile (if it's
4092  * already available), otherwise falls back to plain reducing.
4093  */
4094 static u64 btrfs_reduce_alloc_profile(struct btrfs_fs_info *fs_info, u64 flags)
4095 {
4096         u64 num_devices = fs_info->fs_devices->rw_devices;
4097         u64 target;
4098         u64 raid_type;
4099         u64 allowed = 0;
4100
4101         /*
4102          * see if restripe for this chunk_type is in progress, if so
4103          * try to reduce to the target profile
4104          */
4105         spin_lock(&fs_info->balance_lock);
4106         target = get_restripe_target(fs_info, flags);
4107         if (target) {
4108                 /* pick target profile only if it's already available */
4109                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
4110                         spin_unlock(&fs_info->balance_lock);
4111                         return extended_to_chunk(target);
4112                 }
4113         }
4114         spin_unlock(&fs_info->balance_lock);
4115
4116         /* First, mask out the RAID levels which aren't possible */
4117         for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
4118                 if (num_devices >= btrfs_raid_array[raid_type].devs_min)
4119                         allowed |= btrfs_raid_group[raid_type];
4120         }
4121         allowed &= flags;
4122
4123         if (allowed & BTRFS_BLOCK_GROUP_RAID6)
4124                 allowed = BTRFS_BLOCK_GROUP_RAID6;
4125         else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
4126                 allowed = BTRFS_BLOCK_GROUP_RAID5;
4127         else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
4128                 allowed = BTRFS_BLOCK_GROUP_RAID10;
4129         else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
4130                 allowed = BTRFS_BLOCK_GROUP_RAID1;
4131         else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
4132                 allowed = BTRFS_BLOCK_GROUP_RAID0;
4133
4134         flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
4135
4136         return extended_to_chunk(flags | allowed);
4137 }
4138
4139 static u64 get_alloc_profile(struct btrfs_fs_info *fs_info, u64 orig_flags)
4140 {
4141         unsigned seq;
4142         u64 flags;
4143
4144         do {
4145                 flags = orig_flags;
4146                 seq = read_seqbegin(&fs_info->profiles_lock);
4147
4148                 if (flags & BTRFS_BLOCK_GROUP_DATA)
4149                         flags |= fs_info->avail_data_alloc_bits;
4150                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4151                         flags |= fs_info->avail_system_alloc_bits;
4152                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
4153                         flags |= fs_info->avail_metadata_alloc_bits;
4154         } while (read_seqretry(&fs_info->profiles_lock, seq));
4155
4156         return btrfs_reduce_alloc_profile(fs_info, flags);
4157 }
4158
4159 static u64 get_alloc_profile_by_root(struct btrfs_root *root, int data)
4160 {
4161         struct btrfs_fs_info *fs_info = root->fs_info;
4162         u64 flags;
4163         u64 ret;
4164
4165         if (data)
4166                 flags = BTRFS_BLOCK_GROUP_DATA;
4167         else if (root == fs_info->chunk_root)
4168                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
4169         else
4170                 flags = BTRFS_BLOCK_GROUP_METADATA;
4171
4172         ret = get_alloc_profile(fs_info, flags);
4173         return ret;
4174 }
4175
4176 u64 btrfs_data_alloc_profile(struct btrfs_fs_info *fs_info)
4177 {
4178         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_DATA);
4179 }
4180
4181 u64 btrfs_metadata_alloc_profile(struct btrfs_fs_info *fs_info)
4182 {
4183         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4184 }
4185
4186 u64 btrfs_system_alloc_profile(struct btrfs_fs_info *fs_info)
4187 {
4188         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4189 }
4190
4191 static u64 btrfs_space_info_used(struct btrfs_space_info *s_info,
4192                                  bool may_use_included)
4193 {
4194         ASSERT(s_info);
4195         return s_info->bytes_used + s_info->bytes_reserved +
4196                 s_info->bytes_pinned + s_info->bytes_readonly +
4197                 (may_use_included ? s_info->bytes_may_use : 0);
4198 }
4199
4200 int btrfs_alloc_data_chunk_ondemand(struct btrfs_inode *inode, u64 bytes)
4201 {
4202         struct btrfs_root *root = inode->root;
4203         struct btrfs_fs_info *fs_info = root->fs_info;
4204         struct btrfs_space_info *data_sinfo = fs_info->data_sinfo;
4205         u64 used;
4206         int ret = 0;
4207         int need_commit = 2;
4208         int have_pinned_space;
4209
4210         /* make sure bytes are sectorsize aligned */
4211         bytes = ALIGN(bytes, fs_info->sectorsize);
4212
4213         if (btrfs_is_free_space_inode(inode)) {
4214                 need_commit = 0;
4215                 ASSERT(current->journal_info);
4216         }
4217
4218 again:
4219         /* make sure we have enough space to handle the data first */
4220         spin_lock(&data_sinfo->lock);
4221         used = btrfs_space_info_used(data_sinfo, true);
4222
4223         if (used + bytes > data_sinfo->total_bytes) {
4224                 struct btrfs_trans_handle *trans;
4225
4226                 /*
4227                  * if we don't have enough free bytes in this space then we need
4228                  * to alloc a new chunk.
4229                  */
4230                 if (!data_sinfo->full) {
4231                         u64 alloc_target;
4232
4233                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
4234                         spin_unlock(&data_sinfo->lock);
4235
4236                         alloc_target = btrfs_data_alloc_profile(fs_info);
4237                         /*
4238                          * It is ugly that we don't call nolock join
4239                          * transaction for the free space inode case here.
4240                          * But it is safe because we only do the data space
4241                          * reservation for the free space cache in the
4242                          * transaction context, the common join transaction
4243                          * just increase the counter of the current transaction
4244                          * handler, doesn't try to acquire the trans_lock of
4245                          * the fs.
4246                          */
4247                         trans = btrfs_join_transaction(root);
4248                         if (IS_ERR(trans))
4249                                 return PTR_ERR(trans);
4250
4251                         ret = do_chunk_alloc(trans, fs_info, alloc_target,
4252                                              CHUNK_ALLOC_NO_FORCE);
4253                         btrfs_end_transaction(trans);
4254                         if (ret < 0) {
4255                                 if (ret != -ENOSPC)
4256                                         return ret;
4257                                 else {
4258                                         have_pinned_space = 1;
4259                                         goto commit_trans;
4260                                 }
4261                         }
4262
4263                         goto again;
4264                 }
4265
4266                 /*
4267                  * If we don't have enough pinned space to deal with this
4268                  * allocation, and no removed chunk in current transaction,
4269                  * don't bother committing the transaction.
4270                  */
4271                 have_pinned_space = percpu_counter_compare(
4272                         &data_sinfo->total_bytes_pinned,
4273                         used + bytes - data_sinfo->total_bytes);
4274                 spin_unlock(&data_sinfo->lock);
4275
4276                 /* commit the current transaction and try again */
4277 commit_trans:
4278                 if (need_commit &&
4279                     !atomic_read(&fs_info->open_ioctl_trans)) {
4280                         need_commit--;
4281
4282                         if (need_commit > 0) {
4283                                 btrfs_start_delalloc_roots(fs_info, 0, -1);
4284                                 btrfs_wait_ordered_roots(fs_info, U64_MAX, 0,
4285                                                          (u64)-1);
4286                         }
4287
4288                         trans = btrfs_join_transaction(root);
4289                         if (IS_ERR(trans))
4290                                 return PTR_ERR(trans);
4291                         if (have_pinned_space >= 0 ||
4292                             test_bit(BTRFS_TRANS_HAVE_FREE_BGS,
4293                                      &trans->transaction->flags) ||
4294                             need_commit > 0) {
4295                                 ret = btrfs_commit_transaction(trans);
4296                                 if (ret)
4297                                         return ret;
4298                                 /*
4299                                  * The cleaner kthread might still be doing iput
4300                                  * operations. Wait for it to finish so that
4301                                  * more space is released.
4302                                  */
4303                                 mutex_lock(&fs_info->cleaner_delayed_iput_mutex);
4304                                 mutex_unlock(&fs_info->cleaner_delayed_iput_mutex);
4305                                 goto again;
4306                         } else {
4307                                 btrfs_end_transaction(trans);
4308                         }
4309                 }
4310
4311                 trace_btrfs_space_reservation(fs_info,
4312                                               "space_info:enospc",
4313                                               data_sinfo->flags, bytes, 1);
4314                 return -ENOSPC;
4315         }
4316         data_sinfo->bytes_may_use += bytes;
4317         trace_btrfs_space_reservation(fs_info, "space_info",
4318                                       data_sinfo->flags, bytes, 1);
4319         spin_unlock(&data_sinfo->lock);
4320
4321         return ret;
4322 }
4323
4324 int btrfs_check_data_free_space(struct inode *inode,
4325                         struct extent_changeset **reserved, u64 start, u64 len)
4326 {
4327         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4328         int ret;
4329
4330         /* align the range */
4331         len = round_up(start + len, fs_info->sectorsize) -
4332               round_down(start, fs_info->sectorsize);
4333         start = round_down(start, fs_info->sectorsize);
4334
4335         ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode), len);
4336         if (ret < 0)
4337                 return ret;
4338
4339         /* Use new btrfs_qgroup_reserve_data to reserve precious data space. */
4340         ret = btrfs_qgroup_reserve_data(inode, reserved, start, len);
4341         if (ret < 0)
4342                 btrfs_free_reserved_data_space_noquota(inode, start, len);
4343         else
4344                 ret = 0;
4345         return ret;
4346 }
4347
4348 /*
4349  * Called if we need to clear a data reservation for this inode
4350  * Normally in a error case.
4351  *
4352  * This one will *NOT* use accurate qgroup reserved space API, just for case
4353  * which we can't sleep and is sure it won't affect qgroup reserved space.
4354  * Like clear_bit_hook().
4355  */
4356 void btrfs_free_reserved_data_space_noquota(struct inode *inode, u64 start,
4357                                             u64 len)
4358 {
4359         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4360         struct btrfs_space_info *data_sinfo;
4361
4362         /* Make sure the range is aligned to sectorsize */
4363         len = round_up(start + len, fs_info->sectorsize) -
4364               round_down(start, fs_info->sectorsize);
4365         start = round_down(start, fs_info->sectorsize);
4366
4367         data_sinfo = fs_info->data_sinfo;
4368         spin_lock(&data_sinfo->lock);
4369         if (WARN_ON(data_sinfo->bytes_may_use < len))
4370                 data_sinfo->bytes_may_use = 0;
4371         else
4372                 data_sinfo->bytes_may_use -= len;
4373         trace_btrfs_space_reservation(fs_info, "space_info",
4374                                       data_sinfo->flags, len, 0);
4375         spin_unlock(&data_sinfo->lock);
4376 }
4377
4378 /*
4379  * Called if we need to clear a data reservation for this inode
4380  * Normally in a error case.
4381  *
4382  * This one will handle the per-inode data rsv map for accurate reserved
4383  * space framework.
4384  */
4385 void btrfs_free_reserved_data_space(struct inode *inode,
4386                         struct extent_changeset *reserved, u64 start, u64 len)
4387 {
4388         struct btrfs_root *root = BTRFS_I(inode)->root;
4389
4390         /* Make sure the range is aligned to sectorsize */
4391         len = round_up(start + len, root->fs_info->sectorsize) -
4392               round_down(start, root->fs_info->sectorsize);
4393         start = round_down(start, root->fs_info->sectorsize);
4394
4395         btrfs_free_reserved_data_space_noquota(inode, start, len);
4396         btrfs_qgroup_free_data(inode, reserved, start, len);
4397 }
4398
4399 static void force_metadata_allocation(struct btrfs_fs_info *info)
4400 {
4401         struct list_head *head = &info->space_info;
4402         struct btrfs_space_info *found;
4403
4404         rcu_read_lock();
4405         list_for_each_entry_rcu(found, head, list) {
4406                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
4407                         found->force_alloc = CHUNK_ALLOC_FORCE;
4408         }
4409         rcu_read_unlock();
4410 }
4411
4412 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
4413 {
4414         return (global->size << 1);
4415 }
4416
4417 static int should_alloc_chunk(struct btrfs_fs_info *fs_info,
4418                               struct btrfs_space_info *sinfo, int force)
4419 {
4420         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4421         u64 bytes_used = btrfs_space_info_used(sinfo, false);
4422         u64 thresh;
4423
4424         if (force == CHUNK_ALLOC_FORCE)
4425                 return 1;
4426
4427         /*
4428          * We need to take into account the global rsv because for all intents
4429          * and purposes it's used space.  Don't worry about locking the
4430          * global_rsv, it doesn't change except when the transaction commits.
4431          */
4432         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
4433                 bytes_used += calc_global_rsv_need_space(global_rsv);
4434
4435         /*
4436          * in limited mode, we want to have some free space up to
4437          * about 1% of the FS size.
4438          */
4439         if (force == CHUNK_ALLOC_LIMITED) {
4440                 thresh = btrfs_super_total_bytes(fs_info->super_copy);
4441                 thresh = max_t(u64, SZ_64M, div_factor_fine(thresh, 1));
4442
4443                 if (sinfo->total_bytes - bytes_used < thresh)
4444                         return 1;
4445         }
4446
4447         if (bytes_used + SZ_2M < div_factor(sinfo->total_bytes, 8))
4448                 return 0;
4449         return 1;
4450 }
4451
4452 static u64 get_profile_num_devs(struct btrfs_fs_info *fs_info, u64 type)
4453 {
4454         u64 num_dev;
4455
4456         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
4457                     BTRFS_BLOCK_GROUP_RAID0 |
4458                     BTRFS_BLOCK_GROUP_RAID5 |
4459                     BTRFS_BLOCK_GROUP_RAID6))
4460                 num_dev = fs_info->fs_devices->rw_devices;
4461         else if (type & BTRFS_BLOCK_GROUP_RAID1)
4462                 num_dev = 2;
4463         else
4464                 num_dev = 1;    /* DUP or single */
4465
4466         return num_dev;
4467 }
4468
4469 /*
4470  * If @is_allocation is true, reserve space in the system space info necessary
4471  * for allocating a chunk, otherwise if it's false, reserve space necessary for
4472  * removing a chunk.
4473  */
4474 void check_system_chunk(struct btrfs_trans_handle *trans,
4475                         struct btrfs_fs_info *fs_info, u64 type)
4476 {
4477         struct btrfs_space_info *info;
4478         u64 left;
4479         u64 thresh;
4480         int ret = 0;
4481         u64 num_devs;
4482
4483         /*
4484          * Needed because we can end up allocating a system chunk and for an
4485          * atomic and race free space reservation in the chunk block reserve.
4486          */
4487         ASSERT(mutex_is_locked(&fs_info->chunk_mutex));
4488
4489         info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4490         spin_lock(&info->lock);
4491         left = info->total_bytes - btrfs_space_info_used(info, true);
4492         spin_unlock(&info->lock);
4493
4494         num_devs = get_profile_num_devs(fs_info, type);
4495
4496         /* num_devs device items to update and 1 chunk item to add or remove */
4497         thresh = btrfs_calc_trunc_metadata_size(fs_info, num_devs) +
4498                 btrfs_calc_trans_metadata_size(fs_info, 1);
4499
4500         if (left < thresh && btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
4501                 btrfs_info(fs_info, "left=%llu, need=%llu, flags=%llu",
4502                            left, thresh, type);
4503                 dump_space_info(fs_info, info, 0, 0);
4504         }
4505
4506         if (left < thresh) {
4507                 u64 flags = btrfs_system_alloc_profile(fs_info);
4508
4509                 /*
4510                  * Ignore failure to create system chunk. We might end up not
4511                  * needing it, as we might not need to COW all nodes/leafs from
4512                  * the paths we visit in the chunk tree (they were already COWed
4513                  * or created in the current transaction for example).
4514                  */
4515                 ret = btrfs_alloc_chunk(trans, fs_info, flags);
4516         }
4517
4518         if (!ret) {
4519                 ret = btrfs_block_rsv_add(fs_info->chunk_root,
4520                                           &fs_info->chunk_block_rsv,
4521                                           thresh, BTRFS_RESERVE_NO_FLUSH);
4522                 if (!ret)
4523                         trans->chunk_bytes_reserved += thresh;
4524         }
4525 }
4526
4527 /*
4528  * If force is CHUNK_ALLOC_FORCE:
4529  *    - return 1 if it successfully allocates a chunk,
4530  *    - return errors including -ENOSPC otherwise.
4531  * If force is NOT CHUNK_ALLOC_FORCE:
4532  *    - return 0 if it doesn't need to allocate a new chunk,
4533  *    - return 1 if it successfully allocates a chunk,
4534  *    - return errors including -ENOSPC otherwise.
4535  */
4536 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
4537                           struct btrfs_fs_info *fs_info, u64 flags, int force)
4538 {
4539         struct btrfs_space_info *space_info;
4540         int wait_for_alloc = 0;
4541         int ret = 0;
4542
4543         /* Don't re-enter if we're already allocating a chunk */
4544         if (trans->allocating_chunk)
4545                 return -ENOSPC;
4546
4547         space_info = __find_space_info(fs_info, flags);
4548         if (!space_info) {
4549                 ret = create_space_info(fs_info, flags, &space_info);
4550                 if (ret)
4551                         return ret;
4552         }
4553
4554 again:
4555         spin_lock(&space_info->lock);
4556         if (force < space_info->force_alloc)
4557                 force = space_info->force_alloc;
4558         if (space_info->full) {
4559                 if (should_alloc_chunk(fs_info, space_info, force))
4560                         ret = -ENOSPC;
4561                 else
4562                         ret = 0;
4563                 spin_unlock(&space_info->lock);
4564                 return ret;
4565         }
4566
4567         if (!should_alloc_chunk(fs_info, space_info, force)) {
4568                 spin_unlock(&space_info->lock);
4569                 return 0;
4570         } else if (space_info->chunk_alloc) {
4571                 wait_for_alloc = 1;
4572         } else {
4573                 space_info->chunk_alloc = 1;
4574         }
4575
4576         spin_unlock(&space_info->lock);
4577
4578         mutex_lock(&fs_info->chunk_mutex);
4579
4580         /*
4581          * The chunk_mutex is held throughout the entirety of a chunk
4582          * allocation, so once we've acquired the chunk_mutex we know that the
4583          * other guy is done and we need to recheck and see if we should
4584          * allocate.
4585          */
4586         if (wait_for_alloc) {
4587                 mutex_unlock(&fs_info->chunk_mutex);
4588                 wait_for_alloc = 0;
4589                 goto again;
4590         }
4591
4592         trans->allocating_chunk = true;
4593
4594         /*
4595          * If we have mixed data/metadata chunks we want to make sure we keep
4596          * allocating mixed chunks instead of individual chunks.
4597          */
4598         if (btrfs_mixed_space_info(space_info))
4599                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4600
4601         /*
4602          * if we're doing a data chunk, go ahead and make sure that
4603          * we keep a reasonable number of metadata chunks allocated in the
4604          * FS as well.
4605          */
4606         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4607                 fs_info->data_chunk_allocations++;
4608                 if (!(fs_info->data_chunk_allocations %
4609                       fs_info->metadata_ratio))
4610                         force_metadata_allocation(fs_info);
4611         }
4612
4613         /*
4614          * Check if we have enough space in SYSTEM chunk because we may need
4615          * to update devices.
4616          */
4617         check_system_chunk(trans, fs_info, flags);
4618
4619         ret = btrfs_alloc_chunk(trans, fs_info, flags);
4620         trans->allocating_chunk = false;
4621
4622         spin_lock(&space_info->lock);
4623         if (ret < 0 && ret != -ENOSPC)
4624                 goto out;
4625         if (ret)
4626                 space_info->full = 1;
4627         else
4628                 ret = 1;
4629
4630         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4631 out:
4632         space_info->chunk_alloc = 0;
4633         spin_unlock(&space_info->lock);
4634         mutex_unlock(&fs_info->chunk_mutex);
4635         /*
4636          * When we allocate a new chunk we reserve space in the chunk block
4637          * reserve to make sure we can COW nodes/leafs in the chunk tree or
4638          * add new nodes/leafs to it if we end up needing to do it when
4639          * inserting the chunk item and updating device items as part of the
4640          * second phase of chunk allocation, performed by
4641          * btrfs_finish_chunk_alloc(). So make sure we don't accumulate a
4642          * large number of new block groups to create in our transaction
4643          * handle's new_bgs list to avoid exhausting the chunk block reserve
4644          * in extreme cases - like having a single transaction create many new
4645          * block groups when starting to write out the free space caches of all
4646          * the block groups that were made dirty during the lifetime of the
4647          * transaction.
4648          */
4649         if (trans->can_flush_pending_bgs &&
4650             trans->chunk_bytes_reserved >= (u64)SZ_2M) {
4651                 btrfs_create_pending_block_groups(trans, fs_info);
4652                 btrfs_trans_release_chunk_metadata(trans);
4653         }
4654         return ret;
4655 }
4656
4657 static int can_overcommit(struct btrfs_fs_info *fs_info,
4658                           struct btrfs_space_info *space_info, u64 bytes,
4659                           enum btrfs_reserve_flush_enum flush,
4660                           bool system_chunk)
4661 {
4662         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4663         u64 profile;
4664         u64 space_size;
4665         u64 avail;
4666         u64 used;
4667
4668         /* Don't overcommit when in mixed mode. */
4669         if (space_info->flags & BTRFS_BLOCK_GROUP_DATA)
4670                 return 0;
4671
4672         if (system_chunk)
4673                 profile = btrfs_system_alloc_profile(fs_info);
4674         else
4675                 profile = btrfs_metadata_alloc_profile(fs_info);
4676
4677         used = btrfs_space_info_used(space_info, false);
4678
4679         /*
4680          * We only want to allow over committing if we have lots of actual space
4681          * free, but if we don't have enough space to handle the global reserve
4682          * space then we could end up having a real enospc problem when trying
4683          * to allocate a chunk or some other such important allocation.
4684          */
4685         spin_lock(&global_rsv->lock);
4686         space_size = calc_global_rsv_need_space(global_rsv);
4687         spin_unlock(&global_rsv->lock);
4688         if (used + space_size >= space_info->total_bytes)
4689                 return 0;
4690
4691         used += space_info->bytes_may_use;
4692
4693         avail = atomic64_read(&fs_info->free_chunk_space);
4694
4695         /*
4696          * If we have dup, raid1 or raid10 then only half of the free
4697          * space is actually useable.  For raid56, the space info used
4698          * doesn't include the parity drive, so we don't have to
4699          * change the math
4700          */
4701         if (profile & (BTRFS_BLOCK_GROUP_DUP |
4702                        BTRFS_BLOCK_GROUP_RAID1 |
4703                        BTRFS_BLOCK_GROUP_RAID10))
4704                 avail >>= 1;
4705
4706         /*
4707          * If we aren't flushing all things, let us overcommit up to
4708          * 1/2th of the space. If we can flush, don't let us overcommit
4709          * too much, let it overcommit up to 1/8 of the space.
4710          */
4711         if (flush == BTRFS_RESERVE_FLUSH_ALL)
4712                 avail >>= 3;
4713         else
4714                 avail >>= 1;
4715
4716         if (used + bytes < space_info->total_bytes + avail)
4717                 return 1;
4718         return 0;
4719 }
4720
4721 static void btrfs_writeback_inodes_sb_nr(struct btrfs_fs_info *fs_info,
4722                                          unsigned long nr_pages, int nr_items)
4723 {
4724         struct super_block *sb = fs_info->sb;
4725
4726         if (down_read_trylock(&sb->s_umount)) {
4727                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
4728                 up_read(&sb->s_umount);
4729         } else {
4730                 /*
4731                  * We needn't worry the filesystem going from r/w to r/o though
4732                  * we don't acquire ->s_umount mutex, because the filesystem
4733                  * should guarantee the delalloc inodes list be empty after
4734                  * the filesystem is readonly(all dirty pages are written to
4735                  * the disk).
4736                  */
4737                 btrfs_start_delalloc_roots(fs_info, 0, nr_items);
4738                 if (!current->journal_info)
4739                         btrfs_wait_ordered_roots(fs_info, nr_items, 0, (u64)-1);
4740         }
4741 }
4742
4743 static inline u64 calc_reclaim_items_nr(struct btrfs_fs_info *fs_info,
4744                                         u64 to_reclaim)
4745 {
4746         u64 bytes;
4747         u64 nr;
4748
4749         bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
4750         nr = div64_u64(to_reclaim, bytes);
4751         if (!nr)
4752                 nr = 1;
4753         return nr;
4754 }
4755
4756 #define EXTENT_SIZE_PER_ITEM    SZ_256K
4757
4758 /*
4759  * shrink metadata reservation for delalloc
4760  */
4761 static void shrink_delalloc(struct btrfs_fs_info *fs_info, u64 to_reclaim,
4762                             u64 orig, bool wait_ordered)
4763 {
4764         struct btrfs_block_rsv *block_rsv;
4765         struct btrfs_space_info *space_info;
4766         struct btrfs_trans_handle *trans;
4767         u64 delalloc_bytes;
4768         u64 max_reclaim;
4769         u64 items;
4770         long time_left;
4771         unsigned long nr_pages;
4772         int loops;
4773         enum btrfs_reserve_flush_enum flush;
4774
4775         /* Calc the number of the pages we need flush for space reservation */
4776         items = calc_reclaim_items_nr(fs_info, to_reclaim);
4777         to_reclaim = items * EXTENT_SIZE_PER_ITEM;
4778
4779         trans = (struct btrfs_trans_handle *)current->journal_info;
4780         block_rsv = &fs_info->delalloc_block_rsv;
4781         space_info = block_rsv->space_info;
4782
4783         delalloc_bytes = percpu_counter_sum_positive(
4784                                                 &fs_info->delalloc_bytes);
4785         if (delalloc_bytes == 0) {
4786                 if (trans)
4787                         return;
4788                 if (wait_ordered)
4789                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
4790                 return;
4791         }
4792
4793         loops = 0;
4794         while (delalloc_bytes && loops < 3) {
4795                 max_reclaim = min(delalloc_bytes, to_reclaim);
4796                 nr_pages = max_reclaim >> PAGE_SHIFT;
4797                 btrfs_writeback_inodes_sb_nr(fs_info, nr_pages, items);
4798                 /*
4799                  * We need to wait for the async pages to actually start before
4800                  * we do anything.
4801                  */
4802                 max_reclaim = atomic_read(&fs_info->async_delalloc_pages);
4803                 if (!max_reclaim)
4804                         goto skip_async;
4805
4806                 if (max_reclaim <= nr_pages)
4807                         max_reclaim = 0;
4808                 else
4809                         max_reclaim -= nr_pages;
4810
4811                 wait_event(fs_info->async_submit_wait,
4812                            atomic_read(&fs_info->async_delalloc_pages) <=
4813                            (int)max_reclaim);
4814 skip_async:
4815                 if (!trans)
4816                         flush = BTRFS_RESERVE_FLUSH_ALL;
4817                 else
4818                         flush = BTRFS_RESERVE_NO_FLUSH;
4819                 spin_lock(&space_info->lock);
4820                 if (list_empty(&space_info->tickets) &&
4821                     list_empty(&space_info->priority_tickets)) {
4822                         spin_unlock(&space_info->lock);
4823                         break;
4824                 }
4825                 spin_unlock(&space_info->lock);
4826
4827                 loops++;
4828                 if (wait_ordered && !trans) {
4829                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
4830                 } else {
4831                         time_left = schedule_timeout_killable(1);
4832                         if (time_left)
4833                                 break;
4834                 }
4835                 delalloc_bytes = percpu_counter_sum_positive(
4836                                                 &fs_info->delalloc_bytes);
4837         }
4838 }
4839
4840 /**
4841  * maybe_commit_transaction - possibly commit the transaction if its ok to
4842  * @root - the root we're allocating for
4843  * @bytes - the number of bytes we want to reserve
4844  * @force - force the commit
4845  *
4846  * This will check to make sure that committing the transaction will actually
4847  * get us somewhere and then commit the transaction if it does.  Otherwise it
4848  * will return -ENOSPC.
4849  */
4850 static int may_commit_transaction(struct btrfs_fs_info *fs_info,
4851                                   struct btrfs_space_info *space_info,
4852                                   u64 bytes, int force)
4853 {
4854         struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_block_rsv;
4855         struct btrfs_trans_handle *trans;
4856
4857         trans = (struct btrfs_trans_handle *)current->journal_info;
4858         if (trans)
4859                 return -EAGAIN;
4860
4861         if (force)
4862                 goto commit;
4863
4864         /* See if there is enough pinned space to make this reservation */
4865         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4866                                    bytes) >= 0)
4867                 goto commit;
4868
4869         /*
4870          * See if there is some space in the delayed insertion reservation for
4871          * this reservation.
4872          */
4873         if (space_info != delayed_rsv->space_info)
4874                 return -ENOSPC;
4875
4876         spin_lock(&delayed_rsv->lock);
4877         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4878                                    bytes - delayed_rsv->size) < 0) {
4879                 spin_unlock(&delayed_rsv->lock);
4880                 return -ENOSPC;
4881         }
4882         spin_unlock(&delayed_rsv->lock);
4883
4884 commit:
4885         trans = btrfs_join_transaction(fs_info->extent_root);
4886         if (IS_ERR(trans))
4887                 return -ENOSPC;
4888
4889         return btrfs_commit_transaction(trans);
4890 }
4891
4892 struct reserve_ticket {
4893         u64 bytes;
4894         int error;
4895         struct list_head list;
4896         wait_queue_head_t wait;
4897 };
4898
4899 /*
4900  * Try to flush some data based on policy set by @state. This is only advisory
4901  * and may fail for various reasons. The caller is supposed to examine the
4902  * state of @space_info to detect the outcome.
4903  */
4904 static void flush_space(struct btrfs_fs_info *fs_info,
4905                        struct btrfs_space_info *space_info, u64 num_bytes,
4906                        int state)
4907 {
4908         struct btrfs_root *root = fs_info->extent_root;
4909         struct btrfs_trans_handle *trans;
4910         int nr;
4911         int ret = 0;
4912
4913         switch (state) {
4914         case FLUSH_DELAYED_ITEMS_NR:
4915         case FLUSH_DELAYED_ITEMS:
4916                 if (state == FLUSH_DELAYED_ITEMS_NR)
4917                         nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2;
4918                 else
4919                         nr = -1;
4920
4921                 trans = btrfs_join_transaction(root);
4922                 if (IS_ERR(trans)) {
4923                         ret = PTR_ERR(trans);
4924                         break;
4925                 }
4926                 ret = btrfs_run_delayed_items_nr(trans, fs_info, nr);
4927                 btrfs_end_transaction(trans);
4928                 break;
4929         case FLUSH_DELALLOC:
4930         case FLUSH_DELALLOC_WAIT:
4931                 shrink_delalloc(fs_info, num_bytes * 2, num_bytes,
4932                                 state == FLUSH_DELALLOC_WAIT);
4933                 break;
4934         case ALLOC_CHUNK:
4935                 trans = btrfs_join_transaction(root);
4936                 if (IS_ERR(trans)) {
4937                         ret = PTR_ERR(trans);
4938                         break;
4939                 }
4940                 ret = do_chunk_alloc(trans, fs_info,
4941                                      btrfs_metadata_alloc_profile(fs_info),
4942                                      CHUNK_ALLOC_NO_FORCE);
4943                 btrfs_end_transaction(trans);
4944                 if (ret > 0 || ret == -ENOSPC)
4945                         ret = 0;
4946                 break;
4947         case COMMIT_TRANS:
4948                 ret = may_commit_transaction(fs_info, space_info,
4949                                              num_bytes, 0);
4950                 break;
4951         default:
4952                 ret = -ENOSPC;
4953                 break;
4954         }
4955
4956         trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state,
4957                                 ret);
4958         return;
4959 }
4960
4961 static inline u64
4962 btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info,
4963                                  struct btrfs_space_info *space_info,
4964                                  bool system_chunk)
4965 {
4966         struct reserve_ticket *ticket;
4967         u64 used;
4968         u64 expected;
4969         u64 to_reclaim = 0;
4970
4971         list_for_each_entry(ticket, &space_info->tickets, list)
4972                 to_reclaim += ticket->bytes;
4973         list_for_each_entry(ticket, &space_info->priority_tickets, list)
4974                 to_reclaim += ticket->bytes;
4975         if (to_reclaim)
4976                 return to_reclaim;
4977
4978         to_reclaim = min_t(u64, num_online_cpus() * SZ_1M, SZ_16M);
4979         if (can_overcommit(fs_info, space_info, to_reclaim,
4980                            BTRFS_RESERVE_FLUSH_ALL, system_chunk))
4981                 return 0;
4982
4983         used = btrfs_space_info_used(space_info, true);
4984
4985         if (can_overcommit(fs_info, space_info, SZ_1M,
4986                            BTRFS_RESERVE_FLUSH_ALL, system_chunk))
4987                 expected = div_factor_fine(space_info->total_bytes, 95);
4988         else
4989                 expected = div_factor_fine(space_info->total_bytes, 90);
4990
4991         if (used > expected)
4992                 to_reclaim = used - expected;
4993         else
4994                 to_reclaim = 0;
4995         to_reclaim = min(to_reclaim, space_info->bytes_may_use +
4996                                      space_info->bytes_reserved);
4997         return to_reclaim;
4998 }
4999
5000 static inline int need_do_async_reclaim(struct btrfs_fs_info *fs_info,
5001                                         struct btrfs_space_info *space_info,
5002                                         u64 used, bool system_chunk)
5003 {
5004         u64 thresh = div_factor_fine(space_info->total_bytes, 98);
5005
5006         /* If we're just plain full then async reclaim just slows us down. */
5007         if ((space_info->bytes_used + space_info->bytes_reserved) >= thresh)
5008                 return 0;
5009
5010         if (!btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5011                                               system_chunk))
5012                 return 0;
5013
5014         return (used >= thresh && !btrfs_fs_closing(fs_info) &&
5015                 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
5016 }
5017
5018 static void wake_all_tickets(struct list_head *head)
5019 {
5020         struct reserve_ticket *ticket;
5021
5022         while (!list_empty(head)) {
5023                 ticket = list_first_entry(head, struct reserve_ticket, list);
5024                 list_del_init(&ticket->list);
5025                 ticket->error = -ENOSPC;
5026                 wake_up(&ticket->wait);
5027         }
5028 }
5029
5030 /*
5031  * This is for normal flushers, we can wait all goddamned day if we want to.  We
5032  * will loop and continuously try to flush as long as we are making progress.
5033  * We count progress as clearing off tickets each time we have to loop.
5034  */
5035 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
5036 {
5037         struct btrfs_fs_info *fs_info;
5038         struct btrfs_space_info *space_info;
5039         u64 to_reclaim;
5040         int flush_state;
5041         int commit_cycles = 0;
5042         u64 last_tickets_id;
5043
5044         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
5045         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5046
5047         spin_lock(&space_info->lock);
5048         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5049                                                       false);
5050         if (!to_reclaim) {
5051                 space_info->flush = 0;
5052                 spin_unlock(&space_info->lock);
5053                 return;
5054         }
5055         last_tickets_id = space_info->tickets_id;
5056         spin_unlock(&space_info->lock);
5057
5058         flush_state = FLUSH_DELAYED_ITEMS_NR;
5059         do {
5060                 flush_space(fs_info, space_info, to_reclaim, flush_state);
5061                 spin_lock(&space_info->lock);
5062                 if (list_empty(&space_info->tickets)) {
5063                         space_info->flush = 0;
5064                         spin_unlock(&space_info->lock);
5065                         return;
5066                 }
5067                 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info,
5068                                                               space_info,
5069                                                               false);
5070                 if (last_tickets_id == space_info->tickets_id) {
5071                         flush_state++;
5072                 } else {
5073                         last_tickets_id = space_info->tickets_id;
5074                         flush_state = FLUSH_DELAYED_ITEMS_NR;
5075                         if (commit_cycles)
5076                                 commit_cycles--;
5077                 }
5078
5079                 if (flush_state > COMMIT_TRANS) {
5080                         commit_cycles++;
5081                         if (commit_cycles > 2) {
5082                                 wake_all_tickets(&space_info->tickets);
5083                                 space_info->flush = 0;
5084                         } else {
5085                                 flush_state = FLUSH_DELAYED_ITEMS_NR;
5086                         }
5087                 }
5088                 spin_unlock(&space_info->lock);
5089         } while (flush_state <= COMMIT_TRANS);
5090 }
5091
5092 void btrfs_init_async_reclaim_work(struct work_struct *work)
5093 {
5094         INIT_WORK(work, btrfs_async_reclaim_metadata_space);
5095 }
5096
5097 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info,
5098                                             struct btrfs_space_info *space_info,
5099                                             struct reserve_ticket *ticket)
5100 {
5101         u64 to_reclaim;
5102         int flush_state = FLUSH_DELAYED_ITEMS_NR;
5103
5104         spin_lock(&space_info->lock);
5105         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5106                                                       false);
5107         if (!to_reclaim) {
5108                 spin_unlock(&space_info->lock);
5109                 return;
5110         }
5111         spin_unlock(&space_info->lock);
5112
5113         do {
5114                 flush_space(fs_info, space_info, to_reclaim, flush_state);
5115                 flush_state++;
5116                 spin_lock(&space_info->lock);
5117                 if (ticket->bytes == 0) {
5118                         spin_unlock(&space_info->lock);
5119                         return;
5120                 }
5121                 spin_unlock(&space_info->lock);
5122
5123                 /*
5124                  * Priority flushers can't wait on delalloc without
5125                  * deadlocking.
5126                  */
5127                 if (flush_state == FLUSH_DELALLOC ||
5128                     flush_state == FLUSH_DELALLOC_WAIT)
5129                         flush_state = ALLOC_CHUNK;
5130         } while (flush_state < COMMIT_TRANS);
5131 }
5132
5133 static int wait_reserve_ticket(struct btrfs_fs_info *fs_info,
5134                                struct btrfs_space_info *space_info,
5135                                struct reserve_ticket *ticket, u64 orig_bytes)
5136
5137 {
5138         DEFINE_WAIT(wait);
5139         int ret = 0;
5140
5141         spin_lock(&space_info->lock);
5142         while (ticket->bytes > 0 && ticket->error == 0) {
5143                 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE);
5144                 if (ret) {
5145                         ret = -EINTR;
5146                         break;
5147                 }
5148                 spin_unlock(&space_info->lock);
5149
5150                 schedule();
5151
5152                 finish_wait(&ticket->wait, &wait);
5153                 spin_lock(&space_info->lock);
5154         }
5155         if (!ret)
5156                 ret = ticket->error;
5157         if (!list_empty(&ticket->list))
5158                 list_del_init(&ticket->list);
5159         if (ticket->bytes && ticket->bytes < orig_bytes) {
5160                 u64 num_bytes = orig_bytes - ticket->bytes;
5161                 space_info->bytes_may_use -= num_bytes;
5162                 trace_btrfs_space_reservation(fs_info, "space_info",
5163                                               space_info->flags, num_bytes, 0);
5164         }
5165         spin_unlock(&space_info->lock);
5166
5167         return ret;
5168 }
5169
5170 /**
5171  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5172  * @root - the root we're allocating for
5173  * @space_info - the space info we want to allocate from
5174  * @orig_bytes - the number of bytes we want
5175  * @flush - whether or not we can flush to make our reservation
5176  *
5177  * This will reserve orig_bytes number of bytes from the space info associated
5178  * with the block_rsv.  If there is not enough space it will make an attempt to
5179  * flush out space to make room.  It will do this by flushing delalloc if
5180  * possible or committing the transaction.  If flush is 0 then no attempts to
5181  * regain reservations will be made and this will fail if there is not enough
5182  * space already.
5183  */
5184 static int __reserve_metadata_bytes(struct btrfs_fs_info *fs_info,
5185                                     struct btrfs_space_info *space_info,
5186                                     u64 orig_bytes,
5187                                     enum btrfs_reserve_flush_enum flush,
5188                                     bool system_chunk)
5189 {
5190         struct reserve_ticket ticket;
5191         u64 used;
5192         int ret = 0;
5193
5194         ASSERT(orig_bytes);
5195         ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL);
5196
5197         spin_lock(&space_info->lock);
5198         ret = -ENOSPC;
5199         used = btrfs_space_info_used(space_info, true);
5200
5201         /*
5202          * If we have enough space then hooray, make our reservation and carry
5203          * on.  If not see if we can overcommit, and if we can, hooray carry on.
5204          * If not things get more complicated.
5205          */
5206         if (used + orig_bytes <= space_info->total_bytes) {
5207                 space_info->bytes_may_use += orig_bytes;
5208                 trace_btrfs_space_reservation(fs_info, "space_info",
5209                                               space_info->flags, orig_bytes, 1);
5210                 ret = 0;
5211         } else if (can_overcommit(fs_info, space_info, orig_bytes, flush,
5212                                   system_chunk)) {
5213                 space_info->bytes_may_use += orig_bytes;
5214                 trace_btrfs_space_reservation(fs_info, "space_info",
5215                                               space_info->flags, orig_bytes, 1);
5216                 ret = 0;
5217         }
5218
5219         /*
5220          * If we couldn't make a reservation then setup our reservation ticket
5221          * and kick the async worker if it's not already running.
5222          *
5223          * If we are a priority flusher then we just need to add our ticket to
5224          * the list and we will do our own flushing further down.
5225          */
5226         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
5227                 ticket.bytes = orig_bytes;
5228                 ticket.error = 0;
5229                 init_waitqueue_head(&ticket.wait);
5230                 if (flush == BTRFS_RESERVE_FLUSH_ALL) {
5231                         list_add_tail(&ticket.list, &space_info->tickets);
5232                         if (!space_info->flush) {
5233                                 space_info->flush = 1;
5234                                 trace_btrfs_trigger_flush(fs_info,
5235                                                           space_info->flags,
5236                                                           orig_bytes, flush,
5237                                                           "enospc");
5238                                 queue_work(system_unbound_wq,
5239                                            &fs_info->async_reclaim_work);
5240                         }
5241                 } else {
5242                         list_add_tail(&ticket.list,
5243                                       &space_info->priority_tickets);
5244                 }
5245         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
5246                 used += orig_bytes;
5247                 /*
5248                  * We will do the space reservation dance during log replay,
5249                  * which means we won't have fs_info->fs_root set, so don't do
5250                  * the async reclaim as we will panic.
5251                  */
5252                 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) &&
5253                     need_do_async_reclaim(fs_info, space_info,
5254                                           used, system_chunk) &&
5255                     !work_busy(&fs_info->async_reclaim_work)) {
5256                         trace_btrfs_trigger_flush(fs_info, space_info->flags,
5257                                                   orig_bytes, flush, "preempt");
5258                         queue_work(system_unbound_wq,
5259                                    &fs_info->async_reclaim_work);
5260                 }
5261         }
5262         spin_unlock(&space_info->lock);
5263         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
5264                 return ret;
5265
5266         if (flush == BTRFS_RESERVE_FLUSH_ALL)
5267                 return wait_reserve_ticket(fs_info, space_info, &ticket,
5268                                            orig_bytes);
5269
5270         ret = 0;
5271         priority_reclaim_metadata_space(fs_info, space_info, &ticket);
5272         spin_lock(&space_info->lock);
5273         if (ticket.bytes) {
5274                 if (ticket.bytes < orig_bytes) {
5275                         u64 num_bytes = orig_bytes - ticket.bytes;
5276                         space_info->bytes_may_use -= num_bytes;
5277                         trace_btrfs_space_reservation(fs_info, "space_info",
5278                                                       space_info->flags,
5279                                                       num_bytes, 0);
5280
5281                 }
5282                 list_del_init(&ticket.list);
5283                 ret = -ENOSPC;
5284         }
5285         spin_unlock(&space_info->lock);
5286         ASSERT(list_empty(&ticket.list));
5287         return ret;
5288 }
5289
5290 /**
5291  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5292  * @root - the root we're allocating for
5293  * @block_rsv - the block_rsv we're allocating for
5294  * @orig_bytes - the number of bytes we want
5295  * @flush - whether or not we can flush to make our reservation
5296  *
5297  * This will reserve orgi_bytes number of bytes from the space info associated
5298  * with the block_rsv.  If there is not enough space it will make an attempt to
5299  * flush out space to make room.  It will do this by flushing delalloc if
5300  * possible or committing the transaction.  If flush is 0 then no attempts to
5301  * regain reservations will be made and this will fail if there is not enough
5302  * space already.
5303  */
5304 static int reserve_metadata_bytes(struct btrfs_root *root,
5305                                   struct btrfs_block_rsv *block_rsv,
5306                                   u64 orig_bytes,
5307                                   enum btrfs_reserve_flush_enum flush)
5308 {
5309         struct btrfs_fs_info *fs_info = root->fs_info;
5310         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5311         int ret;
5312         bool system_chunk = (root == fs_info->chunk_root);
5313
5314         ret = __reserve_metadata_bytes(fs_info, block_rsv->space_info,
5315                                        orig_bytes, flush, system_chunk);
5316         if (ret == -ENOSPC &&
5317             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
5318                 if (block_rsv != global_rsv &&
5319                     !block_rsv_use_bytes(global_rsv, orig_bytes))
5320                         ret = 0;
5321         }
5322         if (ret == -ENOSPC)
5323                 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
5324                                               block_rsv->space_info->flags,
5325                                               orig_bytes, 1);
5326         return ret;
5327 }
5328
5329 static struct btrfs_block_rsv *get_block_rsv(
5330                                         const struct btrfs_trans_handle *trans,
5331                                         const struct btrfs_root *root)
5332 {
5333         struct btrfs_fs_info *fs_info = root->fs_info;
5334         struct btrfs_block_rsv *block_rsv = NULL;
5335
5336         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
5337             (root == fs_info->csum_root && trans->adding_csums) ||
5338             (root == fs_info->uuid_root))
5339                 block_rsv = trans->block_rsv;
5340
5341         if (!block_rsv)
5342                 block_rsv = root->block_rsv;
5343
5344         if (!block_rsv)
5345                 block_rsv = &fs_info->empty_block_rsv;
5346
5347         return block_rsv;
5348 }
5349
5350 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
5351                                u64 num_bytes)
5352 {
5353         int ret = -ENOSPC;
5354         spin_lock(&block_rsv->lock);
5355         if (block_rsv->reserved >= num_bytes) {
5356                 block_rsv->reserved -= num_bytes;
5357                 if (block_rsv->reserved < block_rsv->size)
5358                         block_rsv->full = 0;
5359                 ret = 0;
5360         }
5361         spin_unlock(&block_rsv->lock);
5362         return ret;
5363 }
5364
5365 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
5366                                 u64 num_bytes, int update_size)
5367 {
5368         spin_lock(&block_rsv->lock);
5369         block_rsv->reserved += num_bytes;
5370         if (update_size)
5371                 block_rsv->size += num_bytes;
5372         else if (block_rsv->reserved >= block_rsv->size)
5373                 block_rsv->full = 1;
5374         spin_unlock(&block_rsv->lock);
5375 }
5376
5377 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
5378                              struct btrfs_block_rsv *dest, u64 num_bytes,
5379                              int min_factor)
5380 {
5381         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5382         u64 min_bytes;
5383
5384         if (global_rsv->space_info != dest->space_info)
5385                 return -ENOSPC;
5386
5387         spin_lock(&global_rsv->lock);
5388         min_bytes = div_factor(global_rsv->size, min_factor);
5389         if (global_rsv->reserved < min_bytes + num_bytes) {
5390                 spin_unlock(&global_rsv->lock);
5391                 return -ENOSPC;
5392         }
5393         global_rsv->reserved -= num_bytes;
5394         if (global_rsv->reserved < global_rsv->size)
5395                 global_rsv->full = 0;
5396         spin_unlock(&global_rsv->lock);
5397
5398         block_rsv_add_bytes(dest, num_bytes, 1);
5399         return 0;
5400 }
5401
5402 /*
5403  * This is for space we already have accounted in space_info->bytes_may_use, so
5404  * basically when we're returning space from block_rsv's.
5405  */
5406 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
5407                                      struct btrfs_space_info *space_info,
5408                                      u64 num_bytes)
5409 {
5410         struct reserve_ticket *ticket;
5411         struct list_head *head;
5412         u64 used;
5413         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH;
5414         bool check_overcommit = false;
5415
5416         spin_lock(&space_info->lock);
5417         head = &space_info->priority_tickets;
5418
5419         /*
5420          * If we are over our limit then we need to check and see if we can
5421          * overcommit, and if we can't then we just need to free up our space
5422          * and not satisfy any requests.
5423          */
5424         used = btrfs_space_info_used(space_info, true);
5425         if (used - num_bytes >= space_info->total_bytes)
5426                 check_overcommit = true;
5427 again:
5428         while (!list_empty(head) && num_bytes) {
5429                 ticket = list_first_entry(head, struct reserve_ticket,
5430                                           list);
5431                 /*
5432                  * We use 0 bytes because this space is already reserved, so
5433                  * adding the ticket space would be a double count.
5434                  */
5435                 if (check_overcommit &&
5436                     !can_overcommit(fs_info, space_info, 0, flush, false))
5437                         break;
5438                 if (num_bytes >= ticket->bytes) {
5439                         list_del_init(&ticket->list);
5440                         num_bytes -= ticket->bytes;
5441                         ticket->bytes = 0;
5442                         space_info->tickets_id++;
5443                         wake_up(&ticket->wait);
5444                 } else {
5445                         ticket->bytes -= num_bytes;
5446                         num_bytes = 0;
5447                 }
5448         }
5449
5450         if (num_bytes && head == &space_info->priority_tickets) {
5451                 head = &space_info->tickets;
5452                 flush = BTRFS_RESERVE_FLUSH_ALL;
5453                 goto again;
5454         }
5455         space_info->bytes_may_use -= num_bytes;
5456         trace_btrfs_space_reservation(fs_info, "space_info",
5457                                       space_info->flags, num_bytes, 0);
5458         spin_unlock(&space_info->lock);
5459 }
5460
5461 /*
5462  * This is for newly allocated space that isn't accounted in
5463  * space_info->bytes_may_use yet.  So if we allocate a chunk or unpin an extent
5464  * we use this helper.
5465  */
5466 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
5467                                      struct btrfs_space_info *space_info,
5468                                      u64 num_bytes)
5469 {
5470         struct reserve_ticket *ticket;
5471         struct list_head *head = &space_info->priority_tickets;
5472
5473 again:
5474         while (!list_empty(head) && num_bytes) {
5475                 ticket = list_first_entry(head, struct reserve_ticket,
5476                                           list);
5477                 if (num_bytes >= ticket->bytes) {
5478                         trace_btrfs_space_reservation(fs_info, "space_info",
5479                                                       space_info->flags,
5480                                                       ticket->bytes, 1);
5481                         list_del_init(&ticket->list);
5482                         num_bytes -= ticket->bytes;
5483                         space_info->bytes_may_use += ticket->bytes;
5484                         ticket->bytes = 0;
5485                         space_info->tickets_id++;
5486                         wake_up(&ticket->wait);
5487                 } else {
5488                         trace_btrfs_space_reservation(fs_info, "space_info",
5489                                                       space_info->flags,
5490                                                       num_bytes, 1);
5491                         space_info->bytes_may_use += num_bytes;
5492                         ticket->bytes -= num_bytes;
5493                         num_bytes = 0;
5494                 }
5495         }
5496
5497         if (num_bytes && head == &space_info->priority_tickets) {
5498                 head = &space_info->tickets;
5499                 goto again;
5500         }
5501 }
5502
5503 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
5504                                     struct btrfs_block_rsv *block_rsv,
5505                                     struct btrfs_block_rsv *dest, u64 num_bytes)
5506 {
5507         struct btrfs_space_info *space_info = block_rsv->space_info;
5508
5509         spin_lock(&block_rsv->lock);
5510         if (num_bytes == (u64)-1)
5511                 num_bytes = block_rsv->size;
5512         block_rsv->size -= num_bytes;
5513         if (block_rsv->reserved >= block_rsv->size) {
5514                 num_bytes = block_rsv->reserved - block_rsv->size;
5515                 block_rsv->reserved = block_rsv->size;
5516                 block_rsv->full = 1;
5517         } else {
5518                 num_bytes = 0;
5519         }
5520         spin_unlock(&block_rsv->lock);
5521
5522         if (num_bytes > 0) {
5523                 if (dest) {
5524                         spin_lock(&dest->lock);
5525                         if (!dest->full) {
5526                                 u64 bytes_to_add;
5527
5528                                 bytes_to_add = dest->size - dest->reserved;
5529                                 bytes_to_add = min(num_bytes, bytes_to_add);
5530                                 dest->reserved += bytes_to_add;
5531                                 if (dest->reserved >= dest->size)
5532                                         dest->full = 1;
5533                                 num_bytes -= bytes_to_add;
5534                         }
5535                         spin_unlock(&dest->lock);
5536                 }
5537                 if (num_bytes)
5538                         space_info_add_old_bytes(fs_info, space_info,
5539                                                  num_bytes);
5540         }
5541 }
5542
5543 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src,
5544                             struct btrfs_block_rsv *dst, u64 num_bytes,
5545                             int update_size)
5546 {
5547         int ret;
5548
5549         ret = block_rsv_use_bytes(src, num_bytes);
5550         if (ret)
5551                 return ret;
5552
5553         block_rsv_add_bytes(dst, num_bytes, update_size);
5554         return 0;
5555 }
5556
5557 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
5558 {
5559         memset(rsv, 0, sizeof(*rsv));
5560         spin_lock_init(&rsv->lock);
5561         rsv->type = type;
5562 }
5563
5564 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_fs_info *fs_info,
5565                                               unsigned short type)
5566 {
5567         struct btrfs_block_rsv *block_rsv;
5568
5569         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
5570         if (!block_rsv)
5571                 return NULL;
5572
5573         btrfs_init_block_rsv(block_rsv, type);
5574         block_rsv->space_info = __find_space_info(fs_info,
5575                                                   BTRFS_BLOCK_GROUP_METADATA);
5576         return block_rsv;
5577 }
5578
5579 void btrfs_free_block_rsv(struct btrfs_fs_info *fs_info,
5580                           struct btrfs_block_rsv *rsv)
5581 {
5582         if (!rsv)
5583                 return;
5584         btrfs_block_rsv_release(fs_info, rsv, (u64)-1);
5585         kfree(rsv);
5586 }
5587
5588 void __btrfs_free_block_rsv(struct btrfs_block_rsv *rsv)
5589 {
5590         kfree(rsv);
5591 }
5592
5593 int btrfs_block_rsv_add(struct btrfs_root *root,
5594                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
5595                         enum btrfs_reserve_flush_enum flush)
5596 {
5597         int ret;
5598
5599         if (num_bytes == 0)
5600                 return 0;
5601
5602         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5603         if (!ret) {
5604                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
5605                 return 0;
5606         }
5607
5608         return ret;
5609 }
5610
5611 int btrfs_block_rsv_check(struct btrfs_block_rsv *block_rsv, int min_factor)
5612 {
5613         u64 num_bytes = 0;
5614         int ret = -ENOSPC;
5615
5616         if (!block_rsv)
5617                 return 0;
5618
5619         spin_lock(&block_rsv->lock);
5620         num_bytes = div_factor(block_rsv->size, min_factor);
5621         if (block_rsv->reserved >= num_bytes)
5622                 ret = 0;
5623         spin_unlock(&block_rsv->lock);
5624
5625         return ret;
5626 }
5627
5628 int btrfs_block_rsv_refill(struct btrfs_root *root,
5629                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
5630                            enum btrfs_reserve_flush_enum flush)
5631 {
5632         u64 num_bytes = 0;
5633         int ret = -ENOSPC;
5634
5635         if (!block_rsv)
5636                 return 0;
5637
5638         spin_lock(&block_rsv->lock);
5639         num_bytes = min_reserved;
5640         if (block_rsv->reserved >= num_bytes)
5641                 ret = 0;
5642         else
5643                 num_bytes -= block_rsv->reserved;
5644         spin_unlock(&block_rsv->lock);
5645
5646         if (!ret)
5647                 return 0;
5648
5649         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5650         if (!ret) {
5651                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
5652                 return 0;
5653         }
5654
5655         return ret;
5656 }
5657
5658 void btrfs_block_rsv_release(struct btrfs_fs_info *fs_info,
5659                              struct btrfs_block_rsv *block_rsv,
5660                              u64 num_bytes)
5661 {
5662         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5663
5664         if (global_rsv == block_rsv ||
5665             block_rsv->space_info != global_rsv->space_info)
5666                 global_rsv = NULL;
5667         block_rsv_release_bytes(fs_info, block_rsv, global_rsv, num_bytes);
5668 }
5669
5670 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
5671 {
5672         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
5673         struct btrfs_space_info *sinfo = block_rsv->space_info;
5674         u64 num_bytes;
5675
5676         /*
5677          * The global block rsv is based on the size of the extent tree, the
5678          * checksum tree and the root tree.  If the fs is empty we want to set
5679          * it to a minimal amount for safety.
5680          */
5681         num_bytes = btrfs_root_used(&fs_info->extent_root->root_item) +
5682                 btrfs_root_used(&fs_info->csum_root->root_item) +
5683                 btrfs_root_used(&fs_info->tree_root->root_item);
5684         num_bytes = max_t(u64, num_bytes, SZ_16M);
5685
5686         spin_lock(&sinfo->lock);
5687         spin_lock(&block_rsv->lock);
5688
5689         block_rsv->size = min_t(u64, num_bytes, SZ_512M);
5690
5691         if (block_rsv->reserved < block_rsv->size) {
5692                 num_bytes = btrfs_space_info_used(sinfo, true);
5693                 if (sinfo->total_bytes > num_bytes) {
5694                         num_bytes = sinfo->total_bytes - num_bytes;
5695                         num_bytes = min(num_bytes,
5696                                         block_rsv->size - block_rsv->reserved);
5697                         block_rsv->reserved += num_bytes;
5698                         sinfo->bytes_may_use += num_bytes;
5699                         trace_btrfs_space_reservation(fs_info, "space_info",
5700                                                       sinfo->flags, num_bytes,
5701                                                       1);
5702                 }
5703         } else if (block_rsv->reserved > block_rsv->size) {
5704                 num_bytes = block_rsv->reserved - block_rsv->size;
5705                 sinfo->bytes_may_use -= num_bytes;
5706                 trace_btrfs_space_reservation(fs_info, "space_info",
5707                                       sinfo->flags, num_bytes, 0);
5708                 block_rsv->reserved = block_rsv->size;
5709         }
5710
5711         if (block_rsv->reserved == block_rsv->size)
5712                 block_rsv->full = 1;
5713         else
5714                 block_rsv->full = 0;
5715
5716         spin_unlock(&block_rsv->lock);
5717         spin_unlock(&sinfo->lock);
5718 }
5719
5720 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
5721 {
5722         struct btrfs_space_info *space_info;
5723
5724         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
5725         fs_info->chunk_block_rsv.space_info = space_info;
5726
5727         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5728         fs_info->global_block_rsv.space_info = space_info;
5729         fs_info->delalloc_block_rsv.space_info = space_info;
5730         fs_info->trans_block_rsv.space_info = space_info;
5731         fs_info->empty_block_rsv.space_info = space_info;
5732         fs_info->delayed_block_rsv.space_info = space_info;
5733
5734         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
5735         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
5736         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
5737         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
5738         if (fs_info->quota_root)
5739                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
5740         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
5741
5742         update_global_block_rsv(fs_info);
5743 }
5744
5745 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
5746 {
5747         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
5748                                 (u64)-1);
5749         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
5750         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
5751         WARN_ON(fs_info->trans_block_rsv.size > 0);
5752         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
5753         WARN_ON(fs_info->chunk_block_rsv.size > 0);
5754         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
5755         WARN_ON(fs_info->delayed_block_rsv.size > 0);
5756         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
5757 }
5758
5759 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
5760                                   struct btrfs_fs_info *fs_info)
5761 {
5762         if (!trans->block_rsv)
5763                 return;
5764
5765         if (!trans->bytes_reserved)
5766                 return;
5767
5768         trace_btrfs_space_reservation(fs_info, "transaction",
5769                                       trans->transid, trans->bytes_reserved, 0);
5770         btrfs_block_rsv_release(fs_info, trans->block_rsv,
5771                                 trans->bytes_reserved);
5772         trans->bytes_reserved = 0;
5773 }
5774
5775 /*
5776  * To be called after all the new block groups attached to the transaction
5777  * handle have been created (btrfs_create_pending_block_groups()).
5778  */
5779 void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans)
5780 {
5781         struct btrfs_fs_info *fs_info = trans->fs_info;
5782
5783         if (!trans->chunk_bytes_reserved)
5784                 return;
5785
5786         WARN_ON_ONCE(!list_empty(&trans->new_bgs));
5787
5788         block_rsv_release_bytes(fs_info, &fs_info->chunk_block_rsv, NULL,
5789                                 trans->chunk_bytes_reserved);
5790         trans->chunk_bytes_reserved = 0;
5791 }
5792
5793 /* Can only return 0 or -ENOSPC */
5794 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
5795                                   struct btrfs_inode *inode)
5796 {
5797         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
5798         struct btrfs_root *root = inode->root;
5799         /*
5800          * We always use trans->block_rsv here as we will have reserved space
5801          * for our orphan when starting the transaction, using get_block_rsv()
5802          * here will sometimes make us choose the wrong block rsv as we could be
5803          * doing a reloc inode for a non refcounted root.
5804          */
5805         struct btrfs_block_rsv *src_rsv = trans->block_rsv;
5806         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
5807
5808         /*
5809          * We need to hold space in order to delete our orphan item once we've
5810          * added it, so this takes the reservation so we can release it later
5811          * when we are truly done with the orphan item.
5812          */
5813         u64 num_bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
5814
5815         trace_btrfs_space_reservation(fs_info, "orphan", btrfs_ino(inode),
5816                         num_bytes, 1);
5817         return btrfs_block_rsv_migrate(src_rsv, dst_rsv, num_bytes, 1);
5818 }
5819
5820 void btrfs_orphan_release_metadata(struct btrfs_inode *inode)
5821 {
5822         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
5823         struct btrfs_root *root = inode->root;
5824         u64 num_bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
5825
5826         trace_btrfs_space_reservation(fs_info, "orphan", btrfs_ino(inode),
5827                         num_bytes, 0);
5828         btrfs_block_rsv_release(fs_info, root->orphan_block_rsv, num_bytes);
5829 }
5830
5831 /*
5832  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
5833  * root: the root of the parent directory
5834  * rsv: block reservation
5835  * items: the number of items that we need do reservation
5836  * qgroup_reserved: used to return the reserved size in qgroup
5837  *
5838  * This function is used to reserve the space for snapshot/subvolume
5839  * creation and deletion. Those operations are different with the
5840  * common file/directory operations, they change two fs/file trees
5841  * and root tree, the number of items that the qgroup reserves is
5842  * different with the free space reservation. So we can not use
5843  * the space reservation mechanism in start_transaction().
5844  */
5845 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
5846                                      struct btrfs_block_rsv *rsv,
5847                                      int items,
5848                                      u64 *qgroup_reserved,
5849                                      bool use_global_rsv)
5850 {
5851         u64 num_bytes;
5852         int ret;
5853         struct btrfs_fs_info *fs_info = root->fs_info;
5854         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5855
5856         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)) {
5857                 /* One for parent inode, two for dir entries */
5858                 num_bytes = 3 * fs_info->nodesize;
5859                 ret = btrfs_qgroup_reserve_meta(root, num_bytes, true);
5860                 if (ret)
5861                         return ret;
5862         } else {
5863                 num_bytes = 0;
5864         }
5865
5866         *qgroup_reserved = num_bytes;
5867
5868         num_bytes = btrfs_calc_trans_metadata_size(fs_info, items);
5869         rsv->space_info = __find_space_info(fs_info,
5870                                             BTRFS_BLOCK_GROUP_METADATA);
5871         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
5872                                   BTRFS_RESERVE_FLUSH_ALL);
5873
5874         if (ret == -ENOSPC && use_global_rsv)
5875                 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes, 1);
5876
5877         if (ret && *qgroup_reserved)
5878                 btrfs_qgroup_free_meta(root, *qgroup_reserved);
5879
5880         return ret;
5881 }
5882
5883 void btrfs_subvolume_release_metadata(struct btrfs_fs_info *fs_info,
5884                                       struct btrfs_block_rsv *rsv)
5885 {
5886         btrfs_block_rsv_release(fs_info, rsv, (u64)-1);
5887 }
5888
5889 /**
5890  * drop_outstanding_extent - drop an outstanding extent
5891  * @inode: the inode we're dropping the extent for
5892  * @num_bytes: the number of bytes we're releasing.
5893  *
5894  * This is called when we are freeing up an outstanding extent, either called
5895  * after an error or after an extent is written.  This will return the number of
5896  * reserved extents that need to be freed.  This must be called with
5897  * BTRFS_I(inode)->lock held.
5898  */
5899 static unsigned drop_outstanding_extent(struct btrfs_inode *inode,
5900                 u64 num_bytes)
5901 {
5902         unsigned drop_inode_space = 0;
5903         unsigned dropped_extents = 0;
5904         unsigned num_extents;
5905
5906         num_extents = count_max_extents(num_bytes);
5907         ASSERT(num_extents);
5908         ASSERT(inode->outstanding_extents >= num_extents);
5909         inode->outstanding_extents -= num_extents;
5910
5911         if (inode->outstanding_extents == 0 &&
5912             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5913                                &inode->runtime_flags))
5914                 drop_inode_space = 1;
5915
5916         /*
5917          * If we have more or the same amount of outstanding extents than we have
5918          * reserved then we need to leave the reserved extents count alone.
5919          */
5920         if (inode->outstanding_extents >= inode->reserved_extents)
5921                 return drop_inode_space;
5922
5923         dropped_extents = inode->reserved_extents - inode->outstanding_extents;
5924         inode->reserved_extents -= dropped_extents;
5925         return dropped_extents + drop_inode_space;
5926 }
5927
5928 /**
5929  * calc_csum_metadata_size - return the amount of metadata space that must be
5930  *      reserved/freed for the given bytes.
5931  * @inode: the inode we're manipulating
5932  * @num_bytes: the number of bytes in question
5933  * @reserve: 1 if we are reserving space, 0 if we are freeing space
5934  *
5935  * This adjusts the number of csum_bytes in the inode and then returns the
5936  * correct amount of metadata that must either be reserved or freed.  We
5937  * calculate how many checksums we can fit into one leaf and then divide the
5938  * number of bytes that will need to be checksumed by this value to figure out
5939  * how many checksums will be required.  If we are adding bytes then the number
5940  * may go up and we will return the number of additional bytes that must be
5941  * reserved.  If it is going down we will return the number of bytes that must
5942  * be freed.
5943  *
5944  * This must be called with BTRFS_I(inode)->lock held.
5945  */
5946 static u64 calc_csum_metadata_size(struct btrfs_inode *inode, u64 num_bytes,
5947                                    int reserve)
5948 {
5949         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
5950         u64 old_csums, num_csums;
5951
5952         if (inode->flags & BTRFS_INODE_NODATASUM && inode->csum_bytes == 0)
5953                 return 0;
5954
5955         old_csums = btrfs_csum_bytes_to_leaves(fs_info, inode->csum_bytes);
5956         if (reserve)
5957                 inode->csum_bytes += num_bytes;
5958         else
5959                 inode->csum_bytes -= num_bytes;
5960         num_csums = btrfs_csum_bytes_to_leaves(fs_info, inode->csum_bytes);
5961
5962         /* No change, no need to reserve more */
5963         if (old_csums == num_csums)
5964                 return 0;
5965
5966         if (reserve)
5967                 return btrfs_calc_trans_metadata_size(fs_info,
5968                                                       num_csums - old_csums);
5969
5970         return btrfs_calc_trans_metadata_size(fs_info, old_csums - num_csums);
5971 }
5972
5973 int btrfs_delalloc_reserve_metadata(struct btrfs_inode *inode, u64 num_bytes)
5974 {
5975         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
5976         struct btrfs_root *root = inode->root;
5977         struct btrfs_block_rsv *block_rsv = &fs_info->delalloc_block_rsv;
5978         u64 to_reserve = 0;
5979         u64 csum_bytes;
5980         unsigned nr_extents;
5981         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
5982         int ret = 0;
5983         bool delalloc_lock = true;
5984         u64 to_free = 0;
5985         unsigned dropped;
5986         bool release_extra = false;
5987
5988         /* If we are a free space inode we need to not flush since we will be in
5989          * the middle of a transaction commit.  We also don't need the delalloc
5990          * mutex since we won't race with anybody.  We need this mostly to make
5991          * lockdep shut its filthy mouth.
5992          *
5993          * If we have a transaction open (can happen if we call truncate_block
5994          * from truncate), then we need FLUSH_LIMIT so we don't deadlock.
5995          */
5996         if (btrfs_is_free_space_inode(inode)) {
5997                 flush = BTRFS_RESERVE_NO_FLUSH;
5998                 delalloc_lock = false;
5999         } else if (current->journal_info) {
6000                 flush = BTRFS_RESERVE_FLUSH_LIMIT;
6001         }
6002
6003         if (flush != BTRFS_RESERVE_NO_FLUSH &&
6004             btrfs_transaction_in_commit(fs_info))
6005                 schedule_timeout(1);
6006
6007         if (delalloc_lock)
6008                 mutex_lock(&inode->delalloc_mutex);
6009
6010         num_bytes = ALIGN(num_bytes, fs_info->sectorsize);
6011
6012         spin_lock(&inode->lock);
6013         nr_extents = count_max_extents(num_bytes);
6014         inode->outstanding_extents += nr_extents;
6015
6016         nr_extents = 0;
6017         if (inode->outstanding_extents > inode->reserved_extents)
6018                 nr_extents += inode->outstanding_extents -
6019                         inode->reserved_extents;
6020
6021         /* We always want to reserve a slot for updating the inode. */
6022         to_reserve = btrfs_calc_trans_metadata_size(fs_info, nr_extents + 1);
6023         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
6024         csum_bytes = inode->csum_bytes;
6025         spin_unlock(&inode->lock);
6026
6027         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)) {
6028                 ret = btrfs_qgroup_reserve_meta(root,
6029                                 nr_extents * fs_info->nodesize, true);
6030                 if (ret)
6031                         goto out_fail;
6032         }
6033
6034         ret = btrfs_block_rsv_add(root, block_rsv, to_reserve, flush);
6035         if (unlikely(ret)) {
6036                 btrfs_qgroup_free_meta(root,
6037                                        nr_extents * fs_info->nodesize);
6038                 goto out_fail;
6039         }
6040
6041         spin_lock(&inode->lock);
6042         if (test_and_set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
6043                              &inode->runtime_flags)) {
6044                 to_reserve -= btrfs_calc_trans_metadata_size(fs_info, 1);
6045                 release_extra = true;
6046         }
6047         inode->reserved_extents += nr_extents;
6048         spin_unlock(&inode->lock);
6049
6050         if (delalloc_lock)
6051                 mutex_unlock(&inode->delalloc_mutex);
6052
6053         if (to_reserve)
6054                 trace_btrfs_space_reservation(fs_info, "delalloc",
6055                                               btrfs_ino(inode), to_reserve, 1);
6056         if (release_extra)
6057                 btrfs_block_rsv_release(fs_info, block_rsv,
6058                                 btrfs_calc_trans_metadata_size(fs_info, 1));
6059         return 0;
6060
6061 out_fail:
6062         spin_lock(&inode->lock);
6063         dropped = drop_outstanding_extent(inode, num_bytes);
6064         /*
6065          * If the inodes csum_bytes is the same as the original
6066          * csum_bytes then we know we haven't raced with any free()ers
6067          * so we can just reduce our inodes csum bytes and carry on.
6068          */
6069         if (inode->csum_bytes == csum_bytes) {
6070                 calc_csum_metadata_size(inode, num_bytes, 0);
6071         } else {
6072                 u64 orig_csum_bytes = inode->csum_bytes;
6073                 u64 bytes;
6074
6075                 /*
6076                  * This is tricky, but first we need to figure out how much we
6077                  * freed from any free-ers that occurred during this
6078                  * reservation, so we reset ->csum_bytes to the csum_bytes
6079                  * before we dropped our lock, and then call the free for the
6080                  * number of bytes that were freed while we were trying our
6081                  * reservation.
6082                  */
6083                 bytes = csum_bytes - inode->csum_bytes;
6084                 inode->csum_bytes = csum_bytes;
6085                 to_free = calc_csum_metadata_size(inode, bytes, 0);
6086
6087
6088                 /*
6089                  * Now we need to see how much we would have freed had we not
6090                  * been making this reservation and our ->csum_bytes were not
6091                  * artificially inflated.
6092                  */
6093                 inode->csum_bytes = csum_bytes - num_bytes;
6094                 bytes = csum_bytes - orig_csum_bytes;
6095                 bytes = calc_csum_metadata_size(inode, bytes, 0);
6096
6097                 /*
6098                  * Now reset ->csum_bytes to what it should be.  If bytes is
6099                  * more than to_free then we would have freed more space had we
6100                  * not had an artificially high ->csum_bytes, so we need to free
6101                  * the remainder.  If bytes is the same or less then we don't
6102                  * need to do anything, the other free-ers did the correct
6103                  * thing.
6104                  */
6105                 inode->csum_bytes = orig_csum_bytes - num_bytes;
6106                 if (bytes > to_free)
6107                         to_free = bytes - to_free;
6108                 else
6109                         to_free = 0;
6110         }
6111         spin_unlock(&inode->lock);
6112         if (dropped)
6113                 to_free += btrfs_calc_trans_metadata_size(fs_info, dropped);
6114
6115         if (to_free) {
6116                 btrfs_block_rsv_release(fs_info, block_rsv, to_free);
6117                 trace_btrfs_space_reservation(fs_info, "delalloc",
6118                                               btrfs_ino(inode), to_free, 0);
6119         }
6120         if (delalloc_lock)
6121                 mutex_unlock(&inode->delalloc_mutex);
6122         return ret;
6123 }
6124
6125 /**
6126  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
6127  * @inode: the inode to release the reservation for
6128  * @num_bytes: the number of bytes we're releasing
6129  *
6130  * This will release the metadata reservation for an inode.  This can be called
6131  * once we complete IO for a given set of bytes to release their metadata
6132  * reservations.
6133  */
6134 void btrfs_delalloc_release_metadata(struct btrfs_inode *inode, u64 num_bytes)
6135 {
6136         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
6137         u64 to_free = 0;
6138         unsigned dropped;
6139
6140         num_bytes = ALIGN(num_bytes, fs_info->sectorsize);
6141         spin_lock(&inode->lock);
6142         dropped = drop_outstanding_extent(inode, num_bytes);
6143
6144         if (num_bytes)
6145                 to_free = calc_csum_metadata_size(inode, num_bytes, 0);
6146         spin_unlock(&inode->lock);
6147         if (dropped > 0)
6148                 to_free += btrfs_calc_trans_metadata_size(fs_info, dropped);
6149
6150         if (btrfs_is_testing(fs_info))
6151                 return;
6152
6153         trace_btrfs_space_reservation(fs_info, "delalloc", btrfs_ino(inode),
6154                                       to_free, 0);
6155
6156         btrfs_block_rsv_release(fs_info, &fs_info->delalloc_block_rsv, to_free);
6157 }
6158
6159 /**
6160  * btrfs_delalloc_reserve_space - reserve data and metadata space for
6161  * delalloc
6162  * @inode: inode we're writing to
6163  * @start: start range we are writing to
6164  * @len: how long the range we are writing to
6165  * @reserved: mandatory parameter, record actually reserved qgroup ranges of
6166  *            current reservation.
6167  *
6168  * This will do the following things
6169  *
6170  * o reserve space in data space info for num bytes
6171  *   and reserve precious corresponding qgroup space
6172  *   (Done in check_data_free_space)
6173  *
6174  * o reserve space for metadata space, based on the number of outstanding
6175  *   extents and how much csums will be needed
6176  *   also reserve metadata space in a per root over-reserve method.
6177  * o add to the inodes->delalloc_bytes
6178  * o add it to the fs_info's delalloc inodes list.
6179  *   (Above 3 all done in delalloc_reserve_metadata)
6180  *
6181  * Return 0 for success
6182  * Return <0 for error(-ENOSPC or -EQUOT)
6183  */
6184 int btrfs_delalloc_reserve_space(struct inode *inode,
6185                         struct extent_changeset **reserved, u64 start, u64 len)
6186 {
6187         int ret;
6188
6189         ret = btrfs_check_data_free_space(inode, reserved, start, len);
6190         if (ret < 0)
6191                 return ret;
6192         ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode), len);
6193         if (ret < 0)
6194                 btrfs_free_reserved_data_space(inode, *reserved, start, len);
6195         return ret;
6196 }
6197
6198 /**
6199  * btrfs_delalloc_release_space - release data and metadata space for delalloc
6200  * @inode: inode we're releasing space for
6201  * @start: start position of the space already reserved
6202  * @len: the len of the space already reserved
6203  *
6204  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
6205  * called in the case that we don't need the metadata AND data reservations
6206  * anymore.  So if there is an error or we insert an inline extent.
6207  *
6208  * This function will release the metadata space that was not used and will
6209  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
6210  * list if there are no delalloc bytes left.
6211  * Also it will handle the qgroup reserved space.
6212  */
6213 void btrfs_delalloc_release_space(struct inode *inode,
6214                         struct extent_changeset *reserved, u64 start, u64 len)
6215 {
6216         btrfs_delalloc_release_metadata(BTRFS_I(inode), len);
6217         btrfs_free_reserved_data_space(inode, reserved, start, len);
6218 }
6219
6220 static int update_block_group(struct btrfs_trans_handle *trans,
6221                               struct btrfs_fs_info *info, u64 bytenr,
6222                               u64 num_bytes, int alloc)
6223 {
6224         struct btrfs_block_group_cache *cache = NULL;
6225         u64 total = num_bytes;
6226         u64 old_val;
6227         u64 byte_in_group;
6228         int factor;
6229
6230         /* block accounting for super block */
6231         spin_lock(&info->delalloc_root_lock);
6232         old_val = btrfs_super_bytes_used(info->super_copy);
6233         if (alloc)
6234                 old_val += num_bytes;
6235         else
6236                 old_val -= num_bytes;
6237         btrfs_set_super_bytes_used(info->super_copy, old_val);
6238         spin_unlock(&info->delalloc_root_lock);
6239
6240         while (total) {
6241                 cache = btrfs_lookup_block_group(info, bytenr);
6242                 if (!cache)
6243                         return -ENOENT;
6244                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
6245                                     BTRFS_BLOCK_GROUP_RAID1 |
6246                                     BTRFS_BLOCK_GROUP_RAID10))
6247                         factor = 2;
6248                 else
6249                         factor = 1;
6250                 /*
6251                  * If this block group has free space cache written out, we
6252                  * need to make sure to load it if we are removing space.  This
6253                  * is because we need the unpinning stage to actually add the
6254                  * space back to the block group, otherwise we will leak space.
6255                  */
6256                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
6257                         cache_block_group(cache, 1);
6258
6259                 byte_in_group = bytenr - cache->key.objectid;
6260                 WARN_ON(byte_in_group > cache->key.offset);
6261
6262                 spin_lock(&cache->space_info->lock);
6263                 spin_lock(&cache->lock);
6264
6265                 if (btrfs_test_opt(info, SPACE_CACHE) &&
6266                     cache->disk_cache_state < BTRFS_DC_CLEAR)
6267                         cache->disk_cache_state = BTRFS_DC_CLEAR;
6268
6269                 old_val = btrfs_block_group_used(&cache->item);
6270                 num_bytes = min(total, cache->key.offset - byte_in_group);
6271                 if (alloc) {
6272                         old_val += num_bytes;
6273                         btrfs_set_block_group_used(&cache->item, old_val);
6274                         cache->reserved -= num_bytes;
6275                         cache->space_info->bytes_reserved -= num_bytes;
6276                         cache->space_info->bytes_used += num_bytes;
6277                         cache->space_info->disk_used += num_bytes * factor;
6278                         spin_unlock(&cache->lock);
6279                         spin_unlock(&cache->space_info->lock);
6280                 } else {
6281                         old_val -= num_bytes;
6282                         btrfs_set_block_group_used(&cache->item, old_val);
6283                         cache->pinned += num_bytes;
6284                         cache->space_info->bytes_pinned += num_bytes;
6285                         cache->space_info->bytes_used -= num_bytes;
6286                         cache->space_info->disk_used -= num_bytes * factor;
6287                         spin_unlock(&cache->lock);
6288                         spin_unlock(&cache->space_info->lock);
6289
6290                         trace_btrfs_space_reservation(info, "pinned",
6291                                                       cache->space_info->flags,
6292                                                       num_bytes, 1);
6293                         percpu_counter_add(&cache->space_info->total_bytes_pinned,
6294                                            num_bytes);
6295                         set_extent_dirty(info->pinned_extents,
6296                                          bytenr, bytenr + num_bytes - 1,
6297                                          GFP_NOFS | __GFP_NOFAIL);
6298                 }
6299
6300                 spin_lock(&trans->transaction->dirty_bgs_lock);
6301                 if (list_empty(&cache->dirty_list)) {
6302                         list_add_tail(&cache->dirty_list,
6303                                       &trans->transaction->dirty_bgs);
6304                                 trans->transaction->num_dirty_bgs++;
6305                         btrfs_get_block_group(cache);
6306                 }
6307                 spin_unlock(&trans->transaction->dirty_bgs_lock);
6308
6309                 /*
6310                  * No longer have used bytes in this block group, queue it for
6311                  * deletion. We do this after adding the block group to the
6312                  * dirty list to avoid races between cleaner kthread and space
6313                  * cache writeout.
6314                  */
6315                 if (!alloc && old_val == 0) {
6316                         spin_lock(&info->unused_bgs_lock);
6317                         if (list_empty(&cache->bg_list)) {
6318                                 btrfs_get_block_group(cache);
6319                                 list_add_tail(&cache->bg_list,
6320                                               &info->unused_bgs);
6321                         }
6322                         spin_unlock(&info->unused_bgs_lock);
6323                 }
6324
6325                 btrfs_put_block_group(cache);
6326                 total -= num_bytes;
6327                 bytenr += num_bytes;
6328         }
6329         return 0;
6330 }
6331
6332 static u64 first_logical_byte(struct btrfs_fs_info *fs_info, u64 search_start)
6333 {
6334         struct btrfs_block_group_cache *cache;
6335         u64 bytenr;
6336
6337         spin_lock(&fs_info->block_group_cache_lock);
6338         bytenr = fs_info->first_logical_byte;
6339         spin_unlock(&fs_info->block_group_cache_lock);
6340
6341         if (bytenr < (u64)-1)
6342                 return bytenr;
6343
6344         cache = btrfs_lookup_first_block_group(fs_info, search_start);
6345         if (!cache)
6346                 return 0;
6347
6348         bytenr = cache->key.objectid;
6349         btrfs_put_block_group(cache);
6350
6351         return bytenr;
6352 }
6353
6354 static int pin_down_extent(struct btrfs_fs_info *fs_info,
6355                            struct btrfs_block_group_cache *cache,
6356                            u64 bytenr, u64 num_bytes, int reserved)
6357 {
6358         spin_lock(&cache->space_info->lock);
6359         spin_lock(&cache->lock);
6360         cache->pinned += num_bytes;
6361         cache->space_info->bytes_pinned += num_bytes;
6362         if (reserved) {
6363                 cache->reserved -= num_bytes;
6364                 cache->space_info->bytes_reserved -= num_bytes;
6365         }
6366         spin_unlock(&cache->lock);
6367         spin_unlock(&cache->space_info->lock);
6368
6369         trace_btrfs_space_reservation(fs_info, "pinned",
6370                                       cache->space_info->flags, num_bytes, 1);
6371         percpu_counter_add(&cache->space_info->total_bytes_pinned, num_bytes);
6372         set_extent_dirty(fs_info->pinned_extents, bytenr,
6373                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
6374         return 0;
6375 }
6376
6377 /*
6378  * this function must be called within transaction
6379  */
6380 int btrfs_pin_extent(struct btrfs_fs_info *fs_info,
6381                      u64 bytenr, u64 num_bytes, int reserved)
6382 {
6383         struct btrfs_block_group_cache *cache;
6384
6385         cache = btrfs_lookup_block_group(fs_info, bytenr);
6386         BUG_ON(!cache); /* Logic error */
6387
6388         pin_down_extent(fs_info, cache, bytenr, num_bytes, reserved);
6389
6390         btrfs_put_block_group(cache);
6391         return 0;
6392 }
6393
6394 /*
6395  * this function must be called within transaction
6396  */
6397 int btrfs_pin_extent_for_log_replay(struct btrfs_fs_info *fs_info,
6398                                     u64 bytenr, u64 num_bytes)
6399 {
6400         struct btrfs_block_group_cache *cache;
6401         int ret;
6402
6403         cache = btrfs_lookup_block_group(fs_info, bytenr);
6404         if (!cache)
6405                 return -EINVAL;
6406
6407         /*
6408          * pull in the free space cache (if any) so that our pin
6409          * removes the free space from the cache.  We have load_only set
6410          * to one because the slow code to read in the free extents does check
6411          * the pinned extents.
6412          */
6413         cache_block_group(cache, 1);
6414
6415         pin_down_extent(fs_info, cache, bytenr, num_bytes, 0);
6416
6417         /* remove us from the free space cache (if we're there at all) */
6418         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
6419         btrfs_put_block_group(cache);
6420         return ret;
6421 }
6422
6423 static int __exclude_logged_extent(struct btrfs_fs_info *fs_info,
6424                                    u64 start, u64 num_bytes)
6425 {
6426         int ret;
6427         struct btrfs_block_group_cache *block_group;
6428         struct btrfs_caching_control *caching_ctl;
6429
6430         block_group = btrfs_lookup_block_group(fs_info, start);
6431         if (!block_group)
6432                 return -EINVAL;
6433
6434         cache_block_group(block_group, 0);
6435         caching_ctl = get_caching_control(block_group);
6436
6437         if (!caching_ctl) {
6438                 /* Logic error */
6439                 BUG_ON(!block_group_cache_done(block_group));
6440                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6441         } else {
6442                 mutex_lock(&caching_ctl->mutex);
6443
6444                 if (start >= caching_ctl->progress) {
6445                         ret = add_excluded_extent(fs_info, start, num_bytes);
6446                 } else if (start + num_bytes <= caching_ctl->progress) {
6447                         ret = btrfs_remove_free_space(block_group,
6448                                                       start, num_bytes);
6449                 } else {
6450                         num_bytes = caching_ctl->progress - start;
6451                         ret = btrfs_remove_free_space(block_group,
6452                                                       start, num_bytes);
6453                         if (ret)
6454                                 goto out_lock;
6455
6456                         num_bytes = (start + num_bytes) -
6457                                 caching_ctl->progress;
6458                         start = caching_ctl->progress;
6459                         ret = add_excluded_extent(fs_info, start, num_bytes);
6460                 }
6461 out_lock:
6462                 mutex_unlock(&caching_ctl->mutex);
6463                 put_caching_control(caching_ctl);
6464         }
6465         btrfs_put_block_group(block_group);
6466         return ret;
6467 }
6468
6469 int btrfs_exclude_logged_extents(struct btrfs_fs_info *fs_info,
6470                                  struct extent_buffer *eb)
6471 {
6472         struct btrfs_file_extent_item *item;
6473         struct btrfs_key key;
6474         int found_type;
6475         int i;
6476
6477         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS))
6478                 return 0;
6479
6480         for (i = 0; i < btrfs_header_nritems(eb); i++) {
6481                 btrfs_item_key_to_cpu(eb, &key, i);
6482                 if (key.type != BTRFS_EXTENT_DATA_KEY)
6483                         continue;
6484                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
6485                 found_type = btrfs_file_extent_type(eb, item);
6486                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
6487                         continue;
6488                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
6489                         continue;
6490                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
6491                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
6492                 __exclude_logged_extent(fs_info, key.objectid, key.offset);
6493         }
6494
6495         return 0;
6496 }
6497
6498 static void
6499 btrfs_inc_block_group_reservations(struct btrfs_block_group_cache *bg)
6500 {
6501         atomic_inc(&bg->reservations);
6502 }
6503
6504 void btrfs_dec_block_group_reservations(struct btrfs_fs_info *fs_info,
6505                                         const u64 start)
6506 {
6507         struct btrfs_block_group_cache *bg;
6508
6509         bg = btrfs_lookup_block_group(fs_info, start);
6510         ASSERT(bg);
6511         if (atomic_dec_and_test(&bg->reservations))
6512                 wake_up_atomic_t(&bg->reservations);
6513         btrfs_put_block_group(bg);
6514 }
6515
6516 static int btrfs_wait_bg_reservations_atomic_t(atomic_t *a)
6517 {
6518         schedule();
6519         return 0;
6520 }
6521
6522 void btrfs_wait_block_group_reservations(struct btrfs_block_group_cache *bg)
6523 {
6524         struct btrfs_space_info *space_info = bg->space_info;
6525
6526         ASSERT(bg->ro);
6527
6528         if (!(bg->flags & BTRFS_BLOCK_GROUP_DATA))
6529                 return;
6530
6531         /*
6532          * Our block group is read only but before we set it to read only,
6533          * some task might have had allocated an extent from it already, but it
6534          * has not yet created a respective ordered extent (and added it to a
6535          * root's list of ordered extents).
6536          * Therefore wait for any task currently allocating extents, since the
6537          * block group's reservations counter is incremented while a read lock
6538          * on the groups' semaphore is held and decremented after releasing
6539          * the read access on that semaphore and creating the ordered extent.
6540          */
6541         down_write(&space_info->groups_sem);
6542         up_write(&space_info->groups_sem);
6543
6544         wait_on_atomic_t(&bg->reservations,
6545                          btrfs_wait_bg_reservations_atomic_t,
6546                          TASK_UNINTERRUPTIBLE);
6547 }
6548
6549 /**
6550  * btrfs_add_reserved_bytes - update the block_group and space info counters
6551  * @cache:      The cache we are manipulating
6552  * @ram_bytes:  The number of bytes of file content, and will be same to
6553  *              @num_bytes except for the compress path.
6554  * @num_bytes:  The number of bytes in question
6555  * @delalloc:   The blocks are allocated for the delalloc write
6556  *
6557  * This is called by the allocator when it reserves space. If this is a
6558  * reservation and the block group has become read only we cannot make the
6559  * reservation and return -EAGAIN, otherwise this function always succeeds.
6560  */
6561 static int btrfs_add_reserved_bytes(struct btrfs_block_group_cache *cache,
6562                                     u64 ram_bytes, u64 num_bytes, int delalloc)
6563 {
6564         struct btrfs_space_info *space_info = cache->space_info;
6565         int ret = 0;
6566
6567         spin_lock(&space_info->lock);
6568         spin_lock(&cache->lock);
6569         if (cache->ro) {
6570                 ret = -EAGAIN;
6571         } else {
6572                 cache->reserved += num_bytes;
6573                 space_info->bytes_reserved += num_bytes;
6574
6575                 trace_btrfs_space_reservation(cache->fs_info,
6576                                 "space_info", space_info->flags,
6577                                 ram_bytes, 0);
6578                 space_info->bytes_may_use -= ram_bytes;
6579                 if (delalloc)
6580                         cache->delalloc_bytes += num_bytes;
6581         }
6582         spin_unlock(&cache->lock);
6583         spin_unlock(&space_info->lock);
6584         return ret;
6585 }
6586
6587 /**
6588  * btrfs_free_reserved_bytes - update the block_group and space info counters
6589  * @cache:      The cache we are manipulating
6590  * @num_bytes:  The number of bytes in question
6591  * @delalloc:   The blocks are allocated for the delalloc write
6592  *
6593  * This is called by somebody who is freeing space that was never actually used
6594  * on disk.  For example if you reserve some space for a new leaf in transaction
6595  * A and before transaction A commits you free that leaf, you call this with
6596  * reserve set to 0 in order to clear the reservation.
6597  */
6598
6599 static int btrfs_free_reserved_bytes(struct btrfs_block_group_cache *cache,
6600                                      u64 num_bytes, int delalloc)
6601 {
6602         struct btrfs_space_info *space_info = cache->space_info;
6603         int ret = 0;
6604
6605         spin_lock(&space_info->lock);
6606         spin_lock(&cache->lock);
6607         if (cache->ro)
6608                 space_info->bytes_readonly += num_bytes;
6609         cache->reserved -= num_bytes;
6610         space_info->bytes_reserved -= num_bytes;
6611
6612         if (delalloc)
6613                 cache->delalloc_bytes -= num_bytes;
6614         spin_unlock(&cache->lock);
6615         spin_unlock(&space_info->lock);
6616         return ret;
6617 }
6618 void btrfs_prepare_extent_commit(struct btrfs_fs_info *fs_info)
6619 {
6620         struct btrfs_caching_control *next;
6621         struct btrfs_caching_control *caching_ctl;
6622         struct btrfs_block_group_cache *cache;
6623
6624         down_write(&fs_info->commit_root_sem);
6625
6626         list_for_each_entry_safe(caching_ctl, next,
6627                                  &fs_info->caching_block_groups, list) {
6628                 cache = caching_ctl->block_group;
6629                 if (block_group_cache_done(cache)) {
6630                         cache->last_byte_to_unpin = (u64)-1;
6631                         list_del_init(&caching_ctl->list);
6632                         put_caching_control(caching_ctl);
6633                 } else {
6634                         cache->last_byte_to_unpin = caching_ctl->progress;
6635                 }
6636         }
6637
6638         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6639                 fs_info->pinned_extents = &fs_info->freed_extents[1];
6640         else
6641                 fs_info->pinned_extents = &fs_info->freed_extents[0];
6642
6643         up_write(&fs_info->commit_root_sem);
6644
6645         update_global_block_rsv(fs_info);
6646 }
6647
6648 /*
6649  * Returns the free cluster for the given space info and sets empty_cluster to
6650  * what it should be based on the mount options.
6651  */
6652 static struct btrfs_free_cluster *
6653 fetch_cluster_info(struct btrfs_fs_info *fs_info,
6654                    struct btrfs_space_info *space_info, u64 *empty_cluster)
6655 {
6656         struct btrfs_free_cluster *ret = NULL;
6657
6658         *empty_cluster = 0;
6659         if (btrfs_mixed_space_info(space_info))
6660                 return ret;
6661
6662         if (space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
6663                 ret = &fs_info->meta_alloc_cluster;
6664                 if (btrfs_test_opt(fs_info, SSD))
6665                         *empty_cluster = SZ_2M;
6666                 else
6667                         *empty_cluster = SZ_64K;
6668         } else if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) &&
6669                    btrfs_test_opt(fs_info, SSD_SPREAD)) {
6670                 *empty_cluster = SZ_2M;
6671                 ret = &fs_info->data_alloc_cluster;
6672         }
6673
6674         return ret;
6675 }
6676
6677 static int unpin_extent_range(struct btrfs_fs_info *fs_info,
6678                               u64 start, u64 end,
6679                               const bool return_free_space)
6680 {
6681         struct btrfs_block_group_cache *cache = NULL;
6682         struct btrfs_space_info *space_info;
6683         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
6684         struct btrfs_free_cluster *cluster = NULL;
6685         u64 len;
6686         u64 total_unpinned = 0;
6687         u64 empty_cluster = 0;
6688         bool readonly;
6689
6690         while (start <= end) {
6691                 readonly = false;
6692                 if (!cache ||
6693                     start >= cache->key.objectid + cache->key.offset) {
6694                         if (cache)
6695                                 btrfs_put_block_group(cache);
6696                         total_unpinned = 0;
6697                         cache = btrfs_lookup_block_group(fs_info, start);
6698                         BUG_ON(!cache); /* Logic error */
6699
6700                         cluster = fetch_cluster_info(fs_info,
6701                                                      cache->space_info,
6702                                                      &empty_cluster);
6703                         empty_cluster <<= 1;
6704                 }
6705
6706                 len = cache->key.objectid + cache->key.offset - start;
6707                 len = min(len, end + 1 - start);
6708
6709                 if (start < cache->last_byte_to_unpin) {
6710                         len = min(len, cache->last_byte_to_unpin - start);
6711                         if (return_free_space)
6712                                 btrfs_add_free_space(cache, start, len);
6713                 }
6714
6715                 start += len;
6716                 total_unpinned += len;
6717                 space_info = cache->space_info;
6718
6719                 /*
6720                  * If this space cluster has been marked as fragmented and we've
6721                  * unpinned enough in this block group to potentially allow a
6722                  * cluster to be created inside of it go ahead and clear the
6723                  * fragmented check.
6724                  */
6725                 if (cluster && cluster->fragmented &&
6726                     total_unpinned > empty_cluster) {
6727                         spin_lock(&cluster->lock);
6728                         cluster->fragmented = 0;
6729                         spin_unlock(&cluster->lock);
6730                 }
6731
6732                 spin_lock(&space_info->lock);
6733                 spin_lock(&cache->lock);
6734                 cache->pinned -= len;
6735                 space_info->bytes_pinned -= len;
6736
6737                 trace_btrfs_space_reservation(fs_info, "pinned",
6738                                               space_info->flags, len, 0);
6739                 space_info->max_extent_size = 0;
6740                 percpu_counter_add(&space_info->total_bytes_pinned, -len);
6741                 if (cache->ro) {
6742                         space_info->bytes_readonly += len;
6743                         readonly = true;
6744                 }
6745                 spin_unlock(&cache->lock);
6746                 if (!readonly && return_free_space &&
6747                     global_rsv->space_info == space_info) {
6748                         u64 to_add = len;
6749
6750                         spin_lock(&global_rsv->lock);
6751                         if (!global_rsv->full) {
6752                                 to_add = min(len, global_rsv->size -
6753                                              global_rsv->reserved);
6754                                 global_rsv->reserved += to_add;
6755                                 space_info->bytes_may_use += to_add;
6756                                 if (global_rsv->reserved >= global_rsv->size)
6757                                         global_rsv->full = 1;
6758                                 trace_btrfs_space_reservation(fs_info,
6759                                                               "space_info",
6760                                                               space_info->flags,
6761                                                               to_add, 1);
6762                                 len -= to_add;
6763                         }
6764                         spin_unlock(&global_rsv->lock);
6765                         /* Add to any tickets we may have */
6766                         if (len)
6767                                 space_info_add_new_bytes(fs_info, space_info,
6768                                                          len);
6769                 }
6770                 spin_unlock(&space_info->lock);
6771         }
6772
6773         if (cache)
6774                 btrfs_put_block_group(cache);
6775         return 0;
6776 }
6777
6778 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
6779                                struct btrfs_fs_info *fs_info)
6780 {
6781         struct btrfs_block_group_cache *block_group, *tmp;
6782         struct list_head *deleted_bgs;
6783         struct extent_io_tree *unpin;
6784         u64 start;
6785         u64 end;
6786         int ret;
6787
6788         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6789                 unpin = &fs_info->freed_extents[1];
6790         else
6791                 unpin = &fs_info->freed_extents[0];
6792
6793         while (!trans->aborted) {
6794                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
6795                 ret = find_first_extent_bit(unpin, 0, &start, &end,
6796                                             EXTENT_DIRTY, NULL);
6797                 if (ret) {
6798                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6799                         break;
6800                 }
6801
6802                 if (btrfs_test_opt(fs_info, DISCARD))
6803                         ret = btrfs_discard_extent(fs_info, start,
6804                                                    end + 1 - start, NULL);
6805
6806                 clear_extent_dirty(unpin, start, end);
6807                 unpin_extent_range(fs_info, start, end, true);
6808                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6809                 cond_resched();
6810         }
6811
6812         /*
6813          * Transaction is finished.  We don't need the lock anymore.  We
6814          * do need to clean up the block groups in case of a transaction
6815          * abort.
6816          */
6817         deleted_bgs = &trans->transaction->deleted_bgs;
6818         list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
6819                 u64 trimmed = 0;
6820
6821                 ret = -EROFS;
6822                 if (!trans->aborted)
6823                         ret = btrfs_discard_extent(fs_info,
6824                                                    block_group->key.objectid,
6825                                                    block_group->key.offset,
6826                                                    &trimmed);
6827
6828                 list_del_init(&block_group->bg_list);
6829                 btrfs_put_block_group_trimming(block_group);
6830                 btrfs_put_block_group(block_group);
6831
6832                 if (ret) {
6833                         const char *errstr = btrfs_decode_error(ret);
6834                         btrfs_warn(fs_info,
6835                            "discard failed while removing blockgroup: errno=%d %s",
6836                                    ret, errstr);
6837                 }
6838         }
6839
6840         return 0;
6841 }
6842
6843 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
6844                                 struct btrfs_fs_info *info,
6845                                 struct btrfs_delayed_ref_node *node, u64 parent,
6846                                 u64 root_objectid, u64 owner_objectid,
6847                                 u64 owner_offset, int refs_to_drop,
6848                                 struct btrfs_delayed_extent_op *extent_op)
6849 {
6850         struct btrfs_key key;
6851         struct btrfs_path *path;
6852         struct btrfs_root *extent_root = info->extent_root;
6853         struct extent_buffer *leaf;
6854         struct btrfs_extent_item *ei;
6855         struct btrfs_extent_inline_ref *iref;
6856         int ret;
6857         int is_data;
6858         int extent_slot = 0;
6859         int found_extent = 0;
6860         int num_to_del = 1;
6861         u32 item_size;
6862         u64 refs;
6863         u64 bytenr = node->bytenr;
6864         u64 num_bytes = node->num_bytes;
6865         int last_ref = 0;
6866         bool skinny_metadata = btrfs_fs_incompat(info, SKINNY_METADATA);
6867
6868         path = btrfs_alloc_path();
6869         if (!path)
6870                 return -ENOMEM;
6871
6872         path->reada = READA_FORWARD;
6873         path->leave_spinning = 1;
6874
6875         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
6876         BUG_ON(!is_data && refs_to_drop != 1);
6877
6878         if (is_data)
6879                 skinny_metadata = 0;
6880
6881         ret = lookup_extent_backref(trans, info, path, &iref,
6882                                     bytenr, num_bytes, parent,
6883                                     root_objectid, owner_objectid,
6884                                     owner_offset);
6885         if (ret == 0) {
6886                 extent_slot = path->slots[0];
6887                 while (extent_slot >= 0) {
6888                         btrfs_item_key_to_cpu(path->nodes[0], &key,
6889                                               extent_slot);
6890                         if (key.objectid != bytenr)
6891                                 break;
6892                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
6893                             key.offset == num_bytes) {
6894                                 found_extent = 1;
6895                                 break;
6896                         }
6897                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
6898                             key.offset == owner_objectid) {
6899                                 found_extent = 1;
6900                                 break;
6901                         }
6902                         if (path->slots[0] - extent_slot > 5)
6903                                 break;
6904                         extent_slot--;
6905                 }
6906 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6907                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
6908                 if (found_extent && item_size < sizeof(*ei))
6909                         found_extent = 0;
6910 #endif
6911                 if (!found_extent) {
6912                         BUG_ON(iref);
6913                         ret = remove_extent_backref(trans, info, path, NULL,
6914                                                     refs_to_drop,
6915                                                     is_data, &last_ref);
6916                         if (ret) {
6917                                 btrfs_abort_transaction(trans, ret);
6918                                 goto out;
6919                         }
6920                         btrfs_release_path(path);
6921                         path->leave_spinning = 1;
6922
6923                         key.objectid = bytenr;
6924                         key.type = BTRFS_EXTENT_ITEM_KEY;
6925                         key.offset = num_bytes;
6926
6927                         if (!is_data && skinny_metadata) {
6928                                 key.type = BTRFS_METADATA_ITEM_KEY;
6929                                 key.offset = owner_objectid;
6930                         }
6931
6932                         ret = btrfs_search_slot(trans, extent_root,
6933                                                 &key, path, -1, 1);
6934                         if (ret > 0 && skinny_metadata && path->slots[0]) {
6935                                 /*
6936                                  * Couldn't find our skinny metadata item,
6937                                  * see if we have ye olde extent item.
6938                                  */
6939                                 path->slots[0]--;
6940                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
6941                                                       path->slots[0]);
6942                                 if (key.objectid == bytenr &&
6943                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
6944                                     key.offset == num_bytes)
6945                                         ret = 0;
6946                         }
6947
6948                         if (ret > 0 && skinny_metadata) {
6949                                 skinny_metadata = false;
6950                                 key.objectid = bytenr;
6951                                 key.type = BTRFS_EXTENT_ITEM_KEY;
6952                                 key.offset = num_bytes;
6953                                 btrfs_release_path(path);
6954                                 ret = btrfs_search_slot(trans, extent_root,
6955                                                         &key, path, -1, 1);
6956                         }
6957
6958                         if (ret) {
6959                                 btrfs_err(info,
6960                                           "umm, got %d back from search, was looking for %llu",
6961                                           ret, bytenr);
6962                                 if (ret > 0)
6963                                         btrfs_print_leaf(path->nodes[0]);
6964                         }
6965                         if (ret < 0) {
6966                                 btrfs_abort_transaction(trans, ret);
6967                                 goto out;
6968                         }
6969                         extent_slot = path->slots[0];
6970                 }
6971         } else if (WARN_ON(ret == -ENOENT)) {
6972                 btrfs_print_leaf(path->nodes[0]);
6973                 btrfs_err(info,
6974                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
6975                         bytenr, parent, root_objectid, owner_objectid,
6976                         owner_offset);
6977                 btrfs_abort_transaction(trans, ret);
6978                 goto out;
6979         } else {
6980                 btrfs_abort_transaction(trans, ret);
6981                 goto out;
6982         }
6983
6984         leaf = path->nodes[0];
6985         item_size = btrfs_item_size_nr(leaf, extent_slot);
6986 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6987         if (item_size < sizeof(*ei)) {
6988                 BUG_ON(found_extent || extent_slot != path->slots[0]);
6989                 ret = convert_extent_item_v0(trans, info, path, owner_objectid,
6990                                              0);
6991                 if (ret < 0) {
6992                         btrfs_abort_transaction(trans, ret);
6993                         goto out;
6994                 }
6995
6996                 btrfs_release_path(path);
6997                 path->leave_spinning = 1;
6998
6999                 key.objectid = bytenr;
7000                 key.type = BTRFS_EXTENT_ITEM_KEY;
7001                 key.offset = num_bytes;
7002
7003                 ret = btrfs_search_slot(trans, extent_root, &key, path,
7004                                         -1, 1);
7005                 if (ret) {
7006                         btrfs_err(info,
7007                                   "umm, got %d back from search, was looking for %llu",
7008                                 ret, bytenr);
7009                         btrfs_print_leaf(path->nodes[0]);
7010                 }
7011                 if (ret < 0) {
7012                         btrfs_abort_transaction(trans, ret);
7013                         goto out;
7014                 }
7015
7016                 extent_slot = path->slots[0];
7017                 leaf = path->nodes[0];
7018                 item_size = btrfs_item_size_nr(leaf, extent_slot);
7019         }
7020 #endif
7021         BUG_ON(item_size < sizeof(*ei));
7022         ei = btrfs_item_ptr(leaf, extent_slot,
7023                             struct btrfs_extent_item);
7024         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
7025             key.type == BTRFS_EXTENT_ITEM_KEY) {
7026                 struct btrfs_tree_block_info *bi;
7027                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
7028                 bi = (struct btrfs_tree_block_info *)(ei + 1);
7029                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
7030         }
7031
7032         refs = btrfs_extent_refs(leaf, ei);
7033         if (refs < refs_to_drop) {
7034                 btrfs_err(info,
7035                           "trying to drop %d refs but we only have %Lu for bytenr %Lu",
7036                           refs_to_drop, refs, bytenr);
7037                 ret = -EINVAL;
7038                 btrfs_abort_transaction(trans, ret);
7039                 goto out;
7040         }
7041         refs -= refs_to_drop;
7042
7043         if (refs > 0) {
7044                 if (extent_op)
7045                         __run_delayed_extent_op(extent_op, leaf, ei);
7046                 /*
7047                  * In the case of inline back ref, reference count will
7048                  * be updated by remove_extent_backref
7049                  */
7050                 if (iref) {
7051                         BUG_ON(!found_extent);
7052                 } else {
7053                         btrfs_set_extent_refs(leaf, ei, refs);
7054                         btrfs_mark_buffer_dirty(leaf);
7055                 }
7056                 if (found_extent) {
7057                         ret = remove_extent_backref(trans, info, path,
7058                                                     iref, refs_to_drop,
7059                                                     is_data, &last_ref);
7060                         if (ret) {
7061                                 btrfs_abort_transaction(trans, ret);
7062                                 goto out;
7063                         }
7064                 }
7065         } else {
7066                 if (found_extent) {
7067                         BUG_ON(is_data && refs_to_drop !=
7068                                extent_data_ref_count(path, iref));
7069                         if (iref) {
7070                                 BUG_ON(path->slots[0] != extent_slot);
7071                         } else {
7072                                 BUG_ON(path->slots[0] != extent_slot + 1);
7073                                 path->slots[0] = extent_slot;
7074                                 num_to_del = 2;
7075                         }
7076                 }
7077
7078                 last_ref = 1;
7079                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
7080                                       num_to_del);
7081                 if (ret) {
7082                         btrfs_abort_transaction(trans, ret);
7083                         goto out;
7084                 }
7085                 btrfs_release_path(path);
7086
7087                 if (is_data) {
7088                         ret = btrfs_del_csums(trans, info, bytenr, num_bytes);
7089                         if (ret) {
7090                                 btrfs_abort_transaction(trans, ret);
7091                                 goto out;
7092                         }
7093                 }
7094
7095                 ret = add_to_free_space_tree(trans, info, bytenr, num_bytes);
7096                 if (ret) {
7097                         btrfs_abort_transaction(trans, ret);
7098                         goto out;
7099                 }
7100
7101                 ret = update_block_group(trans, info, bytenr, num_bytes, 0);
7102                 if (ret) {
7103                         btrfs_abort_transaction(trans, ret);
7104                         goto out;
7105                 }
7106         }
7107         btrfs_release_path(path);
7108
7109 out:
7110         btrfs_free_path(path);
7111         return ret;
7112 }
7113
7114 /*
7115  * when we free an block, it is possible (and likely) that we free the last
7116  * delayed ref for that extent as well.  This searches the delayed ref tree for
7117  * a given extent, and if there are no other delayed refs to be processed, it
7118  * removes it from the tree.
7119  */
7120 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
7121                                       u64 bytenr)
7122 {
7123         struct btrfs_delayed_ref_head *head;
7124         struct btrfs_delayed_ref_root *delayed_refs;
7125         int ret = 0;
7126
7127         delayed_refs = &trans->transaction->delayed_refs;
7128         spin_lock(&delayed_refs->lock);
7129         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
7130         if (!head)
7131                 goto out_delayed_unlock;
7132
7133         spin_lock(&head->lock);
7134         if (!list_empty(&head->ref_list))
7135                 goto out;
7136
7137         if (head->extent_op) {
7138                 if (!head->must_insert_reserved)
7139                         goto out;
7140                 btrfs_free_delayed_extent_op(head->extent_op);
7141                 head->extent_op = NULL;
7142         }
7143
7144         /*
7145          * waiting for the lock here would deadlock.  If someone else has it
7146          * locked they are already in the process of dropping it anyway
7147          */
7148         if (!mutex_trylock(&head->mutex))
7149                 goto out;
7150
7151         /*
7152          * at this point we have a head with no other entries.  Go
7153          * ahead and process it.
7154          */
7155         head->node.in_tree = 0;
7156         rb_erase(&head->href_node, &delayed_refs->href_root);
7157
7158         atomic_dec(&delayed_refs->num_entries);
7159
7160         /*
7161          * we don't take a ref on the node because we're removing it from the
7162          * tree, so we just steal the ref the tree was holding.
7163          */
7164         delayed_refs->num_heads--;
7165         if (head->processing == 0)
7166                 delayed_refs->num_heads_ready--;
7167         head->processing = 0;
7168         spin_unlock(&head->lock);
7169         spin_unlock(&delayed_refs->lock);
7170
7171         BUG_ON(head->extent_op);
7172         if (head->must_insert_reserved)
7173                 ret = 1;
7174
7175         mutex_unlock(&head->mutex);
7176         btrfs_put_delayed_ref(&head->node);
7177         return ret;
7178 out:
7179         spin_unlock(&head->lock);
7180
7181 out_delayed_unlock:
7182         spin_unlock(&delayed_refs->lock);
7183         return 0;
7184 }
7185
7186 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
7187                            struct btrfs_root *root,
7188                            struct extent_buffer *buf,
7189                            u64 parent, int last_ref)
7190 {
7191         struct btrfs_fs_info *fs_info = root->fs_info;
7192         int pin = 1;
7193         int ret;
7194
7195         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7196                 int old_ref_mod, new_ref_mod;
7197
7198                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, buf->start,
7199                                                  buf->len, parent,
7200                                                  root->root_key.objectid,
7201                                                  btrfs_header_level(buf),
7202                                                  BTRFS_DROP_DELAYED_REF, NULL,
7203                                                  &old_ref_mod, &new_ref_mod);
7204                 BUG_ON(ret); /* -ENOMEM */
7205                 pin = old_ref_mod >= 0 && new_ref_mod < 0;
7206         }
7207
7208         if (last_ref && btrfs_header_generation(buf) == trans->transid) {
7209                 struct btrfs_block_group_cache *cache;
7210
7211                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7212                         ret = check_ref_cleanup(trans, buf->start);
7213                         if (!ret)
7214                                 goto out;
7215                 }
7216
7217                 pin = 0;
7218                 cache = btrfs_lookup_block_group(fs_info, buf->start);
7219
7220                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
7221                         pin_down_extent(fs_info, cache, buf->start,
7222                                         buf->len, 1);
7223                         btrfs_put_block_group(cache);
7224                         goto out;
7225                 }
7226
7227                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
7228
7229                 btrfs_add_free_space(cache, buf->start, buf->len);
7230                 btrfs_free_reserved_bytes(cache, buf->len, 0);
7231                 btrfs_put_block_group(cache);
7232                 trace_btrfs_reserved_extent_free(fs_info, buf->start, buf->len);
7233         }
7234 out:
7235         if (pin)
7236                 add_pinned_bytes(fs_info, buf->len, btrfs_header_level(buf),
7237                                  root->root_key.objectid);
7238
7239         if (last_ref) {
7240                 /*
7241                  * Deleting the buffer, clear the corrupt flag since it doesn't
7242                  * matter anymore.
7243                  */
7244                 clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
7245         }
7246 }
7247
7248 /* Can return -ENOMEM */
7249 int btrfs_free_extent(struct btrfs_trans_handle *trans,
7250                       struct btrfs_fs_info *fs_info,
7251                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
7252                       u64 owner, u64 offset)
7253 {
7254         int old_ref_mod, new_ref_mod;
7255         int ret;
7256
7257         if (btrfs_is_testing(fs_info))
7258                 return 0;
7259
7260
7261         /*
7262          * tree log blocks never actually go into the extent allocation
7263          * tree, just update pinning info and exit early.
7264          */
7265         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
7266                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
7267                 /* unlocks the pinned mutex */
7268                 btrfs_pin_extent(fs_info, bytenr, num_bytes, 1);
7269                 old_ref_mod = new_ref_mod = 0;
7270                 ret = 0;
7271         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
7272                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
7273                                                  num_bytes, parent,
7274                                                  root_objectid, (int)owner,
7275                                                  BTRFS_DROP_DELAYED_REF, NULL,
7276                                                  &old_ref_mod, &new_ref_mod);
7277         } else {
7278                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
7279                                                  num_bytes, parent,
7280                                                  root_objectid, owner, offset,
7281                                                  0, BTRFS_DROP_DELAYED_REF,
7282                                                  &old_ref_mod, &new_ref_mod);
7283         }
7284
7285         if (ret == 0 && old_ref_mod >= 0 && new_ref_mod < 0)
7286                 add_pinned_bytes(fs_info, num_bytes, owner, root_objectid);
7287
7288         return ret;
7289 }
7290
7291 /*
7292  * when we wait for progress in the block group caching, its because
7293  * our allocation attempt failed at least once.  So, we must sleep
7294  * and let some progress happen before we try again.
7295  *
7296  * This function will sleep at least once waiting for new free space to
7297  * show up, and then it will check the block group free space numbers
7298  * for our min num_bytes.  Another option is to have it go ahead
7299  * and look in the rbtree for a free extent of a given size, but this
7300  * is a good start.
7301  *
7302  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
7303  * any of the information in this block group.
7304  */
7305 static noinline void
7306 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
7307                                 u64 num_bytes)
7308 {
7309         struct btrfs_caching_control *caching_ctl;
7310
7311         caching_ctl = get_caching_control(cache);
7312         if (!caching_ctl)
7313                 return;
7314
7315         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
7316                    (cache->free_space_ctl->free_space >= num_bytes));
7317
7318         put_caching_control(caching_ctl);
7319 }
7320
7321 static noinline int
7322 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
7323 {
7324         struct btrfs_caching_control *caching_ctl;
7325         int ret = 0;
7326
7327         caching_ctl = get_caching_control(cache);
7328         if (!caching_ctl)
7329                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
7330
7331         wait_event(caching_ctl->wait, block_group_cache_done(cache));
7332         if (cache->cached == BTRFS_CACHE_ERROR)
7333                 ret = -EIO;
7334         put_caching_control(caching_ctl);
7335         return ret;
7336 }
7337
7338 int __get_raid_index(u64 flags)
7339 {
7340         if (flags & BTRFS_BLOCK_GROUP_RAID10)
7341                 return BTRFS_RAID_RAID10;
7342         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
7343                 return BTRFS_RAID_RAID1;
7344         else if (flags & BTRFS_BLOCK_GROUP_DUP)
7345                 return BTRFS_RAID_DUP;
7346         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
7347                 return BTRFS_RAID_RAID0;
7348         else if (flags & BTRFS_BLOCK_GROUP_RAID5)
7349                 return BTRFS_RAID_RAID5;
7350         else if (flags & BTRFS_BLOCK_GROUP_RAID6)
7351                 return BTRFS_RAID_RAID6;
7352
7353         return BTRFS_RAID_SINGLE; /* BTRFS_BLOCK_GROUP_SINGLE */
7354 }
7355
7356 int get_block_group_index(struct btrfs_block_group_cache *cache)
7357 {
7358         return __get_raid_index(cache->flags);
7359 }
7360
7361 static const char *btrfs_raid_type_names[BTRFS_NR_RAID_TYPES] = {
7362         [BTRFS_RAID_RAID10]     = "raid10",
7363         [BTRFS_RAID_RAID1]      = "raid1",
7364         [BTRFS_RAID_DUP]        = "dup",
7365         [BTRFS_RAID_RAID0]      = "raid0",
7366         [BTRFS_RAID_SINGLE]     = "single",
7367         [BTRFS_RAID_RAID5]      = "raid5",
7368         [BTRFS_RAID_RAID6]      = "raid6",
7369 };
7370
7371 static const char *get_raid_name(enum btrfs_raid_types type)
7372 {
7373         if (type >= BTRFS_NR_RAID_TYPES)
7374                 return NULL;
7375
7376         return btrfs_raid_type_names[type];
7377 }
7378
7379 enum btrfs_loop_type {
7380         LOOP_CACHING_NOWAIT = 0,
7381         LOOP_CACHING_WAIT = 1,
7382         LOOP_ALLOC_CHUNK = 2,
7383         LOOP_NO_EMPTY_SIZE = 3,
7384 };
7385
7386 static inline void
7387 btrfs_lock_block_group(struct btrfs_block_group_cache *cache,
7388                        int delalloc)
7389 {
7390         if (delalloc)
7391                 down_read(&cache->data_rwsem);
7392 }
7393
7394 static inline void
7395 btrfs_grab_block_group(struct btrfs_block_group_cache *cache,
7396                        int delalloc)
7397 {
7398         btrfs_get_block_group(cache);
7399         if (delalloc)
7400                 down_read(&cache->data_rwsem);
7401 }
7402
7403 static struct btrfs_block_group_cache *
7404 btrfs_lock_cluster(struct btrfs_block_group_cache *block_group,
7405                    struct btrfs_free_cluster *cluster,
7406                    int delalloc)
7407 {
7408         struct btrfs_block_group_cache *used_bg = NULL;
7409
7410         spin_lock(&cluster->refill_lock);
7411         while (1) {
7412                 used_bg = cluster->block_group;
7413                 if (!used_bg)
7414                         return NULL;
7415
7416                 if (used_bg == block_group)
7417                         return used_bg;
7418
7419                 btrfs_get_block_group(used_bg);
7420
7421                 if (!delalloc)
7422                         return used_bg;
7423
7424                 if (down_read_trylock(&used_bg->data_rwsem))
7425                         return used_bg;
7426
7427                 spin_unlock(&cluster->refill_lock);
7428
7429                 /* We should only have one-level nested. */
7430                 down_read_nested(&used_bg->data_rwsem, SINGLE_DEPTH_NESTING);
7431
7432                 spin_lock(&cluster->refill_lock);
7433                 if (used_bg == cluster->block_group)
7434                         return used_bg;
7435
7436                 up_read(&used_bg->data_rwsem);
7437                 btrfs_put_block_group(used_bg);
7438         }
7439 }
7440
7441 static inline void
7442 btrfs_release_block_group(struct btrfs_block_group_cache *cache,
7443                          int delalloc)
7444 {
7445         if (delalloc)
7446                 up_read(&cache->data_rwsem);
7447         btrfs_put_block_group(cache);
7448 }
7449
7450 /*
7451  * walks the btree of allocated extents and find a hole of a given size.
7452  * The key ins is changed to record the hole:
7453  * ins->objectid == start position
7454  * ins->flags = BTRFS_EXTENT_ITEM_KEY
7455  * ins->offset == the size of the hole.
7456  * Any available blocks before search_start are skipped.
7457  *
7458  * If there is no suitable free space, we will record the max size of
7459  * the free space extent currently.
7460  */
7461 static noinline int find_free_extent(struct btrfs_fs_info *fs_info,
7462                                 u64 ram_bytes, u64 num_bytes, u64 empty_size,
7463                                 u64 hint_byte, struct btrfs_key *ins,
7464                                 u64 flags, int delalloc)
7465 {
7466         int ret = 0;
7467         struct btrfs_root *root = fs_info->extent_root;
7468         struct btrfs_free_cluster *last_ptr = NULL;
7469         struct btrfs_block_group_cache *block_group = NULL;
7470         u64 search_start = 0;
7471         u64 max_extent_size = 0;
7472         u64 empty_cluster = 0;
7473         struct btrfs_space_info *space_info;
7474         int loop = 0;
7475         int index = __get_raid_index(flags);
7476         bool failed_cluster_refill = false;
7477         bool failed_alloc = false;
7478         bool use_cluster = true;
7479         bool have_caching_bg = false;
7480         bool orig_have_caching_bg = false;
7481         bool full_search = false;
7482
7483         WARN_ON(num_bytes < fs_info->sectorsize);
7484         ins->type = BTRFS_EXTENT_ITEM_KEY;
7485         ins->objectid = 0;
7486         ins->offset = 0;
7487
7488         trace_find_free_extent(fs_info, num_bytes, empty_size, flags);
7489
7490         space_info = __find_space_info(fs_info, flags);
7491         if (!space_info) {
7492                 btrfs_err(fs_info, "No space info for %llu", flags);
7493                 return -ENOSPC;
7494         }
7495
7496         /*
7497          * If our free space is heavily fragmented we may not be able to make
7498          * big contiguous allocations, so instead of doing the expensive search
7499          * for free space, simply return ENOSPC with our max_extent_size so we
7500          * can go ahead and search for a more manageable chunk.
7501          *
7502          * If our max_extent_size is large enough for our allocation simply
7503          * disable clustering since we will likely not be able to find enough
7504          * space to create a cluster and induce latency trying.
7505          */
7506         if (unlikely(space_info->max_extent_size)) {
7507                 spin_lock(&space_info->lock);
7508                 if (space_info->max_extent_size &&
7509                     num_bytes > space_info->max_extent_size) {
7510                         ins->offset = space_info->max_extent_size;
7511                         spin_unlock(&space_info->lock);
7512                         return -ENOSPC;
7513                 } else if (space_info->max_extent_size) {
7514                         use_cluster = false;
7515                 }
7516                 spin_unlock(&space_info->lock);
7517         }
7518
7519         last_ptr = fetch_cluster_info(fs_info, space_info, &empty_cluster);
7520         if (last_ptr) {
7521                 spin_lock(&last_ptr->lock);
7522                 if (last_ptr->block_group)
7523                         hint_byte = last_ptr->window_start;
7524                 if (last_ptr->fragmented) {
7525                         /*
7526                          * We still set window_start so we can keep track of the
7527                          * last place we found an allocation to try and save
7528                          * some time.
7529                          */
7530                         hint_byte = last_ptr->window_start;
7531                         use_cluster = false;
7532                 }
7533                 spin_unlock(&last_ptr->lock);
7534         }
7535
7536         search_start = max(search_start, first_logical_byte(fs_info, 0));
7537         search_start = max(search_start, hint_byte);
7538         if (search_start == hint_byte) {
7539                 block_group = btrfs_lookup_block_group(fs_info, search_start);
7540                 /*
7541                  * we don't want to use the block group if it doesn't match our
7542                  * allocation bits, or if its not cached.
7543                  *
7544                  * However if we are re-searching with an ideal block group
7545                  * picked out then we don't care that the block group is cached.
7546                  */
7547                 if (block_group && block_group_bits(block_group, flags) &&
7548                     block_group->cached != BTRFS_CACHE_NO) {
7549                         down_read(&space_info->groups_sem);
7550                         if (list_empty(&block_group->list) ||
7551                             block_group->ro) {
7552                                 /*
7553                                  * someone is removing this block group,
7554                                  * we can't jump into the have_block_group
7555                                  * target because our list pointers are not
7556                                  * valid
7557                                  */
7558                                 btrfs_put_block_group(block_group);
7559                                 up_read(&space_info->groups_sem);
7560                         } else {
7561                                 index = get_block_group_index(block_group);
7562                                 btrfs_lock_block_group(block_group, delalloc);
7563                                 goto have_block_group;
7564                         }
7565                 } else if (block_group) {
7566                         btrfs_put_block_group(block_group);
7567                 }
7568         }
7569 search:
7570         have_caching_bg = false;
7571         if (index == 0 || index == __get_raid_index(flags))
7572                 full_search = true;
7573         down_read(&space_info->groups_sem);
7574         list_for_each_entry(block_group, &space_info->block_groups[index],
7575                             list) {
7576                 u64 offset;
7577                 int cached;
7578
7579                 /* If the block group is read-only, we can skip it entirely. */
7580                 if (unlikely(block_group->ro))
7581                         continue;
7582
7583                 btrfs_grab_block_group(block_group, delalloc);
7584                 search_start = block_group->key.objectid;
7585
7586                 /*
7587                  * this can happen if we end up cycling through all the
7588                  * raid types, but we want to make sure we only allocate
7589                  * for the proper type.
7590                  */
7591                 if (!block_group_bits(block_group, flags)) {
7592                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
7593                                 BTRFS_BLOCK_GROUP_RAID1 |
7594                                 BTRFS_BLOCK_GROUP_RAID5 |
7595                                 BTRFS_BLOCK_GROUP_RAID6 |
7596                                 BTRFS_BLOCK_GROUP_RAID10;
7597
7598                         /*
7599                          * if they asked for extra copies and this block group
7600                          * doesn't provide them, bail.  This does allow us to
7601                          * fill raid0 from raid1.
7602                          */
7603                         if ((flags & extra) && !(block_group->flags & extra))
7604                                 goto loop;
7605                 }
7606
7607 have_block_group:
7608                 cached = block_group_cache_done(block_group);
7609                 if (unlikely(!cached)) {
7610                         have_caching_bg = true;
7611                         ret = cache_block_group(block_group, 0);
7612                         BUG_ON(ret < 0);
7613                         ret = 0;
7614                 }
7615
7616                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
7617                         goto loop;
7618
7619                 /*
7620                  * Ok we want to try and use the cluster allocator, so
7621                  * lets look there
7622                  */
7623                 if (last_ptr && use_cluster) {
7624                         struct btrfs_block_group_cache *used_block_group;
7625                         unsigned long aligned_cluster;
7626                         /*
7627                          * the refill lock keeps out other
7628                          * people trying to start a new cluster
7629                          */
7630                         used_block_group = btrfs_lock_cluster(block_group,
7631                                                               last_ptr,
7632                                                               delalloc);
7633                         if (!used_block_group)
7634                                 goto refill_cluster;
7635
7636                         if (used_block_group != block_group &&
7637                             (used_block_group->ro ||
7638                              !block_group_bits(used_block_group, flags)))
7639                                 goto release_cluster;
7640
7641                         offset = btrfs_alloc_from_cluster(used_block_group,
7642                                                 last_ptr,
7643                                                 num_bytes,
7644                                                 used_block_group->key.objectid,
7645                                                 &max_extent_size);
7646                         if (offset) {
7647                                 /* we have a block, we're done */
7648                                 spin_unlock(&last_ptr->refill_lock);
7649                                 trace_btrfs_reserve_extent_cluster(fs_info,
7650                                                 used_block_group,
7651                                                 search_start, num_bytes);
7652                                 if (used_block_group != block_group) {
7653                                         btrfs_release_block_group(block_group,
7654                                                                   delalloc);
7655                                         block_group = used_block_group;
7656                                 }
7657                                 goto checks;
7658                         }
7659
7660                         WARN_ON(last_ptr->block_group != used_block_group);
7661 release_cluster:
7662                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
7663                          * set up a new clusters, so lets just skip it
7664                          * and let the allocator find whatever block
7665                          * it can find.  If we reach this point, we
7666                          * will have tried the cluster allocator
7667                          * plenty of times and not have found
7668                          * anything, so we are likely way too
7669                          * fragmented for the clustering stuff to find
7670                          * anything.
7671                          *
7672                          * However, if the cluster is taken from the
7673                          * current block group, release the cluster
7674                          * first, so that we stand a better chance of
7675                          * succeeding in the unclustered
7676                          * allocation.  */
7677                         if (loop >= LOOP_NO_EMPTY_SIZE &&
7678                             used_block_group != block_group) {
7679                                 spin_unlock(&last_ptr->refill_lock);
7680                                 btrfs_release_block_group(used_block_group,
7681                                                           delalloc);
7682                                 goto unclustered_alloc;
7683                         }
7684
7685                         /*
7686                          * this cluster didn't work out, free it and
7687                          * start over
7688                          */
7689                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7690
7691                         if (used_block_group != block_group)
7692                                 btrfs_release_block_group(used_block_group,
7693                                                           delalloc);
7694 refill_cluster:
7695                         if (loop >= LOOP_NO_EMPTY_SIZE) {
7696                                 spin_unlock(&last_ptr->refill_lock);
7697                                 goto unclustered_alloc;
7698                         }
7699
7700                         aligned_cluster = max_t(unsigned long,
7701                                                 empty_cluster + empty_size,
7702                                               block_group->full_stripe_len);
7703
7704                         /* allocate a cluster in this block group */
7705                         ret = btrfs_find_space_cluster(fs_info, block_group,
7706                                                        last_ptr, search_start,
7707                                                        num_bytes,
7708                                                        aligned_cluster);
7709                         if (ret == 0) {
7710                                 /*
7711                                  * now pull our allocation out of this
7712                                  * cluster
7713                                  */
7714                                 offset = btrfs_alloc_from_cluster(block_group,
7715                                                         last_ptr,
7716                                                         num_bytes,
7717                                                         search_start,
7718                                                         &max_extent_size);
7719                                 if (offset) {
7720                                         /* we found one, proceed */
7721                                         spin_unlock(&last_ptr->refill_lock);
7722                                         trace_btrfs_reserve_extent_cluster(fs_info,
7723                                                 block_group, search_start,
7724                                                 num_bytes);
7725                                         goto checks;
7726                                 }
7727                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
7728                                    && !failed_cluster_refill) {
7729                                 spin_unlock(&last_ptr->refill_lock);
7730
7731                                 failed_cluster_refill = true;
7732                                 wait_block_group_cache_progress(block_group,
7733                                        num_bytes + empty_cluster + empty_size);
7734                                 goto have_block_group;
7735                         }
7736
7737                         /*
7738                          * at this point we either didn't find a cluster
7739                          * or we weren't able to allocate a block from our
7740                          * cluster.  Free the cluster we've been trying
7741                          * to use, and go to the next block group
7742                          */
7743                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7744                         spin_unlock(&last_ptr->refill_lock);
7745                         goto loop;
7746                 }
7747
7748 unclustered_alloc:
7749                 /*
7750                  * We are doing an unclustered alloc, set the fragmented flag so
7751                  * we don't bother trying to setup a cluster again until we get
7752                  * more space.
7753                  */
7754                 if (unlikely(last_ptr)) {
7755                         spin_lock(&last_ptr->lock);
7756                         last_ptr->fragmented = 1;
7757                         spin_unlock(&last_ptr->lock);
7758                 }
7759                 if (cached) {
7760                         struct btrfs_free_space_ctl *ctl =
7761                                 block_group->free_space_ctl;
7762
7763                         spin_lock(&ctl->tree_lock);
7764                         if (ctl->free_space <
7765                             num_bytes + empty_cluster + empty_size) {
7766                                 if (ctl->free_space > max_extent_size)
7767                                         max_extent_size = ctl->free_space;
7768                                 spin_unlock(&ctl->tree_lock);
7769                                 goto loop;
7770                         }
7771                         spin_unlock(&ctl->tree_lock);
7772                 }
7773
7774                 offset = btrfs_find_space_for_alloc(block_group, search_start,
7775                                                     num_bytes, empty_size,
7776                                                     &max_extent_size);
7777                 /*
7778                  * If we didn't find a chunk, and we haven't failed on this
7779                  * block group before, and this block group is in the middle of
7780                  * caching and we are ok with waiting, then go ahead and wait
7781                  * for progress to be made, and set failed_alloc to true.
7782                  *
7783                  * If failed_alloc is true then we've already waited on this
7784                  * block group once and should move on to the next block group.
7785                  */
7786                 if (!offset && !failed_alloc && !cached &&
7787                     loop > LOOP_CACHING_NOWAIT) {
7788                         wait_block_group_cache_progress(block_group,
7789                                                 num_bytes + empty_size);
7790                         failed_alloc = true;
7791                         goto have_block_group;
7792                 } else if (!offset) {
7793                         goto loop;
7794                 }
7795 checks:
7796                 search_start = ALIGN(offset, fs_info->stripesize);
7797
7798                 /* move on to the next group */
7799                 if (search_start + num_bytes >
7800                     block_group->key.objectid + block_group->key.offset) {
7801                         btrfs_add_free_space(block_group, offset, num_bytes);
7802                         goto loop;
7803                 }
7804
7805                 if (offset < search_start)
7806                         btrfs_add_free_space(block_group, offset,
7807                                              search_start - offset);
7808                 BUG_ON(offset > search_start);
7809
7810                 ret = btrfs_add_reserved_bytes(block_group, ram_bytes,
7811                                 num_bytes, delalloc);
7812                 if (ret == -EAGAIN) {
7813                         btrfs_add_free_space(block_group, offset, num_bytes);
7814                         goto loop;
7815                 }
7816                 btrfs_inc_block_group_reservations(block_group);
7817
7818                 /* we are all good, lets return */
7819                 ins->objectid = search_start;
7820                 ins->offset = num_bytes;
7821
7822                 trace_btrfs_reserve_extent(fs_info, block_group,
7823                                            search_start, num_bytes);
7824                 btrfs_release_block_group(block_group, delalloc);
7825                 break;
7826 loop:
7827                 failed_cluster_refill = false;
7828                 failed_alloc = false;
7829                 BUG_ON(index != get_block_group_index(block_group));
7830                 btrfs_release_block_group(block_group, delalloc);
7831                 cond_resched();
7832         }
7833         up_read(&space_info->groups_sem);
7834
7835         if ((loop == LOOP_CACHING_NOWAIT) && have_caching_bg
7836                 && !orig_have_caching_bg)
7837                 orig_have_caching_bg = true;
7838
7839         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
7840                 goto search;
7841
7842         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
7843                 goto search;
7844
7845         /*
7846          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
7847          *                      caching kthreads as we move along
7848          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
7849          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
7850          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
7851          *                      again
7852          */
7853         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
7854                 index = 0;
7855                 if (loop == LOOP_CACHING_NOWAIT) {
7856                         /*
7857                          * We want to skip the LOOP_CACHING_WAIT step if we
7858                          * don't have any uncached bgs and we've already done a
7859                          * full search through.
7860                          */
7861                         if (orig_have_caching_bg || !full_search)
7862                                 loop = LOOP_CACHING_WAIT;
7863                         else
7864                                 loop = LOOP_ALLOC_CHUNK;
7865                 } else {
7866                         loop++;
7867                 }
7868
7869                 if (loop == LOOP_ALLOC_CHUNK) {
7870                         struct btrfs_trans_handle *trans;
7871                         int exist = 0;
7872
7873                         trans = current->journal_info;
7874                         if (trans)
7875                                 exist = 1;
7876                         else
7877                                 trans = btrfs_join_transaction(root);
7878
7879                         if (IS_ERR(trans)) {
7880                                 ret = PTR_ERR(trans);
7881                                 goto out;
7882                         }
7883
7884                         ret = do_chunk_alloc(trans, fs_info, flags,
7885                                              CHUNK_ALLOC_FORCE);
7886
7887                         /*
7888                          * If we can't allocate a new chunk we've already looped
7889                          * through at least once, move on to the NO_EMPTY_SIZE
7890                          * case.
7891                          */
7892                         if (ret == -ENOSPC)
7893                                 loop = LOOP_NO_EMPTY_SIZE;
7894
7895                         /*
7896                          * Do not bail out on ENOSPC since we
7897                          * can do more things.
7898                          */
7899                         if (ret < 0 && ret != -ENOSPC)
7900                                 btrfs_abort_transaction(trans, ret);
7901                         else
7902                                 ret = 0;
7903                         if (!exist)
7904                                 btrfs_end_transaction(trans);
7905                         if (ret)
7906                                 goto out;
7907                 }
7908
7909                 if (loop == LOOP_NO_EMPTY_SIZE) {
7910                         /*
7911                          * Don't loop again if we already have no empty_size and
7912                          * no empty_cluster.
7913                          */
7914                         if (empty_size == 0 &&
7915                             empty_cluster == 0) {
7916                                 ret = -ENOSPC;
7917                                 goto out;
7918                         }
7919                         empty_size = 0;
7920                         empty_cluster = 0;
7921                 }
7922
7923                 goto search;
7924         } else if (!ins->objectid) {
7925                 ret = -ENOSPC;
7926         } else if (ins->objectid) {
7927                 if (!use_cluster && last_ptr) {
7928                         spin_lock(&last_ptr->lock);
7929                         last_ptr->window_start = ins->objectid;
7930                         spin_unlock(&last_ptr->lock);
7931                 }
7932                 ret = 0;
7933         }
7934 out:
7935         if (ret == -ENOSPC) {
7936                 spin_lock(&space_info->lock);
7937                 space_info->max_extent_size = max_extent_size;
7938                 spin_unlock(&space_info->lock);
7939                 ins->offset = max_extent_size;
7940         }
7941         return ret;
7942 }
7943
7944 static void dump_space_info(struct btrfs_fs_info *fs_info,
7945                             struct btrfs_space_info *info, u64 bytes,
7946                             int dump_block_groups)
7947 {
7948         struct btrfs_block_group_cache *cache;
7949         int index = 0;
7950
7951         spin_lock(&info->lock);
7952         btrfs_info(fs_info, "space_info %llu has %llu free, is %sfull",
7953                    info->flags,
7954                    info->total_bytes - btrfs_space_info_used(info, true),
7955                    info->full ? "" : "not ");
7956         btrfs_info(fs_info,
7957                 "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu",
7958                 info->total_bytes, info->bytes_used, info->bytes_pinned,
7959                 info->bytes_reserved, info->bytes_may_use,
7960                 info->bytes_readonly);
7961         spin_unlock(&info->lock);
7962
7963         if (!dump_block_groups)
7964                 return;
7965
7966         down_read(&info->groups_sem);
7967 again:
7968         list_for_each_entry(cache, &info->block_groups[index], list) {
7969                 spin_lock(&cache->lock);
7970                 btrfs_info(fs_info,
7971                         "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s",
7972                         cache->key.objectid, cache->key.offset,
7973                         btrfs_block_group_used(&cache->item), cache->pinned,
7974                         cache->reserved, cache->ro ? "[readonly]" : "");
7975                 btrfs_dump_free_space(cache, bytes);
7976                 spin_unlock(&cache->lock);
7977         }
7978         if (++index < BTRFS_NR_RAID_TYPES)
7979                 goto again;
7980         up_read(&info->groups_sem);
7981 }
7982
7983 int btrfs_reserve_extent(struct btrfs_root *root, u64 ram_bytes,
7984                          u64 num_bytes, u64 min_alloc_size,
7985                          u64 empty_size, u64 hint_byte,
7986                          struct btrfs_key *ins, int is_data, int delalloc)
7987 {
7988         struct btrfs_fs_info *fs_info = root->fs_info;
7989         bool final_tried = num_bytes == min_alloc_size;
7990         u64 flags;
7991         int ret;
7992
7993         flags = get_alloc_profile_by_root(root, is_data);
7994 again:
7995         WARN_ON(num_bytes < fs_info->sectorsize);
7996         ret = find_free_extent(fs_info, ram_bytes, num_bytes, empty_size,
7997                                hint_byte, ins, flags, delalloc);
7998         if (!ret && !is_data) {
7999                 btrfs_dec_block_group_reservations(fs_info, ins->objectid);
8000         } else if (ret == -ENOSPC) {
8001                 if (!final_tried && ins->offset) {
8002                         num_bytes = min(num_bytes >> 1, ins->offset);
8003                         num_bytes = round_down(num_bytes,
8004                                                fs_info->sectorsize);
8005                         num_bytes = max(num_bytes, min_alloc_size);
8006                         ram_bytes = num_bytes;
8007                         if (num_bytes == min_alloc_size)
8008                                 final_tried = true;
8009                         goto again;
8010                 } else if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
8011                         struct btrfs_space_info *sinfo;
8012
8013                         sinfo = __find_space_info(fs_info, flags);
8014                         btrfs_err(fs_info,
8015                                   "allocation failed flags %llu, wanted %llu",
8016                                   flags, num_bytes);
8017                         if (sinfo)
8018                                 dump_space_info(fs_info, sinfo, num_bytes, 1);
8019                 }
8020         }
8021
8022         return ret;
8023 }
8024
8025 static int __btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
8026                                         u64 start, u64 len,
8027                                         int pin, int delalloc)
8028 {
8029         struct btrfs_block_group_cache *cache;
8030         int ret = 0;
8031
8032         cache = btrfs_lookup_block_group(fs_info, start);
8033         if (!cache) {
8034                 btrfs_err(fs_info, "Unable to find block group for %llu",
8035                           start);
8036                 return -ENOSPC;
8037         }
8038
8039         if (pin)
8040                 pin_down_extent(fs_info, cache, start, len, 1);
8041         else {
8042                 if (btrfs_test_opt(fs_info, DISCARD))
8043                         ret = btrfs_discard_extent(fs_info, start, len, NULL);
8044                 btrfs_add_free_space(cache, start, len);
8045                 btrfs_free_reserved_bytes(cache, len, delalloc);
8046                 trace_btrfs_reserved_extent_free(fs_info, start, len);
8047         }
8048
8049         btrfs_put_block_group(cache);
8050         return ret;
8051 }
8052
8053 int btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
8054                                u64 start, u64 len, int delalloc)
8055 {
8056         return __btrfs_free_reserved_extent(fs_info, start, len, 0, delalloc);
8057 }
8058
8059 int btrfs_free_and_pin_reserved_extent(struct btrfs_fs_info *fs_info,
8060                                        u64 start, u64 len)
8061 {
8062         return __btrfs_free_reserved_extent(fs_info, start, len, 1, 0);
8063 }
8064
8065 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8066                                       struct btrfs_fs_info *fs_info,
8067                                       u64 parent, u64 root_objectid,
8068                                       u64 flags, u64 owner, u64 offset,
8069                                       struct btrfs_key *ins, int ref_mod)
8070 {
8071         int ret;
8072         struct btrfs_extent_item *extent_item;
8073         struct btrfs_extent_inline_ref *iref;
8074         struct btrfs_path *path;
8075         struct extent_buffer *leaf;
8076         int type;
8077         u32 size;
8078
8079         if (parent > 0)
8080                 type = BTRFS_SHARED_DATA_REF_KEY;
8081         else
8082                 type = BTRFS_EXTENT_DATA_REF_KEY;
8083
8084         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
8085
8086         path = btrfs_alloc_path();
8087         if (!path)
8088                 return -ENOMEM;
8089
8090         path->leave_spinning = 1;
8091         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
8092                                       ins, size);
8093         if (ret) {
8094                 btrfs_free_path(path);
8095                 return ret;
8096         }
8097
8098         leaf = path->nodes[0];
8099         extent_item = btrfs_item_ptr(leaf, path->slots[0],
8100                                      struct btrfs_extent_item);
8101         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
8102         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8103         btrfs_set_extent_flags(leaf, extent_item,
8104                                flags | BTRFS_EXTENT_FLAG_DATA);
8105
8106         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8107         btrfs_set_extent_inline_ref_type(leaf, iref, type);
8108         if (parent > 0) {
8109                 struct btrfs_shared_data_ref *ref;
8110                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
8111                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
8112                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
8113         } else {
8114                 struct btrfs_extent_data_ref *ref;
8115                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
8116                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
8117                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
8118                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
8119                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
8120         }
8121
8122         btrfs_mark_buffer_dirty(path->nodes[0]);
8123         btrfs_free_path(path);
8124
8125         ret = remove_from_free_space_tree(trans, fs_info, ins->objectid,
8126                                           ins->offset);
8127         if (ret)
8128                 return ret;
8129
8130         ret = update_block_group(trans, fs_info, ins->objectid, ins->offset, 1);
8131         if (ret) { /* -ENOENT, logic error */
8132                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8133                         ins->objectid, ins->offset);
8134                 BUG();
8135         }
8136         trace_btrfs_reserved_extent_alloc(fs_info, ins->objectid, ins->offset);
8137         return ret;
8138 }
8139
8140 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
8141                                      struct btrfs_fs_info *fs_info,
8142                                      u64 parent, u64 root_objectid,
8143                                      u64 flags, struct btrfs_disk_key *key,
8144                                      int level, struct btrfs_key *ins)
8145 {
8146         int ret;
8147         struct btrfs_extent_item *extent_item;
8148         struct btrfs_tree_block_info *block_info;
8149         struct btrfs_extent_inline_ref *iref;
8150         struct btrfs_path *path;
8151         struct extent_buffer *leaf;
8152         u32 size = sizeof(*extent_item) + sizeof(*iref);
8153         u64 num_bytes = ins->offset;
8154         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
8155
8156         if (!skinny_metadata)
8157                 size += sizeof(*block_info);
8158
8159         path = btrfs_alloc_path();
8160         if (!path) {
8161                 btrfs_free_and_pin_reserved_extent(fs_info, ins->objectid,
8162                                                    fs_info->nodesize);
8163                 return -ENOMEM;
8164         }
8165
8166         path->leave_spinning = 1;
8167         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
8168                                       ins, size);
8169         if (ret) {
8170                 btrfs_free_path(path);
8171                 btrfs_free_and_pin_reserved_extent(fs_info, ins->objectid,
8172                                                    fs_info->nodesize);
8173                 return ret;
8174         }
8175
8176         leaf = path->nodes[0];
8177         extent_item = btrfs_item_ptr(leaf, path->slots[0],
8178                                      struct btrfs_extent_item);
8179         btrfs_set_extent_refs(leaf, extent_item, 1);
8180         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8181         btrfs_set_extent_flags(leaf, extent_item,
8182                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
8183
8184         if (skinny_metadata) {
8185                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8186                 num_bytes = fs_info->nodesize;
8187         } else {
8188                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
8189                 btrfs_set_tree_block_key(leaf, block_info, key);
8190                 btrfs_set_tree_block_level(leaf, block_info, level);
8191                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
8192         }
8193
8194         if (parent > 0) {
8195                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
8196                 btrfs_set_extent_inline_ref_type(leaf, iref,
8197                                                  BTRFS_SHARED_BLOCK_REF_KEY);
8198                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
8199         } else {
8200                 btrfs_set_extent_inline_ref_type(leaf, iref,
8201                                                  BTRFS_TREE_BLOCK_REF_KEY);
8202                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
8203         }
8204
8205         btrfs_mark_buffer_dirty(leaf);
8206         btrfs_free_path(path);
8207
8208         ret = remove_from_free_space_tree(trans, fs_info, ins->objectid,
8209                                           num_bytes);
8210         if (ret)
8211                 return ret;
8212
8213         ret = update_block_group(trans, fs_info, ins->objectid,
8214                                  fs_info->nodesize, 1);
8215         if (ret) { /* -ENOENT, logic error */
8216                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8217                         ins->objectid, ins->offset);
8218                 BUG();
8219         }
8220
8221         trace_btrfs_reserved_extent_alloc(fs_info, ins->objectid,
8222                                           fs_info->nodesize);
8223         return ret;
8224 }
8225
8226 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8227                                      u64 root_objectid, u64 owner,
8228                                      u64 offset, u64 ram_bytes,
8229                                      struct btrfs_key *ins)
8230 {
8231         struct btrfs_fs_info *fs_info = trans->fs_info;
8232         int ret;
8233
8234         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
8235
8236         ret = btrfs_add_delayed_data_ref(fs_info, trans, ins->objectid,
8237                                          ins->offset, 0, root_objectid, owner,
8238                                          offset, ram_bytes,
8239                                          BTRFS_ADD_DELAYED_EXTENT, NULL, NULL);
8240         return ret;
8241 }
8242
8243 /*
8244  * this is used by the tree logging recovery code.  It records that
8245  * an extent has been allocated and makes sure to clear the free
8246  * space cache bits as well
8247  */
8248 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
8249                                    struct btrfs_fs_info *fs_info,
8250                                    u64 root_objectid, u64 owner, u64 offset,
8251                                    struct btrfs_key *ins)
8252 {
8253         int ret;
8254         struct btrfs_block_group_cache *block_group;
8255         struct btrfs_space_info *space_info;
8256
8257         /*
8258          * Mixed block groups will exclude before processing the log so we only
8259          * need to do the exclude dance if this fs isn't mixed.
8260          */
8261         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS)) {
8262                 ret = __exclude_logged_extent(fs_info, ins->objectid,
8263                                               ins->offset);
8264                 if (ret)
8265                         return ret;
8266         }
8267
8268         block_group = btrfs_lookup_block_group(fs_info, ins->objectid);
8269         if (!block_group)
8270                 return -EINVAL;
8271
8272         space_info = block_group->space_info;
8273         spin_lock(&space_info->lock);
8274         spin_lock(&block_group->lock);
8275         space_info->bytes_reserved += ins->offset;
8276         block_group->reserved += ins->offset;
8277         spin_unlock(&block_group->lock);
8278         spin_unlock(&space_info->lock);
8279
8280         ret = alloc_reserved_file_extent(trans, fs_info, 0, root_objectid,
8281                                          0, owner, offset, ins, 1);
8282         btrfs_put_block_group(block_group);
8283         return ret;
8284 }
8285
8286 static struct extent_buffer *
8287 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
8288                       u64 bytenr, int level)
8289 {
8290         struct btrfs_fs_info *fs_info = root->fs_info;
8291         struct extent_buffer *buf;
8292
8293         buf = btrfs_find_create_tree_block(fs_info, bytenr);
8294         if (IS_ERR(buf))
8295                 return buf;
8296
8297         btrfs_set_header_generation(buf, trans->transid);
8298         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
8299         btrfs_tree_lock(buf);
8300         clean_tree_block(fs_info, buf);
8301         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
8302
8303         btrfs_set_lock_blocking(buf);
8304         set_extent_buffer_uptodate(buf);
8305
8306         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
8307                 buf->log_index = root->log_transid % 2;
8308                 /*
8309                  * we allow two log transactions at a time, use different
8310                  * EXENT bit to differentiate dirty pages.
8311                  */
8312                 if (buf->log_index == 0)
8313                         set_extent_dirty(&root->dirty_log_pages, buf->start,
8314                                         buf->start + buf->len - 1, GFP_NOFS);
8315                 else
8316                         set_extent_new(&root->dirty_log_pages, buf->start,
8317                                         buf->start + buf->len - 1);
8318         } else {
8319                 buf->log_index = -1;
8320                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
8321                          buf->start + buf->len - 1, GFP_NOFS);
8322         }
8323         trans->dirty = true;
8324         /* this returns a buffer locked for blocking */
8325         return buf;
8326 }
8327
8328 static struct btrfs_block_rsv *
8329 use_block_rsv(struct btrfs_trans_handle *trans,
8330               struct btrfs_root *root, u32 blocksize)
8331 {
8332         struct btrfs_fs_info *fs_info = root->fs_info;
8333         struct btrfs_block_rsv *block_rsv;
8334         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
8335         int ret;
8336         bool global_updated = false;
8337
8338         block_rsv = get_block_rsv(trans, root);
8339
8340         if (unlikely(block_rsv->size == 0))
8341                 goto try_reserve;
8342 again:
8343         ret = block_rsv_use_bytes(block_rsv, blocksize);
8344         if (!ret)
8345                 return block_rsv;
8346
8347         if (block_rsv->failfast)
8348                 return ERR_PTR(ret);
8349
8350         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
8351                 global_updated = true;
8352                 update_global_block_rsv(fs_info);
8353                 goto again;
8354         }
8355
8356         if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
8357                 static DEFINE_RATELIMIT_STATE(_rs,
8358                                 DEFAULT_RATELIMIT_INTERVAL * 10,
8359                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
8360                 if (__ratelimit(&_rs))
8361                         WARN(1, KERN_DEBUG
8362                                 "BTRFS: block rsv returned %d\n", ret);
8363         }
8364 try_reserve:
8365         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
8366                                      BTRFS_RESERVE_NO_FLUSH);
8367         if (!ret)
8368                 return block_rsv;
8369         /*
8370          * If we couldn't reserve metadata bytes try and use some from
8371          * the global reserve if its space type is the same as the global
8372          * reservation.
8373          */
8374         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
8375             block_rsv->space_info == global_rsv->space_info) {
8376                 ret = block_rsv_use_bytes(global_rsv, blocksize);
8377                 if (!ret)
8378                         return global_rsv;
8379         }
8380         return ERR_PTR(ret);
8381 }
8382
8383 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
8384                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
8385 {
8386         block_rsv_add_bytes(block_rsv, blocksize, 0);
8387         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
8388 }
8389
8390 /*
8391  * finds a free extent and does all the dirty work required for allocation
8392  * returns the tree buffer or an ERR_PTR on error.
8393  */
8394 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
8395                                              struct btrfs_root *root,
8396                                              u64 parent, u64 root_objectid,
8397                                              const struct btrfs_disk_key *key,
8398                                              int level, u64 hint,
8399                                              u64 empty_size)
8400 {
8401         struct btrfs_fs_info *fs_info = root->fs_info;
8402         struct btrfs_key ins;
8403         struct btrfs_block_rsv *block_rsv;
8404         struct extent_buffer *buf;
8405         struct btrfs_delayed_extent_op *extent_op;
8406         u64 flags = 0;
8407         int ret;
8408         u32 blocksize = fs_info->nodesize;
8409         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
8410
8411 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
8412         if (btrfs_is_testing(fs_info)) {
8413                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
8414                                             level);
8415                 if (!IS_ERR(buf))
8416                         root->alloc_bytenr += blocksize;
8417                 return buf;
8418         }
8419 #endif
8420
8421         block_rsv = use_block_rsv(trans, root, blocksize);
8422         if (IS_ERR(block_rsv))
8423                 return ERR_CAST(block_rsv);
8424
8425         ret = btrfs_reserve_extent(root, blocksize, blocksize, blocksize,
8426                                    empty_size, hint, &ins, 0, 0);
8427         if (ret)
8428                 goto out_unuse;
8429
8430         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level);
8431         if (IS_ERR(buf)) {
8432                 ret = PTR_ERR(buf);
8433                 goto out_free_reserved;
8434         }
8435
8436         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
8437                 if (parent == 0)
8438                         parent = ins.objectid;
8439                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
8440         } else
8441                 BUG_ON(parent > 0);
8442
8443         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
8444                 extent_op = btrfs_alloc_delayed_extent_op();
8445                 if (!extent_op) {
8446                         ret = -ENOMEM;
8447                         goto out_free_buf;
8448                 }
8449                 if (key)
8450                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
8451                 else
8452                         memset(&extent_op->key, 0, sizeof(extent_op->key));
8453                 extent_op->flags_to_set = flags;
8454                 extent_op->update_key = skinny_metadata ? false : true;
8455                 extent_op->update_flags = true;
8456                 extent_op->is_data = false;
8457                 extent_op->level = level;
8458
8459                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, ins.objectid,
8460                                                  ins.offset, parent,
8461                                                  root_objectid, level,
8462                                                  BTRFS_ADD_DELAYED_EXTENT,
8463                                                  extent_op, NULL, NULL);
8464                 if (ret)
8465                         goto out_free_delayed;
8466         }
8467         return buf;
8468
8469 out_free_delayed:
8470         btrfs_free_delayed_extent_op(extent_op);
8471 out_free_buf:
8472         free_extent_buffer(buf);
8473 out_free_reserved:
8474         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 0);
8475 out_unuse:
8476         unuse_block_rsv(fs_info, block_rsv, blocksize);
8477         return ERR_PTR(ret);
8478 }
8479
8480 struct walk_control {
8481         u64 refs[BTRFS_MAX_LEVEL];
8482         u64 flags[BTRFS_MAX_LEVEL];
8483         struct btrfs_key update_progress;
8484         int stage;
8485         int level;
8486         int shared_level;
8487         int update_ref;
8488         int keep_locks;
8489         int reada_slot;
8490         int reada_count;
8491         int for_reloc;
8492 };
8493
8494 #define DROP_REFERENCE  1
8495 #define UPDATE_BACKREF  2
8496
8497 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
8498                                      struct btrfs_root *root,
8499                                      struct walk_control *wc,
8500                                      struct btrfs_path *path)
8501 {
8502         struct btrfs_fs_info *fs_info = root->fs_info;
8503         u64 bytenr;
8504         u64 generation;
8505         u64 refs;
8506         u64 flags;
8507         u32 nritems;
8508         struct btrfs_key key;
8509         struct extent_buffer *eb;
8510         int ret;
8511         int slot;
8512         int nread = 0;
8513
8514         if (path->slots[wc->level] < wc->reada_slot) {
8515                 wc->reada_count = wc->reada_count * 2 / 3;
8516                 wc->reada_count = max(wc->reada_count, 2);
8517         } else {
8518                 wc->reada_count = wc->reada_count * 3 / 2;
8519                 wc->reada_count = min_t(int, wc->reada_count,
8520                                         BTRFS_NODEPTRS_PER_BLOCK(fs_info));
8521         }
8522
8523         eb = path->nodes[wc->level];
8524         nritems = btrfs_header_nritems(eb);
8525
8526         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
8527                 if (nread >= wc->reada_count)
8528                         break;
8529
8530                 cond_resched();
8531                 bytenr = btrfs_node_blockptr(eb, slot);
8532                 generation = btrfs_node_ptr_generation(eb, slot);
8533
8534                 if (slot == path->slots[wc->level])
8535                         goto reada;
8536
8537                 if (wc->stage == UPDATE_BACKREF &&
8538                     generation <= root->root_key.offset)
8539                         continue;
8540
8541                 /* We don't lock the tree block, it's OK to be racy here */
8542                 ret = btrfs_lookup_extent_info(trans, fs_info, bytenr,
8543                                                wc->level - 1, 1, &refs,
8544                                                &flags);
8545                 /* We don't care about errors in readahead. */
8546                 if (ret < 0)
8547                         continue;
8548                 BUG_ON(refs == 0);
8549
8550                 if (wc->stage == DROP_REFERENCE) {
8551                         if (refs == 1)
8552                                 goto reada;
8553
8554                         if (wc->level == 1 &&
8555                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8556                                 continue;
8557                         if (!wc->update_ref ||
8558                             generation <= root->root_key.offset)
8559                                 continue;
8560                         btrfs_node_key_to_cpu(eb, &key, slot);
8561                         ret = btrfs_comp_cpu_keys(&key,
8562                                                   &wc->update_progress);
8563                         if (ret < 0)
8564                                 continue;
8565                 } else {
8566                         if (wc->level == 1 &&
8567                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8568                                 continue;
8569                 }
8570 reada:
8571                 readahead_tree_block(fs_info, bytenr);
8572                 nread++;
8573         }
8574         wc->reada_slot = slot;
8575 }
8576
8577 /*
8578  * helper to process tree block while walking down the tree.
8579  *
8580  * when wc->stage == UPDATE_BACKREF, this function updates
8581  * back refs for pointers in the block.
8582  *
8583  * NOTE: return value 1 means we should stop walking down.
8584  */
8585 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
8586                                    struct btrfs_root *root,
8587                                    struct btrfs_path *path,
8588                                    struct walk_control *wc, int lookup_info)
8589 {
8590         struct btrfs_fs_info *fs_info = root->fs_info;
8591         int level = wc->level;
8592         struct extent_buffer *eb = path->nodes[level];
8593         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8594         int ret;
8595
8596         if (wc->stage == UPDATE_BACKREF &&
8597             btrfs_header_owner(eb) != root->root_key.objectid)
8598                 return 1;
8599
8600         /*
8601          * when reference count of tree block is 1, it won't increase
8602          * again. once full backref flag is set, we never clear it.
8603          */
8604         if (lookup_info &&
8605             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
8606              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
8607                 BUG_ON(!path->locks[level]);
8608                 ret = btrfs_lookup_extent_info(trans, fs_info,
8609                                                eb->start, level, 1,
8610                                                &wc->refs[level],
8611                                                &wc->flags[level]);
8612                 BUG_ON(ret == -ENOMEM);
8613                 if (ret)
8614                         return ret;
8615                 BUG_ON(wc->refs[level] == 0);
8616         }
8617
8618         if (wc->stage == DROP_REFERENCE) {
8619                 if (wc->refs[level] > 1)
8620                         return 1;
8621
8622                 if (path->locks[level] && !wc->keep_locks) {
8623                         btrfs_tree_unlock_rw(eb, path->locks[level]);
8624                         path->locks[level] = 0;
8625                 }
8626                 return 0;
8627         }
8628
8629         /* wc->stage == UPDATE_BACKREF */
8630         if (!(wc->flags[level] & flag)) {
8631                 BUG_ON(!path->locks[level]);
8632                 ret = btrfs_inc_ref(trans, root, eb, 1);
8633                 BUG_ON(ret); /* -ENOMEM */
8634                 ret = btrfs_dec_ref(trans, root, eb, 0);
8635                 BUG_ON(ret); /* -ENOMEM */
8636                 ret = btrfs_set_disk_extent_flags(trans, fs_info, eb->start,
8637                                                   eb->len, flag,
8638                                                   btrfs_header_level(eb), 0);
8639                 BUG_ON(ret); /* -ENOMEM */
8640                 wc->flags[level] |= flag;
8641         }
8642
8643         /*
8644          * the block is shared by multiple trees, so it's not good to
8645          * keep the tree lock
8646          */
8647         if (path->locks[level] && level > 0) {
8648                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8649                 path->locks[level] = 0;
8650         }
8651         return 0;
8652 }
8653
8654 /*
8655  * helper to process tree block pointer.
8656  *
8657  * when wc->stage == DROP_REFERENCE, this function checks
8658  * reference count of the block pointed to. if the block
8659  * is shared and we need update back refs for the subtree
8660  * rooted at the block, this function changes wc->stage to
8661  * UPDATE_BACKREF. if the block is shared and there is no
8662  * need to update back, this function drops the reference
8663  * to the block.
8664  *
8665  * NOTE: return value 1 means we should stop walking down.
8666  */
8667 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
8668                                  struct btrfs_root *root,
8669                                  struct btrfs_path *path,
8670                                  struct walk_control *wc, int *lookup_info)
8671 {
8672         struct btrfs_fs_info *fs_info = root->fs_info;
8673         u64 bytenr;
8674         u64 generation;
8675         u64 parent;
8676         u32 blocksize;
8677         struct btrfs_key key;
8678         struct extent_buffer *next;
8679         int level = wc->level;
8680         int reada = 0;
8681         int ret = 0;
8682         bool need_account = false;
8683
8684         generation = btrfs_node_ptr_generation(path->nodes[level],
8685                                                path->slots[level]);
8686         /*
8687          * if the lower level block was created before the snapshot
8688          * was created, we know there is no need to update back refs
8689          * for the subtree
8690          */
8691         if (wc->stage == UPDATE_BACKREF &&
8692             generation <= root->root_key.offset) {
8693                 *lookup_info = 1;
8694                 return 1;
8695         }
8696
8697         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
8698         blocksize = fs_info->nodesize;
8699
8700         next = find_extent_buffer(fs_info, bytenr);
8701         if (!next) {
8702                 next = btrfs_find_create_tree_block(fs_info, bytenr);
8703                 if (IS_ERR(next))
8704                         return PTR_ERR(next);
8705
8706                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
8707                                                level - 1);
8708                 reada = 1;
8709         }
8710         btrfs_tree_lock(next);
8711         btrfs_set_lock_blocking(next);
8712
8713         ret = btrfs_lookup_extent_info(trans, fs_info, bytenr, level - 1, 1,
8714                                        &wc->refs[level - 1],
8715                                        &wc->flags[level - 1]);
8716         if (ret < 0)
8717                 goto out_unlock;
8718
8719         if (unlikely(wc->refs[level - 1] == 0)) {
8720                 btrfs_err(fs_info, "Missing references.");
8721                 ret = -EIO;
8722                 goto out_unlock;
8723         }
8724         *lookup_info = 0;
8725
8726         if (wc->stage == DROP_REFERENCE) {
8727                 if (wc->refs[level - 1] > 1) {
8728                         need_account = true;
8729                         if (level == 1 &&
8730                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8731                                 goto skip;
8732
8733                         if (!wc->update_ref ||
8734                             generation <= root->root_key.offset)
8735                                 goto skip;
8736
8737                         btrfs_node_key_to_cpu(path->nodes[level], &key,
8738                                               path->slots[level]);
8739                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
8740                         if (ret < 0)
8741                                 goto skip;
8742
8743                         wc->stage = UPDATE_BACKREF;
8744                         wc->shared_level = level - 1;
8745                 }
8746         } else {
8747                 if (level == 1 &&
8748                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8749                         goto skip;
8750         }
8751
8752         if (!btrfs_buffer_uptodate(next, generation, 0)) {
8753                 btrfs_tree_unlock(next);
8754                 free_extent_buffer(next);
8755                 next = NULL;
8756                 *lookup_info = 1;
8757         }
8758
8759         if (!next) {
8760                 if (reada && level == 1)
8761                         reada_walk_down(trans, root, wc, path);
8762                 next = read_tree_block(fs_info, bytenr, generation);
8763                 if (IS_ERR(next)) {
8764                         return PTR_ERR(next);
8765                 } else if (!extent_buffer_uptodate(next)) {
8766                         free_extent_buffer(next);
8767                         return -EIO;
8768                 }
8769                 btrfs_tree_lock(next);
8770                 btrfs_set_lock_blocking(next);
8771         }
8772
8773         level--;
8774         ASSERT(level == btrfs_header_level(next));
8775         if (level != btrfs_header_level(next)) {
8776                 btrfs_err(root->fs_info, "mismatched level");
8777                 ret = -EIO;
8778                 goto out_unlock;
8779         }
8780         path->nodes[level] = next;
8781         path->slots[level] = 0;
8782         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8783         wc->level = level;
8784         if (wc->level == 1)
8785                 wc->reada_slot = 0;
8786         return 0;
8787 skip:
8788         wc->refs[level - 1] = 0;
8789         wc->flags[level - 1] = 0;
8790         if (wc->stage == DROP_REFERENCE) {
8791                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
8792                         parent = path->nodes[level]->start;
8793                 } else {
8794                         ASSERT(root->root_key.objectid ==
8795                                btrfs_header_owner(path->nodes[level]));
8796                         if (root->root_key.objectid !=
8797                             btrfs_header_owner(path->nodes[level])) {
8798                                 btrfs_err(root->fs_info,
8799                                                 "mismatched block owner");
8800                                 ret = -EIO;
8801                                 goto out_unlock;
8802                         }
8803                         parent = 0;
8804                 }
8805
8806                 if (need_account) {
8807                         ret = btrfs_qgroup_trace_subtree(trans, root, next,
8808                                                          generation, level - 1);
8809                         if (ret) {
8810                                 btrfs_err_rl(fs_info,
8811                                              "Error %d accounting shared subtree. Quota is out of sync, rescan required.",
8812                                              ret);
8813                         }
8814                 }
8815                 ret = btrfs_free_extent(trans, fs_info, bytenr, blocksize,
8816                                         parent, root->root_key.objectid,
8817                                         level - 1, 0);
8818                 if (ret)
8819                         goto out_unlock;
8820         }
8821
8822         *lookup_info = 1;
8823         ret = 1;
8824
8825 out_unlock:
8826         btrfs_tree_unlock(next);
8827         free_extent_buffer(next);
8828
8829         return ret;
8830 }
8831
8832 /*
8833  * helper to process tree block while walking up the tree.
8834  *
8835  * when wc->stage == DROP_REFERENCE, this function drops
8836  * reference count on the block.
8837  *
8838  * when wc->stage == UPDATE_BACKREF, this function changes
8839  * wc->stage back to DROP_REFERENCE if we changed wc->stage
8840  * to UPDATE_BACKREF previously while processing the block.
8841  *
8842  * NOTE: return value 1 means we should stop walking up.
8843  */
8844 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
8845                                  struct btrfs_root *root,
8846                                  struct btrfs_path *path,
8847                                  struct walk_control *wc)
8848 {
8849         struct btrfs_fs_info *fs_info = root->fs_info;
8850         int ret;
8851         int level = wc->level;
8852         struct extent_buffer *eb = path->nodes[level];
8853         u64 parent = 0;
8854
8855         if (wc->stage == UPDATE_BACKREF) {
8856                 BUG_ON(wc->shared_level < level);
8857                 if (level < wc->shared_level)
8858                         goto out;
8859
8860                 ret = find_next_key(path, level + 1, &wc->update_progress);
8861                 if (ret > 0)
8862                         wc->update_ref = 0;
8863
8864                 wc->stage = DROP_REFERENCE;
8865                 wc->shared_level = -1;
8866                 path->slots[level] = 0;
8867
8868                 /*
8869                  * check reference count again if the block isn't locked.
8870                  * we should start walking down the tree again if reference
8871                  * count is one.
8872                  */
8873                 if (!path->locks[level]) {
8874                         BUG_ON(level == 0);
8875                         btrfs_tree_lock(eb);
8876                         btrfs_set_lock_blocking(eb);
8877                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8878
8879                         ret = btrfs_lookup_extent_info(trans, fs_info,
8880                                                        eb->start, level, 1,
8881                                                        &wc->refs[level],
8882                                                        &wc->flags[level]);
8883                         if (ret < 0) {
8884                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8885                                 path->locks[level] = 0;
8886                                 return ret;
8887                         }
8888                         BUG_ON(wc->refs[level] == 0);
8889                         if (wc->refs[level] == 1) {
8890                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8891                                 path->locks[level] = 0;
8892                                 return 1;
8893                         }
8894                 }
8895         }
8896
8897         /* wc->stage == DROP_REFERENCE */
8898         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
8899
8900         if (wc->refs[level] == 1) {
8901                 if (level == 0) {
8902                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8903                                 ret = btrfs_dec_ref(trans, root, eb, 1);
8904                         else
8905                                 ret = btrfs_dec_ref(trans, root, eb, 0);
8906                         BUG_ON(ret); /* -ENOMEM */
8907                         ret = btrfs_qgroup_trace_leaf_items(trans, fs_info, eb);
8908                         if (ret) {
8909                                 btrfs_err_rl(fs_info,
8910                                              "error %d accounting leaf items. Quota is out of sync, rescan required.",
8911                                              ret);
8912                         }
8913                 }
8914                 /* make block locked assertion in clean_tree_block happy */
8915                 if (!path->locks[level] &&
8916                     btrfs_header_generation(eb) == trans->transid) {
8917                         btrfs_tree_lock(eb);
8918                         btrfs_set_lock_blocking(eb);
8919                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8920                 }
8921                 clean_tree_block(fs_info, eb);
8922         }
8923
8924         if (eb == root->node) {
8925                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8926                         parent = eb->start;
8927                 else
8928                         BUG_ON(root->root_key.objectid !=
8929                                btrfs_header_owner(eb));
8930         } else {
8931                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8932                         parent = path->nodes[level + 1]->start;
8933                 else
8934                         BUG_ON(root->root_key.objectid !=
8935                                btrfs_header_owner(path->nodes[level + 1]));
8936         }
8937
8938         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
8939 out:
8940         wc->refs[level] = 0;
8941         wc->flags[level] = 0;
8942         return 0;
8943 }
8944
8945 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
8946                                    struct btrfs_root *root,
8947                                    struct btrfs_path *path,
8948                                    struct walk_control *wc)
8949 {
8950         int level = wc->level;
8951         int lookup_info = 1;
8952         int ret;
8953
8954         while (level >= 0) {
8955                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
8956                 if (ret > 0)
8957                         break;
8958
8959                 if (level == 0)
8960                         break;
8961
8962                 if (path->slots[level] >=
8963                     btrfs_header_nritems(path->nodes[level]))
8964                         break;
8965
8966                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
8967                 if (ret > 0) {
8968                         path->slots[level]++;
8969                         continue;
8970                 } else if (ret < 0)
8971                         return ret;
8972                 level = wc->level;
8973         }
8974         return 0;
8975 }
8976
8977 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
8978                                  struct btrfs_root *root,
8979                                  struct btrfs_path *path,
8980                                  struct walk_control *wc, int max_level)
8981 {
8982         int level = wc->level;
8983         int ret;
8984
8985         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
8986         while (level < max_level && path->nodes[level]) {
8987                 wc->level = level;
8988                 if (path->slots[level] + 1 <
8989                     btrfs_header_nritems(path->nodes[level])) {
8990                         path->slots[level]++;
8991                         return 0;
8992                 } else {
8993                         ret = walk_up_proc(trans, root, path, wc);
8994                         if (ret > 0)
8995                                 return 0;
8996
8997                         if (path->locks[level]) {
8998                                 btrfs_tree_unlock_rw(path->nodes[level],
8999                                                      path->locks[level]);
9000                                 path->locks[level] = 0;
9001                         }
9002                         free_extent_buffer(path->nodes[level]);
9003                         path->nodes[level] = NULL;
9004                         level++;
9005                 }
9006         }
9007         return 1;
9008 }
9009
9010 /*
9011  * drop a subvolume tree.
9012  *
9013  * this function traverses the tree freeing any blocks that only
9014  * referenced by the tree.
9015  *
9016  * when a shared tree block is found. this function decreases its
9017  * reference count by one. if update_ref is true, this function
9018  * also make sure backrefs for the shared block and all lower level
9019  * blocks are properly updated.
9020  *
9021  * If called with for_reloc == 0, may exit early with -EAGAIN
9022  */
9023 int btrfs_drop_snapshot(struct btrfs_root *root,
9024                          struct btrfs_block_rsv *block_rsv, int update_ref,
9025                          int for_reloc)
9026 {
9027         struct btrfs_fs_info *fs_info = root->fs_info;
9028         struct btrfs_path *path;
9029         struct btrfs_trans_handle *trans;
9030         struct btrfs_root *tree_root = fs_info->tree_root;
9031         struct btrfs_root_item *root_item = &root->root_item;
9032         struct walk_control *wc;
9033         struct btrfs_key key;
9034         int err = 0;
9035         int ret;
9036         int level;
9037         bool root_dropped = false;
9038
9039         btrfs_debug(fs_info, "Drop subvolume %llu", root->objectid);
9040
9041         path = btrfs_alloc_path();
9042         if (!path) {
9043                 err = -ENOMEM;
9044                 goto out;
9045         }
9046
9047         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9048         if (!wc) {
9049                 btrfs_free_path(path);
9050                 err = -ENOMEM;
9051                 goto out;
9052         }
9053
9054         trans = btrfs_start_transaction(tree_root, 0);
9055         if (IS_ERR(trans)) {
9056                 err = PTR_ERR(trans);
9057                 goto out_free;
9058         }
9059
9060         if (block_rsv)
9061                 trans->block_rsv = block_rsv;
9062
9063         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
9064                 level = btrfs_header_level(root->node);
9065                 path->nodes[level] = btrfs_lock_root_node(root);
9066                 btrfs_set_lock_blocking(path->nodes[level]);
9067                 path->slots[level] = 0;
9068                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9069                 memset(&wc->update_progress, 0,
9070                        sizeof(wc->update_progress));
9071         } else {
9072                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
9073                 memcpy(&wc->update_progress, &key,
9074                        sizeof(wc->update_progress));
9075
9076                 level = root_item->drop_level;
9077                 BUG_ON(level == 0);
9078                 path->lowest_level = level;
9079                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
9080                 path->lowest_level = 0;
9081                 if (ret < 0) {
9082                         err = ret;
9083                         goto out_end_trans;
9084                 }
9085                 WARN_ON(ret > 0);
9086
9087                 /*
9088                  * unlock our path, this is safe because only this
9089                  * function is allowed to delete this snapshot
9090                  */
9091                 btrfs_unlock_up_safe(path, 0);
9092
9093                 level = btrfs_header_level(root->node);
9094                 while (1) {
9095                         btrfs_tree_lock(path->nodes[level]);
9096                         btrfs_set_lock_blocking(path->nodes[level]);
9097                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9098
9099                         ret = btrfs_lookup_extent_info(trans, fs_info,
9100                                                 path->nodes[level]->start,
9101                                                 level, 1, &wc->refs[level],
9102                                                 &wc->flags[level]);
9103                         if (ret < 0) {
9104                                 err = ret;
9105                                 goto out_end_trans;
9106                         }
9107                         BUG_ON(wc->refs[level] == 0);
9108
9109                         if (level == root_item->drop_level)
9110                                 break;
9111
9112                         btrfs_tree_unlock(path->nodes[level]);
9113                         path->locks[level] = 0;
9114                         WARN_ON(wc->refs[level] != 1);
9115                         level--;
9116                 }
9117         }
9118
9119         wc->level = level;
9120         wc->shared_level = -1;
9121         wc->stage = DROP_REFERENCE;
9122         wc->update_ref = update_ref;
9123         wc->keep_locks = 0;
9124         wc->for_reloc = for_reloc;
9125         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
9126
9127         while (1) {
9128
9129                 ret = walk_down_tree(trans, root, path, wc);
9130                 if (ret < 0) {
9131                         err = ret;
9132                         break;
9133                 }
9134
9135                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
9136                 if (ret < 0) {
9137                         err = ret;
9138                         break;
9139                 }
9140
9141                 if (ret > 0) {
9142                         BUG_ON(wc->stage != DROP_REFERENCE);
9143                         break;
9144                 }
9145
9146                 if (wc->stage == DROP_REFERENCE) {
9147                         level = wc->level;
9148                         btrfs_node_key(path->nodes[level],
9149                                        &root_item->drop_progress,
9150                                        path->slots[level]);
9151                         root_item->drop_level = level;
9152                 }
9153
9154                 BUG_ON(wc->level == 0);
9155                 if (btrfs_should_end_transaction(trans) ||
9156                     (!for_reloc && btrfs_need_cleaner_sleep(fs_info))) {
9157                         ret = btrfs_update_root(trans, tree_root,
9158                                                 &root->root_key,
9159                                                 root_item);
9160                         if (ret) {
9161                                 btrfs_abort_transaction(trans, ret);
9162                                 err = ret;
9163                                 goto out_end_trans;
9164                         }
9165
9166                         btrfs_end_transaction_throttle(trans);
9167                         if (!for_reloc && btrfs_need_cleaner_sleep(fs_info)) {
9168                                 btrfs_debug(fs_info,
9169                                             "drop snapshot early exit");
9170                                 err = -EAGAIN;
9171                                 goto out_free;
9172                         }
9173
9174                         trans = btrfs_start_transaction(tree_root, 0);
9175                         if (IS_ERR(trans)) {
9176                                 err = PTR_ERR(trans);
9177                                 goto out_free;
9178                         }
9179                         if (block_rsv)
9180                                 trans->block_rsv = block_rsv;
9181                 }
9182         }
9183         btrfs_release_path(path);
9184         if (err)
9185                 goto out_end_trans;
9186
9187         ret = btrfs_del_root(trans, tree_root, &root->root_key);
9188         if (ret) {
9189                 btrfs_abort_transaction(trans, ret);
9190                 goto out_end_trans;
9191         }
9192
9193         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
9194                 ret = btrfs_find_root(tree_root, &root->root_key, path,
9195                                       NULL, NULL);
9196                 if (ret < 0) {
9197                         btrfs_abort_transaction(trans, ret);
9198                         err = ret;
9199                         goto out_end_trans;
9200                 } else if (ret > 0) {
9201                         /* if we fail to delete the orphan item this time
9202                          * around, it'll get picked up the next time.
9203                          *
9204                          * The most common failure here is just -ENOENT.
9205                          */
9206                         btrfs_del_orphan_item(trans, tree_root,
9207                                               root->root_key.objectid);
9208                 }
9209         }
9210
9211         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
9212                 btrfs_add_dropped_root(trans, root);
9213         } else {
9214                 free_extent_buffer(root->node);
9215                 free_extent_buffer(root->commit_root);
9216                 btrfs_put_fs_root(root);
9217         }
9218         root_dropped = true;
9219 out_end_trans:
9220         btrfs_end_transaction_throttle(trans);
9221 out_free:
9222         kfree(wc);
9223         btrfs_free_path(path);
9224 out:
9225         /*
9226          * So if we need to stop dropping the snapshot for whatever reason we
9227          * need to make sure to add it back to the dead root list so that we
9228          * keep trying to do the work later.  This also cleans up roots if we
9229          * don't have it in the radix (like when we recover after a power fail
9230          * or unmount) so we don't leak memory.
9231          */
9232         if (!for_reloc && root_dropped == false)
9233                 btrfs_add_dead_root(root);
9234         if (err && err != -EAGAIN)
9235                 btrfs_handle_fs_error(fs_info, err, NULL);
9236         return err;
9237 }
9238
9239 /*
9240  * drop subtree rooted at tree block 'node'.
9241  *
9242  * NOTE: this function will unlock and release tree block 'node'
9243  * only used by relocation code
9244  */
9245 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
9246                         struct btrfs_root *root,
9247                         struct extent_buffer *node,
9248                         struct extent_buffer *parent)
9249 {
9250         struct btrfs_fs_info *fs_info = root->fs_info;
9251         struct btrfs_path *path;
9252         struct walk_control *wc;
9253         int level;
9254         int parent_level;
9255         int ret = 0;
9256         int wret;
9257
9258         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
9259
9260         path = btrfs_alloc_path();
9261         if (!path)
9262                 return -ENOMEM;
9263
9264         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9265         if (!wc) {
9266                 btrfs_free_path(path);
9267                 return -ENOMEM;
9268         }
9269
9270         btrfs_assert_tree_locked(parent);
9271         parent_level = btrfs_header_level(parent);
9272         extent_buffer_get(parent);
9273         path->nodes[parent_level] = parent;
9274         path->slots[parent_level] = btrfs_header_nritems(parent);
9275
9276         btrfs_assert_tree_locked(node);
9277         level = btrfs_header_level(node);
9278         path->nodes[level] = node;
9279         path->slots[level] = 0;
9280         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9281
9282         wc->refs[parent_level] = 1;
9283         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
9284         wc->level = level;
9285         wc->shared_level = -1;
9286         wc->stage = DROP_REFERENCE;
9287         wc->update_ref = 0;
9288         wc->keep_locks = 1;
9289         wc->for_reloc = 1;
9290         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
9291
9292         while (1) {
9293                 wret = walk_down_tree(trans, root, path, wc);
9294                 if (wret < 0) {
9295                         ret = wret;
9296                         break;
9297                 }
9298
9299                 wret = walk_up_tree(trans, root, path, wc, parent_level);
9300                 if (wret < 0)
9301                         ret = wret;
9302                 if (wret != 0)
9303                         break;
9304         }
9305
9306         kfree(wc);
9307         btrfs_free_path(path);
9308         return ret;
9309 }
9310
9311 static u64 update_block_group_flags(struct btrfs_fs_info *fs_info, u64 flags)
9312 {
9313         u64 num_devices;
9314         u64 stripped;
9315
9316         /*
9317          * if restripe for this chunk_type is on pick target profile and
9318          * return, otherwise do the usual balance
9319          */
9320         stripped = get_restripe_target(fs_info, flags);
9321         if (stripped)
9322                 return extended_to_chunk(stripped);
9323
9324         num_devices = fs_info->fs_devices->rw_devices;
9325
9326         stripped = BTRFS_BLOCK_GROUP_RAID0 |
9327                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
9328                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
9329
9330         if (num_devices == 1) {
9331                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9332                 stripped = flags & ~stripped;
9333
9334                 /* turn raid0 into single device chunks */
9335                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
9336                         return stripped;
9337
9338                 /* turn mirroring into duplication */
9339                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
9340                              BTRFS_BLOCK_GROUP_RAID10))
9341                         return stripped | BTRFS_BLOCK_GROUP_DUP;
9342         } else {
9343                 /* they already had raid on here, just return */
9344                 if (flags & stripped)
9345                         return flags;
9346
9347                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9348                 stripped = flags & ~stripped;
9349
9350                 /* switch duplicated blocks with raid1 */
9351                 if (flags & BTRFS_BLOCK_GROUP_DUP)
9352                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
9353
9354                 /* this is drive concat, leave it alone */
9355         }
9356
9357         return flags;
9358 }
9359
9360 static int inc_block_group_ro(struct btrfs_block_group_cache *cache, int force)
9361 {
9362         struct btrfs_space_info *sinfo = cache->space_info;
9363         u64 num_bytes;
9364         u64 min_allocable_bytes;
9365         int ret = -ENOSPC;
9366
9367         /*
9368          * We need some metadata space and system metadata space for
9369          * allocating chunks in some corner cases until we force to set
9370          * it to be readonly.
9371          */
9372         if ((sinfo->flags &
9373              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
9374             !force)
9375                 min_allocable_bytes = SZ_1M;
9376         else
9377                 min_allocable_bytes = 0;
9378
9379         spin_lock(&sinfo->lock);
9380         spin_lock(&cache->lock);
9381
9382         if (cache->ro) {
9383                 cache->ro++;
9384                 ret = 0;
9385                 goto out;
9386         }
9387
9388         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
9389                     cache->bytes_super - btrfs_block_group_used(&cache->item);
9390
9391         if (btrfs_space_info_used(sinfo, true) + num_bytes +
9392             min_allocable_bytes <= sinfo->total_bytes) {
9393                 sinfo->bytes_readonly += num_bytes;
9394                 cache->ro++;
9395                 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
9396                 ret = 0;
9397         }
9398 out:
9399         spin_unlock(&cache->lock);
9400         spin_unlock(&sinfo->lock);
9401         return ret;
9402 }
9403
9404 int btrfs_inc_block_group_ro(struct btrfs_fs_info *fs_info,
9405                              struct btrfs_block_group_cache *cache)
9406
9407 {
9408         struct btrfs_trans_handle *trans;
9409         u64 alloc_flags;
9410         int ret;
9411
9412 again:
9413         trans = btrfs_join_transaction(fs_info->extent_root);
9414         if (IS_ERR(trans))
9415                 return PTR_ERR(trans);
9416
9417         /*
9418          * we're not allowed to set block groups readonly after the dirty
9419          * block groups cache has started writing.  If it already started,
9420          * back off and let this transaction commit
9421          */
9422         mutex_lock(&fs_info->ro_block_group_mutex);
9423         if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
9424                 u64 transid = trans->transid;
9425
9426                 mutex_unlock(&fs_info->ro_block_group_mutex);
9427                 btrfs_end_transaction(trans);
9428
9429                 ret = btrfs_wait_for_commit(fs_info, transid);
9430                 if (ret)
9431                         return ret;
9432                 goto again;
9433         }
9434
9435         /*
9436          * if we are changing raid levels, try to allocate a corresponding
9437          * block group with the new raid level.
9438          */
9439         alloc_flags = update_block_group_flags(fs_info, cache->flags);
9440         if (alloc_flags != cache->flags) {
9441                 ret = do_chunk_alloc(trans, fs_info, alloc_flags,
9442                                      CHUNK_ALLOC_FORCE);
9443                 /*
9444                  * ENOSPC is allowed here, we may have enough space
9445                  * already allocated at the new raid level to
9446                  * carry on
9447                  */
9448                 if (ret == -ENOSPC)
9449                         ret = 0;
9450                 if (ret < 0)
9451                         goto out;
9452         }
9453
9454         ret = inc_block_group_ro(cache, 0);
9455         if (!ret)
9456                 goto out;
9457         alloc_flags = get_alloc_profile(fs_info, cache->space_info->flags);
9458         ret = do_chunk_alloc(trans, fs_info, alloc_flags,
9459                              CHUNK_ALLOC_FORCE);
9460         if (ret < 0)
9461                 goto out;
9462         ret = inc_block_group_ro(cache, 0);
9463 out:
9464         if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
9465                 alloc_flags = update_block_group_flags(fs_info, cache->flags);
9466                 mutex_lock(&fs_info->chunk_mutex);
9467                 check_system_chunk(trans, fs_info, alloc_flags);
9468                 mutex_unlock(&fs_info->chunk_mutex);
9469         }
9470         mutex_unlock(&fs_info->ro_block_group_mutex);
9471
9472         btrfs_end_transaction(trans);
9473         return ret;
9474 }
9475
9476 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
9477                             struct btrfs_fs_info *fs_info, u64 type)
9478 {
9479         u64 alloc_flags = get_alloc_profile(fs_info, type);
9480
9481         return do_chunk_alloc(trans, fs_info, alloc_flags, CHUNK_ALLOC_FORCE);
9482 }
9483
9484 /*
9485  * helper to account the unused space of all the readonly block group in the
9486  * space_info. takes mirrors into account.
9487  */
9488 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
9489 {
9490         struct btrfs_block_group_cache *block_group;
9491         u64 free_bytes = 0;
9492         int factor;
9493
9494         /* It's df, we don't care if it's racy */
9495         if (list_empty(&sinfo->ro_bgs))
9496                 return 0;
9497
9498         spin_lock(&sinfo->lock);
9499         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
9500                 spin_lock(&block_group->lock);
9501
9502                 if (!block_group->ro) {
9503                         spin_unlock(&block_group->lock);
9504                         continue;
9505                 }
9506
9507                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
9508                                           BTRFS_BLOCK_GROUP_RAID10 |
9509                                           BTRFS_BLOCK_GROUP_DUP))
9510                         factor = 2;
9511                 else
9512                         factor = 1;
9513
9514                 free_bytes += (block_group->key.offset -
9515                                btrfs_block_group_used(&block_group->item)) *
9516                                factor;
9517
9518                 spin_unlock(&block_group->lock);
9519         }
9520         spin_unlock(&sinfo->lock);
9521
9522         return free_bytes;
9523 }
9524
9525 void btrfs_dec_block_group_ro(struct btrfs_block_group_cache *cache)
9526 {
9527         struct btrfs_space_info *sinfo = cache->space_info;
9528         u64 num_bytes;
9529
9530         BUG_ON(!cache->ro);
9531
9532         spin_lock(&sinfo->lock);
9533         spin_lock(&cache->lock);
9534         if (!--cache->ro) {
9535                 num_bytes = cache->key.offset - cache->reserved -
9536                             cache->pinned - cache->bytes_super -
9537                             btrfs_block_group_used(&cache->item);
9538                 sinfo->bytes_readonly -= num_bytes;
9539                 list_del_init(&cache->ro_list);
9540         }
9541         spin_unlock(&cache->lock);
9542         spin_unlock(&sinfo->lock);
9543 }
9544
9545 /*
9546  * checks to see if its even possible to relocate this block group.
9547  *
9548  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
9549  * ok to go ahead and try.
9550  */
9551 int btrfs_can_relocate(struct btrfs_fs_info *fs_info, u64 bytenr)
9552 {
9553         struct btrfs_root *root = fs_info->extent_root;
9554         struct btrfs_block_group_cache *block_group;
9555         struct btrfs_space_info *space_info;
9556         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
9557         struct btrfs_device *device;
9558         struct btrfs_trans_handle *trans;
9559         u64 min_free;
9560         u64 dev_min = 1;
9561         u64 dev_nr = 0;
9562         u64 target;
9563         int debug;
9564         int index;
9565         int full = 0;
9566         int ret = 0;
9567
9568         debug = btrfs_test_opt(fs_info, ENOSPC_DEBUG);
9569
9570         block_group = btrfs_lookup_block_group(fs_info, bytenr);
9571
9572         /* odd, couldn't find the block group, leave it alone */
9573         if (!block_group) {
9574                 if (debug)
9575                         btrfs_warn(fs_info,
9576                                    "can't find block group for bytenr %llu",
9577                                    bytenr);
9578                 return -1;
9579         }
9580
9581         min_free = btrfs_block_group_used(&block_group->item);
9582
9583         /* no bytes used, we're good */
9584         if (!min_free)
9585                 goto out;
9586
9587         space_info = block_group->space_info;
9588         spin_lock(&space_info->lock);
9589
9590         full = space_info->full;
9591
9592         /*
9593          * if this is the last block group we have in this space, we can't
9594          * relocate it unless we're able to allocate a new chunk below.
9595          *
9596          * Otherwise, we need to make sure we have room in the space to handle
9597          * all of the extents from this block group.  If we can, we're good
9598          */
9599         if ((space_info->total_bytes != block_group->key.offset) &&
9600             (btrfs_space_info_used(space_info, false) + min_free <
9601              space_info->total_bytes)) {
9602                 spin_unlock(&space_info->lock);
9603                 goto out;
9604         }
9605         spin_unlock(&space_info->lock);
9606
9607         /*
9608          * ok we don't have enough space, but maybe we have free space on our
9609          * devices to allocate new chunks for relocation, so loop through our
9610          * alloc devices and guess if we have enough space.  if this block
9611          * group is going to be restriped, run checks against the target
9612          * profile instead of the current one.
9613          */
9614         ret = -1;
9615
9616         /*
9617          * index:
9618          *      0: raid10
9619          *      1: raid1
9620          *      2: dup
9621          *      3: raid0
9622          *      4: single
9623          */
9624         target = get_restripe_target(fs_info, block_group->flags);
9625         if (target) {
9626                 index = __get_raid_index(extended_to_chunk(target));
9627         } else {
9628                 /*
9629                  * this is just a balance, so if we were marked as full
9630                  * we know there is no space for a new chunk
9631                  */
9632                 if (full) {
9633                         if (debug)
9634                                 btrfs_warn(fs_info,
9635                                            "no space to alloc new chunk for block group %llu",
9636                                            block_group->key.objectid);
9637                         goto out;
9638                 }
9639
9640                 index = get_block_group_index(block_group);
9641         }
9642
9643         if (index == BTRFS_RAID_RAID10) {
9644                 dev_min = 4;
9645                 /* Divide by 2 */
9646                 min_free >>= 1;
9647         } else if (index == BTRFS_RAID_RAID1) {
9648                 dev_min = 2;
9649         } else if (index == BTRFS_RAID_DUP) {
9650                 /* Multiply by 2 */
9651                 min_free <<= 1;
9652         } else if (index == BTRFS_RAID_RAID0) {
9653                 dev_min = fs_devices->rw_devices;
9654                 min_free = div64_u64(min_free, dev_min);
9655         }
9656
9657         /* We need to do this so that we can look at pending chunks */
9658         trans = btrfs_join_transaction(root);
9659         if (IS_ERR(trans)) {
9660                 ret = PTR_ERR(trans);
9661                 goto out;
9662         }
9663
9664         mutex_lock(&fs_info->chunk_mutex);
9665         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
9666                 u64 dev_offset;
9667
9668                 /*
9669                  * check to make sure we can actually find a chunk with enough
9670                  * space to fit our block group in.
9671                  */
9672                 if (device->total_bytes > device->bytes_used + min_free &&
9673                     !device->is_tgtdev_for_dev_replace) {
9674                         ret = find_free_dev_extent(trans, device, min_free,
9675                                                    &dev_offset, NULL);
9676                         if (!ret)
9677                                 dev_nr++;
9678
9679                         if (dev_nr >= dev_min)
9680                                 break;
9681
9682                         ret = -1;
9683                 }
9684         }
9685         if (debug && ret == -1)
9686                 btrfs_warn(fs_info,
9687                            "no space to allocate a new chunk for block group %llu",
9688                            block_group->key.objectid);
9689         mutex_unlock(&fs_info->chunk_mutex);
9690         btrfs_end_transaction(trans);
9691 out:
9692         btrfs_put_block_group(block_group);
9693         return ret;
9694 }
9695
9696 static int find_first_block_group(struct btrfs_fs_info *fs_info,
9697                                   struct btrfs_path *path,
9698                                   struct btrfs_key *key)
9699 {
9700         struct btrfs_root *root = fs_info->extent_root;
9701         int ret = 0;
9702         struct btrfs_key found_key;
9703         struct extent_buffer *leaf;
9704         int slot;
9705
9706         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
9707         if (ret < 0)
9708                 goto out;
9709
9710         while (1) {
9711                 slot = path->slots[0];
9712                 leaf = path->nodes[0];
9713                 if (slot >= btrfs_header_nritems(leaf)) {
9714                         ret = btrfs_next_leaf(root, path);
9715                         if (ret == 0)
9716                                 continue;
9717                         if (ret < 0)
9718                                 goto out;
9719                         break;
9720                 }
9721                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
9722
9723                 if (found_key.objectid >= key->objectid &&
9724                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
9725                         struct extent_map_tree *em_tree;
9726                         struct extent_map *em;
9727
9728                         em_tree = &root->fs_info->mapping_tree.map_tree;
9729                         read_lock(&em_tree->lock);
9730                         em = lookup_extent_mapping(em_tree, found_key.objectid,
9731                                                    found_key.offset);
9732                         read_unlock(&em_tree->lock);
9733                         if (!em) {
9734                                 btrfs_err(fs_info,
9735                         "logical %llu len %llu found bg but no related chunk",
9736                                           found_key.objectid, found_key.offset);
9737                                 ret = -ENOENT;
9738                         } else {
9739                                 ret = 0;
9740                         }
9741                         free_extent_map(em);
9742                         goto out;
9743                 }
9744                 path->slots[0]++;
9745         }
9746 out:
9747         return ret;
9748 }
9749
9750 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
9751 {
9752         struct btrfs_block_group_cache *block_group;
9753         u64 last = 0;
9754
9755         while (1) {
9756                 struct inode *inode;
9757
9758                 block_group = btrfs_lookup_first_block_group(info, last);
9759                 while (block_group) {
9760                         spin_lock(&block_group->lock);
9761                         if (block_group->iref)
9762                                 break;
9763                         spin_unlock(&block_group->lock);
9764                         block_group = next_block_group(info, block_group);
9765                 }
9766                 if (!block_group) {
9767                         if (last == 0)
9768                                 break;
9769                         last = 0;
9770                         continue;
9771                 }
9772
9773                 inode = block_group->inode;
9774                 block_group->iref = 0;
9775                 block_group->inode = NULL;
9776                 spin_unlock(&block_group->lock);
9777                 ASSERT(block_group->io_ctl.inode == NULL);
9778                 iput(inode);
9779                 last = block_group->key.objectid + block_group->key.offset;
9780                 btrfs_put_block_group(block_group);
9781         }
9782 }
9783
9784 /*
9785  * Must be called only after stopping all workers, since we could have block
9786  * group caching kthreads running, and therefore they could race with us if we
9787  * freed the block groups before stopping them.
9788  */
9789 int btrfs_free_block_groups(struct btrfs_fs_info *info)
9790 {
9791         struct btrfs_block_group_cache *block_group;
9792         struct btrfs_space_info *space_info;
9793         struct btrfs_caching_control *caching_ctl;
9794         struct rb_node *n;
9795
9796         down_write(&info->commit_root_sem);
9797         while (!list_empty(&info->caching_block_groups)) {
9798                 caching_ctl = list_entry(info->caching_block_groups.next,
9799                                          struct btrfs_caching_control, list);
9800                 list_del(&caching_ctl->list);
9801                 put_caching_control(caching_ctl);
9802         }
9803         up_write(&info->commit_root_sem);
9804
9805         spin_lock(&info->unused_bgs_lock);
9806         while (!list_empty(&info->unused_bgs)) {
9807                 block_group = list_first_entry(&info->unused_bgs,
9808                                                struct btrfs_block_group_cache,
9809                                                bg_list);
9810                 list_del_init(&block_group->bg_list);
9811                 btrfs_put_block_group(block_group);
9812         }
9813         spin_unlock(&info->unused_bgs_lock);
9814
9815         spin_lock(&info->block_group_cache_lock);
9816         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
9817                 block_group = rb_entry(n, struct btrfs_block_group_cache,
9818                                        cache_node);
9819                 rb_erase(&block_group->cache_node,
9820                          &info->block_group_cache_tree);
9821                 RB_CLEAR_NODE(&block_group->cache_node);
9822                 spin_unlock(&info->block_group_cache_lock);
9823
9824                 down_write(&block_group->space_info->groups_sem);
9825                 list_del(&block_group->list);
9826                 up_write(&block_group->space_info->groups_sem);
9827
9828                 /*
9829                  * We haven't cached this block group, which means we could
9830                  * possibly have excluded extents on this block group.
9831                  */
9832                 if (block_group->cached == BTRFS_CACHE_NO ||
9833                     block_group->cached == BTRFS_CACHE_ERROR)
9834                         free_excluded_extents(info, block_group);
9835
9836                 btrfs_remove_free_space_cache(block_group);
9837                 ASSERT(block_group->cached != BTRFS_CACHE_STARTED);
9838                 ASSERT(list_empty(&block_group->dirty_list));
9839                 ASSERT(list_empty(&block_group->io_list));
9840                 ASSERT(list_empty(&block_group->bg_list));
9841                 ASSERT(atomic_read(&block_group->count) == 1);
9842                 btrfs_put_block_group(block_group);
9843
9844                 spin_lock(&info->block_group_cache_lock);
9845         }
9846         spin_unlock(&info->block_group_cache_lock);
9847
9848         /* now that all the block groups are freed, go through and
9849          * free all the space_info structs.  This is only called during
9850          * the final stages of unmount, and so we know nobody is
9851          * using them.  We call synchronize_rcu() once before we start,
9852          * just to be on the safe side.
9853          */
9854         synchronize_rcu();
9855
9856         release_global_block_rsv(info);
9857
9858         while (!list_empty(&info->space_info)) {
9859                 int i;
9860
9861                 space_info = list_entry(info->space_info.next,
9862                                         struct btrfs_space_info,
9863                                         list);
9864
9865                 /*
9866                  * Do not hide this behind enospc_debug, this is actually
9867                  * important and indicates a real bug if this happens.
9868                  */
9869                 if (WARN_ON(space_info->bytes_pinned > 0 ||
9870                             space_info->bytes_reserved > 0 ||
9871                             space_info->bytes_may_use > 0))
9872                         dump_space_info(info, space_info, 0, 0);
9873                 list_del(&space_info->list);
9874                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
9875                         struct kobject *kobj;
9876                         kobj = space_info->block_group_kobjs[i];
9877                         space_info->block_group_kobjs[i] = NULL;
9878                         if (kobj) {
9879                                 kobject_del(kobj);
9880                                 kobject_put(kobj);
9881                         }
9882                 }
9883                 kobject_del(&space_info->kobj);
9884                 kobject_put(&space_info->kobj);
9885         }
9886         return 0;
9887 }
9888
9889 static void __link_block_group(struct btrfs_space_info *space_info,
9890                                struct btrfs_block_group_cache *cache)
9891 {
9892         int index = get_block_group_index(cache);
9893         bool first = false;
9894
9895         down_write(&space_info->groups_sem);
9896         if (list_empty(&space_info->block_groups[index]))
9897                 first = true;
9898         list_add_tail(&cache->list, &space_info->block_groups[index]);
9899         up_write(&space_info->groups_sem);
9900
9901         if (first) {
9902                 struct raid_kobject *rkobj;
9903                 int ret;
9904
9905                 rkobj = kzalloc(sizeof(*rkobj), GFP_NOFS);
9906                 if (!rkobj)
9907                         goto out_err;
9908                 rkobj->raid_type = index;
9909                 kobject_init(&rkobj->kobj, &btrfs_raid_ktype);
9910                 ret = kobject_add(&rkobj->kobj, &space_info->kobj,
9911                                   "%s", get_raid_name(index));
9912                 if (ret) {
9913                         kobject_put(&rkobj->kobj);
9914                         goto out_err;
9915                 }
9916                 space_info->block_group_kobjs[index] = &rkobj->kobj;
9917         }
9918
9919         return;
9920 out_err:
9921         btrfs_warn(cache->fs_info,
9922                    "failed to add kobject for block cache, ignoring");
9923 }
9924
9925 static struct btrfs_block_group_cache *
9926 btrfs_create_block_group_cache(struct btrfs_fs_info *fs_info,
9927                                u64 start, u64 size)
9928 {
9929         struct btrfs_block_group_cache *cache;
9930
9931         cache = kzalloc(sizeof(*cache), GFP_NOFS);
9932         if (!cache)
9933                 return NULL;
9934
9935         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
9936                                         GFP_NOFS);
9937         if (!cache->free_space_ctl) {
9938                 kfree(cache);
9939                 return NULL;
9940         }
9941
9942         cache->key.objectid = start;
9943         cache->key.offset = size;
9944         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
9945
9946         cache->fs_info = fs_info;
9947         cache->full_stripe_len = btrfs_full_stripe_len(fs_info, start);
9948         set_free_space_tree_thresholds(cache);
9949
9950         atomic_set(&cache->count, 1);
9951         spin_lock_init(&cache->lock);
9952         init_rwsem(&cache->data_rwsem);
9953         INIT_LIST_HEAD(&cache->list);
9954         INIT_LIST_HEAD(&cache->cluster_list);
9955         INIT_LIST_HEAD(&cache->bg_list);
9956         INIT_LIST_HEAD(&cache->ro_list);
9957         INIT_LIST_HEAD(&cache->dirty_list);
9958         INIT_LIST_HEAD(&cache->io_list);
9959         btrfs_init_free_space_ctl(cache);
9960         atomic_set(&cache->trimming, 0);
9961         mutex_init(&cache->free_space_lock);
9962         btrfs_init_full_stripe_locks_tree(&cache->full_stripe_locks_root);
9963
9964         return cache;
9965 }
9966
9967 int btrfs_read_block_groups(struct btrfs_fs_info *info)
9968 {
9969         struct btrfs_path *path;
9970         int ret;
9971         struct btrfs_block_group_cache *cache;
9972         struct btrfs_space_info *space_info;
9973         struct btrfs_key key;
9974         struct btrfs_key found_key;
9975         struct extent_buffer *leaf;
9976         int need_clear = 0;
9977         u64 cache_gen;
9978         u64 feature;
9979         int mixed;
9980
9981         feature = btrfs_super_incompat_flags(info->super_copy);
9982         mixed = !!(feature & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS);
9983
9984         key.objectid = 0;
9985         key.offset = 0;
9986         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
9987         path = btrfs_alloc_path();
9988         if (!path)
9989                 return -ENOMEM;
9990         path->reada = READA_FORWARD;
9991
9992         cache_gen = btrfs_super_cache_generation(info->super_copy);
9993         if (btrfs_test_opt(info, SPACE_CACHE) &&
9994             btrfs_super_generation(info->super_copy) != cache_gen)
9995                 need_clear = 1;
9996         if (btrfs_test_opt(info, CLEAR_CACHE))
9997                 need_clear = 1;
9998
9999         while (1) {
10000                 ret = find_first_block_group(info, path, &key);
10001                 if (ret > 0)
10002                         break;
10003                 if (ret != 0)
10004                         goto error;
10005
10006                 leaf = path->nodes[0];
10007                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
10008
10009                 cache = btrfs_create_block_group_cache(info, found_key.objectid,
10010                                                        found_key.offset);
10011                 if (!cache) {
10012                         ret = -ENOMEM;
10013                         goto error;
10014                 }
10015
10016                 if (need_clear) {
10017                         /*
10018                          * When we mount with old space cache, we need to
10019                          * set BTRFS_DC_CLEAR and set dirty flag.
10020                          *
10021                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
10022                          *    truncate the old free space cache inode and
10023                          *    setup a new one.
10024                          * b) Setting 'dirty flag' makes sure that we flush
10025                          *    the new space cache info onto disk.
10026                          */
10027                         if (btrfs_test_opt(info, SPACE_CACHE))
10028                                 cache->disk_cache_state = BTRFS_DC_CLEAR;
10029                 }
10030
10031                 read_extent_buffer(leaf, &cache->item,
10032                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
10033                                    sizeof(cache->item));
10034                 cache->flags = btrfs_block_group_flags(&cache->item);
10035                 if (!mixed &&
10036                     ((cache->flags & BTRFS_BLOCK_GROUP_METADATA) &&
10037                     (cache->flags & BTRFS_BLOCK_GROUP_DATA))) {
10038                         btrfs_err(info,
10039 "bg %llu is a mixed block group but filesystem hasn't enabled mixed block groups",
10040                                   cache->key.objectid);
10041                         ret = -EINVAL;
10042                         goto error;
10043                 }
10044
10045                 key.objectid = found_key.objectid + found_key.offset;
10046                 btrfs_release_path(path);
10047
10048                 /*
10049                  * We need to exclude the super stripes now so that the space
10050                  * info has super bytes accounted for, otherwise we'll think
10051                  * we have more space than we actually do.
10052                  */
10053                 ret = exclude_super_stripes(info, cache);
10054                 if (ret) {
10055                         /*
10056                          * We may have excluded something, so call this just in
10057                          * case.
10058                          */
10059                         free_excluded_extents(info, cache);
10060                         btrfs_put_block_group(cache);
10061                         goto error;
10062                 }
10063
10064                 /*
10065                  * check for two cases, either we are full, and therefore
10066                  * don't need to bother with the caching work since we won't
10067                  * find any space, or we are empty, and we can just add all
10068                  * the space in and be done with it.  This saves us _alot_ of
10069                  * time, particularly in the full case.
10070                  */
10071                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
10072                         cache->last_byte_to_unpin = (u64)-1;
10073                         cache->cached = BTRFS_CACHE_FINISHED;
10074                         free_excluded_extents(info, cache);
10075                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10076                         cache->last_byte_to_unpin = (u64)-1;
10077                         cache->cached = BTRFS_CACHE_FINISHED;
10078                         add_new_free_space(cache, info,
10079                                            found_key.objectid,
10080                                            found_key.objectid +
10081                                            found_key.offset);
10082                         free_excluded_extents(info, cache);
10083                 }
10084
10085                 ret = btrfs_add_block_group_cache(info, cache);
10086                 if (ret) {
10087                         btrfs_remove_free_space_cache(cache);
10088                         btrfs_put_block_group(cache);
10089                         goto error;
10090                 }
10091
10092                 trace_btrfs_add_block_group(info, cache, 0);
10093                 update_space_info(info, cache->flags, found_key.offset,
10094                                   btrfs_block_group_used(&cache->item),
10095                                   cache->bytes_super, &space_info);
10096
10097                 cache->space_info = space_info;
10098
10099                 __link_block_group(space_info, cache);
10100
10101                 set_avail_alloc_bits(info, cache->flags);
10102                 if (btrfs_chunk_readonly(info, cache->key.objectid)) {
10103                         inc_block_group_ro(cache, 1);
10104                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10105                         spin_lock(&info->unused_bgs_lock);
10106                         /* Should always be true but just in case. */
10107                         if (list_empty(&cache->bg_list)) {
10108                                 btrfs_get_block_group(cache);
10109                                 list_add_tail(&cache->bg_list,
10110                                               &info->unused_bgs);
10111                         }
10112                         spin_unlock(&info->unused_bgs_lock);
10113                 }
10114         }
10115
10116         list_for_each_entry_rcu(space_info, &info->space_info, list) {
10117                 if (!(get_alloc_profile(info, space_info->flags) &
10118                       (BTRFS_BLOCK_GROUP_RAID10 |
10119                        BTRFS_BLOCK_GROUP_RAID1 |
10120                        BTRFS_BLOCK_GROUP_RAID5 |
10121                        BTRFS_BLOCK_GROUP_RAID6 |
10122                        BTRFS_BLOCK_GROUP_DUP)))
10123                         continue;
10124                 /*
10125                  * avoid allocating from un-mirrored block group if there are
10126                  * mirrored block groups.
10127                  */
10128                 list_for_each_entry(cache,
10129                                 &space_info->block_groups[BTRFS_RAID_RAID0],
10130                                 list)
10131                         inc_block_group_ro(cache, 1);
10132                 list_for_each_entry(cache,
10133                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
10134                                 list)
10135                         inc_block_group_ro(cache, 1);
10136         }
10137
10138         init_global_block_rsv(info);
10139         ret = 0;
10140 error:
10141         btrfs_free_path(path);
10142         return ret;
10143 }
10144
10145 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans,
10146                                        struct btrfs_fs_info *fs_info)
10147 {
10148         struct btrfs_block_group_cache *block_group, *tmp;
10149         struct btrfs_root *extent_root = fs_info->extent_root;
10150         struct btrfs_block_group_item item;
10151         struct btrfs_key key;
10152         int ret = 0;
10153         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
10154
10155         trans->can_flush_pending_bgs = false;
10156         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) {
10157                 if (ret)
10158                         goto next;
10159
10160                 spin_lock(&block_group->lock);
10161                 memcpy(&item, &block_group->item, sizeof(item));
10162                 memcpy(&key, &block_group->key, sizeof(key));
10163                 spin_unlock(&block_group->lock);
10164
10165                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
10166                                         sizeof(item));
10167                 if (ret)
10168                         btrfs_abort_transaction(trans, ret);
10169                 ret = btrfs_finish_chunk_alloc(trans, fs_info, key.objectid,
10170                                                key.offset);
10171                 if (ret)
10172                         btrfs_abort_transaction(trans, ret);
10173                 add_block_group_free_space(trans, fs_info, block_group);
10174                 /* already aborted the transaction if it failed. */
10175 next:
10176                 list_del_init(&block_group->bg_list);
10177         }
10178         trans->can_flush_pending_bgs = can_flush_pending_bgs;
10179 }
10180
10181 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
10182                            struct btrfs_fs_info *fs_info, u64 bytes_used,
10183                            u64 type, u64 chunk_offset, u64 size)
10184 {
10185         struct btrfs_block_group_cache *cache;
10186         int ret;
10187
10188         btrfs_set_log_full_commit(fs_info, trans);
10189
10190         cache = btrfs_create_block_group_cache(fs_info, chunk_offset, size);
10191         if (!cache)
10192                 return -ENOMEM;
10193
10194         btrfs_set_block_group_used(&cache->item, bytes_used);
10195         btrfs_set_block_group_chunk_objectid(&cache->item,
10196                                              BTRFS_FIRST_CHUNK_TREE_OBJECTID);
10197         btrfs_set_block_group_flags(&cache->item, type);
10198
10199         cache->flags = type;
10200         cache->last_byte_to_unpin = (u64)-1;
10201         cache->cached = BTRFS_CACHE_FINISHED;
10202         cache->needs_free_space = 1;
10203         ret = exclude_super_stripes(fs_info, cache);
10204         if (ret) {
10205                 /*
10206                  * We may have excluded something, so call this just in
10207                  * case.
10208                  */
10209                 free_excluded_extents(fs_info, cache);
10210                 btrfs_put_block_group(cache);
10211                 return ret;
10212         }
10213
10214         add_new_free_space(cache, fs_info, chunk_offset, chunk_offset + size);
10215
10216         free_excluded_extents(fs_info, cache);
10217
10218 #ifdef CONFIG_BTRFS_DEBUG
10219         if (btrfs_should_fragment_free_space(cache)) {
10220                 u64 new_bytes_used = size - bytes_used;
10221
10222                 bytes_used += new_bytes_used >> 1;
10223                 fragment_free_space(cache);
10224         }
10225 #endif
10226         /*
10227          * Ensure the corresponding space_info object is created and
10228          * assigned to our block group. We want our bg to be added to the rbtree
10229          * with its ->space_info set.
10230          */
10231         cache->space_info = __find_space_info(fs_info, cache->flags);
10232         if (!cache->space_info) {
10233                 ret = create_space_info(fs_info, cache->flags,
10234                                        &cache->space_info);
10235                 if (ret) {
10236                         btrfs_remove_free_space_cache(cache);
10237                         btrfs_put_block_group(cache);
10238                         return ret;
10239                 }
10240         }
10241
10242         ret = btrfs_add_block_group_cache(fs_info, cache);
10243         if (ret) {
10244                 btrfs_remove_free_space_cache(cache);
10245                 btrfs_put_block_group(cache);
10246                 return ret;
10247         }
10248
10249         /*
10250          * Now that our block group has its ->space_info set and is inserted in
10251          * the rbtree, update the space info's counters.
10252          */
10253         trace_btrfs_add_block_group(fs_info, cache, 1);
10254         update_space_info(fs_info, cache->flags, size, bytes_used,
10255                                 cache->bytes_super, &cache->space_info);
10256         update_global_block_rsv(fs_info);
10257
10258         __link_block_group(cache->space_info, cache);
10259
10260         list_add_tail(&cache->bg_list, &trans->new_bgs);
10261
10262         set_avail_alloc_bits(fs_info, type);
10263         return 0;
10264 }
10265
10266 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
10267 {
10268         u64 extra_flags = chunk_to_extended(flags) &
10269                                 BTRFS_EXTENDED_PROFILE_MASK;
10270
10271         write_seqlock(&fs_info->profiles_lock);
10272         if (flags & BTRFS_BLOCK_GROUP_DATA)
10273                 fs_info->avail_data_alloc_bits &= ~extra_flags;
10274         if (flags & BTRFS_BLOCK_GROUP_METADATA)
10275                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
10276         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
10277                 fs_info->avail_system_alloc_bits &= ~extra_flags;
10278         write_sequnlock(&fs_info->profiles_lock);
10279 }
10280
10281 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
10282                              struct btrfs_fs_info *fs_info, u64 group_start,
10283                              struct extent_map *em)
10284 {
10285         struct btrfs_root *root = fs_info->extent_root;
10286         struct btrfs_path *path;
10287         struct btrfs_block_group_cache *block_group;
10288         struct btrfs_free_cluster *cluster;
10289         struct btrfs_root *tree_root = fs_info->tree_root;
10290         struct btrfs_key key;
10291         struct inode *inode;
10292         struct kobject *kobj = NULL;
10293         int ret;
10294         int index;
10295         int factor;
10296         struct btrfs_caching_control *caching_ctl = NULL;
10297         bool remove_em;
10298
10299         block_group = btrfs_lookup_block_group(fs_info, group_start);
10300         BUG_ON(!block_group);
10301         BUG_ON(!block_group->ro);
10302
10303         /*
10304          * Free the reserved super bytes from this block group before
10305          * remove it.
10306          */
10307         free_excluded_extents(fs_info, block_group);
10308
10309         memcpy(&key, &block_group->key, sizeof(key));
10310         index = get_block_group_index(block_group);
10311         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
10312                                   BTRFS_BLOCK_GROUP_RAID1 |
10313                                   BTRFS_BLOCK_GROUP_RAID10))
10314                 factor = 2;
10315         else
10316                 factor = 1;
10317
10318         /* make sure this block group isn't part of an allocation cluster */
10319         cluster = &fs_info->data_alloc_cluster;
10320         spin_lock(&cluster->refill_lock);
10321         btrfs_return_cluster_to_free_space(block_group, cluster);
10322         spin_unlock(&cluster->refill_lock);
10323
10324         /*
10325          * make sure this block group isn't part of a metadata
10326          * allocation cluster
10327          */
10328         cluster = &fs_info->meta_alloc_cluster;
10329         spin_lock(&cluster->refill_lock);
10330         btrfs_return_cluster_to_free_space(block_group, cluster);
10331         spin_unlock(&cluster->refill_lock);
10332
10333         path = btrfs_alloc_path();
10334         if (!path) {
10335                 ret = -ENOMEM;
10336                 goto out;
10337         }
10338
10339         /*
10340          * get the inode first so any iput calls done for the io_list
10341          * aren't the final iput (no unlinks allowed now)
10342          */
10343         inode = lookup_free_space_inode(fs_info, block_group, path);
10344
10345         mutex_lock(&trans->transaction->cache_write_mutex);
10346         /*
10347          * make sure our free spache cache IO is done before remove the
10348          * free space inode
10349          */
10350         spin_lock(&trans->transaction->dirty_bgs_lock);
10351         if (!list_empty(&block_group->io_list)) {
10352                 list_del_init(&block_group->io_list);
10353
10354                 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
10355
10356                 spin_unlock(&trans->transaction->dirty_bgs_lock);
10357                 btrfs_wait_cache_io(trans, block_group, path);
10358                 btrfs_put_block_group(block_group);
10359                 spin_lock(&trans->transaction->dirty_bgs_lock);
10360         }
10361
10362         if (!list_empty(&block_group->dirty_list)) {
10363                 list_del_init(&block_group->dirty_list);
10364                 btrfs_put_block_group(block_group);
10365         }
10366         spin_unlock(&trans->transaction->dirty_bgs_lock);
10367         mutex_unlock(&trans->transaction->cache_write_mutex);
10368
10369         if (!IS_ERR(inode)) {
10370                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
10371                 if (ret) {
10372                         btrfs_add_delayed_iput(inode);
10373                         goto out;
10374                 }
10375                 clear_nlink(inode);
10376                 /* One for the block groups ref */
10377                 spin_lock(&block_group->lock);
10378                 if (block_group->iref) {
10379                         block_group->iref = 0;
10380                         block_group->inode = NULL;
10381                         spin_unlock(&block_group->lock);
10382                         iput(inode);
10383                 } else {
10384                         spin_unlock(&block_group->lock);
10385                 }
10386                 /* One for our lookup ref */
10387                 btrfs_add_delayed_iput(inode);
10388         }
10389
10390         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
10391         key.offset = block_group->key.objectid;
10392         key.type = 0;
10393
10394         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
10395         if (ret < 0)
10396                 goto out;
10397         if (ret > 0)
10398                 btrfs_release_path(path);
10399         if (ret == 0) {
10400                 ret = btrfs_del_item(trans, tree_root, path);
10401                 if (ret)
10402                         goto out;
10403                 btrfs_release_path(path);
10404         }
10405
10406         spin_lock(&fs_info->block_group_cache_lock);
10407         rb_erase(&block_group->cache_node,
10408                  &fs_info->block_group_cache_tree);
10409         RB_CLEAR_NODE(&block_group->cache_node);
10410
10411         if (fs_info->first_logical_byte == block_group->key.objectid)
10412                 fs_info->first_logical_byte = (u64)-1;
10413         spin_unlock(&fs_info->block_group_cache_lock);
10414
10415         down_write(&block_group->space_info->groups_sem);
10416         /*
10417          * we must use list_del_init so people can check to see if they
10418          * are still on the list after taking the semaphore
10419          */
10420         list_del_init(&block_group->list);
10421         if (list_empty(&block_group->space_info->block_groups[index])) {
10422                 kobj = block_group->space_info->block_group_kobjs[index];
10423                 block_group->space_info->block_group_kobjs[index] = NULL;
10424                 clear_avail_alloc_bits(fs_info, block_group->flags);
10425         }
10426         up_write(&block_group->space_info->groups_sem);
10427         if (kobj) {
10428                 kobject_del(kobj);
10429                 kobject_put(kobj);
10430         }
10431
10432         if (block_group->has_caching_ctl)
10433                 caching_ctl = get_caching_control(block_group);
10434         if (block_group->cached == BTRFS_CACHE_STARTED)
10435                 wait_block_group_cache_done(block_group);
10436         if (block_group->has_caching_ctl) {
10437                 down_write(&fs_info->commit_root_sem);
10438                 if (!caching_ctl) {
10439                         struct btrfs_caching_control *ctl;
10440
10441                         list_for_each_entry(ctl,
10442                                     &fs_info->caching_block_groups, list)
10443                                 if (ctl->block_group == block_group) {
10444                                         caching_ctl = ctl;
10445                                         refcount_inc(&caching_ctl->count);
10446                                         break;
10447                                 }
10448                 }
10449                 if (caching_ctl)
10450                         list_del_init(&caching_ctl->list);
10451                 up_write(&fs_info->commit_root_sem);
10452                 if (caching_ctl) {
10453                         /* Once for the caching bgs list and once for us. */
10454                         put_caching_control(caching_ctl);
10455                         put_caching_control(caching_ctl);
10456                 }
10457         }
10458
10459         spin_lock(&trans->transaction->dirty_bgs_lock);
10460         if (!list_empty(&block_group->dirty_list)) {
10461                 WARN_ON(1);
10462         }
10463         if (!list_empty(&block_group->io_list)) {
10464                 WARN_ON(1);
10465         }
10466         spin_unlock(&trans->transaction->dirty_bgs_lock);
10467         btrfs_remove_free_space_cache(block_group);
10468
10469         spin_lock(&block_group->space_info->lock);
10470         list_del_init(&block_group->ro_list);
10471
10472         if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
10473                 WARN_ON(block_group->space_info->total_bytes
10474                         < block_group->key.offset);
10475                 WARN_ON(block_group->space_info->bytes_readonly
10476                         < block_group->key.offset);
10477                 WARN_ON(block_group->space_info->disk_total
10478                         < block_group->key.offset * factor);
10479         }
10480         block_group->space_info->total_bytes -= block_group->key.offset;
10481         block_group->space_info->bytes_readonly -= block_group->key.offset;
10482         block_group->space_info->disk_total -= block_group->key.offset * factor;
10483
10484         spin_unlock(&block_group->space_info->lock);
10485
10486         memcpy(&key, &block_group->key, sizeof(key));
10487
10488         mutex_lock(&fs_info->chunk_mutex);
10489         if (!list_empty(&em->list)) {
10490                 /* We're in the transaction->pending_chunks list. */
10491                 free_extent_map(em);
10492         }
10493         spin_lock(&block_group->lock);
10494         block_group->removed = 1;
10495         /*
10496          * At this point trimming can't start on this block group, because we
10497          * removed the block group from the tree fs_info->block_group_cache_tree
10498          * so no one can't find it anymore and even if someone already got this
10499          * block group before we removed it from the rbtree, they have already
10500          * incremented block_group->trimming - if they didn't, they won't find
10501          * any free space entries because we already removed them all when we
10502          * called btrfs_remove_free_space_cache().
10503          *
10504          * And we must not remove the extent map from the fs_info->mapping_tree
10505          * to prevent the same logical address range and physical device space
10506          * ranges from being reused for a new block group. This is because our
10507          * fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
10508          * completely transactionless, so while it is trimming a range the
10509          * currently running transaction might finish and a new one start,
10510          * allowing for new block groups to be created that can reuse the same
10511          * physical device locations unless we take this special care.
10512          *
10513          * There may also be an implicit trim operation if the file system
10514          * is mounted with -odiscard. The same protections must remain
10515          * in place until the extents have been discarded completely when
10516          * the transaction commit has completed.
10517          */
10518         remove_em = (atomic_read(&block_group->trimming) == 0);
10519         /*
10520          * Make sure a trimmer task always sees the em in the pinned_chunks list
10521          * if it sees block_group->removed == 1 (needs to lock block_group->lock
10522          * before checking block_group->removed).
10523          */
10524         if (!remove_em) {
10525                 /*
10526                  * Our em might be in trans->transaction->pending_chunks which
10527                  * is protected by fs_info->chunk_mutex ([lock|unlock]_chunks),
10528                  * and so is the fs_info->pinned_chunks list.
10529                  *
10530                  * So at this point we must be holding the chunk_mutex to avoid
10531                  * any races with chunk allocation (more specifically at
10532                  * volumes.c:contains_pending_extent()), to ensure it always
10533                  * sees the em, either in the pending_chunks list or in the
10534                  * pinned_chunks list.
10535                  */
10536                 list_move_tail(&em->list, &fs_info->pinned_chunks);
10537         }
10538         spin_unlock(&block_group->lock);
10539
10540         if (remove_em) {
10541                 struct extent_map_tree *em_tree;
10542
10543                 em_tree = &fs_info->mapping_tree.map_tree;
10544                 write_lock(&em_tree->lock);
10545                 /*
10546                  * The em might be in the pending_chunks list, so make sure the
10547                  * chunk mutex is locked, since remove_extent_mapping() will
10548                  * delete us from that list.
10549                  */
10550                 remove_extent_mapping(em_tree, em);
10551                 write_unlock(&em_tree->lock);
10552                 /* once for the tree */
10553                 free_extent_map(em);
10554         }
10555
10556         mutex_unlock(&fs_info->chunk_mutex);
10557
10558         ret = remove_block_group_free_space(trans, fs_info, block_group);
10559         if (ret)
10560                 goto out;
10561
10562         btrfs_put_block_group(block_group);
10563         btrfs_put_block_group(block_group);
10564
10565         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
10566         if (ret > 0)
10567                 ret = -EIO;
10568         if (ret < 0)
10569                 goto out;
10570
10571         ret = btrfs_del_item(trans, root, path);
10572 out:
10573         btrfs_free_path(path);
10574         return ret;
10575 }
10576
10577 struct btrfs_trans_handle *
10578 btrfs_start_trans_remove_block_group(struct btrfs_fs_info *fs_info,
10579                                      const u64 chunk_offset)
10580 {
10581         struct extent_map_tree *em_tree = &fs_info->mapping_tree.map_tree;
10582         struct extent_map *em;
10583         struct map_lookup *map;
10584         unsigned int num_items;
10585
10586         read_lock(&em_tree->lock);
10587         em = lookup_extent_mapping(em_tree, chunk_offset, 1);
10588         read_unlock(&em_tree->lock);
10589         ASSERT(em && em->start == chunk_offset);
10590
10591         /*
10592          * We need to reserve 3 + N units from the metadata space info in order
10593          * to remove a block group (done at btrfs_remove_chunk() and at
10594          * btrfs_remove_block_group()), which are used for:
10595          *
10596          * 1 unit for adding the free space inode's orphan (located in the tree
10597          * of tree roots).
10598          * 1 unit for deleting the block group item (located in the extent
10599          * tree).
10600          * 1 unit for deleting the free space item (located in tree of tree
10601          * roots).
10602          * N units for deleting N device extent items corresponding to each
10603          * stripe (located in the device tree).
10604          *
10605          * In order to remove a block group we also need to reserve units in the
10606          * system space info in order to update the chunk tree (update one or
10607          * more device items and remove one chunk item), but this is done at
10608          * btrfs_remove_chunk() through a call to check_system_chunk().
10609          */
10610         map = em->map_lookup;
10611         num_items = 3 + map->num_stripes;
10612         free_extent_map(em);
10613
10614         return btrfs_start_transaction_fallback_global_rsv(fs_info->extent_root,
10615                                                            num_items, 1);
10616 }
10617
10618 /*
10619  * Process the unused_bgs list and remove any that don't have any allocated
10620  * space inside of them.
10621  */
10622 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
10623 {
10624         struct btrfs_block_group_cache *block_group;
10625         struct btrfs_space_info *space_info;
10626         struct btrfs_trans_handle *trans;
10627         int ret = 0;
10628
10629         if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
10630                 return;
10631
10632         spin_lock(&fs_info->unused_bgs_lock);
10633         while (!list_empty(&fs_info->unused_bgs)) {
10634                 u64 start, end;
10635                 int trimming;
10636
10637                 block_group = list_first_entry(&fs_info->unused_bgs,
10638                                                struct btrfs_block_group_cache,
10639                                                bg_list);
10640                 list_del_init(&block_group->bg_list);
10641
10642                 space_info = block_group->space_info;
10643
10644                 if (ret || btrfs_mixed_space_info(space_info)) {
10645                         btrfs_put_block_group(block_group);
10646                         continue;
10647                 }
10648                 spin_unlock(&fs_info->unused_bgs_lock);
10649
10650                 mutex_lock(&fs_info->delete_unused_bgs_mutex);
10651
10652                 /* Don't want to race with allocators so take the groups_sem */
10653                 down_write(&space_info->groups_sem);
10654                 spin_lock(&block_group->lock);
10655                 if (block_group->reserved ||
10656                     btrfs_block_group_used(&block_group->item) ||
10657                     block_group->ro ||
10658                     list_is_singular(&block_group->list)) {
10659                         /*
10660                          * We want to bail if we made new allocations or have
10661                          * outstanding allocations in this block group.  We do
10662                          * the ro check in case balance is currently acting on
10663                          * this block group.
10664                          */
10665                         spin_unlock(&block_group->lock);
10666                         up_write(&space_info->groups_sem);
10667                         goto next;
10668                 }
10669                 spin_unlock(&block_group->lock);
10670
10671                 /* We don't want to force the issue, only flip if it's ok. */
10672                 ret = inc_block_group_ro(block_group, 0);
10673                 up_write(&space_info->groups_sem);
10674                 if (ret < 0) {
10675                         ret = 0;
10676                         goto next;
10677                 }
10678
10679                 /*
10680                  * Want to do this before we do anything else so we can recover
10681                  * properly if we fail to join the transaction.
10682                  */
10683                 trans = btrfs_start_trans_remove_block_group(fs_info,
10684                                                      block_group->key.objectid);
10685                 if (IS_ERR(trans)) {
10686                         btrfs_dec_block_group_ro(block_group);
10687                         ret = PTR_ERR(trans);
10688                         goto next;
10689                 }
10690
10691                 /*
10692                  * We could have pending pinned extents for this block group,
10693                  * just delete them, we don't care about them anymore.
10694                  */
10695                 start = block_group->key.objectid;
10696                 end = start + block_group->key.offset - 1;
10697                 /*
10698                  * Hold the unused_bg_unpin_mutex lock to avoid racing with
10699                  * btrfs_finish_extent_commit(). If we are at transaction N,
10700                  * another task might be running finish_extent_commit() for the
10701                  * previous transaction N - 1, and have seen a range belonging
10702                  * to the block group in freed_extents[] before we were able to
10703                  * clear the whole block group range from freed_extents[]. This
10704                  * means that task can lookup for the block group after we
10705                  * unpinned it from freed_extents[] and removed it, leading to
10706                  * a BUG_ON() at btrfs_unpin_extent_range().
10707                  */
10708                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
10709                 ret = clear_extent_bits(&fs_info->freed_extents[0], start, end,
10710                                   EXTENT_DIRTY);
10711                 if (ret) {
10712                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10713                         btrfs_dec_block_group_ro(block_group);
10714                         goto end_trans;
10715                 }
10716                 ret = clear_extent_bits(&fs_info->freed_extents[1], start, end,
10717                                   EXTENT_DIRTY);
10718                 if (ret) {
10719                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10720                         btrfs_dec_block_group_ro(block_group);
10721                         goto end_trans;
10722                 }
10723                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10724
10725                 /* Reset pinned so btrfs_put_block_group doesn't complain */
10726                 spin_lock(&space_info->lock);
10727                 spin_lock(&block_group->lock);
10728
10729                 space_info->bytes_pinned -= block_group->pinned;
10730                 space_info->bytes_readonly += block_group->pinned;
10731                 percpu_counter_add(&space_info->total_bytes_pinned,
10732                                    -block_group->pinned);
10733                 block_group->pinned = 0;
10734
10735                 spin_unlock(&block_group->lock);
10736                 spin_unlock(&space_info->lock);
10737
10738                 /* DISCARD can flip during remount */
10739                 trimming = btrfs_test_opt(fs_info, DISCARD);
10740
10741                 /* Implicit trim during transaction commit. */
10742                 if (trimming)
10743                         btrfs_get_block_group_trimming(block_group);
10744
10745                 /*
10746                  * Btrfs_remove_chunk will abort the transaction if things go
10747                  * horribly wrong.
10748                  */
10749                 ret = btrfs_remove_chunk(trans, fs_info,
10750                                          block_group->key.objectid);
10751
10752                 if (ret) {
10753                         if (trimming)
10754                                 btrfs_put_block_group_trimming(block_group);
10755                         goto end_trans;
10756                 }
10757
10758                 /*
10759                  * If we're not mounted with -odiscard, we can just forget
10760                  * about this block group. Otherwise we'll need to wait
10761                  * until transaction commit to do the actual discard.
10762                  */
10763                 if (trimming) {
10764                         spin_lock(&fs_info->unused_bgs_lock);
10765                         /*
10766                          * A concurrent scrub might have added us to the list
10767                          * fs_info->unused_bgs, so use a list_move operation
10768                          * to add the block group to the deleted_bgs list.
10769                          */
10770                         list_move(&block_group->bg_list,
10771                                   &trans->transaction->deleted_bgs);
10772                         spin_unlock(&fs_info->unused_bgs_lock);
10773                         btrfs_get_block_group(block_group);
10774                 }
10775 end_trans:
10776                 btrfs_end_transaction(trans);
10777 next:
10778                 mutex_unlock(&fs_info->delete_unused_bgs_mutex);
10779                 btrfs_put_block_group(block_group);
10780                 spin_lock(&fs_info->unused_bgs_lock);
10781         }
10782         spin_unlock(&fs_info->unused_bgs_lock);
10783 }
10784
10785 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
10786 {
10787         struct btrfs_space_info *space_info;
10788         struct btrfs_super_block *disk_super;
10789         u64 features;
10790         u64 flags;
10791         int mixed = 0;
10792         int ret;
10793
10794         disk_super = fs_info->super_copy;
10795         if (!btrfs_super_root(disk_super))
10796                 return -EINVAL;
10797
10798         features = btrfs_super_incompat_flags(disk_super);
10799         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
10800                 mixed = 1;
10801
10802         flags = BTRFS_BLOCK_GROUP_SYSTEM;
10803         ret = create_space_info(fs_info, flags, &space_info);
10804         if (ret)
10805                 goto out;
10806
10807         if (mixed) {
10808                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
10809                 ret = create_space_info(fs_info, flags, &space_info);
10810         } else {
10811                 flags = BTRFS_BLOCK_GROUP_METADATA;
10812                 ret = create_space_info(fs_info, flags, &space_info);
10813                 if (ret)
10814                         goto out;
10815
10816                 flags = BTRFS_BLOCK_GROUP_DATA;
10817                 ret = create_space_info(fs_info, flags, &space_info);
10818         }
10819 out:
10820         return ret;
10821 }
10822
10823 int btrfs_error_unpin_extent_range(struct btrfs_fs_info *fs_info,
10824                                    u64 start, u64 end)
10825 {
10826         return unpin_extent_range(fs_info, start, end, false);
10827 }
10828
10829 /*
10830  * It used to be that old block groups would be left around forever.
10831  * Iterating over them would be enough to trim unused space.  Since we
10832  * now automatically remove them, we also need to iterate over unallocated
10833  * space.
10834  *
10835  * We don't want a transaction for this since the discard may take a
10836  * substantial amount of time.  We don't require that a transaction be
10837  * running, but we do need to take a running transaction into account
10838  * to ensure that we're not discarding chunks that were released in
10839  * the current transaction.
10840  *
10841  * Holding the chunks lock will prevent other threads from allocating
10842  * or releasing chunks, but it won't prevent a running transaction
10843  * from committing and releasing the memory that the pending chunks
10844  * list head uses.  For that, we need to take a reference to the
10845  * transaction.
10846  */
10847 static int btrfs_trim_free_extents(struct btrfs_device *device,
10848                                    u64 minlen, u64 *trimmed)
10849 {
10850         u64 start = 0, len = 0;
10851         int ret;
10852
10853         *trimmed = 0;
10854
10855         /* Not writeable = nothing to do. */
10856         if (!device->writeable)
10857                 return 0;
10858
10859         /* No free space = nothing to do. */
10860         if (device->total_bytes <= device->bytes_used)
10861                 return 0;
10862
10863         ret = 0;
10864
10865         while (1) {
10866                 struct btrfs_fs_info *fs_info = device->fs_info;
10867                 struct btrfs_transaction *trans;
10868                 u64 bytes;
10869
10870                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
10871                 if (ret)
10872                         return ret;
10873
10874                 down_read(&fs_info->commit_root_sem);
10875
10876                 spin_lock(&fs_info->trans_lock);
10877                 trans = fs_info->running_transaction;
10878                 if (trans)
10879                         refcount_inc(&trans->use_count);
10880                 spin_unlock(&fs_info->trans_lock);
10881
10882                 ret = find_free_dev_extent_start(trans, device, minlen, start,
10883                                                  &start, &len);
10884                 if (trans)
10885                         btrfs_put_transaction(trans);
10886
10887                 if (ret) {
10888                         up_read(&fs_info->commit_root_sem);
10889                         mutex_unlock(&fs_info->chunk_mutex);
10890                         if (ret == -ENOSPC)
10891                                 ret = 0;
10892                         break;
10893                 }
10894
10895                 ret = btrfs_issue_discard(device->bdev, start, len, &bytes);
10896                 up_read(&fs_info->commit_root_sem);
10897                 mutex_unlock(&fs_info->chunk_mutex);
10898
10899                 if (ret)
10900                         break;
10901
10902                 start += len;
10903                 *trimmed += bytes;
10904
10905                 if (fatal_signal_pending(current)) {
10906                         ret = -ERESTARTSYS;
10907                         break;
10908                 }
10909
10910                 cond_resched();
10911         }
10912
10913         return ret;
10914 }
10915
10916 int btrfs_trim_fs(struct btrfs_fs_info *fs_info, struct fstrim_range *range)
10917 {
10918         struct btrfs_block_group_cache *cache = NULL;
10919         struct btrfs_device *device;
10920         struct list_head *devices;
10921         u64 group_trimmed;
10922         u64 start;
10923         u64 end;
10924         u64 trimmed = 0;
10925         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
10926         int ret = 0;
10927
10928         /*
10929          * try to trim all FS space, our block group may start from non-zero.
10930          */
10931         if (range->len == total_bytes)
10932                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
10933         else
10934                 cache = btrfs_lookup_block_group(fs_info, range->start);
10935
10936         while (cache) {
10937                 if (cache->key.objectid >= (range->start + range->len)) {
10938                         btrfs_put_block_group(cache);
10939                         break;
10940                 }
10941
10942                 start = max(range->start, cache->key.objectid);
10943                 end = min(range->start + range->len,
10944                                 cache->key.objectid + cache->key.offset);
10945
10946                 if (end - start >= range->minlen) {
10947                         if (!block_group_cache_done(cache)) {
10948                                 ret = cache_block_group(cache, 0);
10949                                 if (ret) {
10950                                         btrfs_put_block_group(cache);
10951                                         break;
10952                                 }
10953                                 ret = wait_block_group_cache_done(cache);
10954                                 if (ret) {
10955                                         btrfs_put_block_group(cache);
10956                                         break;
10957                                 }
10958                         }
10959                         ret = btrfs_trim_block_group(cache,
10960                                                      &group_trimmed,
10961                                                      start,
10962                                                      end,
10963                                                      range->minlen);
10964
10965                         trimmed += group_trimmed;
10966                         if (ret) {
10967                                 btrfs_put_block_group(cache);
10968                                 break;
10969                         }
10970                 }
10971
10972                 cache = next_block_group(fs_info, cache);
10973         }
10974
10975         mutex_lock(&fs_info->fs_devices->device_list_mutex);
10976         devices = &fs_info->fs_devices->alloc_list;
10977         list_for_each_entry(device, devices, dev_alloc_list) {
10978                 ret = btrfs_trim_free_extents(device, range->minlen,
10979                                               &group_trimmed);
10980                 if (ret)
10981                         break;
10982
10983                 trimmed += group_trimmed;
10984         }
10985         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
10986
10987         range->len = trimmed;
10988         return ret;
10989 }
10990
10991 /*
10992  * btrfs_{start,end}_write_no_snapshotting() are similar to
10993  * mnt_{want,drop}_write(), they are used to prevent some tasks from writing
10994  * data into the page cache through nocow before the subvolume is snapshoted,
10995  * but flush the data into disk after the snapshot creation, or to prevent
10996  * operations while snapshotting is ongoing and that cause the snapshot to be
10997  * inconsistent (writes followed by expanding truncates for example).
10998  */
10999 void btrfs_end_write_no_snapshotting(struct btrfs_root *root)
11000 {
11001         percpu_counter_dec(&root->subv_writers->counter);
11002         /*
11003          * Make sure counter is updated before we wake up waiters.
11004          */
11005         smp_mb();
11006         if (waitqueue_active(&root->subv_writers->wait))
11007                 wake_up(&root->subv_writers->wait);
11008 }
11009
11010 int btrfs_start_write_no_snapshotting(struct btrfs_root *root)
11011 {
11012         if (atomic_read(&root->will_be_snapshotted))
11013                 return 0;
11014
11015         percpu_counter_inc(&root->subv_writers->counter);
11016         /*
11017          * Make sure counter is updated before we check for snapshot creation.
11018          */
11019         smp_mb();
11020         if (atomic_read(&root->will_be_snapshotted)) {
11021                 btrfs_end_write_no_snapshotting(root);
11022                 return 0;
11023         }
11024         return 1;
11025 }
11026
11027 static int wait_snapshotting_atomic_t(atomic_t *a)
11028 {
11029         schedule();
11030         return 0;
11031 }
11032
11033 void btrfs_wait_for_snapshot_creation(struct btrfs_root *root)
11034 {
11035         while (true) {
11036                 int ret;
11037
11038                 ret = btrfs_start_write_no_snapshotting(root);
11039                 if (ret)
11040                         break;
11041                 wait_on_atomic_t(&root->will_be_snapshotted,
11042                                  wait_snapshotting_atomic_t,
11043                                  TASK_UNINTERRUPTIBLE);
11044         }
11045 }