fs/btrfs/extent-tree.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2007 Oracle.  All rights reserved.
   4  */
   5
   6 #include <linux/sched.h>
   7 #include <linux/sched/signal.h>
   8 #include <linux/pagemap.h>
   9 #include <linux/writeback.h>
  10 #include <linux/blkdev.h>
  11 #include <linux/sort.h>
  12 #include <linux/rcupdate.h>
  13 #include <linux/kthread.h>
  14 #include <linux/slab.h>
  15 #include <linux/ratelimit.h>
  16 #include <linux/percpu_counter.h>
  17 #include <linux/lockdep.h>
  18 #include <linux/crc32c.h>
  19 #include "tree-log.h"
  20 #include "disk-io.h"
  21 #include "print-tree.h"
  22 #include "volumes.h"
  23 #include "raid56.h"
  24 #include "locking.h"
  25 #include "free-space-cache.h"
  26 #include "free-space-tree.h"
  27 #include "math.h"
  28 #include "sysfs.h"
  29 #include "qgroup.h"
  30 #include "ref-verify.h"
  31
  32 #undef SCRAMBLE_DELAYED_REFS
  33
  34 /*
  35  * control flags for do_chunk_alloc's force field
  36  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
  37  * if we really need one.
  38  *
  39  * CHUNK_ALLOC_LIMITED means to only try and allocate one
  40  * if we have very few chunks already allocated.  This is
  41  * used as part of the clustering code to help make sure
  42  * we have a good pool of storage to cluster in, without
  43  * filling the FS with empty chunks
  44  *
  45  * CHUNK_ALLOC_FORCE means it must try to allocate one
  46  *
  47  */
  48 enum {
  49         CHUNK_ALLOC_NO_FORCE = 0,
  50         CHUNK_ALLOC_LIMITED = 1,
  51         CHUNK_ALLOC_FORCE = 2,
  52 };
  53
  54 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  55                                struct btrfs_delayed_ref_node *node, u64 parent,
  56                                u64 root_objectid, u64 owner_objectid,
  57                                u64 owner_offset, int refs_to_drop,
  58                                struct btrfs_delayed_extent_op *extra_op);
  59 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  60                                     struct extent_buffer *leaf,
  61                                     struct btrfs_extent_item *ei);
  62 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  63                                       u64 parent, u64 root_objectid,
  64                                       u64 flags, u64 owner, u64 offset,
  65                                       struct btrfs_key *ins, int ref_mod);
  66 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  67                                      struct btrfs_delayed_ref_node *node,
  68                                      struct btrfs_delayed_extent_op *extent_op);
  69 static int do_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags,
  70                           int force);
  71 static int find_next_key(struct btrfs_path *path, int level,
  72                          struct btrfs_key *key);
  73 static void dump_space_info(struct btrfs_fs_info *fs_info,
  74                             struct btrfs_space_info *info, u64 bytes,
  75                             int dump_block_groups);
  76 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
  77                                u64 num_bytes);
  78 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
  79                                      struct btrfs_space_info *space_info,
  80                                      u64 num_bytes);
  81 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
  82                                      struct btrfs_space_info *space_info,
  83                                      u64 num_bytes);
  84
  85 static noinline int
  86 block_group_cache_done(struct btrfs_block_group_cache *cache)
  87 {
  88         smp_mb();
  89         return cache->cached == BTRFS_CACHE_FINISHED ||
  90                 cache->cached == BTRFS_CACHE_ERROR;
  91 }
  92
  93 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
  94 {
  95         return (cache->flags & bits) == bits;
  96 }
  97
  98 void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
  99 {
 100         atomic_inc(&cache->count);
 101 }
 102
 103 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 104 {
 105         if (atomic_dec_and_test(&cache->count)) {
 106                 WARN_ON(cache->pinned > 0);
 107                 WARN_ON(cache->reserved > 0);
 108
 109                 /*
 110                  * If not empty, someone is still holding mutex of
 111                  * full_stripe_lock, which can only be released by caller.
 112                  * And it will definitely cause use-after-free when caller
 113                  * tries to release full stripe lock.
 114                  *
 115                  * No better way to resolve, but only to warn.
 116                  */
 117                 WARN_ON(!RB_EMPTY_ROOT(&cache->full_stripe_locks_root.root));
 118                 kfree(cache->free_space_ctl);
 119                 kfree(cache);
 120         }
 121 }
 122
 123 /*
 124  * this adds the block group to the fs_info rb tree for the block group
 125  * cache
 126  */
 127 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 128                                 struct btrfs_block_group_cache *block_group)
 129 {
 130         struct rb_node **p;
 131         struct rb_node *parent = NULL;
 132         struct btrfs_block_group_cache *cache;
 133
 134         spin_lock(&info->block_group_cache_lock);
 135         p = &info->block_group_cache_tree.rb_node;
 136
 137         while (*p) {
 138                 parent = *p;
 139                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 140                                  cache_node);
 141                 if (block_group->key.objectid < cache->key.objectid) {
 142                         p = &(*p)->rb_left;
 143                 } else if (block_group->key.objectid > cache->key.objectid) {
 144                         p = &(*p)->rb_right;
 145                 } else {
 146                         spin_unlock(&info->block_group_cache_lock);
 147                         return -EEXIST;
 148                 }
 149         }
 150
 151         rb_link_node(&block_group->cache_node, parent, p);
 152         rb_insert_color(&block_group->cache_node,
 153                         &info->block_group_cache_tree);
 154
 155         if (info->first_logical_byte > block_group->key.objectid)
 156                 info->first_logical_byte = block_group->key.objectid;
 157
 158         spin_unlock(&info->block_group_cache_lock);
 159
 160         return 0;
 161 }
 162
 163 /*
 164  * This will return the block group at or after bytenr if contains is 0, else
 165  * it will return the block group that contains the bytenr
 166  */
 167 static struct btrfs_block_group_cache *
 168 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 169                               int contains)
 170 {
 171         struct btrfs_block_group_cache *cache, *ret = NULL;
 172         struct rb_node *n;
 173         u64 end, start;
 174
 175         spin_lock(&info->block_group_cache_lock);
 176         n = info->block_group_cache_tree.rb_node;
 177
 178         while (n) {
 179                 cache = rb_entry(n, struct btrfs_block_group_cache,
 180                                  cache_node);
 181                 end = cache->key.objectid + cache->key.offset - 1;
 182                 start = cache->key.objectid;
 183
 184                 if (bytenr < start) {
 185                         if (!contains && (!ret || start < ret->key.objectid))
 186                                 ret = cache;
 187                         n = n->rb_left;
 188                 } else if (bytenr > start) {
 189                         if (contains && bytenr <= end) {
 190                                 ret = cache;
 191                                 break;
 192                         }
 193                         n = n->rb_right;
 194                 } else {
 195                         ret = cache;
 196                         break;
 197                 }
 198         }
 199         if (ret) {
 200                 btrfs_get_block_group(ret);
 201                 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
 202                         info->first_logical_byte = ret->key.objectid;
 203         }
 204         spin_unlock(&info->block_group_cache_lock);
 205
 206         return ret;
 207 }
 208
 209 static int add_excluded_extent(struct btrfs_fs_info *fs_info,
 210                                u64 start, u64 num_bytes)
 211 {
 212         u64 end = start + num_bytes - 1;
 213         set_extent_bits(&fs_info->freed_extents[0],
 214                         start, end, EXTENT_UPTODATE);
 215         set_extent_bits(&fs_info->freed_extents[1],
 216                         start, end, EXTENT_UPTODATE);
 217         return 0;
 218 }
 219
 220 static void free_excluded_extents(struct btrfs_block_group_cache *cache)
 221 {
 222         struct btrfs_fs_info *fs_info = cache->fs_info;
 223         u64 start, end;
 224
 225         start = cache->key.objectid;
 226         end = start + cache->key.offset - 1;
 227
 228         clear_extent_bits(&fs_info->freed_extents[0],
 229                           start, end, EXTENT_UPTODATE);
 230         clear_extent_bits(&fs_info->freed_extents[1],
 231                           start, end, EXTENT_UPTODATE);
 232 }
 233
 234 static int exclude_super_stripes(struct btrfs_block_group_cache *cache)
 235 {
 236         struct btrfs_fs_info *fs_info = cache->fs_info;
 237         u64 bytenr;
 238         u64 *logical;
 239         int stripe_len;
 240         int i, nr, ret;
 241
 242         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 243                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 244                 cache->bytes_super += stripe_len;
 245                 ret = add_excluded_extent(fs_info, cache->key.objectid,
 246                                           stripe_len);
 247                 if (ret)
 248                         return ret;
 249         }
 250
 251         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 252                 bytenr = btrfs_sb_offset(i);
 253                 ret = btrfs_rmap_block(fs_info, cache->key.objectid,
 254                                        bytenr, &logical, &nr, &stripe_len);
 255                 if (ret)
 256                         return ret;
 257
 258                 while (nr--) {
 259                         u64 start, len;
 260
 261                         if (logical[nr] > cache->key.objectid +
 262                             cache->key.offset)
 263                                 continue;
 264
 265                         if (logical[nr] + stripe_len <= cache->key.objectid)
 266                                 continue;
 267
 268                         start = logical[nr];
 269                         if (start < cache->key.objectid) {
 270                                 start = cache->key.objectid;
 271                                 len = (logical[nr] + stripe_len) - start;
 272                         } else {
 273                                 len = min_t(u64, stripe_len,
 274                                             cache->key.objectid +
 275                                             cache->key.offset - start);
 276                         }
 277
 278                         cache->bytes_super += len;
 279                         ret = add_excluded_extent(fs_info, start, len);
 280                         if (ret) {
 281                                 kfree(logical);
 282                                 return ret;
 283                         }
 284                 }
 285
 286                 kfree(logical);
 287         }
 288         return 0;
 289 }
 290
 291 static struct btrfs_caching_control *
 292 get_caching_control(struct btrfs_block_group_cache *cache)
 293 {
 294         struct btrfs_caching_control *ctl;
 295
 296         spin_lock(&cache->lock);
 297         if (!cache->caching_ctl) {
 298                 spin_unlock(&cache->lock);
 299                 return NULL;
 300         }
 301
 302         ctl = cache->caching_ctl;
 303         refcount_inc(&ctl->count);
 304         spin_unlock(&cache->lock);
 305         return ctl;
 306 }
 307
 308 static void put_caching_control(struct btrfs_caching_control *ctl)
 309 {
 310         if (refcount_dec_and_test(&ctl->count))
 311                 kfree(ctl);
 312 }
 313
 314 #ifdef CONFIG_BTRFS_DEBUG
 315 static void fragment_free_space(struct btrfs_block_group_cache *block_group)
 316 {
 317         struct btrfs_fs_info *fs_info = block_group->fs_info;
 318         u64 start = block_group->key.objectid;
 319         u64 len = block_group->key.offset;
 320         u64 chunk = block_group->flags & BTRFS_BLOCK_GROUP_METADATA ?
 321                 fs_info->nodesize : fs_info->sectorsize;
 322         u64 step = chunk << 1;
 323
 324         while (len > chunk) {
 325                 btrfs_remove_free_space(block_group, start, chunk);
 326                 start += step;
 327                 if (len < step)
 328                         len = 0;
 329                 else
 330                         len -= step;
 331         }
 332 }
 333 #endif
 334
 335 /*
 336  * this is only called by cache_block_group, since we could have freed extents
 337  * we need to check the pinned_extents for any extents that can't be used yet
 338  * since their free space will be released as soon as the transaction commits.
 339  */
 340 u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 341                        u64 start, u64 end)
 342 {
 343         struct btrfs_fs_info *info = block_group->fs_info;
 344         u64 extent_start, extent_end, size, total_added = 0;
 345         int ret;
 346
 347         while (start < end) {
 348                 ret = find_first_extent_bit(info->pinned_extents, start,
 349                                             &extent_start, &extent_end,
 350                                             EXTENT_DIRTY | EXTENT_UPTODATE,
 351                                             NULL);
 352                 if (ret)
 353                         break;
 354
 355                 if (extent_start <= start) {
 356                         start = extent_end + 1;
 357                 } else if (extent_start > start && extent_start < end) {
 358                         size = extent_start - start;
 359                         total_added += size;
 360                         ret = btrfs_add_free_space(block_group, start,
 361                                                    size);
 362                         BUG_ON(ret); /* -ENOMEM or logic error */
 363                         start = extent_end + 1;
 364                 } else {
 365                         break;
 366                 }
 367         }
 368
 369         if (start < end) {
 370                 size = end - start;
 371                 total_added += size;
 372                 ret = btrfs_add_free_space(block_group, start, size);
 373                 BUG_ON(ret); /* -ENOMEM or logic error */
 374         }
 375
 376         return total_added;
 377 }
 378
 379 static int load_extent_tree_free(struct btrfs_caching_control *caching_ctl)
 380 {
 381         struct btrfs_block_group_cache *block_group = caching_ctl->block_group;
 382         struct btrfs_fs_info *fs_info = block_group->fs_info;
 383         struct btrfs_root *extent_root = fs_info->extent_root;
 384         struct btrfs_path *path;
 385         struct extent_buffer *leaf;
 386         struct btrfs_key key;
 387         u64 total_found = 0;
 388         u64 last = 0;
 389         u32 nritems;
 390         int ret;
 391         bool wakeup = true;
 392
 393         path = btrfs_alloc_path();
 394         if (!path)
 395                 return -ENOMEM;
 396
 397         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 398
 399 #ifdef CONFIG_BTRFS_DEBUG
 400         /*
 401          * If we're fragmenting we don't want to make anybody think we can
 402          * allocate from this block group until we've had a chance to fragment
 403          * the free space.
 404          */
 405         if (btrfs_should_fragment_free_space(block_group))
 406                 wakeup = false;
 407 #endif
 408         /*
 409          * We don't want to deadlock with somebody trying to allocate a new
 410          * extent for the extent root while also trying to search the extent
 411          * root to add free space.  So we skip locking and search the commit
 412          * root, since its read-only
 413          */
 414         path->skip_locking = 1;
 415         path->search_commit_root = 1;
 416         path->reada = READA_FORWARD;
 417
 418         key.objectid = last;
 419         key.offset = 0;
 420         key.type = BTRFS_EXTENT_ITEM_KEY;
 421
 422 next:
 423         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 424         if (ret < 0)
 425                 goto out;
 426
 427         leaf = path->nodes[0];
 428         nritems = btrfs_header_nritems(leaf);
 429
 430         while (1) {
 431                 if (btrfs_fs_closing(fs_info) > 1) {
 432                         last = (u64)-1;
 433                         break;
 434                 }
 435
 436                 if (path->slots[0] < nritems) {
 437                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 438                 } else {
 439                         ret = find_next_key(path, 0, &key);
 440                         if (ret)
 441                                 break;
 442
 443                         if (need_resched() ||
 444                             rwsem_is_contended(&fs_info->commit_root_sem)) {
 445                                 if (wakeup)
 446                                         caching_ctl->progress = last;
 447                                 btrfs_release_path(path);
 448                                 up_read(&fs_info->commit_root_sem);
 449                                 mutex_unlock(&caching_ctl->mutex);
 450                                 cond_resched();
 451                                 mutex_lock(&caching_ctl->mutex);
 452                                 down_read(&fs_info->commit_root_sem);
 453                                 goto next;
 454                         }
 455
 456                         ret = btrfs_next_leaf(extent_root, path);
 457                         if (ret < 0)
 458                                 goto out;
 459                         if (ret)
 460                                 break;
 461                         leaf = path->nodes[0];
 462                         nritems = btrfs_header_nritems(leaf);
 463                         continue;
 464                 }
 465
 466                 if (key.objectid < last) {
 467                         key.objectid = last;
 468                         key.offset = 0;
 469                         key.type = BTRFS_EXTENT_ITEM_KEY;
 470
 471                         if (wakeup)
 472                                 caching_ctl->progress = last;
 473                         btrfs_release_path(path);
 474                         goto next;
 475                 }
 476
 477                 if (key.objectid < block_group->key.objectid) {
 478                         path->slots[0]++;
 479                         continue;
 480                 }
 481
 482                 if (key.objectid >= block_group->key.objectid +
 483                     block_group->key.offset)
 484                         break;
 485
 486                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 487                     key.type == BTRFS_METADATA_ITEM_KEY) {
 488                         total_found += add_new_free_space(block_group, last,
 489                                                           key.objectid);
 490                         if (key.type == BTRFS_METADATA_ITEM_KEY)
 491                                 last = key.objectid +
 492                                         fs_info->nodesize;
 493                         else
 494                                 last = key.objectid + key.offset;
 495
 496                         if (total_found > CACHING_CTL_WAKE_UP) {
 497                                 total_found = 0;
 498                                 if (wakeup)
 499                                         wake_up(&caching_ctl->wait);
 500                         }
 501                 }
 502                 path->slots[0]++;
 503         }
 504         ret = 0;
 505
 506         total_found += add_new_free_space(block_group, last,
 507                                           block_group->key.objectid +
 508                                           block_group->key.offset);
 509         caching_ctl->progress = (u64)-1;
 510
 511 out:
 512         btrfs_free_path(path);
 513         return ret;
 514 }
 515
 516 static noinline void caching_thread(struct btrfs_work *work)
 517 {
 518         struct btrfs_block_group_cache *block_group;
 519         struct btrfs_fs_info *fs_info;
 520         struct btrfs_caching_control *caching_ctl;
 521         int ret;
 522
 523         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 524         block_group = caching_ctl->block_group;
 525         fs_info = block_group->fs_info;
 526
 527         mutex_lock(&caching_ctl->mutex);
 528         down_read(&fs_info->commit_root_sem);
 529
 530         if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
 531                 ret = load_free_space_tree(caching_ctl);
 532         else
 533                 ret = load_extent_tree_free(caching_ctl);
 534
 535         spin_lock(&block_group->lock);
 536         block_group->caching_ctl = NULL;
 537         block_group->cached = ret ? BTRFS_CACHE_ERROR : BTRFS_CACHE_FINISHED;
 538         spin_unlock(&block_group->lock);
 539
 540 #ifdef CONFIG_BTRFS_DEBUG
 541         if (btrfs_should_fragment_free_space(block_group)) {
 542                 u64 bytes_used;
 543
 544                 spin_lock(&block_group->space_info->lock);
 545                 spin_lock(&block_group->lock);
 546                 bytes_used = block_group->key.offset -
 547                         btrfs_block_group_used(&block_group->item);
 548                 block_group->space_info->bytes_used += bytes_used >> 1;
 549                 spin_unlock(&block_group->lock);
 550                 spin_unlock(&block_group->space_info->lock);
 551                 fragment_free_space(block_group);
 552         }
 553 #endif
 554
 555         caching_ctl->progress = (u64)-1;
 556
 557         up_read(&fs_info->commit_root_sem);
 558         free_excluded_extents(block_group);
 559         mutex_unlock(&caching_ctl->mutex);
 560
 561         wake_up(&caching_ctl->wait);
 562
 563         put_caching_control(caching_ctl);
 564         btrfs_put_block_group(block_group);
 565 }
 566
 567 static int cache_block_group(struct btrfs_block_group_cache *cache,
 568                              int load_cache_only)
 569 {
 570         DEFINE_WAIT(wait);
 571         struct btrfs_fs_info *fs_info = cache->fs_info;
 572         struct btrfs_caching_control *caching_ctl;
 573         int ret = 0;
 574
 575         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 576         if (!caching_ctl)
 577                 return -ENOMEM;
 578
 579         INIT_LIST_HEAD(&caching_ctl->list);
 580         mutex_init(&caching_ctl->mutex);
 581         init_waitqueue_head(&caching_ctl->wait);
 582         caching_ctl->block_group = cache;
 583         caching_ctl->progress = cache->key.objectid;
 584         refcount_set(&caching_ctl->count, 1);
 585         btrfs_init_work(&caching_ctl->work, btrfs_cache_helper,
 586                         caching_thread, NULL, NULL);
 587
 588         spin_lock(&cache->lock);
 589         /*
 590          * This should be a rare occasion, but this could happen I think in the
 591          * case where one thread starts to load the space cache info, and then
 592          * some other thread starts a transaction commit which tries to do an
 593          * allocation while the other thread is still loading the space cache
 594          * info.  The previous loop should have kept us from choosing this block
 595          * group, but if we've moved to the state where we will wait on caching
 596          * block groups we need to first check if we're doing a fast load here,
 597          * so we can wait for it to finish, otherwise we could end up allocating
 598          * from a block group who's cache gets evicted for one reason or
 599          * another.
 600          */
 601         while (cache->cached == BTRFS_CACHE_FAST) {
 602                 struct btrfs_caching_control *ctl;
 603
 604                 ctl = cache->caching_ctl;
 605                 refcount_inc(&ctl->count);
 606                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 607                 spin_unlock(&cache->lock);
 608
 609                 schedule();
 610
 611                 finish_wait(&ctl->wait, &wait);
 612                 put_caching_control(ctl);
 613                 spin_lock(&cache->lock);
 614         }
 615
 616         if (cache->cached != BTRFS_CACHE_NO) {
 617                 spin_unlock(&cache->lock);
 618                 kfree(caching_ctl);
 619                 return 0;
 620         }
 621         WARN_ON(cache->caching_ctl);
 622         cache->caching_ctl = caching_ctl;
 623         cache->cached = BTRFS_CACHE_FAST;
 624         spin_unlock(&cache->lock);
 625
 626         if (btrfs_test_opt(fs_info, SPACE_CACHE)) {
 627                 mutex_lock(&caching_ctl->mutex);
 628                 ret = load_free_space_cache(fs_info, cache);
 629
 630                 spin_lock(&cache->lock);
 631                 if (ret == 1) {
 632                         cache->caching_ctl = NULL;
 633                         cache->cached = BTRFS_CACHE_FINISHED;
 634                         cache->last_byte_to_unpin = (u64)-1;
 635                         caching_ctl->progress = (u64)-1;
 636                 } else {
 637                         if (load_cache_only) {
 638                                 cache->caching_ctl = NULL;
 639                                 cache->cached = BTRFS_CACHE_NO;
 640                         } else {
 641                                 cache->cached = BTRFS_CACHE_STARTED;
 642                                 cache->has_caching_ctl = 1;
 643                         }
 644                 }
 645                 spin_unlock(&cache->lock);
 646 #ifdef CONFIG_BTRFS_DEBUG
 647                 if (ret == 1 &&
 648                     btrfs_should_fragment_free_space(cache)) {
 649                         u64 bytes_used;
 650
 651                         spin_lock(&cache->space_info->lock);
 652                         spin_lock(&cache->lock);
 653                         bytes_used = cache->key.offset -
 654                                 btrfs_block_group_used(&cache->item);
 655                         cache->space_info->bytes_used += bytes_used >> 1;
 656                         spin_unlock(&cache->lock);
 657                         spin_unlock(&cache->space_info->lock);
 658                         fragment_free_space(cache);
 659                 }
 660 #endif
 661                 mutex_unlock(&caching_ctl->mutex);
 662
 663                 wake_up(&caching_ctl->wait);
 664                 if (ret == 1) {
 665                         put_caching_control(caching_ctl);
 666                         free_excluded_extents(cache);
 667                         return 0;
 668                 }
 669         } else {
 670                 /*
 671                  * We're either using the free space tree or no caching at all.
 672                  * Set cached to the appropriate value and wakeup any waiters.
 673                  */
 674                 spin_lock(&cache->lock);
 675                 if (load_cache_only) {
 676                         cache->caching_ctl = NULL;
 677                         cache->cached = BTRFS_CACHE_NO;
 678                 } else {
 679                         cache->cached = BTRFS_CACHE_STARTED;
 680                         cache->has_caching_ctl = 1;
 681                 }
 682                 spin_unlock(&cache->lock);
 683                 wake_up(&caching_ctl->wait);
 684         }
 685
 686         if (load_cache_only) {
 687                 put_caching_control(caching_ctl);
 688                 return 0;
 689         }
 690
 691         down_write(&fs_info->commit_root_sem);
 692         refcount_inc(&caching_ctl->count);
 693         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 694         up_write(&fs_info->commit_root_sem);
 695
 696         btrfs_get_block_group(cache);
 697
 698         btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
 699
 700         return ret;
 701 }
 702
 703 /*
 704  * return the block group that starts at or after bytenr
 705  */
 706 static struct btrfs_block_group_cache *
 707 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 708 {
 709         return block_group_cache_tree_search(info, bytenr, 0);
 710 }
 711
 712 /*
 713  * return the block group that contains the given bytenr
 714  */
 715 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 716                                                  struct btrfs_fs_info *info,
 717                                                  u64 bytenr)
 718 {
 719         return block_group_cache_tree_search(info, bytenr, 1);
 720 }
 721
 722 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 723                                                   u64 flags)
 724 {
 725         struct list_head *head = &info->space_info;
 726         struct btrfs_space_info *found;
 727
 728         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 729
 730         rcu_read_lock();
 731         list_for_each_entry_rcu(found, head, list) {
 732                 if (found->flags & flags) {
 733                         rcu_read_unlock();
 734                         return found;
 735                 }
 736         }
 737         rcu_read_unlock();
 738         return NULL;
 739 }
 740
 741 static void add_pinned_bytes(struct btrfs_fs_info *fs_info, s64 num_bytes,
 742                              bool metadata, u64 root_objectid)
 743 {
 744         struct btrfs_space_info *space_info;
 745         u64 flags;
 746
 747         if (metadata) {
 748                 if (root_objectid == BTRFS_CHUNK_TREE_OBJECTID)
 749                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
 750                 else
 751                         flags = BTRFS_BLOCK_GROUP_METADATA;
 752         } else {
 753                 flags = BTRFS_BLOCK_GROUP_DATA;
 754         }
 755
 756         space_info = __find_space_info(fs_info, flags);
 757         ASSERT(space_info);
 758         percpu_counter_add_batch(&space_info->total_bytes_pinned, num_bytes,
 759                     BTRFS_TOTAL_BYTES_PINNED_BATCH);
 760 }
 761
 762 /*
 763  * after adding space to the filesystem, we need to clear the full flags
 764  * on all the space infos.
 765  */
 766 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 767 {
 768         struct list_head *head = &info->space_info;
 769         struct btrfs_space_info *found;
 770
 771         rcu_read_lock();
 772         list_for_each_entry_rcu(found, head, list)
 773                 found->full = 0;
 774         rcu_read_unlock();
 775 }
 776
 777 /* simple helper to search for an existing data extent at a given offset */
 778 int btrfs_lookup_data_extent(struct btrfs_fs_info *fs_info, u64 start, u64 len)
 779 {
 780         int ret;
 781         struct btrfs_key key;
 782         struct btrfs_path *path;
 783
 784         path = btrfs_alloc_path();
 785         if (!path)
 786                 return -ENOMEM;
 787
 788         key.objectid = start;
 789         key.offset = len;
 790         key.type = BTRFS_EXTENT_ITEM_KEY;
 791         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
 792         btrfs_free_path(path);
 793         return ret;
 794 }
 795
 796 /*
 797  * helper function to lookup reference count and flags of a tree block.
 798  *
 799  * the head node for delayed ref is used to store the sum of all the
 800  * reference count modifications queued up in the rbtree. the head
 801  * node may also store the extent flags to set. This way you can check
 802  * to see what the reference count and extent flags would be if all of
 803  * the delayed refs are not processed.
 804  */
 805 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 806                              struct btrfs_fs_info *fs_info, u64 bytenr,
 807                              u64 offset, int metadata, u64 *refs, u64 *flags)
 808 {
 809         struct btrfs_delayed_ref_head *head;
 810         struct btrfs_delayed_ref_root *delayed_refs;
 811         struct btrfs_path *path;
 812         struct btrfs_extent_item *ei;
 813         struct extent_buffer *leaf;
 814         struct btrfs_key key;
 815         u32 item_size;
 816         u64 num_refs;
 817         u64 extent_flags;
 818         int ret;
 819
 820         /*
 821          * If we don't have skinny metadata, don't bother doing anything
 822          * different
 823          */
 824         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA)) {
 825                 offset = fs_info->nodesize;
 826                 metadata = 0;
 827         }
 828
 829         path = btrfs_alloc_path();
 830         if (!path)
 831                 return -ENOMEM;
 832
 833         if (!trans) {
 834                 path->skip_locking = 1;
 835                 path->search_commit_root = 1;
 836         }
 837
 838 search_again:
 839         key.objectid = bytenr;
 840         key.offset = offset;
 841         if (metadata)
 842                 key.type = BTRFS_METADATA_ITEM_KEY;
 843         else
 844                 key.type = BTRFS_EXTENT_ITEM_KEY;
 845
 846         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 0);
 847         if (ret < 0)
 848                 goto out_free;
 849
 850         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 851                 if (path->slots[0]) {
 852                         path->slots[0]--;
 853                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 854                                               path->slots[0]);
 855                         if (key.objectid == bytenr &&
 856                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 857                             key.offset == fs_info->nodesize)
 858                                 ret = 0;
 859                 }
 860         }
 861
 862         if (ret == 0) {
 863                 leaf = path->nodes[0];
 864                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 865                 if (item_size >= sizeof(*ei)) {
 866                         ei = btrfs_item_ptr(leaf, path->slots[0],
 867                                             struct btrfs_extent_item);
 868                         num_refs = btrfs_extent_refs(leaf, ei);
 869                         extent_flags = btrfs_extent_flags(leaf, ei);
 870                 } else {
 871                         ret = -EINVAL;
 872                         btrfs_print_v0_err(fs_info);
 873                         if (trans)
 874                                 btrfs_abort_transaction(trans, ret);
 875                         else
 876                                 btrfs_handle_fs_error(fs_info, ret, NULL);
 877
 878                         goto out_free;
 879                 }
 880
 881                 BUG_ON(num_refs == 0);
 882         } else {
 883                 num_refs = 0;
 884                 extent_flags = 0;
 885                 ret = 0;
 886         }
 887
 888         if (!trans)
 889                 goto out;
 890
 891         delayed_refs = &trans->transaction->delayed_refs;
 892         spin_lock(&delayed_refs->lock);
 893         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
 894         if (head) {
 895                 if (!mutex_trylock(&head->mutex)) {
 896                         refcount_inc(&head->refs);
 897                         spin_unlock(&delayed_refs->lock);
 898
 899                         btrfs_release_path(path);
 900
 901                         /*
 902                          * Mutex was contended, block until it's released and try
 903                          * again
 904                          */
 905                         mutex_lock(&head->mutex);
 906                         mutex_unlock(&head->mutex);
 907                         btrfs_put_delayed_ref_head(head);
 908                         goto search_again;
 909                 }
 910                 spin_lock(&head->lock);
 911                 if (head->extent_op && head->extent_op->update_flags)
 912                         extent_flags |= head->extent_op->flags_to_set;
 913                 else
 914                         BUG_ON(num_refs == 0);
 915
 916                 num_refs += head->ref_mod;
 917                 spin_unlock(&head->lock);
 918                 mutex_unlock(&head->mutex);
 919         }
 920         spin_unlock(&delayed_refs->lock);
 921 out:
 922         WARN_ON(num_refs == 0);
 923         if (refs)
 924                 *refs = num_refs;
 925         if (flags)
 926                 *flags = extent_flags;
 927 out_free:
 928         btrfs_free_path(path);
 929         return ret;
 930 }
 931
 932 /*
 933  * Back reference rules.  Back refs have three main goals:
 934  *
 935  * 1) differentiate between all holders of references to an extent so that
 936  *    when a reference is dropped we can make sure it was a valid reference
 937  *    before freeing the extent.
 938  *
 939  * 2) Provide enough information to quickly find the holders of an extent
 940  *    if we notice a given block is corrupted or bad.
 941  *
 942  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 943  *    maintenance.  This is actually the same as #2, but with a slightly
 944  *    different use case.
 945  *
 946  * There are two kinds of back refs. The implicit back refs is optimized
 947  * for pointers in non-shared tree blocks. For a given pointer in a block,
 948  * back refs of this kind provide information about the block's owner tree
 949  * and the pointer's key. These information allow us to find the block by
 950  * b-tree searching. The full back refs is for pointers in tree blocks not
 951  * referenced by their owner trees. The location of tree block is recorded
 952  * in the back refs. Actually the full back refs is generic, and can be
 953  * used in all cases the implicit back refs is used. The major shortcoming
 954  * of the full back refs is its overhead. Every time a tree block gets
 955  * COWed, we have to update back refs entry for all pointers in it.
 956  *
 957  * For a newly allocated tree block, we use implicit back refs for
 958  * pointers in it. This means most tree related operations only involve
 959  * implicit back refs. For a tree block created in old transaction, the
 960  * only way to drop a reference to it is COW it. So we can detect the
 961  * event that tree block loses its owner tree's reference and do the
 962  * back refs conversion.
 963  *
 964  * When a tree block is COWed through a tree, there are four cases:
 965  *
 966  * The reference count of the block is one and the tree is the block's
 967  * owner tree. Nothing to do in this case.
 968  *
 969  * The reference count of the block is one and the tree is not the
 970  * block's owner tree. In this case, full back refs is used for pointers
 971  * in the block. Remove these full back refs, add implicit back refs for
 972  * every pointers in the new block.
 973  *
 974  * The reference count of the block is greater than one and the tree is
 975  * the block's owner tree. In this case, implicit back refs is used for
 976  * pointers in the block. Add full back refs for every pointers in the
 977  * block, increase lower level extents' reference counts. The original
 978  * implicit back refs are entailed to the new block.
 979  *
 980  * The reference count of the block is greater than one and the tree is
 981  * not the block's owner tree. Add implicit back refs for every pointer in
 982  * the new block, increase lower level extents' reference count.
 983  *
 984  * Back Reference Key composing:
 985  *
 986  * The key objectid corresponds to the first byte in the extent,
 987  * The key type is used to differentiate between types of back refs.
 988  * There are different meanings of the key offset for different types
 989  * of back refs.
 990  *
 991  * File extents can be referenced by:
 992  *
 993  * - multiple snapshots, subvolumes, or different generations in one subvol
 994  * - different files inside a single subvolume
 995  * - different offsets inside a file (bookend extents in file.c)
 996  *
 997  * The extent ref structure for the implicit back refs has fields for:
 998  *
 999  * - Objectid of the subvolume root
1000  * - objectid of the file holding the reference
1001  * - original offset in the file
1002  * - how many bookend extents
1003  *
1004  * The key offset for the implicit back refs is hash of the first
1005  * three fields.
1006  *
1007  * The extent ref structure for the full back refs has field for:
1008  *
1009  * - number of pointers in the tree leaf
1010  *
1011  * The key offset for the implicit back refs is the first byte of
1012  * the tree leaf
1013  *
1014  * When a file extent is allocated, The implicit back refs is used.
1015  * the fields are filled in:
1016  *
1017  *     (root_key.objectid, inode objectid, offset in file, 1)
1018  *
1019  * When a file extent is removed file truncation, we find the
1020  * corresponding implicit back refs and check the following fields:
1021  *
1022  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
1023  *
1024  * Btree extents can be referenced by:
1025  *
1026  * - Different subvolumes
1027  *
1028  * Both the implicit back refs and the full back refs for tree blocks
1029  * only consist of key. The key offset for the implicit back refs is
1030  * objectid of block's owner tree. The key offset for the full back refs
1031  * is the first byte of parent block.
1032  *
1033  * When implicit back refs is used, information about the lowest key and
1034  * level of the tree block are required. These information are stored in
1035  * tree block info structure.
1036  */
1037
1038 /*
1039  * is_data == BTRFS_REF_TYPE_BLOCK, tree block type is required,
1040  * is_data == BTRFS_REF_TYPE_DATA, data type is requried,
1041  * is_data == BTRFS_REF_TYPE_ANY, either type is OK.
1042  */
1043 int btrfs_get_extent_inline_ref_type(const struct extent_buffer *eb,
1044                                      struct btrfs_extent_inline_ref *iref,
1045                                      enum btrfs_inline_ref_type is_data)
1046 {
1047         int type = btrfs_extent_inline_ref_type(eb, iref);
1048         u64 offset = btrfs_extent_inline_ref_offset(eb, iref);
1049
1050         if (type == BTRFS_TREE_BLOCK_REF_KEY ||
1051             type == BTRFS_SHARED_BLOCK_REF_KEY ||
1052             type == BTRFS_SHARED_DATA_REF_KEY ||
1053             type == BTRFS_EXTENT_DATA_REF_KEY) {
1054                 if (is_data == BTRFS_REF_TYPE_BLOCK) {
1055                         if (type == BTRFS_TREE_BLOCK_REF_KEY)
1056                                 return type;
1057                         if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1058                                 ASSERT(eb->fs_info);
1059                                 /*
1060                                  * Every shared one has parent tree
1061                                  * block, which must be aligned to
1062                                  * nodesize.
1063                                  */
1064                                 if (offset &&
1065                                     IS_ALIGNED(offset, eb->fs_info->nodesize))
1066                                         return type;
1067                         }
1068                 } else if (is_data == BTRFS_REF_TYPE_DATA) {
1069                         if (type == BTRFS_EXTENT_DATA_REF_KEY)
1070                                 return type;
1071                         if (type == BTRFS_SHARED_DATA_REF_KEY) {
1072                                 ASSERT(eb->fs_info);
1073                                 /*
1074                                  * Every shared one has parent tree
1075                                  * block, which must be aligned to
1076                                  * nodesize.
1077                                  */
1078                                 if (offset &&
1079                                     IS_ALIGNED(offset, eb->fs_info->nodesize))
1080                                         return type;
1081                         }
1082                 } else {
1083                         ASSERT(is_data == BTRFS_REF_TYPE_ANY);
1084                         return type;
1085                 }
1086         }
1087
1088         btrfs_print_leaf((struct extent_buffer *)eb);
1089         btrfs_err(eb->fs_info, "eb %llu invalid extent inline ref type %d",
1090                   eb->start, type);
1091         WARN_ON(1);
1092
1093         return BTRFS_REF_TYPE_INVALID;
1094 }
1095
1096 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1097 {
1098         u32 high_crc = ~(u32)0;
1099         u32 low_crc = ~(u32)0;
1100         __le64 lenum;
1101
1102         lenum = cpu_to_le64(root_objectid);
1103         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
1104         lenum = cpu_to_le64(owner);
1105         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1106         lenum = cpu_to_le64(offset);
1107         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1108
1109         return ((u64)high_crc << 31) ^ (u64)low_crc;
1110 }
1111
1112 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1113                                      struct btrfs_extent_data_ref *ref)
1114 {
1115         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1116                                     btrfs_extent_data_ref_objectid(leaf, ref),
1117                                     btrfs_extent_data_ref_offset(leaf, ref));
1118 }
1119
1120 static int match_extent_data_ref(struct extent_buffer *leaf,
1121                                  struct btrfs_extent_data_ref *ref,
1122                                  u64 root_objectid, u64 owner, u64 offset)
1123 {
1124         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1125             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1126             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1127                 return 0;
1128         return 1;
1129 }
1130
1131 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1132                                            struct btrfs_path *path,
1133                                            u64 bytenr, u64 parent,
1134                                            u64 root_objectid,
1135                                            u64 owner, u64 offset)
1136 {
1137         struct btrfs_root *root = trans->fs_info->extent_root;
1138         struct btrfs_key key;
1139         struct btrfs_extent_data_ref *ref;
1140         struct extent_buffer *leaf;
1141         u32 nritems;
1142         int ret;
1143         int recow;
1144         int err = -ENOENT;
1145
1146         key.objectid = bytenr;
1147         if (parent) {
1148                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1149                 key.offset = parent;
1150         } else {
1151                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1152                 key.offset = hash_extent_data_ref(root_objectid,
1153                                                   owner, offset);
1154         }
1155 again:
1156         recow = 0;
1157         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1158         if (ret < 0) {
1159                 err = ret;
1160                 goto fail;
1161         }
1162
1163         if (parent) {
1164                 if (!ret)
1165                         return 0;
1166                 goto fail;
1167         }
1168
1169         leaf = path->nodes[0];
1170         nritems = btrfs_header_nritems(leaf);
1171         while (1) {
1172                 if (path->slots[0] >= nritems) {
1173                         ret = btrfs_next_leaf(root, path);
1174                         if (ret < 0)
1175                                 err = ret;
1176                         if (ret)
1177                                 goto fail;
1178
1179                         leaf = path->nodes[0];
1180                         nritems = btrfs_header_nritems(leaf);
1181                         recow = 1;
1182                 }
1183
1184                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1185                 if (key.objectid != bytenr ||
1186                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1187                         goto fail;
1188
1189                 ref = btrfs_item_ptr(leaf, path->slots[0],
1190                                      struct btrfs_extent_data_ref);
1191
1192                 if (match_extent_data_ref(leaf, ref, root_objectid,
1193                                           owner, offset)) {
1194                         if (recow) {
1195                                 btrfs_release_path(path);
1196                                 goto again;
1197                         }
1198                         err = 0;
1199                         break;
1200                 }
1201                 path->slots[0]++;
1202         }
1203 fail:
1204         return err;
1205 }
1206
1207 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1208                                            struct btrfs_path *path,
1209                                            u64 bytenr, u64 parent,
1210                                            u64 root_objectid, u64 owner,
1211                                            u64 offset, int refs_to_add)
1212 {
1213         struct btrfs_root *root = trans->fs_info->extent_root;
1214         struct btrfs_key key;
1215         struct extent_buffer *leaf;
1216         u32 size;
1217         u32 num_refs;
1218         int ret;
1219
1220         key.objectid = bytenr;
1221         if (parent) {
1222                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1223                 key.offset = parent;
1224                 size = sizeof(struct btrfs_shared_data_ref);
1225         } else {
1226                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1227                 key.offset = hash_extent_data_ref(root_objectid,
1228                                                   owner, offset);
1229                 size = sizeof(struct btrfs_extent_data_ref);
1230         }
1231
1232         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1233         if (ret && ret != -EEXIST)
1234                 goto fail;
1235
1236         leaf = path->nodes[0];
1237         if (parent) {
1238                 struct btrfs_shared_data_ref *ref;
1239                 ref = btrfs_item_ptr(leaf, path->slots[0],
1240                                      struct btrfs_shared_data_ref);
1241                 if (ret == 0) {
1242                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1243                 } else {
1244                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1245                         num_refs += refs_to_add;
1246                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1247                 }
1248         } else {
1249                 struct btrfs_extent_data_ref *ref;
1250                 while (ret == -EEXIST) {
1251                         ref = btrfs_item_ptr(leaf, path->slots[0],
1252                                              struct btrfs_extent_data_ref);
1253                         if (match_extent_data_ref(leaf, ref, root_objectid,
1254                                                   owner, offset))
1255                                 break;
1256                         btrfs_release_path(path);
1257                         key.offset++;
1258                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1259                                                       size);
1260                         if (ret && ret != -EEXIST)
1261                                 goto fail;
1262
1263                         leaf = path->nodes[0];
1264                 }
1265                 ref = btrfs_item_ptr(leaf, path->slots[0],
1266                                      struct btrfs_extent_data_ref);
1267                 if (ret == 0) {
1268                         btrfs_set_extent_data_ref_root(leaf, ref,
1269                                                        root_objectid);
1270                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1271                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1272                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1273                 } else {
1274                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1275                         num_refs += refs_to_add;
1276                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1277                 }
1278         }
1279         btrfs_mark_buffer_dirty(leaf);
1280         ret = 0;
1281 fail:
1282         btrfs_release_path(path);
1283         return ret;
1284 }
1285
1286 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1287                                            struct btrfs_path *path,
1288                                            int refs_to_drop, int *last_ref)
1289 {
1290         struct btrfs_key key;
1291         struct btrfs_extent_data_ref *ref1 = NULL;
1292         struct btrfs_shared_data_ref *ref2 = NULL;
1293         struct extent_buffer *leaf;
1294         u32 num_refs = 0;
1295         int ret = 0;
1296
1297         leaf = path->nodes[0];
1298         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1299
1300         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1301                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1302                                       struct btrfs_extent_data_ref);
1303                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1304         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1305                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1306                                       struct btrfs_shared_data_ref);
1307                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1308         } else if (unlikely(key.type == BTRFS_EXTENT_REF_V0_KEY)) {
1309                 btrfs_print_v0_err(trans->fs_info);
1310                 btrfs_abort_transaction(trans, -EINVAL);
1311                 return -EINVAL;
1312         } else {
1313                 BUG();
1314         }
1315
1316         BUG_ON(num_refs < refs_to_drop);
1317         num_refs -= refs_to_drop;
1318
1319         if (num_refs == 0) {
1320                 ret = btrfs_del_item(trans, trans->fs_info->extent_root, path);
1321                 *last_ref = 1;
1322         } else {
1323                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1324                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1325                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1326                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1327                 btrfs_mark_buffer_dirty(leaf);
1328         }
1329         return ret;
1330 }
1331
1332 static noinline u32 extent_data_ref_count(struct btrfs_path *path,
1333                                           struct btrfs_extent_inline_ref *iref)
1334 {
1335         struct btrfs_key key;
1336         struct extent_buffer *leaf;
1337         struct btrfs_extent_data_ref *ref1;
1338         struct btrfs_shared_data_ref *ref2;
1339         u32 num_refs = 0;
1340         int type;
1341
1342         leaf = path->nodes[0];
1343         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1344
1345         BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
1346         if (iref) {
1347                 /*
1348                  * If type is invalid, we should have bailed out earlier than
1349                  * this call.
1350                  */
1351                 type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
1352                 ASSERT(type != BTRFS_REF_TYPE_INVALID);
1353                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1354                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1355                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1356                 } else {
1357                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1358                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1359                 }
1360         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1361                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1362                                       struct btrfs_extent_data_ref);
1363                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1364         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1365                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1366                                       struct btrfs_shared_data_ref);
1367                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1368         } else {
1369                 WARN_ON(1);
1370         }
1371         return num_refs;
1372 }
1373
1374 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1375                                           struct btrfs_path *path,
1376                                           u64 bytenr, u64 parent,
1377                                           u64 root_objectid)
1378 {
1379         struct btrfs_root *root = trans->fs_info->extent_root;
1380         struct btrfs_key key;
1381         int ret;
1382
1383         key.objectid = bytenr;
1384         if (parent) {
1385                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1386                 key.offset = parent;
1387         } else {
1388                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1389                 key.offset = root_objectid;
1390         }
1391
1392         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1393         if (ret > 0)
1394                 ret = -ENOENT;
1395         return ret;
1396 }
1397
1398 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1399                                           struct btrfs_path *path,
1400                                           u64 bytenr, u64 parent,
1401                                           u64 root_objectid)
1402 {
1403         struct btrfs_key key;
1404         int ret;
1405
1406         key.objectid = bytenr;
1407         if (parent) {
1408                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1409                 key.offset = parent;
1410         } else {
1411                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1412                 key.offset = root_objectid;
1413         }
1414
1415         ret = btrfs_insert_empty_item(trans, trans->fs_info->extent_root,
1416                                       path, &key, 0);
1417         btrfs_release_path(path);
1418         return ret;
1419 }
1420
1421 static inline int extent_ref_type(u64 parent, u64 owner)
1422 {
1423         int type;
1424         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1425                 if (parent > 0)
1426                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1427                 else
1428                         type = BTRFS_TREE_BLOCK_REF_KEY;
1429         } else {
1430                 if (parent > 0)
1431                         type = BTRFS_SHARED_DATA_REF_KEY;
1432                 else
1433                         type = BTRFS_EXTENT_DATA_REF_KEY;
1434         }
1435         return type;
1436 }
1437
1438 static int find_next_key(struct btrfs_path *path, int level,
1439                          struct btrfs_key *key)
1440
1441 {
1442         for (; level < BTRFS_MAX_LEVEL; level++) {
1443                 if (!path->nodes[level])
1444                         break;
1445                 if (path->slots[level] + 1 >=
1446                     btrfs_header_nritems(path->nodes[level]))
1447                         continue;
1448                 if (level == 0)
1449                         btrfs_item_key_to_cpu(path->nodes[level], key,
1450                                               path->slots[level] + 1);
1451                 else
1452                         btrfs_node_key_to_cpu(path->nodes[level], key,
1453                                               path->slots[level] + 1);
1454                 return 0;
1455         }
1456         return 1;
1457 }
1458
1459 /*
1460  * look for inline back ref. if back ref is found, *ref_ret is set
1461  * to the address of inline back ref, and 0 is returned.
1462  *
1463  * if back ref isn't found, *ref_ret is set to the address where it
1464  * should be inserted, and -ENOENT is returned.
1465  *
1466  * if insert is true and there are too many inline back refs, the path
1467  * points to the extent item, and -EAGAIN is returned.
1468  *
1469  * NOTE: inline back refs are ordered in the same way that back ref
1470  *       items in the tree are ordered.
1471  */
1472 static noinline_for_stack
1473 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1474                                  struct btrfs_path *path,
1475                                  struct btrfs_extent_inline_ref **ref_ret,
1476                                  u64 bytenr, u64 num_bytes,
1477                                  u64 parent, u64 root_objectid,
1478                                  u64 owner, u64 offset, int insert)
1479 {
1480         struct btrfs_fs_info *fs_info = trans->fs_info;
1481         struct btrfs_root *root = fs_info->extent_root;
1482         struct btrfs_key key;
1483         struct extent_buffer *leaf;
1484         struct btrfs_extent_item *ei;
1485         struct btrfs_extent_inline_ref *iref;
1486         u64 flags;
1487         u64 item_size;
1488         unsigned long ptr;
1489         unsigned long end;
1490         int extra_size;
1491         int type;
1492         int want;
1493         int ret;
1494         int err = 0;
1495         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
1496         int needed;
1497
1498         key.objectid = bytenr;
1499         key.type = BTRFS_EXTENT_ITEM_KEY;
1500         key.offset = num_bytes;
1501
1502         want = extent_ref_type(parent, owner);
1503         if (insert) {
1504                 extra_size = btrfs_extent_inline_ref_size(want);
1505                 path->keep_locks = 1;
1506         } else
1507                 extra_size = -1;
1508
1509         /*
1510          * Owner is our level, so we can just add one to get the level for the
1511          * block we are interested in.
1512          */
1513         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
1514                 key.type = BTRFS_METADATA_ITEM_KEY;
1515                 key.offset = owner;
1516         }
1517
1518 again:
1519         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1520         if (ret < 0) {
1521                 err = ret;
1522                 goto out;
1523         }
1524
1525         /*
1526          * We may be a newly converted file system which still has the old fat
1527          * extent entries for metadata, so try and see if we have one of those.
1528          */
1529         if (ret > 0 && skinny_metadata) {
1530                 skinny_metadata = false;
1531                 if (path->slots[0]) {
1532                         path->slots[0]--;
1533                         btrfs_item_key_to_cpu(path->nodes[0], &key,
1534                                               path->slots[0]);
1535                         if (key.objectid == bytenr &&
1536                             key.type == BTRFS_EXTENT_ITEM_KEY &&
1537                             key.offset == num_bytes)
1538                                 ret = 0;
1539                 }
1540                 if (ret) {
1541                         key.objectid = bytenr;
1542                         key.type = BTRFS_EXTENT_ITEM_KEY;
1543                         key.offset = num_bytes;
1544                         btrfs_release_path(path);
1545                         goto again;
1546                 }
1547         }
1548
1549         if (ret && !insert) {
1550                 err = -ENOENT;
1551                 goto out;
1552         } else if (WARN_ON(ret)) {
1553                 err = -EIO;
1554                 goto out;
1555         }
1556
1557         leaf = path->nodes[0];
1558         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1559         if (unlikely(item_size < sizeof(*ei))) {
1560                 err = -EINVAL;
1561                 btrfs_print_v0_err(fs_info);
1562                 btrfs_abort_transaction(trans, err);
1563                 goto out;
1564         }
1565
1566         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1567         flags = btrfs_extent_flags(leaf, ei);
1568
1569         ptr = (unsigned long)(ei + 1);
1570         end = (unsigned long)ei + item_size;
1571
1572         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
1573                 ptr += sizeof(struct btrfs_tree_block_info);
1574                 BUG_ON(ptr > end);
1575         }
1576
1577         if (owner >= BTRFS_FIRST_FREE_OBJECTID)
1578                 needed = BTRFS_REF_TYPE_DATA;
1579         else
1580                 needed = BTRFS_REF_TYPE_BLOCK;
1581
1582         err = -ENOENT;
1583         while (1) {
1584                 if (ptr >= end) {
1585                         WARN_ON(ptr > end);
1586                         break;
1587                 }
1588                 iref = (struct btrfs_extent_inline_ref *)ptr;
1589                 type = btrfs_get_extent_inline_ref_type(leaf, iref, needed);
1590                 if (type == BTRFS_REF_TYPE_INVALID) {
1591                         err = -EUCLEAN;
1592                         goto out;
1593                 }
1594
1595                 if (want < type)
1596                         break;
1597                 if (want > type) {
1598                         ptr += btrfs_extent_inline_ref_size(type);
1599                         continue;
1600                 }
1601
1602                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1603                         struct btrfs_extent_data_ref *dref;
1604                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1605                         if (match_extent_data_ref(leaf, dref, root_objectid,
1606                                                   owner, offset)) {
1607                                 err = 0;
1608                                 break;
1609                         }
1610                         if (hash_extent_data_ref_item(leaf, dref) <
1611                             hash_extent_data_ref(root_objectid, owner, offset))
1612                                 break;
1613                 } else {
1614                         u64 ref_offset;
1615                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1616                         if (parent > 0) {
1617                                 if (parent == ref_offset) {
1618                                         err = 0;
1619                                         break;
1620                                 }
1621                                 if (ref_offset < parent)
1622                                         break;
1623                         } else {
1624                                 if (root_objectid == ref_offset) {
1625                                         err = 0;
1626                                         break;
1627                                 }
1628                                 if (ref_offset < root_objectid)
1629                                         break;
1630                         }
1631                 }
1632                 ptr += btrfs_extent_inline_ref_size(type);
1633         }
1634         if (err == -ENOENT && insert) {
1635                 if (item_size + extra_size >=
1636                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1637                         err = -EAGAIN;
1638                         goto out;
1639                 }
1640                 /*
1641                  * To add new inline back ref, we have to make sure
1642                  * there is no corresponding back ref item.
1643                  * For simplicity, we just do not add new inline back
1644                  * ref if there is any kind of item for this block
1645                  */
1646                 if (find_next_key(path, 0, &key) == 0 &&
1647                     key.objectid == bytenr &&
1648                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1649                         err = -EAGAIN;
1650                         goto out;
1651                 }
1652         }
1653         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1654 out:
1655         if (insert) {
1656                 path->keep_locks = 0;
1657                 btrfs_unlock_up_safe(path, 1);
1658         }
1659         return err;
1660 }
1661
1662 /*
1663  * helper to add new inline back ref
1664  */
1665 static noinline_for_stack
1666 void setup_inline_extent_backref(struct btrfs_fs_info *fs_info,
1667                                  struct btrfs_path *path,
1668                                  struct btrfs_extent_inline_ref *iref,
1669                                  u64 parent, u64 root_objectid,
1670                                  u64 owner, u64 offset, int refs_to_add,
1671                                  struct btrfs_delayed_extent_op *extent_op)
1672 {
1673         struct extent_buffer *leaf;
1674         struct btrfs_extent_item *ei;
1675         unsigned long ptr;
1676         unsigned long end;
1677         unsigned long item_offset;
1678         u64 refs;
1679         int size;
1680         int type;
1681
1682         leaf = path->nodes[0];
1683         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1684         item_offset = (unsigned long)iref - (unsigned long)ei;
1685
1686         type = extent_ref_type(parent, owner);
1687         size = btrfs_extent_inline_ref_size(type);
1688
1689         btrfs_extend_item(fs_info, path, size);
1690
1691         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1692         refs = btrfs_extent_refs(leaf, ei);
1693         refs += refs_to_add;
1694         btrfs_set_extent_refs(leaf, ei, refs);
1695         if (extent_op)
1696                 __run_delayed_extent_op(extent_op, leaf, ei);
1697
1698         ptr = (unsigned long)ei + item_offset;
1699         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1700         if (ptr < end - size)
1701                 memmove_extent_buffer(leaf, ptr + size, ptr,
1702                                       end - size - ptr);
1703
1704         iref = (struct btrfs_extent_inline_ref *)ptr;
1705         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1706         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1707                 struct btrfs_extent_data_ref *dref;
1708                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1709                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1710                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1711                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1712                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1713         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1714                 struct btrfs_shared_data_ref *sref;
1715                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1716                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1717                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1718         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1719                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1720         } else {
1721                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1722         }
1723         btrfs_mark_buffer_dirty(leaf);
1724 }
1725
1726 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1727                                  struct btrfs_path *path,
1728                                  struct btrfs_extent_inline_ref **ref_ret,
1729                                  u64 bytenr, u64 num_bytes, u64 parent,
1730                                  u64 root_objectid, u64 owner, u64 offset)
1731 {
1732         int ret;
1733
1734         ret = lookup_inline_extent_backref(trans, path, ref_ret, bytenr,
1735                                            num_bytes, parent, root_objectid,
1736                                            owner, offset, 0);
1737         if (ret != -ENOENT)
1738                 return ret;
1739
1740         btrfs_release_path(path);
1741         *ref_ret = NULL;
1742
1743         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1744                 ret = lookup_tree_block_ref(trans, path, bytenr, parent,
1745                                             root_objectid);
1746         } else {
1747                 ret = lookup_extent_data_ref(trans, path, bytenr, parent,
1748                                              root_objectid, owner, offset);
1749         }
1750         return ret;
1751 }
1752
1753 /*
1754  * helper to update/remove inline back ref
1755  */
1756 static noinline_for_stack
1757 void update_inline_extent_backref(struct btrfs_path *path,
1758                                   struct btrfs_extent_inline_ref *iref,
1759                                   int refs_to_mod,
1760                                   struct btrfs_delayed_extent_op *extent_op,
1761                                   int *last_ref)
1762 {
1763         struct extent_buffer *leaf = path->nodes[0];
1764         struct btrfs_fs_info *fs_info = leaf->fs_info;
1765         struct btrfs_extent_item *ei;
1766         struct btrfs_extent_data_ref *dref = NULL;
1767         struct btrfs_shared_data_ref *sref = NULL;
1768         unsigned long ptr;
1769         unsigned long end;
1770         u32 item_size;
1771         int size;
1772         int type;
1773         u64 refs;
1774
1775         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1776         refs = btrfs_extent_refs(leaf, ei);
1777         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1778         refs += refs_to_mod;
1779         btrfs_set_extent_refs(leaf, ei, refs);
1780         if (extent_op)
1781                 __run_delayed_extent_op(extent_op, leaf, ei);
1782
1783         /*
1784          * If type is invalid, we should have bailed out after
1785          * lookup_inline_extent_backref().
1786          */
1787         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_ANY);
1788         ASSERT(type != BTRFS_REF_TYPE_INVALID);
1789
1790         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1791                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1792                 refs = btrfs_extent_data_ref_count(leaf, dref);
1793         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1794                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1795                 refs = btrfs_shared_data_ref_count(leaf, sref);
1796         } else {
1797                 refs = 1;
1798                 BUG_ON(refs_to_mod != -1);
1799         }
1800
1801         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1802         refs += refs_to_mod;
1803
1804         if (refs > 0) {
1805                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1806                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1807                 else
1808                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1809         } else {
1810                 *last_ref = 1;
1811                 size =  btrfs_extent_inline_ref_size(type);
1812                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1813                 ptr = (unsigned long)iref;
1814                 end = (unsigned long)ei + item_size;
1815                 if (ptr + size < end)
1816                         memmove_extent_buffer(leaf, ptr, ptr + size,
1817                                               end - ptr - size);
1818                 item_size -= size;
1819                 btrfs_truncate_item(fs_info, path, item_size, 1);
1820         }
1821         btrfs_mark_buffer_dirty(leaf);
1822 }
1823
1824 static noinline_for_stack
1825 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1826                                  struct btrfs_path *path,
1827                                  u64 bytenr, u64 num_bytes, u64 parent,
1828                                  u64 root_objectid, u64 owner,
1829                                  u64 offset, int refs_to_add,
1830                                  struct btrfs_delayed_extent_op *extent_op)
1831 {
1832         struct btrfs_extent_inline_ref *iref;
1833         int ret;
1834
1835         ret = lookup_inline_extent_backref(trans, path, &iref, bytenr,
1836                                            num_bytes, parent, root_objectid,
1837                                            owner, offset, 1);
1838         if (ret == 0) {
1839                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1840                 update_inline_extent_backref(path, iref, refs_to_add,
1841                                              extent_op, NULL);
1842         } else if (ret == -ENOENT) {
1843                 setup_inline_extent_backref(trans->fs_info, path, iref, parent,
1844                                             root_objectid, owner, offset,
1845                                             refs_to_add, extent_op);
1846                 ret = 0;
1847         }
1848         return ret;
1849 }
1850
1851 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1852                                  struct btrfs_path *path,
1853                                  u64 bytenr, u64 parent, u64 root_objectid,
1854                                  u64 owner, u64 offset, int refs_to_add)
1855 {
1856         int ret;
1857         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1858                 BUG_ON(refs_to_add != 1);
1859                 ret = insert_tree_block_ref(trans, path, bytenr, parent,
1860                                             root_objectid);
1861         } else {
1862                 ret = insert_extent_data_ref(trans, path, bytenr, parent,
1863                                              root_objectid, owner, offset,
1864                                              refs_to_add);
1865         }
1866         return ret;
1867 }
1868
1869 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1870                                  struct btrfs_path *path,
1871                                  struct btrfs_extent_inline_ref *iref,
1872                                  int refs_to_drop, int is_data, int *last_ref)
1873 {
1874         int ret = 0;
1875
1876         BUG_ON(!is_data && refs_to_drop != 1);
1877         if (iref) {
1878                 update_inline_extent_backref(path, iref, -refs_to_drop, NULL,
1879                                              last_ref);
1880         } else if (is_data) {
1881                 ret = remove_extent_data_ref(trans, path, refs_to_drop,
1882                                              last_ref);
1883         } else {
1884                 *last_ref = 1;
1885                 ret = btrfs_del_item(trans, trans->fs_info->extent_root, path);
1886         }
1887         return ret;
1888 }
1889
1890 #define in_range(b, first, len)        ((b) >= (first) && (b) < (first) + (len))
1891 static int btrfs_issue_discard(struct block_device *bdev, u64 start, u64 len,
1892                                u64 *discarded_bytes)
1893 {
1894         int j, ret = 0;
1895         u64 bytes_left, end;
1896         u64 aligned_start = ALIGN(start, 1 << 9);
1897
1898         if (WARN_ON(start != aligned_start)) {
1899                 len -= aligned_start - start;
1900                 len = round_down(len, 1 << 9);
1901                 start = aligned_start;
1902         }
1903
1904         *discarded_bytes = 0;
1905
1906         if (!len)
1907                 return 0;
1908
1909         end = start + len;
1910         bytes_left = len;
1911
1912         /* Skip any superblocks on this device. */
1913         for (j = 0; j < BTRFS_SUPER_MIRROR_MAX; j++) {
1914                 u64 sb_start = btrfs_sb_offset(j);
1915                 u64 sb_end = sb_start + BTRFS_SUPER_INFO_SIZE;
1916                 u64 size = sb_start - start;
1917
1918                 if (!in_range(sb_start, start, bytes_left) &&
1919                     !in_range(sb_end, start, bytes_left) &&
1920                     !in_range(start, sb_start, BTRFS_SUPER_INFO_SIZE))
1921                         continue;
1922
1923                 /*
1924                  * Superblock spans beginning of range.  Adjust start and
1925                  * try again.
1926                  */
1927                 if (sb_start <= start) {
1928                         start += sb_end - start;
1929                         if (start > end) {
1930                                 bytes_left = 0;
1931                                 break;
1932                         }
1933                         bytes_left = end - start;
1934                         continue;
1935                 }
1936
1937                 if (size) {
1938                         ret = blkdev_issue_discard(bdev, start >> 9, size >> 9,
1939                                                    GFP_NOFS, 0);
1940                         if (!ret)
1941                                 *discarded_bytes += size;
1942                         else if (ret != -EOPNOTSUPP)
1943                                 return ret;
1944                 }
1945
1946                 start = sb_end;
1947                 if (start > end) {
1948                         bytes_left = 0;
1949                         break;
1950                 }
1951                 bytes_left = end - start;
1952         }
1953
1954         if (bytes_left) {
1955                 ret = blkdev_issue_discard(bdev, start >> 9, bytes_left >> 9,
1956                                            GFP_NOFS, 0);
1957                 if (!ret)
1958                         *discarded_bytes += bytes_left;
1959         }
1960         return ret;
1961 }
1962
1963 int btrfs_discard_extent(struct btrfs_fs_info *fs_info, u64 bytenr,
1964                          u64 num_bytes, u64 *actual_bytes)
1965 {
1966         int ret;
1967         u64 discarded_bytes = 0;
1968         struct btrfs_bio *bbio = NULL;
1969
1970
1971         /*
1972          * Avoid races with device replace and make sure our bbio has devices
1973          * associated to its stripes that don't go away while we are discarding.
1974          */
1975         btrfs_bio_counter_inc_blocked(fs_info);
1976         /* Tell the block device(s) that the sectors can be discarded */
1977         ret = btrfs_map_block(fs_info, BTRFS_MAP_DISCARD, bytenr, &num_bytes,
1978                               &bbio, 0);
1979         /* Error condition is -ENOMEM */
1980         if (!ret) {
1981                 struct btrfs_bio_stripe *stripe = bbio->stripes;
1982                 int i;
1983
1984
1985                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
1986                         u64 bytes;
1987                         struct request_queue *req_q;
1988
1989                         if (!stripe->dev->bdev) {
1990                                 ASSERT(btrfs_test_opt(fs_info, DEGRADED));
1991                                 continue;
1992                         }
1993                         req_q = bdev_get_queue(stripe->dev->bdev);
1994                         if (!blk_queue_discard(req_q))
1995                                 continue;
1996
1997                         ret = btrfs_issue_discard(stripe->dev->bdev,
1998                                                   stripe->physical,
1999                                                   stripe->length,
2000                                                   &bytes);
2001                         if (!ret)
2002                                 discarded_bytes += bytes;
2003                         else if (ret != -EOPNOTSUPP)
2004                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
2005
2006                         /*
2007                          * Just in case we get back EOPNOTSUPP for some reason,
2008                          * just ignore the return value so we don't screw up
2009                          * people calling discard_extent.
2010                          */
2011                         ret = 0;
2012                 }
2013                 btrfs_put_bbio(bbio);
2014         }
2015         btrfs_bio_counter_dec(fs_info);
2016
2017         if (actual_bytes)
2018                 *actual_bytes = discarded_bytes;
2019
2020
2021         if (ret == -EOPNOTSUPP)
2022                 ret = 0;
2023         return ret;
2024 }
2025
2026 /* Can return -ENOMEM */
2027 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2028                          struct btrfs_root *root,
2029                          u64 bytenr, u64 num_bytes, u64 parent,
2030                          u64 root_objectid, u64 owner, u64 offset)
2031 {
2032         struct btrfs_fs_info *fs_info = root->fs_info;
2033         int old_ref_mod, new_ref_mod;
2034         int ret;
2035
2036         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
2037                root_objectid == BTRFS_TREE_LOG_OBJECTID);
2038
2039         btrfs_ref_tree_mod(root, bytenr, num_bytes, parent, root_objectid,
2040                            owner, offset, BTRFS_ADD_DELAYED_REF);
2041
2042         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
2043                 ret = btrfs_add_delayed_tree_ref(trans, bytenr,
2044                                                  num_bytes, parent,
2045                                                  root_objectid, (int)owner,
2046                                                  BTRFS_ADD_DELAYED_REF, NULL,
2047                                                  &old_ref_mod, &new_ref_mod);
2048         } else {
2049                 ret = btrfs_add_delayed_data_ref(trans, bytenr,
2050                                                  num_bytes, parent,
2051                                                  root_objectid, owner, offset,
2052                                                  0, BTRFS_ADD_DELAYED_REF,
2053                                                  &old_ref_mod, &new_ref_mod);
2054         }
2055
2056         if (ret == 0 && old_ref_mod < 0 && new_ref_mod >= 0) {
2057                 bool metadata = owner < BTRFS_FIRST_FREE_OBJECTID;
2058
2059                 add_pinned_bytes(fs_info, -num_bytes, metadata, root_objectid);
2060         }
2061
2062         return ret;
2063 }
2064
2065 /*
2066  * __btrfs_inc_extent_ref - insert backreference for a given extent
2067  *
2068  * @trans:          Handle of transaction
2069  *
2070  * @node:           The delayed ref node used to get the bytenr/length for
2071  *                  extent whose references are incremented.
2072  *
2073  * @parent:         If this is a shared extent (BTRFS_SHARED_DATA_REF_KEY/
2074  *                  BTRFS_SHARED_BLOCK_REF_KEY) then it holds the logical
2075  *                  bytenr of the parent block. Since new extents are always
2076  *                  created with indirect references, this will only be the case
2077  *                  when relocating a shared extent. In that case, root_objectid
2078  *                  will be BTRFS_TREE_RELOC_OBJECTID. Otheriwse, parent must
2079  *                  be 0
2080  *
2081  * @root_objectid:  The id of the root where this modification has originated,
2082  *                  this can be either one of the well-known metadata trees or
2083  *                  the subvolume id which references this extent.
2084  *
2085  * @owner:          For data extents it is the inode number of the owning file.
2086  *                  For metadata extents this parameter holds the level in the
2087  *                  tree of the extent.
2088  *
2089  * @offset:         For metadata extents the offset is ignored and is currently
2090  *                  always passed as 0. For data extents it is the fileoffset
2091  *                  this extent belongs to.
2092  *
2093  * @refs_to_add     Number of references to add
2094  *
2095  * @extent_op       Pointer to a structure, holding information necessary when
2096  *                  updating a tree block's flags
2097  *
2098  */
2099 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2100                                   struct btrfs_delayed_ref_node *node,
2101                                   u64 parent, u64 root_objectid,
2102                                   u64 owner, u64 offset, int refs_to_add,
2103                                   struct btrfs_delayed_extent_op *extent_op)
2104 {
2105         struct btrfs_path *path;
2106         struct extent_buffer *leaf;
2107         struct btrfs_extent_item *item;
2108         struct btrfs_key key;
2109         u64 bytenr = node->bytenr;
2110         u64 num_bytes = node->num_bytes;
2111         u64 refs;
2112         int ret;
2113
2114         path = btrfs_alloc_path();
2115         if (!path)
2116                 return -ENOMEM;
2117
2118         path->reada = READA_FORWARD;
2119         path->leave_spinning = 1;
2120         /* this will setup the path even if it fails to insert the back ref */
2121         ret = insert_inline_extent_backref(trans, path, bytenr, num_bytes,
2122                                            parent, root_objectid, owner,
2123                                            offset, refs_to_add, extent_op);
2124         if ((ret < 0 && ret != -EAGAIN) || !ret)
2125                 goto out;
2126
2127         /*
2128          * Ok we had -EAGAIN which means we didn't have space to insert and
2129          * inline extent ref, so just update the reference count and add a
2130          * normal backref.
2131          */
2132         leaf = path->nodes[0];
2133         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2134         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2135         refs = btrfs_extent_refs(leaf, item);
2136         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
2137         if (extent_op)
2138                 __run_delayed_extent_op(extent_op, leaf, item);
2139
2140         btrfs_mark_buffer_dirty(leaf);
2141         btrfs_release_path(path);
2142
2143         path->reada = READA_FORWARD;
2144         path->leave_spinning = 1;
2145         /* now insert the actual backref */
2146         ret = insert_extent_backref(trans, path, bytenr, parent, root_objectid,
2147                                     owner, offset, refs_to_add);
2148         if (ret)
2149                 btrfs_abort_transaction(trans, ret);
2150 out:
2151         btrfs_free_path(path);
2152         return ret;
2153 }
2154
2155 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
2156                                 struct btrfs_delayed_ref_node *node,
2157                                 struct btrfs_delayed_extent_op *extent_op,
2158                                 int insert_reserved)
2159 {
2160         int ret = 0;
2161         struct btrfs_delayed_data_ref *ref;
2162         struct btrfs_key ins;
2163         u64 parent = 0;
2164         u64 ref_root = 0;
2165         u64 flags = 0;
2166
2167         ins.objectid = node->bytenr;
2168         ins.offset = node->num_bytes;
2169         ins.type = BTRFS_EXTENT_ITEM_KEY;
2170
2171         ref = btrfs_delayed_node_to_data_ref(node);
2172         trace_run_delayed_data_ref(trans->fs_info, node, ref, node->action);
2173
2174         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2175                 parent = ref->parent;
2176         ref_root = ref->root;
2177
2178         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2179                 if (extent_op)
2180                         flags |= extent_op->flags_to_set;
2181                 ret = alloc_reserved_file_extent(trans, parent, ref_root,
2182                                                  flags, ref->objectid,
2183                                                  ref->offset, &ins,
2184                                                  node->ref_mod);
2185         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2186                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
2187                                              ref->objectid, ref->offset,
2188                                              node->ref_mod, extent_op);
2189         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2190                 ret = __btrfs_free_extent(trans, node, parent,
2191                                           ref_root, ref->objectid,
2192                                           ref->offset, node->ref_mod,
2193                                           extent_op);
2194         } else {
2195                 BUG();
2196         }
2197         return ret;
2198 }
2199
2200 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2201                                     struct extent_buffer *leaf,
2202                                     struct btrfs_extent_item *ei)
2203 {
2204         u64 flags = btrfs_extent_flags(leaf, ei);
2205         if (extent_op->update_flags) {
2206                 flags |= extent_op->flags_to_set;
2207                 btrfs_set_extent_flags(leaf, ei, flags);
2208         }
2209
2210         if (extent_op->update_key) {
2211                 struct btrfs_tree_block_info *bi;
2212                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2213                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2214                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2215         }
2216 }
2217
2218 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2219                                  struct btrfs_delayed_ref_head *head,
2220                                  struct btrfs_delayed_extent_op *extent_op)
2221 {
2222         struct btrfs_fs_info *fs_info = trans->fs_info;
2223         struct btrfs_key key;
2224         struct btrfs_path *path;
2225         struct btrfs_extent_item *ei;
2226         struct extent_buffer *leaf;
2227         u32 item_size;
2228         int ret;
2229         int err = 0;
2230         int metadata = !extent_op->is_data;
2231
2232         if (trans->aborted)
2233                 return 0;
2234
2235         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2236                 metadata = 0;
2237
2238         path = btrfs_alloc_path();
2239         if (!path)
2240                 return -ENOMEM;
2241
2242         key.objectid = head->bytenr;
2243
2244         if (metadata) {
2245                 key.type = BTRFS_METADATA_ITEM_KEY;
2246                 key.offset = extent_op->level;
2247         } else {
2248                 key.type = BTRFS_EXTENT_ITEM_KEY;
2249                 key.offset = head->num_bytes;
2250         }
2251
2252 again:
2253         path->reada = READA_FORWARD;
2254         path->leave_spinning = 1;
2255         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 1);
2256         if (ret < 0) {
2257                 err = ret;
2258                 goto out;
2259         }
2260         if (ret > 0) {
2261                 if (metadata) {
2262                         if (path->slots[0] > 0) {
2263                                 path->slots[0]--;
2264                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
2265                                                       path->slots[0]);
2266                                 if (key.objectid == head->bytenr &&
2267                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
2268                                     key.offset == head->num_bytes)
2269                                         ret = 0;
2270                         }
2271                         if (ret > 0) {
2272                                 btrfs_release_path(path);
2273                                 metadata = 0;
2274
2275                                 key.objectid = head->bytenr;
2276                                 key.offset = head->num_bytes;
2277                                 key.type = BTRFS_EXTENT_ITEM_KEY;
2278                                 goto again;
2279                         }
2280                 } else {
2281                         err = -EIO;
2282                         goto out;
2283                 }
2284         }
2285
2286         leaf = path->nodes[0];
2287         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2288
2289         if (unlikely(item_size < sizeof(*ei))) {
2290                 err = -EINVAL;
2291                 btrfs_print_v0_err(fs_info);
2292                 btrfs_abort_transaction(trans, err);
2293                 goto out;
2294         }
2295
2296         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2297         __run_delayed_extent_op(extent_op, leaf, ei);
2298
2299         btrfs_mark_buffer_dirty(leaf);
2300 out:
2301         btrfs_free_path(path);
2302         return err;
2303 }
2304
2305 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2306                                 struct btrfs_delayed_ref_node *node,
2307                                 struct btrfs_delayed_extent_op *extent_op,
2308                                 int insert_reserved)
2309 {
2310         int ret = 0;
2311         struct btrfs_delayed_tree_ref *ref;
2312         u64 parent = 0;
2313         u64 ref_root = 0;
2314
2315         ref = btrfs_delayed_node_to_tree_ref(node);
2316         trace_run_delayed_tree_ref(trans->fs_info, node, ref, node->action);
2317
2318         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2319                 parent = ref->parent;
2320         ref_root = ref->root;
2321
2322         if (node->ref_mod != 1) {
2323                 btrfs_err(trans->fs_info,
2324         "btree block(%llu) has %d references rather than 1: action %d ref_root %llu parent %llu",
2325                           node->bytenr, node->ref_mod, node->action, ref_root,
2326                           parent);
2327                 return -EIO;
2328         }
2329         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2330                 BUG_ON(!extent_op || !extent_op->update_flags);
2331                 ret = alloc_reserved_tree_block(trans, node, extent_op);
2332         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2333                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
2334                                              ref->level, 0, 1, extent_op);
2335         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2336                 ret = __btrfs_free_extent(trans, node, parent, ref_root,
2337                                           ref->level, 0, 1, extent_op);
2338         } else {
2339                 BUG();
2340         }
2341         return ret;
2342 }
2343
2344 /* helper function to actually process a single delayed ref entry */
2345 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2346                                struct btrfs_delayed_ref_node *node,
2347                                struct btrfs_delayed_extent_op *extent_op,
2348                                int insert_reserved)
2349 {
2350         int ret = 0;
2351
2352         if (trans->aborted) {
2353                 if (insert_reserved)
2354                         btrfs_pin_extent(trans->fs_info, node->bytenr,
2355                                          node->num_bytes, 1);
2356                 return 0;
2357         }
2358
2359         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2360             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2361                 ret = run_delayed_tree_ref(trans, node, extent_op,
2362                                            insert_reserved);
2363         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2364                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2365                 ret = run_delayed_data_ref(trans, node, extent_op,
2366                                            insert_reserved);
2367         else
2368                 BUG();
2369         return ret;
2370 }
2371
2372 static inline struct btrfs_delayed_ref_node *
2373 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2374 {
2375         struct btrfs_delayed_ref_node *ref;
2376
2377         if (RB_EMPTY_ROOT(&head->ref_tree))
2378                 return NULL;
2379
2380         /*
2381          * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
2382          * This is to prevent a ref count from going down to zero, which deletes
2383          * the extent item from the extent tree, when there still are references
2384          * to add, which would fail because they would not find the extent item.
2385          */
2386         if (!list_empty(&head->ref_add_list))
2387                 return list_first_entry(&head->ref_add_list,
2388                                 struct btrfs_delayed_ref_node, add_list);
2389
2390         ref = rb_entry(rb_first(&head->ref_tree),
2391                        struct btrfs_delayed_ref_node, ref_node);
2392         ASSERT(list_empty(&ref->add_list));
2393         return ref;
2394 }
2395
2396 static void unselect_delayed_ref_head(struct btrfs_delayed_ref_root *delayed_refs,
2397                                       struct btrfs_delayed_ref_head *head)
2398 {
2399         spin_lock(&delayed_refs->lock);
2400         head->processing = 0;
2401         delayed_refs->num_heads_ready++;
2402         spin_unlock(&delayed_refs->lock);
2403         btrfs_delayed_ref_unlock(head);
2404 }
2405
2406 static int cleanup_extent_op(struct btrfs_trans_handle *trans,
2407                              struct btrfs_delayed_ref_head *head)
2408 {
2409         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
2410         int ret;
2411
2412         if (!extent_op)
2413                 return 0;
2414         head->extent_op = NULL;
2415         if (head->must_insert_reserved) {
2416                 btrfs_free_delayed_extent_op(extent_op);
2417                 return 0;
2418         }
2419         spin_unlock(&head->lock);
2420         ret = run_delayed_extent_op(trans, head, extent_op);
2421         btrfs_free_delayed_extent_op(extent_op);
2422         return ret ? ret : 1;
2423 }
2424
2425 static int cleanup_ref_head(struct btrfs_trans_handle *trans,
2426                             struct btrfs_delayed_ref_head *head)
2427 {
2428
2429         struct btrfs_fs_info *fs_info = trans->fs_info;
2430         struct btrfs_delayed_ref_root *delayed_refs;
2431         int ret;
2432
2433         delayed_refs = &trans->transaction->delayed_refs;
2434
2435         ret = cleanup_extent_op(trans, head);
2436         if (ret < 0) {
2437                 unselect_delayed_ref_head(delayed_refs, head);
2438                 btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
2439                 return ret;
2440         } else if (ret) {
2441                 return ret;
2442         }
2443
2444         /*
2445          * Need to drop our head ref lock and re-acquire the delayed ref lock
2446          * and then re-check to make sure nobody got added.
2447          */
2448         spin_unlock(&head->lock);
2449         spin_lock(&delayed_refs->lock);
2450         spin_lock(&head->lock);
2451         if (!RB_EMPTY_ROOT(&head->ref_tree) || head->extent_op) {
2452                 spin_unlock(&head->lock);
2453                 spin_unlock(&delayed_refs->lock);
2454                 return 1;
2455         }
2456         delayed_refs->num_heads--;
2457         rb_erase(&head->href_node, &delayed_refs->href_root);
2458         RB_CLEAR_NODE(&head->href_node);
2459         spin_unlock(&head->lock);
2460         spin_unlock(&delayed_refs->lock);
2461         atomic_dec(&delayed_refs->num_entries);
2462
2463         trace_run_delayed_ref_head(fs_info, head, 0);
2464
2465         if (head->total_ref_mod < 0) {
2466                 struct btrfs_space_info *space_info;
2467                 u64 flags;
2468
2469                 if (head->is_data)
2470                         flags = BTRFS_BLOCK_GROUP_DATA;
2471                 else if (head->is_system)
2472                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
2473                 else
2474                         flags = BTRFS_BLOCK_GROUP_METADATA;
2475                 space_info = __find_space_info(fs_info, flags);
2476                 ASSERT(space_info);
2477                 percpu_counter_add_batch(&space_info->total_bytes_pinned,
2478                                    -head->num_bytes,
2479                                    BTRFS_TOTAL_BYTES_PINNED_BATCH);
2480
2481                 if (head->is_data) {
2482                         spin_lock(&delayed_refs->lock);
2483                         delayed_refs->pending_csums -= head->num_bytes;
2484                         spin_unlock(&delayed_refs->lock);
2485                 }
2486         }
2487
2488         if (head->must_insert_reserved) {
2489                 btrfs_pin_extent(fs_info, head->bytenr,
2490                                  head->num_bytes, 1);
2491                 if (head->is_data) {
2492                         ret = btrfs_del_csums(trans, fs_info, head->bytenr,
2493                                               head->num_bytes);
2494                 }
2495         }
2496
2497         /* Also free its reserved qgroup space */
2498         btrfs_qgroup_free_delayed_ref(fs_info, head->qgroup_ref_root,
2499                                       head->qgroup_reserved);
2500         btrfs_delayed_ref_unlock(head);
2501         btrfs_put_delayed_ref_head(head);
2502         return 0;
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                                              unsigned long nr)
2511 {
2512         struct btrfs_fs_info *fs_info = trans->fs_info;
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, delayed_refs, locked_ref);
2567
2568                 ref = select_delayed_ref(locked_ref);
2569
2570                 if (ref && ref->seq &&
2571                     btrfs_check_delayed_seq(fs_info, ref->seq)) {
2572                         spin_unlock(&locked_ref->lock);
2573                         unselect_delayed_ref_head(delayed_refs, locked_ref);
2574                         locked_ref = NULL;
2575                         cond_resched();
2576                         count++;
2577                         continue;
2578                 }
2579
2580                 /*
2581                  * We're done processing refs in this ref_head, clean everything
2582                  * up and move on to the next ref_head.
2583                  */
2584                 if (!ref) {
2585                         ret = cleanup_ref_head(trans, locked_ref);
2586                         if (ret > 0 ) {
2587                                 /* We dropped our lock, we need to loop. */
2588                                 ret = 0;
2589                                 continue;
2590                         } else if (ret) {
2591                                 return ret;
2592                         }
2593                         locked_ref = NULL;
2594                         count++;
2595                         continue;
2596                 }
2597
2598                 actual_count++;
2599                 ref->in_tree = 0;
2600                 rb_erase(&ref->ref_node, &locked_ref->ref_tree);
2601                 RB_CLEAR_NODE(&ref->ref_node);
2602                 if (!list_empty(&ref->add_list))
2603                         list_del(&ref->add_list);
2604                 /*
2605                  * When we play the delayed ref, also correct the ref_mod on
2606                  * head
2607                  */
2608                 switch (ref->action) {
2609                 case BTRFS_ADD_DELAYED_REF:
2610                 case BTRFS_ADD_DELAYED_EXTENT:
2611                         locked_ref->ref_mod -= ref->ref_mod;
2612                         break;
2613                 case BTRFS_DROP_DELAYED_REF:
2614                         locked_ref->ref_mod += ref->ref_mod;
2615                         break;
2616                 default:
2617                         WARN_ON(1);
2618                 }
2619                 atomic_dec(&delayed_refs->num_entries);
2620
2621                 /*
2622                  * Record the must-insert_reserved flag before we drop the spin
2623                  * lock.
2624                  */
2625                 must_insert_reserved = locked_ref->must_insert_reserved;
2626                 locked_ref->must_insert_reserved = 0;
2627
2628                 extent_op = locked_ref->extent_op;
2629                 locked_ref->extent_op = NULL;
2630                 spin_unlock(&locked_ref->lock);
2631
2632                 ret = run_one_delayed_ref(trans, ref, extent_op,
2633                                           must_insert_reserved);
2634
2635                 btrfs_free_delayed_extent_op(extent_op);
2636                 if (ret) {
2637                         unselect_delayed_ref_head(delayed_refs, locked_ref);
2638                         btrfs_put_delayed_ref(ref);
2639                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d",
2640                                     ret);
2641                         return ret;
2642                 }
2643
2644                 btrfs_put_delayed_ref(ref);
2645                 count++;
2646                 cond_resched();
2647         }
2648
2649         /*
2650          * We don't want to include ref heads since we can have empty ref heads
2651          * and those will drastically skew our runtime down since we just do
2652          * accounting, no actual extent tree updates.
2653          */
2654         if (actual_count > 0) {
2655                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2656                 u64 avg;
2657
2658                 /*
2659                  * We weigh the current average higher than our current runtime
2660                  * to avoid large swings in the average.
2661                  */
2662                 spin_lock(&delayed_refs->lock);
2663                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2664                 fs_info->avg_delayed_ref_runtime = avg >> 2;    /* div by 4 */
2665                 spin_unlock(&delayed_refs->lock);
2666         }
2667         return 0;
2668 }
2669
2670 #ifdef SCRAMBLE_DELAYED_REFS
2671 /*
2672  * Normally delayed refs get processed in ascending bytenr order. This
2673  * correlates in most cases to the order added. To expose dependencies on this
2674  * order, we start to process the tree in the middle instead of the beginning
2675  */
2676 static u64 find_middle(struct rb_root *root)
2677 {
2678         struct rb_node *n = root->rb_node;
2679         struct btrfs_delayed_ref_node *entry;
2680         int alt = 1;
2681         u64 middle;
2682         u64 first = 0, last = 0;
2683
2684         n = rb_first(root);
2685         if (n) {
2686                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2687                 first = entry->bytenr;
2688         }
2689         n = rb_last(root);
2690         if (n) {
2691                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2692                 last = entry->bytenr;
2693         }
2694         n = root->rb_node;
2695
2696         while (n) {
2697                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2698                 WARN_ON(!entry->in_tree);
2699
2700                 middle = entry->bytenr;
2701
2702                 if (alt)
2703                         n = n->rb_left;
2704                 else
2705                         n = n->rb_right;
2706
2707                 alt = 1 - alt;
2708         }
2709         return middle;
2710 }
2711 #endif
2712
2713 static inline u64 heads_to_leaves(struct btrfs_fs_info *fs_info, u64 heads)
2714 {
2715         u64 num_bytes;
2716
2717         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2718                              sizeof(struct btrfs_extent_inline_ref));
2719         if (!btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2720                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2721
2722         /*
2723          * We don't ever fill up leaves all the way so multiply by 2 just to be
2724          * closer to what we're really going to want to use.
2725          */
2726         return div_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(fs_info));
2727 }
2728
2729 /*
2730  * Takes the number of bytes to be csumm'ed and figures out how many leaves it
2731  * would require to store the csums for that many bytes.
2732  */
2733 u64 btrfs_csum_bytes_to_leaves(struct btrfs_fs_info *fs_info, u64 csum_bytes)
2734 {
2735         u64 csum_size;
2736         u64 num_csums_per_leaf;
2737         u64 num_csums;
2738
2739         csum_size = BTRFS_MAX_ITEM_SIZE(fs_info);
2740         num_csums_per_leaf = div64_u64(csum_size,
2741                         (u64)btrfs_super_csum_size(fs_info->super_copy));
2742         num_csums = div64_u64(csum_bytes, fs_info->sectorsize);
2743         num_csums += num_csums_per_leaf - 1;
2744         num_csums = div64_u64(num_csums, num_csums_per_leaf);
2745         return num_csums;
2746 }
2747
2748 int btrfs_check_space_for_delayed_refs(struct btrfs_trans_handle *trans,
2749                                        struct btrfs_fs_info *fs_info)
2750 {
2751         struct btrfs_block_rsv *global_rsv;
2752         u64 num_heads = trans->transaction->delayed_refs.num_heads_ready;
2753         u64 csum_bytes = trans->transaction->delayed_refs.pending_csums;
2754         unsigned int num_dirty_bgs = trans->transaction->num_dirty_bgs;
2755         u64 num_bytes, num_dirty_bgs_bytes;
2756         int ret = 0;
2757
2758         num_bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
2759         num_heads = heads_to_leaves(fs_info, num_heads);
2760         if (num_heads > 1)
2761                 num_bytes += (num_heads - 1) * fs_info->nodesize;
2762         num_bytes <<= 1;
2763         num_bytes += btrfs_csum_bytes_to_leaves(fs_info, csum_bytes) *
2764                                                         fs_info->nodesize;
2765         num_dirty_bgs_bytes = btrfs_calc_trans_metadata_size(fs_info,
2766                                                              num_dirty_bgs);
2767         global_rsv = &fs_info->global_block_rsv;
2768
2769         /*
2770          * If we can't allocate any more chunks lets make sure we have _lots_ of
2771          * wiggle room since running delayed refs can create more delayed refs.
2772          */
2773         if (global_rsv->space_info->full) {
2774                 num_dirty_bgs_bytes <<= 1;
2775                 num_bytes <<= 1;
2776         }
2777
2778         spin_lock(&global_rsv->lock);
2779         if (global_rsv->reserved <= num_bytes + num_dirty_bgs_bytes)
2780                 ret = 1;
2781         spin_unlock(&global_rsv->lock);
2782         return ret;
2783 }
2784
2785 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans,
2786                                        struct btrfs_fs_info *fs_info)
2787 {
2788         u64 num_entries =
2789                 atomic_read(&trans->transaction->delayed_refs.num_entries);
2790         u64 avg_runtime;
2791         u64 val;
2792
2793         smp_mb();
2794         avg_runtime = fs_info->avg_delayed_ref_runtime;
2795         val = num_entries * avg_runtime;
2796         if (val >= NSEC_PER_SEC)
2797                 return 1;
2798         if (val >= NSEC_PER_SEC / 2)
2799                 return 2;
2800
2801         return btrfs_check_space_for_delayed_refs(trans, fs_info);
2802 }
2803
2804 struct async_delayed_refs {
2805         struct btrfs_root *root;
2806         u64 transid;
2807         int count;
2808         int error;
2809         int sync;
2810         struct completion wait;
2811         struct btrfs_work work;
2812 };
2813
2814 static inline struct async_delayed_refs *
2815 to_async_delayed_refs(struct btrfs_work *work)
2816 {
2817         return container_of(work, struct async_delayed_refs, work);
2818 }
2819
2820 static void delayed_ref_async_start(struct btrfs_work *work)
2821 {
2822         struct async_delayed_refs *async = to_async_delayed_refs(work);
2823         struct btrfs_trans_handle *trans;
2824         struct btrfs_fs_info *fs_info = async->root->fs_info;
2825         int ret;
2826
2827         /* if the commit is already started, we don't need to wait here */
2828         if (btrfs_transaction_blocked(fs_info))
2829                 goto done;
2830
2831         trans = btrfs_join_transaction(async->root);
2832         if (IS_ERR(trans)) {
2833                 async->error = PTR_ERR(trans);
2834                 goto done;
2835         }
2836
2837         /*
2838          * trans->sync means that when we call end_transaction, we won't
2839          * wait on delayed refs
2840          */
2841         trans->sync = true;
2842
2843         /* Don't bother flushing if we got into a different transaction */
2844         if (trans->transid > async->transid)
2845                 goto end;
2846
2847         ret = btrfs_run_delayed_refs(trans, async->count);
2848         if (ret)
2849                 async->error = ret;
2850 end:
2851         ret = btrfs_end_transaction(trans);
2852         if (ret && !async->error)
2853                 async->error = ret;
2854 done:
2855         if (async->sync)
2856                 complete(&async->wait);
2857         else
2858                 kfree(async);
2859 }
2860
2861 int btrfs_async_run_delayed_refs(struct btrfs_fs_info *fs_info,
2862                                  unsigned long count, u64 transid, int wait)
2863 {
2864         struct async_delayed_refs *async;
2865         int ret;
2866
2867         async = kmalloc(sizeof(*async), GFP_NOFS);
2868         if (!async)
2869                 return -ENOMEM;
2870
2871         async->root = fs_info->tree_root;
2872         async->count = count;
2873         async->error = 0;
2874         async->transid = transid;
2875         if (wait)
2876                 async->sync = 1;
2877         else
2878                 async->sync = 0;
2879         init_completion(&async->wait);
2880
2881         btrfs_init_work(&async->work, btrfs_extent_refs_helper,
2882                         delayed_ref_async_start, NULL, NULL);
2883
2884         btrfs_queue_work(fs_info->extent_workers, &async->work);
2885
2886         if (wait) {
2887                 wait_for_completion(&async->wait);
2888                 ret = async->error;
2889                 kfree(async);
2890                 return ret;
2891         }
2892         return 0;
2893 }
2894
2895 /*
2896  * this starts processing the delayed reference count updates and
2897  * extent insertions we have queued up so far.  count can be
2898  * 0, which means to process everything in the tree at the start
2899  * of the run (but not newly added entries), or it can be some target
2900  * number you'd like to process.
2901  *
2902  * Returns 0 on success or if called with an aborted transaction
2903  * Returns <0 on error and aborts the transaction
2904  */
2905 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2906                            unsigned long count)
2907 {
2908         struct btrfs_fs_info *fs_info = trans->fs_info;
2909         struct rb_node *node;
2910         struct btrfs_delayed_ref_root *delayed_refs;
2911         struct btrfs_delayed_ref_head *head;
2912         int ret;
2913         int run_all = count == (unsigned long)-1;
2914         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
2915
2916         /* We'll clean this up in btrfs_cleanup_transaction */
2917         if (trans->aborted)
2918                 return 0;
2919
2920         if (test_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags))
2921                 return 0;
2922
2923         delayed_refs = &trans->transaction->delayed_refs;
2924         if (count == 0)
2925                 count = atomic_read(&delayed_refs->num_entries) * 2;
2926
2927 again:
2928 #ifdef SCRAMBLE_DELAYED_REFS
2929         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2930 #endif
2931         trans->can_flush_pending_bgs = false;
2932         ret = __btrfs_run_delayed_refs(trans, count);
2933         if (ret < 0) {
2934                 btrfs_abort_transaction(trans, ret);
2935                 return ret;
2936         }
2937
2938         if (run_all) {
2939                 if (!list_empty(&trans->new_bgs))
2940                         btrfs_create_pending_block_groups(trans);
2941
2942                 spin_lock(&delayed_refs->lock);
2943                 node = rb_first(&delayed_refs->href_root);
2944                 if (!node) {
2945                         spin_unlock(&delayed_refs->lock);
2946                         goto out;
2947                 }
2948                 head = rb_entry(node, struct btrfs_delayed_ref_head,
2949                                 href_node);
2950                 refcount_inc(&head->refs);
2951                 spin_unlock(&delayed_refs->lock);
2952
2953                 /* Mutex was contended, block until it's released and retry. */
2954                 mutex_lock(&head->mutex);
2955                 mutex_unlock(&head->mutex);
2956
2957                 btrfs_put_delayed_ref_head(head);
2958                 cond_resched();
2959                 goto again;
2960         }
2961 out:
2962         trans->can_flush_pending_bgs = can_flush_pending_bgs;
2963         return 0;
2964 }
2965
2966 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2967                                 struct btrfs_fs_info *fs_info,
2968                                 u64 bytenr, u64 num_bytes, u64 flags,
2969                                 int level, int is_data)
2970 {
2971         struct btrfs_delayed_extent_op *extent_op;
2972         int ret;
2973
2974         extent_op = btrfs_alloc_delayed_extent_op();
2975         if (!extent_op)
2976                 return -ENOMEM;
2977
2978         extent_op->flags_to_set = flags;
2979         extent_op->update_flags = true;
2980         extent_op->update_key = false;
2981         extent_op->is_data = is_data ? true : false;
2982         extent_op->level = level;
2983
2984         ret = btrfs_add_delayed_extent_op(fs_info, trans, bytenr,
2985                                           num_bytes, extent_op);
2986         if (ret)
2987                 btrfs_free_delayed_extent_op(extent_op);
2988         return ret;
2989 }
2990
2991 static noinline int check_delayed_ref(struct btrfs_root *root,
2992                                       struct btrfs_path *path,
2993                                       u64 objectid, u64 offset, u64 bytenr)
2994 {
2995         struct btrfs_delayed_ref_head *head;
2996         struct btrfs_delayed_ref_node *ref;
2997         struct btrfs_delayed_data_ref *data_ref;
2998         struct btrfs_delayed_ref_root *delayed_refs;
2999         struct btrfs_transaction *cur_trans;
3000         struct rb_node *node;
3001         int ret = 0;
3002
3003         spin_lock(&root->fs_info->trans_lock);
3004         cur_trans = root->fs_info->running_transaction;
3005         if (cur_trans)
3006                 refcount_inc(&cur_trans->use_count);
3007         spin_unlock(&root->fs_info->trans_lock);
3008         if (!cur_trans)
3009                 return 0;
3010
3011         delayed_refs = &cur_trans->delayed_refs;
3012         spin_lock(&delayed_refs->lock);
3013         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
3014         if (!head) {
3015                 spin_unlock(&delayed_refs->lock);
3016                 btrfs_put_transaction(cur_trans);
3017                 return 0;
3018         }
3019
3020         if (!mutex_trylock(&head->mutex)) {
3021                 refcount_inc(&head->refs);
3022                 spin_unlock(&delayed_refs->lock);
3023
3024                 btrfs_release_path(path);
3025
3026                 /*
3027                  * Mutex was contended, block until it's released and let
3028                  * caller try again
3029                  */
3030                 mutex_lock(&head->mutex);
3031                 mutex_unlock(&head->mutex);
3032                 btrfs_put_delayed_ref_head(head);
3033                 btrfs_put_transaction(cur_trans);
3034                 return -EAGAIN;
3035         }
3036         spin_unlock(&delayed_refs->lock);
3037
3038         spin_lock(&head->lock);
3039         /*
3040          * XXX: We should replace this with a proper search function in the
3041          * future.
3042          */
3043         for (node = rb_first(&head->ref_tree); node; node = rb_next(node)) {
3044                 ref = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
3045                 /* If it's a shared ref we know a cross reference exists */
3046                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
3047                         ret = 1;
3048                         break;
3049                 }
3050
3051                 data_ref = btrfs_delayed_node_to_data_ref(ref);
3052
3053                 /*
3054                  * If our ref doesn't match the one we're currently looking at
3055                  * then we have a cross reference.
3056                  */
3057                 if (data_ref->root != root->root_key.objectid ||
3058                     data_ref->objectid != objectid ||
3059                     data_ref->offset != offset) {
3060                         ret = 1;
3061                         break;
3062                 }
3063         }
3064         spin_unlock(&head->lock);
3065         mutex_unlock(&head->mutex);
3066         btrfs_put_transaction(cur_trans);
3067         return ret;
3068 }
3069
3070 static noinline int check_committed_ref(struct btrfs_root *root,
3071                                         struct btrfs_path *path,
3072                                         u64 objectid, u64 offset, u64 bytenr)
3073 {
3074         struct btrfs_fs_info *fs_info = root->fs_info;
3075         struct btrfs_root *extent_root = fs_info->extent_root;
3076         struct extent_buffer *leaf;
3077         struct btrfs_extent_data_ref *ref;
3078         struct btrfs_extent_inline_ref *iref;
3079         struct btrfs_extent_item *ei;
3080         struct btrfs_key key;
3081         u32 item_size;
3082         int type;
3083         int ret;
3084
3085         key.objectid = bytenr;
3086         key.offset = (u64)-1;
3087         key.type = BTRFS_EXTENT_ITEM_KEY;
3088
3089         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
3090         if (ret < 0)
3091                 goto out;
3092         BUG_ON(ret == 0); /* Corruption */
3093
3094         ret = -ENOENT;
3095         if (path->slots[0] == 0)
3096                 goto out;
3097
3098         path->slots[0]--;
3099         leaf = path->nodes[0];
3100         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3101
3102         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
3103                 goto out;
3104
3105         ret = 1;
3106         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
3107         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
3108
3109         if (item_size != sizeof(*ei) +
3110             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
3111                 goto out;
3112
3113         if (btrfs_extent_generation(leaf, ei) <=
3114             btrfs_root_last_snapshot(&root->root_item))
3115                 goto out;
3116
3117         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
3118
3119         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
3120         if (type != BTRFS_EXTENT_DATA_REF_KEY)
3121                 goto out;
3122
3123         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
3124         if (btrfs_extent_refs(leaf, ei) !=
3125             btrfs_extent_data_ref_count(leaf, ref) ||
3126             btrfs_extent_data_ref_root(leaf, ref) !=
3127             root->root_key.objectid ||
3128             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
3129             btrfs_extent_data_ref_offset(leaf, ref) != offset)
3130                 goto out;
3131
3132         ret = 0;
3133 out:
3134         return ret;
3135 }
3136
3137 int btrfs_cross_ref_exist(struct btrfs_root *root, u64 objectid, u64 offset,
3138                           u64 bytenr)
3139 {
3140         struct btrfs_path *path;
3141         int ret;
3142         int ret2;
3143
3144         path = btrfs_alloc_path();
3145         if (!path)
3146                 return -ENOMEM;
3147
3148         do {
3149                 ret = check_committed_ref(root, path, objectid,
3150                                           offset, bytenr);
3151                 if (ret && ret != -ENOENT)
3152                         goto out;
3153
3154                 ret2 = check_delayed_ref(root, path, objectid,
3155                                          offset, bytenr);
3156         } while (ret2 == -EAGAIN);
3157
3158         if (ret2 && ret2 != -ENOENT) {
3159                 ret = ret2;
3160                 goto out;
3161         }
3162
3163         if (ret != -ENOENT || ret2 != -ENOENT)
3164                 ret = 0;
3165 out:
3166         btrfs_free_path(path);
3167         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
3168                 WARN_ON(ret > 0);
3169         return ret;
3170 }
3171
3172 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
3173                            struct btrfs_root *root,
3174                            struct extent_buffer *buf,
3175                            int full_backref, int inc)
3176 {
3177         struct btrfs_fs_info *fs_info = root->fs_info;
3178         u64 bytenr;
3179         u64 num_bytes;
3180         u64 parent;
3181         u64 ref_root;
3182         u32 nritems;
3183         struct btrfs_key key;
3184         struct btrfs_file_extent_item *fi;
3185         int i;
3186         int level;
3187         int ret = 0;
3188         int (*process_func)(struct btrfs_trans_handle *,
3189                             struct btrfs_root *,
3190                             u64, u64, u64, u64, u64, u64);
3191
3192
3193         if (btrfs_is_testing(fs_info))
3194                 return 0;
3195
3196         ref_root = btrfs_header_owner(buf);
3197         nritems = btrfs_header_nritems(buf);
3198         level = btrfs_header_level(buf);
3199
3200         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state) && level == 0)
3201                 return 0;
3202
3203         if (inc)
3204                 process_func = btrfs_inc_extent_ref;
3205         else
3206                 process_func = btrfs_free_extent;
3207
3208         if (full_backref)
3209                 parent = buf->start;
3210         else
3211                 parent = 0;
3212
3213         for (i = 0; i < nritems; i++) {
3214                 if (level == 0) {
3215                         btrfs_item_key_to_cpu(buf, &key, i);
3216                         if (key.type != BTRFS_EXTENT_DATA_KEY)
3217                                 continue;
3218                         fi = btrfs_item_ptr(buf, i,
3219                                             struct btrfs_file_extent_item);
3220                         if (btrfs_file_extent_type(buf, fi) ==
3221                             BTRFS_FILE_EXTENT_INLINE)
3222                                 continue;
3223                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3224                         if (bytenr == 0)
3225                                 continue;
3226
3227                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3228                         key.offset -= btrfs_file_extent_offset(buf, fi);
3229                         ret = process_func(trans, root, bytenr, num_bytes,
3230                                            parent, ref_root, key.objectid,
3231                                            key.offset);
3232                         if (ret)
3233                                 goto fail;
3234                 } else {
3235                         bytenr = btrfs_node_blockptr(buf, i);
3236                         num_bytes = fs_info->nodesize;
3237                         ret = process_func(trans, root, bytenr, num_bytes,
3238                                            parent, ref_root, level - 1, 0);
3239                         if (ret)
3240                                 goto fail;
3241                 }
3242         }
3243         return 0;
3244 fail:
3245         return ret;
3246 }
3247
3248 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3249                   struct extent_buffer *buf, int full_backref)
3250 {
3251         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
3252 }
3253
3254 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3255                   struct extent_buffer *buf, int full_backref)
3256 {
3257         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
3258 }
3259
3260 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3261                                  struct btrfs_fs_info *fs_info,
3262                                  struct btrfs_path *path,
3263                                  struct btrfs_block_group_cache *cache)
3264 {
3265         int ret;
3266         struct btrfs_root *extent_root = fs_info->extent_root;
3267         unsigned long bi;
3268         struct extent_buffer *leaf;
3269
3270         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3271         if (ret) {
3272                 if (ret > 0)
3273                         ret = -ENOENT;
3274                 goto fail;
3275         }
3276
3277         leaf = path->nodes[0];
3278         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3279         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3280         btrfs_mark_buffer_dirty(leaf);
3281 fail:
3282         btrfs_release_path(path);
3283         return ret;
3284
3285 }
3286
3287 static struct btrfs_block_group_cache *
3288 next_block_group(struct btrfs_fs_info *fs_info,
3289                  struct btrfs_block_group_cache *cache)
3290 {
3291         struct rb_node *node;
3292
3293         spin_lock(&fs_info->block_group_cache_lock);
3294
3295         /* If our block group was removed, we need a full search. */
3296         if (RB_EMPTY_NODE(&cache->cache_node)) {
3297                 const u64 next_bytenr = cache->key.objectid + cache->key.offset;
3298
3299                 spin_unlock(&fs_info->block_group_cache_lock);
3300                 btrfs_put_block_group(cache);
3301                 cache = btrfs_lookup_first_block_group(fs_info, next_bytenr); return cache;
3302         }
3303         node = rb_next(&cache->cache_node);
3304         btrfs_put_block_group(cache);
3305         if (node) {
3306                 cache = rb_entry(node, struct btrfs_block_group_cache,
3307                                  cache_node);
3308                 btrfs_get_block_group(cache);
3309         } else
3310                 cache = NULL;
3311         spin_unlock(&fs_info->block_group_cache_lock);
3312         return cache;
3313 }
3314
3315 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3316                             struct btrfs_trans_handle *trans,
3317                             struct btrfs_path *path)
3318 {
3319         struct btrfs_fs_info *fs_info = block_group->fs_info;
3320         struct btrfs_root *root = fs_info->tree_root;
3321         struct inode *inode = NULL;
3322         struct extent_changeset *data_reserved = NULL;
3323         u64 alloc_hint = 0;
3324         int dcs = BTRFS_DC_ERROR;
3325         u64 num_pages = 0;
3326         int retries = 0;
3327         int ret = 0;
3328
3329         /*
3330          * If this block group is smaller than 100 megs don't bother caching the
3331          * block group.
3332          */
3333         if (block_group->key.offset < (100 * SZ_1M)) {
3334                 spin_lock(&block_group->lock);
3335                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3336                 spin_unlock(&block_group->lock);
3337                 return 0;
3338         }
3339
3340         if (trans->aborted)
3341                 return 0;
3342 again:
3343         inode = lookup_free_space_inode(fs_info, block_group, path);
3344         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3345                 ret = PTR_ERR(inode);
3346                 btrfs_release_path(path);
3347                 goto out;
3348         }
3349
3350         if (IS_ERR(inode)) {
3351                 BUG_ON(retries);
3352                 retries++;
3353
3354                 if (block_group->ro)
3355                         goto out_free;
3356
3357                 ret = create_free_space_inode(fs_info, trans, block_group,
3358                                               path);
3359                 if (ret)
3360                         goto out_free;
3361                 goto again;
3362         }
3363
3364         /*
3365          * We want to set the generation to 0, that way if anything goes wrong
3366          * from here on out we know not to trust this cache when we load up next
3367          * time.
3368          */
3369         BTRFS_I(inode)->generation = 0;
3370         ret = btrfs_update_inode(trans, root, inode);
3371         if (ret) {
3372                 /*
3373                  * So theoretically we could recover from this, simply set the
3374                  * super cache generation to 0 so we know to invalidate the
3375                  * cache, but then we'd have to keep track of the block groups
3376                  * that fail this way so we know we _have_ to reset this cache
3377                  * before the next commit or risk reading stale cache.  So to
3378                  * limit our exposure to horrible edge cases lets just abort the
3379                  * transaction, this only happens in really bad situations
3380                  * anyway.
3381                  */
3382                 btrfs_abort_transaction(trans, ret);
3383                 goto out_put;
3384         }
3385         WARN_ON(ret);
3386
3387         /* We've already setup this transaction, go ahead and exit */
3388         if (block_group->cache_generation == trans->transid &&
3389             i_size_read(inode)) {
3390                 dcs = BTRFS_DC_SETUP;
3391                 goto out_put;
3392         }
3393
3394         if (i_size_read(inode) > 0) {
3395                 ret = btrfs_check_trunc_cache_free_space(fs_info,
3396                                         &fs_info->global_block_rsv);
3397                 if (ret)
3398                         goto out_put;
3399
3400                 ret = btrfs_truncate_free_space_cache(trans, NULL, inode);
3401                 if (ret)
3402                         goto out_put;
3403         }
3404
3405         spin_lock(&block_group->lock);
3406         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3407             !btrfs_test_opt(fs_info, SPACE_CACHE)) {
3408                 /*
3409                  * don't bother trying to write stuff out _if_
3410                  * a) we're not cached,
3411                  * b) we're with nospace_cache mount option,
3412                  * c) we're with v2 space_cache (FREE_SPACE_TREE).
3413                  */
3414                 dcs = BTRFS_DC_WRITTEN;
3415                 spin_unlock(&block_group->lock);
3416                 goto out_put;
3417         }
3418         spin_unlock(&block_group->lock);
3419
3420         /*
3421          * We hit an ENOSPC when setting up the cache in this transaction, just
3422          * skip doing the setup, we've already cleared the cache so we're safe.
3423          */
3424         if (test_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags)) {
3425                 ret = -ENOSPC;
3426                 goto out_put;
3427         }
3428
3429         /*
3430          * Try to preallocate enough space based on how big the block group is.
3431          * Keep in mind this has to include any pinned space which could end up
3432          * taking up quite a bit since it's not folded into the other space
3433          * cache.
3434          */
3435         num_pages = div_u64(block_group->key.offset, SZ_256M);
3436         if (!num_pages)
3437                 num_pages = 1;
3438
3439         num_pages *= 16;
3440         num_pages *= PAGE_SIZE;
3441
3442         ret = btrfs_check_data_free_space(inode, &data_reserved, 0, num_pages);
3443         if (ret)
3444                 goto out_put;
3445
3446         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3447                                               num_pages, num_pages,
3448                                               &alloc_hint);
3449         /*
3450          * Our cache requires contiguous chunks so that we don't modify a bunch
3451          * of metadata or split extents when writing the cache out, which means
3452          * we can enospc if we are heavily fragmented in addition to just normal
3453          * out of space conditions.  So if we hit this just skip setting up any
3454          * other block groups for this transaction, maybe we'll unpin enough
3455          * space the next time around.
3456          */
3457         if (!ret)
3458                 dcs = BTRFS_DC_SETUP;
3459         else if (ret == -ENOSPC)
3460                 set_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags);
3461
3462 out_put:
3463         iput(inode);
3464 out_free:
3465         btrfs_release_path(path);
3466 out:
3467         spin_lock(&block_group->lock);
3468         if (!ret && dcs == BTRFS_DC_SETUP)
3469                 block_group->cache_generation = trans->transid;
3470         block_group->disk_cache_state = dcs;
3471         spin_unlock(&block_group->lock);
3472
3473         extent_changeset_free(data_reserved);
3474         return ret;
3475 }
3476
3477 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans,
3478                             struct btrfs_fs_info *fs_info)
3479 {
3480         struct btrfs_block_group_cache *cache, *tmp;
3481         struct btrfs_transaction *cur_trans = trans->transaction;
3482         struct btrfs_path *path;
3483
3484         if (list_empty(&cur_trans->dirty_bgs) ||
3485             !btrfs_test_opt(fs_info, SPACE_CACHE))
3486                 return 0;
3487
3488         path = btrfs_alloc_path();
3489         if (!path)
3490                 return -ENOMEM;
3491
3492         /* Could add new block groups, use _safe just in case */
3493         list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3494                                  dirty_list) {
3495                 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3496                         cache_save_setup(cache, trans, path);
3497         }
3498
3499         btrfs_free_path(path);
3500         return 0;
3501 }
3502
3503 /*
3504  * transaction commit does final block group cache writeback during a
3505  * critical section where nothing is allowed to change the FS.  This is
3506  * required in order for the cache to actually match the block group,
3507  * but can introduce a lot of latency into the commit.
3508  *
3509  * So, btrfs_start_dirty_block_groups is here to kick off block group
3510  * cache IO.  There's a chance we'll have to redo some of it if the
3511  * block group changes again during the commit, but it greatly reduces
3512  * the commit latency by getting rid of the easy block groups while
3513  * we're still allowing others to join the commit.
3514  */
3515 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans)
3516 {
3517         struct btrfs_fs_info *fs_info = trans->fs_info;
3518         struct btrfs_block_group_cache *cache;
3519         struct btrfs_transaction *cur_trans = trans->transaction;
3520         int ret = 0;
3521         int should_put;
3522         struct btrfs_path *path = NULL;
3523         LIST_HEAD(dirty);
3524         struct list_head *io = &cur_trans->io_bgs;
3525         int num_started = 0;
3526         int loops = 0;
3527
3528         spin_lock(&cur_trans->dirty_bgs_lock);
3529         if (list_empty(&cur_trans->dirty_bgs)) {
3530                 spin_unlock(&cur_trans->dirty_bgs_lock);
3531                 return 0;
3532         }
3533         list_splice_init(&cur_trans->dirty_bgs, &dirty);
3534         spin_unlock(&cur_trans->dirty_bgs_lock);
3535
3536 again:
3537         /*
3538          * make sure all the block groups on our dirty list actually
3539          * exist
3540          */
3541         btrfs_create_pending_block_groups(trans);
3542
3543         if (!path) {
3544                 path = btrfs_alloc_path();
3545                 if (!path)
3546                         return -ENOMEM;
3547         }
3548
3549         /*
3550          * cache_write_mutex is here only to save us from balance or automatic
3551          * removal of empty block groups deleting this block group while we are
3552          * writing out the cache
3553          */
3554         mutex_lock(&trans->transaction->cache_write_mutex);
3555         while (!list_empty(&dirty)) {
3556                 cache = list_first_entry(&dirty,
3557                                          struct btrfs_block_group_cache,
3558                                          dirty_list);
3559                 /*
3560                  * this can happen if something re-dirties a block
3561                  * group that is already under IO.  Just wait for it to
3562                  * finish and then do it all again
3563                  */
3564                 if (!list_empty(&cache->io_list)) {
3565                         list_del_init(&cache->io_list);
3566                         btrfs_wait_cache_io(trans, cache, path);
3567                         btrfs_put_block_group(cache);
3568                 }
3569
3570
3571                 /*
3572                  * btrfs_wait_cache_io uses the cache->dirty_list to decide
3573                  * if it should update the cache_state.  Don't delete
3574                  * until after we wait.
3575                  *
3576                  * Since we're not running in the commit critical section
3577                  * we need the dirty_bgs_lock to protect from update_block_group
3578                  */
3579                 spin_lock(&cur_trans->dirty_bgs_lock);
3580                 list_del_init(&cache->dirty_list);
3581                 spin_unlock(&cur_trans->dirty_bgs_lock);
3582
3583                 should_put = 1;
3584
3585                 cache_save_setup(cache, trans, path);
3586
3587                 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
3588                         cache->io_ctl.inode = NULL;
3589                         ret = btrfs_write_out_cache(fs_info, trans,
3590                                                     cache, path);
3591                         if (ret == 0 && cache->io_ctl.inode) {
3592                                 num_started++;
3593                                 should_put = 0;
3594
3595                                 /*
3596                                  * The cache_write_mutex is protecting the
3597                                  * io_list, also refer to the definition of
3598                                  * btrfs_transaction::io_bgs for more details
3599                                  */
3600                                 list_add_tail(&cache->io_list, io);
3601                         } else {
3602                                 /*
3603                                  * if we failed to write the cache, the
3604                                  * generation will be bad and life goes on
3605                                  */
3606                                 ret = 0;
3607                         }
3608                 }
3609                 if (!ret) {
3610                         ret = write_one_cache_group(trans, fs_info,
3611                                                     path, cache);
3612                         /*
3613                          * Our block group might still be attached to the list
3614                          * of new block groups in the transaction handle of some
3615                          * other task (struct btrfs_trans_handle->new_bgs). This
3616                          * means its block group item isn't yet in the extent
3617                          * tree. If this happens ignore the error, as we will
3618                          * try again later in the critical section of the
3619                          * transaction commit.
3620                          */
3621                         if (ret == -ENOENT) {
3622                                 ret = 0;
3623                                 spin_lock(&cur_trans->dirty_bgs_lock);
3624                                 if (list_empty(&cache->dirty_list)) {
3625                                         list_add_tail(&cache->dirty_list,
3626                                                       &cur_trans->dirty_bgs);
3627                                         btrfs_get_block_group(cache);
3628                                 }
3629                                 spin_unlock(&cur_trans->dirty_bgs_lock);
3630                         } else if (ret) {
3631                                 btrfs_abort_transaction(trans, ret);
3632                         }
3633                 }
3634
3635                 /* if its not on the io list, we need to put the block group */
3636                 if (should_put)
3637                         btrfs_put_block_group(cache);
3638
3639                 if (ret)
3640                         break;
3641
3642                 /*
3643                  * Avoid blocking other tasks for too long. It might even save
3644                  * us from writing caches for block groups that are going to be
3645                  * removed.
3646                  */
3647                 mutex_unlock(&trans->transaction->cache_write_mutex);
3648                 mutex_lock(&trans->transaction->cache_write_mutex);
3649         }
3650         mutex_unlock(&trans->transaction->cache_write_mutex);
3651
3652         /*
3653          * go through delayed refs for all the stuff we've just kicked off
3654          * and then loop back (just once)
3655          */
3656         ret = btrfs_run_delayed_refs(trans, 0);
3657         if (!ret && loops == 0) {
3658                 loops++;
3659                 spin_lock(&cur_trans->dirty_bgs_lock);
3660                 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3661                 /*
3662                  * dirty_bgs_lock protects us from concurrent block group
3663                  * deletes too (not just cache_write_mutex).
3664                  */
3665                 if (!list_empty(&dirty)) {
3666                         spin_unlock(&cur_trans->dirty_bgs_lock);
3667                         goto again;
3668                 }
3669                 spin_unlock(&cur_trans->dirty_bgs_lock);
3670         } else if (ret < 0) {
3671                 btrfs_cleanup_dirty_bgs(cur_trans, fs_info);
3672         }
3673
3674         btrfs_free_path(path);
3675         return ret;
3676 }
3677
3678 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3679                                    struct btrfs_fs_info *fs_info)
3680 {
3681         struct btrfs_block_group_cache *cache;
3682         struct btrfs_transaction *cur_trans = trans->transaction;
3683         int ret = 0;
3684         int should_put;
3685         struct btrfs_path *path;
3686         struct list_head *io = &cur_trans->io_bgs;
3687         int num_started = 0;
3688
3689         path = btrfs_alloc_path();
3690         if (!path)
3691                 return -ENOMEM;
3692
3693         /*
3694          * Even though we are in the critical section of the transaction commit,
3695          * we can still have concurrent tasks adding elements to this
3696          * transaction's list of dirty block groups. These tasks correspond to
3697          * endio free space workers started when writeback finishes for a
3698          * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
3699          * allocate new block groups as a result of COWing nodes of the root
3700          * tree when updating the free space inode. The writeback for the space
3701          * caches is triggered by an earlier call to
3702          * btrfs_start_dirty_block_groups() and iterations of the following
3703          * loop.
3704          * Also we want to do the cache_save_setup first and then run the
3705          * delayed refs to make sure we have the best chance at doing this all
3706          * in one shot.
3707          */
3708         spin_lock(&cur_trans->dirty_bgs_lock);
3709         while (!list_empty(&cur_trans->dirty_bgs)) {
3710                 cache = list_first_entry(&cur_trans->dirty_bgs,
3711                                          struct btrfs_block_group_cache,
3712                                          dirty_list);
3713
3714                 /*
3715                  * this can happen if cache_save_setup re-dirties a block
3716                  * group that is already under IO.  Just wait for it to
3717                  * finish and then do it all again
3718                  */
3719                 if (!list_empty(&cache->io_list)) {
3720                         spin_unlock(&cur_trans->dirty_bgs_lock);
3721                         list_del_init(&cache->io_list);
3722                         btrfs_wait_cache_io(trans, cache, path);
3723                         btrfs_put_block_group(cache);
3724                         spin_lock(&cur_trans->dirty_bgs_lock);
3725                 }
3726
3727                 /*
3728                  * don't remove from the dirty list until after we've waited
3729                  * on any pending IO
3730                  */
3731                 list_del_init(&cache->dirty_list);
3732                 spin_unlock(&cur_trans->dirty_bgs_lock);
3733                 should_put = 1;
3734
3735                 cache_save_setup(cache, trans, path);
3736
3737                 if (!ret)
3738                         ret = btrfs_run_delayed_refs(trans,
3739                                                      (unsigned long) -1);
3740
3741                 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3742                         cache->io_ctl.inode = NULL;
3743                         ret = btrfs_write_out_cache(fs_info, trans,
3744                                                     cache, path);
3745                         if (ret == 0 && cache->io_ctl.inode) {
3746                                 num_started++;
3747                                 should_put = 0;
3748                                 list_add_tail(&cache->io_list, io);
3749                         } else {
3750                                 /*
3751                                  * if we failed to write the cache, the
3752                                  * generation will be bad and life goes on
3753                                  */
3754                                 ret = 0;
3755                         }
3756                 }
3757                 if (!ret) {
3758                         ret = write_one_cache_group(trans, fs_info,
3759                                                     path, cache);
3760                         /*
3761                          * One of the free space endio workers might have
3762                          * created a new block group while updating a free space
3763                          * cache's inode (at inode.c:btrfs_finish_ordered_io())
3764                          * and hasn't released its transaction handle yet, in
3765                          * which case the new block group is still attached to
3766                          * its transaction handle and its creation has not
3767                          * finished yet (no block group item in the extent tree
3768                          * yet, etc). If this is the case, wait for all free
3769                          * space endio workers to finish and retry. This is a
3770                          * a very rare case so no need for a more efficient and
3771                          * complex approach.
3772                          */
3773                         if (ret == -ENOENT) {
3774                                 wait_event(cur_trans->writer_wait,
3775                                    atomic_read(&cur_trans->num_writers) == 1);
3776                                 ret = write_one_cache_group(trans, fs_info,
3777                                                             path, cache);
3778                         }
3779                         if (ret)
3780                                 btrfs_abort_transaction(trans, ret);
3781                 }
3782
3783                 /* if its not on the io list, we need to put the block group */
3784                 if (should_put)
3785                         btrfs_put_block_group(cache);
3786                 spin_lock(&cur_trans->dirty_bgs_lock);
3787         }
3788         spin_unlock(&cur_trans->dirty_bgs_lock);
3789
3790         /*
3791          * Refer to the definition of io_bgs member for details why it's safe
3792          * to use it without any locking
3793          */
3794         while (!list_empty(io)) {
3795                 cache = list_first_entry(io, struct btrfs_block_group_cache,
3796                                          io_list);
3797                 list_del_init(&cache->io_list);
3798                 btrfs_wait_cache_io(trans, cache, path);
3799                 btrfs_put_block_group(cache);
3800         }
3801
3802         btrfs_free_path(path);
3803         return ret;
3804 }
3805
3806 int btrfs_extent_readonly(struct btrfs_fs_info *fs_info, u64 bytenr)
3807 {
3808         struct btrfs_block_group_cache *block_group;
3809         int readonly = 0;
3810
3811         block_group = btrfs_lookup_block_group(fs_info, bytenr);
3812         if (!block_group || block_group->ro)
3813                 readonly = 1;
3814         if (block_group)
3815                 btrfs_put_block_group(block_group);
3816         return readonly;
3817 }
3818
3819 bool btrfs_inc_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3820 {
3821         struct btrfs_block_group_cache *bg;
3822         bool ret = true;
3823
3824         bg = btrfs_lookup_block_group(fs_info, bytenr);
3825         if (!bg)
3826                 return false;
3827
3828         spin_lock(&bg->lock);
3829         if (bg->ro)
3830                 ret = false;
3831         else
3832                 atomic_inc(&bg->nocow_writers);
3833         spin_unlock(&bg->lock);
3834
3835         /* no put on block group, done by btrfs_dec_nocow_writers */
3836         if (!ret)
3837                 btrfs_put_block_group(bg);
3838
3839         return ret;
3840
3841 }
3842
3843 void btrfs_dec_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3844 {
3845         struct btrfs_block_group_cache *bg;
3846
3847         bg = btrfs_lookup_block_group(fs_info, bytenr);
3848         ASSERT(bg);
3849         if (atomic_dec_and_test(&bg->nocow_writers))
3850                 wake_up_var(&bg->nocow_writers);
3851         /*
3852          * Once for our lookup and once for the lookup done by a previous call
3853          * to btrfs_inc_nocow_writers()
3854          */
3855         btrfs_put_block_group(bg);
3856         btrfs_put_block_group(bg);
3857 }
3858
3859 void btrfs_wait_nocow_writers(struct btrfs_block_group_cache *bg)
3860 {
3861         wait_var_event(&bg->nocow_writers, !atomic_read(&bg->nocow_writers));
3862 }
3863
3864 static const char *alloc_name(u64 flags)
3865 {
3866         switch (flags) {
3867         case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA:
3868                 return "mixed";
3869         case BTRFS_BLOCK_GROUP_METADATA:
3870                 return "metadata";
3871         case BTRFS_BLOCK_GROUP_DATA:
3872                 return "data";
3873         case BTRFS_BLOCK_GROUP_SYSTEM:
3874                 return "system";
3875         default:
3876                 WARN_ON(1);
3877                 return "invalid-combination";
3878         };
3879 }
3880
3881 static int create_space_info(struct btrfs_fs_info *info, u64 flags)
3882 {
3883
3884         struct btrfs_space_info *space_info;
3885         int i;
3886         int ret;
3887
3888         space_info = kzalloc(sizeof(*space_info), GFP_NOFS);
3889         if (!space_info)
3890                 return -ENOMEM;
3891
3892         ret = percpu_counter_init(&space_info->total_bytes_pinned, 0,
3893                                  GFP_KERNEL);
3894         if (ret) {
3895                 kfree(space_info);
3896                 return ret;
3897         }
3898
3899         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3900                 INIT_LIST_HEAD(&space_info->block_groups[i]);
3901         init_rwsem(&space_info->groups_sem);
3902         spin_lock_init(&space_info->lock);
3903         space_info->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3904         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3905         init_waitqueue_head(&space_info->wait);
3906         INIT_LIST_HEAD(&space_info->ro_bgs);
3907         INIT_LIST_HEAD(&space_info->tickets);
3908         INIT_LIST_HEAD(&space_info->priority_tickets);
3909
3910         ret = kobject_init_and_add(&space_info->kobj, &space_info_ktype,
3911                                     info->space_info_kobj, "%s",
3912                                     alloc_name(space_info->flags));
3913         if (ret) {
3914                 percpu_counter_destroy(&space_info->total_bytes_pinned);
3915                 kfree(space_info);
3916                 return ret;
3917         }
3918
3919         list_add_rcu(&space_info->list, &info->space_info);
3920         if (flags & BTRFS_BLOCK_GROUP_DATA)
3921                 info->data_sinfo = space_info;
3922
3923         return ret;
3924 }
3925
3926 static void update_space_info(struct btrfs_fs_info *info, u64 flags,
3927                              u64 total_bytes, u64 bytes_used,
3928                              u64 bytes_readonly,
3929                              struct btrfs_space_info **space_info)
3930 {
3931         struct btrfs_space_info *found;
3932         int factor;
3933
3934         factor = btrfs_bg_type_to_factor(flags);
3935
3936         found = __find_space_info(info, flags);
3937         ASSERT(found);
3938         spin_lock(&found->lock);
3939         found->total_bytes += total_bytes;
3940         found->disk_total += total_bytes * factor;
3941         found->bytes_used += bytes_used;
3942         found->disk_used += bytes_used * factor;
3943         found->bytes_readonly += bytes_readonly;
3944         if (total_bytes > 0)
3945                 found->full = 0;
3946         space_info_add_new_bytes(info, found, total_bytes -
3947                                  bytes_used - bytes_readonly);
3948         spin_unlock(&found->lock);
3949         *space_info = found;
3950 }
3951
3952 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3953 {
3954         u64 extra_flags = chunk_to_extended(flags) &
3955                                 BTRFS_EXTENDED_PROFILE_MASK;
3956
3957         write_seqlock(&fs_info->profiles_lock);
3958         if (flags & BTRFS_BLOCK_GROUP_DATA)
3959                 fs_info->avail_data_alloc_bits |= extra_flags;
3960         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3961                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3962         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3963                 fs_info->avail_system_alloc_bits |= extra_flags;
3964         write_sequnlock(&fs_info->profiles_lock);
3965 }
3966
3967 /*
3968  * returns target flags in extended format or 0 if restripe for this
3969  * chunk_type is not in progress
3970  *
3971  * should be called with balance_lock held
3972  */
3973 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
3974 {
3975         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3976         u64 target = 0;
3977
3978         if (!bctl)
3979                 return 0;
3980
3981         if (flags & BTRFS_BLOCK_GROUP_DATA &&
3982             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3983                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
3984         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
3985                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3986                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
3987         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
3988                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3989                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
3990         }
3991
3992         return target;
3993 }
3994
3995 /*
3996  * @flags: available profiles in extended format (see ctree.h)
3997  *
3998  * Returns reduced profile in chunk format.  If profile changing is in
3999  * progress (either running or paused) picks the target profile (if it's
4000  * already available), otherwise falls back to plain reducing.
4001  */
4002 static u64 btrfs_reduce_alloc_profile(struct btrfs_fs_info *fs_info, u64 flags)
4003 {
4004         u64 num_devices = fs_info->fs_devices->rw_devices;
4005         u64 target;
4006         u64 raid_type;
4007         u64 allowed = 0;
4008
4009         /*
4010          * see if restripe for this chunk_type is in progress, if so
4011          * try to reduce to the target profile
4012          */
4013         spin_lock(&fs_info->balance_lock);
4014         target = get_restripe_target(fs_info, flags);
4015         if (target) {
4016                 /* pick target profile only if it's already available */
4017                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
4018                         spin_unlock(&fs_info->balance_lock);
4019                         return extended_to_chunk(target);
4020                 }
4021         }
4022         spin_unlock(&fs_info->balance_lock);
4023
4024         /* First, mask out the RAID levels which aren't possible */
4025         for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
4026                 if (num_devices >= btrfs_raid_array[raid_type].devs_min)
4027                         allowed |= btrfs_raid_array[raid_type].bg_flag;
4028         }
4029         allowed &= flags;
4030
4031         if (allowed & BTRFS_BLOCK_GROUP_RAID6)
4032                 allowed = BTRFS_BLOCK_GROUP_RAID6;
4033         else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
4034                 allowed = BTRFS_BLOCK_GROUP_RAID5;
4035         else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
4036                 allowed = BTRFS_BLOCK_GROUP_RAID10;
4037         else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
4038                 allowed = BTRFS_BLOCK_GROUP_RAID1;
4039         else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
4040                 allowed = BTRFS_BLOCK_GROUP_RAID0;
4041
4042         flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
4043
4044         return extended_to_chunk(flags | allowed);
4045 }
4046
4047 static u64 get_alloc_profile(struct btrfs_fs_info *fs_info, u64 orig_flags)
4048 {
4049         unsigned seq;
4050         u64 flags;
4051
4052         do {
4053                 flags = orig_flags;
4054                 seq = read_seqbegin(&fs_info->profiles_lock);
4055
4056                 if (flags & BTRFS_BLOCK_GROUP_DATA)
4057                         flags |= fs_info->avail_data_alloc_bits;
4058                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4059                         flags |= fs_info->avail_system_alloc_bits;
4060                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
4061                         flags |= fs_info->avail_metadata_alloc_bits;
4062         } while (read_seqretry(&fs_info->profiles_lock, seq));
4063
4064         return btrfs_reduce_alloc_profile(fs_info, flags);
4065 }
4066
4067 static u64 get_alloc_profile_by_root(struct btrfs_root *root, int data)
4068 {
4069         struct btrfs_fs_info *fs_info = root->fs_info;
4070         u64 flags;
4071         u64 ret;
4072
4073         if (data)
4074                 flags = BTRFS_BLOCK_GROUP_DATA;
4075         else if (root == fs_info->chunk_root)
4076                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
4077         else
4078                 flags = BTRFS_BLOCK_GROUP_METADATA;
4079
4080         ret = get_alloc_profile(fs_info, flags);
4081         return ret;
4082 }
4083
4084 u64 btrfs_data_alloc_profile(struct btrfs_fs_info *fs_info)
4085 {
4086         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_DATA);
4087 }
4088
4089 u64 btrfs_metadata_alloc_profile(struct btrfs_fs_info *fs_info)
4090 {
4091         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4092 }
4093
4094 u64 btrfs_system_alloc_profile(struct btrfs_fs_info *fs_info)
4095 {
4096         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4097 }
4098
4099 static u64 btrfs_space_info_used(struct btrfs_space_info *s_info,
4100                                  bool may_use_included)
4101 {
4102         ASSERT(s_info);
4103         return s_info->bytes_used + s_info->bytes_reserved +
4104                 s_info->bytes_pinned + s_info->bytes_readonly +
4105                 (may_use_included ? s_info->bytes_may_use : 0);
4106 }
4107
4108 int btrfs_alloc_data_chunk_ondemand(struct btrfs_inode *inode, u64 bytes)
4109 {
4110         struct btrfs_root *root = inode->root;
4111         struct btrfs_fs_info *fs_info = root->fs_info;
4112         struct btrfs_space_info *data_sinfo = fs_info->data_sinfo;
4113         u64 used;
4114         int ret = 0;
4115         int need_commit = 2;
4116         int have_pinned_space;
4117
4118         /* make sure bytes are sectorsize aligned */
4119         bytes = ALIGN(bytes, fs_info->sectorsize);
4120
4121         if (btrfs_is_free_space_inode(inode)) {
4122                 need_commit = 0;
4123                 ASSERT(current->journal_info);
4124         }
4125
4126 again:
4127         /* make sure we have enough space to handle the data first */
4128         spin_lock(&data_sinfo->lock);
4129         used = btrfs_space_info_used(data_sinfo, true);
4130
4131         if (used + bytes > data_sinfo->total_bytes) {
4132                 struct btrfs_trans_handle *trans;
4133
4134                 /*
4135                  * if we don't have enough free bytes in this space then we need
4136                  * to alloc a new chunk.
4137                  */
4138                 if (!data_sinfo->full) {
4139                         u64 alloc_target;
4140
4141                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
4142                         spin_unlock(&data_sinfo->lock);
4143
4144                         alloc_target = btrfs_data_alloc_profile(fs_info);
4145                         /*
4146                          * It is ugly that we don't call nolock join
4147                          * transaction for the free space inode case here.
4148                          * But it is safe because we only do the data space
4149                          * reservation for the free space cache in the
4150                          * transaction context, the common join transaction
4151                          * just increase the counter of the current transaction
4152                          * handler, doesn't try to acquire the trans_lock of
4153                          * the fs.
4154                          */
4155                         trans = btrfs_join_transaction(root);
4156                         if (IS_ERR(trans))
4157                                 return PTR_ERR(trans);
4158
4159                         ret = do_chunk_alloc(trans, alloc_target,
4160                                              CHUNK_ALLOC_NO_FORCE);
4161                         btrfs_end_transaction(trans);
4162                         if (ret < 0) {
4163                                 if (ret != -ENOSPC)
4164                                         return ret;
4165                                 else {
4166                                         have_pinned_space = 1;
4167                                         goto commit_trans;
4168                                 }
4169                         }
4170
4171                         goto again;
4172                 }
4173
4174                 /*
4175                  * If we don't have enough pinned space to deal with this
4176                  * allocation, and no removed chunk in current transaction,
4177                  * don't bother committing the transaction.
4178                  */
4179                 have_pinned_space = __percpu_counter_compare(
4180                         &data_sinfo->total_bytes_pinned,
4181                         used + bytes - data_sinfo->total_bytes,
4182                         BTRFS_TOTAL_BYTES_PINNED_BATCH);
4183                 spin_unlock(&data_sinfo->lock);
4184
4185                 /* commit the current transaction and try again */
4186 commit_trans:
4187                 if (need_commit) {
4188                         need_commit--;
4189
4190                         if (need_commit > 0) {
4191                                 btrfs_start_delalloc_roots(fs_info, -1);
4192                                 btrfs_wait_ordered_roots(fs_info, U64_MAX, 0,
4193                                                          (u64)-1);
4194                         }
4195
4196                         trans = btrfs_join_transaction(root);
4197                         if (IS_ERR(trans))
4198                                 return PTR_ERR(trans);
4199                         if (have_pinned_space >= 0 ||
4200                             test_bit(BTRFS_TRANS_HAVE_FREE_BGS,
4201                                      &trans->transaction->flags) ||
4202                             need_commit > 0) {
4203                                 ret = btrfs_commit_transaction(trans);
4204                                 if (ret)
4205                                         return ret;
4206                                 /*
4207                                  * The cleaner kthread might still be doing iput
4208                                  * operations. Wait for it to finish so that
4209                                  * more space is released.
4210                                  */
4211                                 mutex_lock(&fs_info->cleaner_delayed_iput_mutex);
4212                                 mutex_unlock(&fs_info->cleaner_delayed_iput_mutex);
4213                                 goto again;
4214                         } else {
4215                                 btrfs_end_transaction(trans);
4216                         }
4217                 }
4218
4219                 trace_btrfs_space_reservation(fs_info,
4220                                               "space_info:enospc",
4221                                               data_sinfo->flags, bytes, 1);
4222                 return -ENOSPC;
4223         }
4224         data_sinfo->bytes_may_use += bytes;
4225         trace_btrfs_space_reservation(fs_info, "space_info",
4226                                       data_sinfo->flags, bytes, 1);
4227         spin_unlock(&data_sinfo->lock);
4228
4229         return 0;
4230 }
4231
4232 int btrfs_check_data_free_space(struct inode *inode,
4233                         struct extent_changeset **reserved, u64 start, u64 len)
4234 {
4235         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4236         int ret;
4237
4238         /* align the range */
4239         len = round_up(start + len, fs_info->sectorsize) -
4240               round_down(start, fs_info->sectorsize);
4241         start = round_down(start, fs_info->sectorsize);
4242
4243         ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode), len);
4244         if (ret < 0)
4245                 return ret;
4246
4247         /* Use new btrfs_qgroup_reserve_data to reserve precious data space. */
4248         ret = btrfs_qgroup_reserve_data(inode, reserved, start, len);
4249         if (ret < 0)
4250                 btrfs_free_reserved_data_space_noquota(inode, start, len);
4251         else
4252                 ret = 0;
4253         return ret;
4254 }
4255
4256 /*
4257  * Called if we need to clear a data reservation for this inode
4258  * Normally in a error case.
4259  *
4260  * This one will *NOT* use accurate qgroup reserved space API, just for case
4261  * which we can't sleep and is sure it won't affect qgroup reserved space.
4262  * Like clear_bit_hook().
4263  */
4264 void btrfs_free_reserved_data_space_noquota(struct inode *inode, u64 start,
4265                                             u64 len)
4266 {
4267         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4268         struct btrfs_space_info *data_sinfo;
4269
4270         /* Make sure the range is aligned to sectorsize */
4271         len = round_up(start + len, fs_info->sectorsize) -
4272               round_down(start, fs_info->sectorsize);
4273         start = round_down(start, fs_info->sectorsize);
4274
4275         data_sinfo = fs_info->data_sinfo;
4276         spin_lock(&data_sinfo->lock);
4277         if (WARN_ON(data_sinfo->bytes_may_use < len))
4278                 data_sinfo->bytes_may_use = 0;
4279         else
4280                 data_sinfo->bytes_may_use -= len;
4281         trace_btrfs_space_reservation(fs_info, "space_info",
4282                                       data_sinfo->flags, len, 0);
4283         spin_unlock(&data_sinfo->lock);
4284 }
4285
4286 /*
4287  * Called if we need to clear a data reservation for this inode
4288  * Normally in a error case.
4289  *
4290  * This one will handle the per-inode data rsv map for accurate reserved
4291  * space framework.
4292  */
4293 void btrfs_free_reserved_data_space(struct inode *inode,
4294                         struct extent_changeset *reserved, u64 start, u64 len)
4295 {
4296         struct btrfs_root *root = BTRFS_I(inode)->root;
4297
4298         /* Make sure the range is aligned to sectorsize */
4299         len = round_up(start + len, root->fs_info->sectorsize) -
4300               round_down(start, root->fs_info->sectorsize);
4301         start = round_down(start, root->fs_info->sectorsize);
4302
4303         btrfs_free_reserved_data_space_noquota(inode, start, len);
4304         btrfs_qgroup_free_data(inode, reserved, start, len);
4305 }
4306
4307 static void force_metadata_allocation(struct btrfs_fs_info *info)
4308 {
4309         struct list_head *head = &info->space_info;
4310         struct btrfs_space_info *found;
4311
4312         rcu_read_lock();
4313         list_for_each_entry_rcu(found, head, list) {
4314                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
4315                         found->force_alloc = CHUNK_ALLOC_FORCE;
4316         }
4317         rcu_read_unlock();
4318 }
4319
4320 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
4321 {
4322         return (global->size << 1);
4323 }
4324
4325 static int should_alloc_chunk(struct btrfs_fs_info *fs_info,
4326                               struct btrfs_space_info *sinfo, int force)
4327 {
4328         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4329         u64 bytes_used = btrfs_space_info_used(sinfo, false);
4330         u64 thresh;
4331
4332         if (force == CHUNK_ALLOC_FORCE)
4333                 return 1;
4334
4335         /*
4336          * We need to take into account the global rsv because for all intents
4337          * and purposes it's used space.  Don't worry about locking the
4338          * global_rsv, it doesn't change except when the transaction commits.
4339          */
4340         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
4341                 bytes_used += calc_global_rsv_need_space(global_rsv);
4342
4343         /*
4344          * in limited mode, we want to have some free space up to
4345          * about 1% of the FS size.
4346          */
4347         if (force == CHUNK_ALLOC_LIMITED) {
4348                 thresh = btrfs_super_total_bytes(fs_info->super_copy);
4349                 thresh = max_t(u64, SZ_64M, div_factor_fine(thresh, 1));
4350
4351                 if (sinfo->total_bytes - bytes_used < thresh)
4352                         return 1;
4353         }
4354
4355         if (bytes_used + SZ_2M < div_factor(sinfo->total_bytes, 8))
4356                 return 0;
4357         return 1;
4358 }
4359
4360 static u64 get_profile_num_devs(struct btrfs_fs_info *fs_info, u64 type)
4361 {
4362         u64 num_dev;
4363
4364         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
4365                     BTRFS_BLOCK_GROUP_RAID0 |
4366                     BTRFS_BLOCK_GROUP_RAID5 |
4367                     BTRFS_BLOCK_GROUP_RAID6))
4368                 num_dev = fs_info->fs_devices->rw_devices;
4369         else if (type & BTRFS_BLOCK_GROUP_RAID1)
4370                 num_dev = 2;
4371         else
4372                 num_dev = 1;    /* DUP or single */
4373
4374         return num_dev;
4375 }
4376
4377 /*
4378  * If @is_allocation is true, reserve space in the system space info necessary
4379  * for allocating a chunk, otherwise if it's false, reserve space necessary for
4380  * removing a chunk.
4381  */
4382 void check_system_chunk(struct btrfs_trans_handle *trans, u64 type)
4383 {
4384         struct btrfs_fs_info *fs_info = trans->fs_info;
4385         struct btrfs_space_info *info;
4386         u64 left;
4387         u64 thresh;
4388         int ret = 0;
4389         u64 num_devs;
4390
4391         /*
4392          * Needed because we can end up allocating a system chunk and for an
4393          * atomic and race free space reservation in the chunk block reserve.
4394          */
4395         lockdep_assert_held(&fs_info->chunk_mutex);
4396
4397         info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4398         spin_lock(&info->lock);
4399         left = info->total_bytes - btrfs_space_info_used(info, true);
4400         spin_unlock(&info->lock);
4401
4402         num_devs = get_profile_num_devs(fs_info, type);
4403
4404         /* num_devs device items to update and 1 chunk item to add or remove */
4405         thresh = btrfs_calc_trunc_metadata_size(fs_info, num_devs) +
4406                 btrfs_calc_trans_metadata_size(fs_info, 1);
4407
4408         if (left < thresh && btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
4409                 btrfs_info(fs_info, "left=%llu, need=%llu, flags=%llu",
4410                            left, thresh, type);
4411                 dump_space_info(fs_info, info, 0, 0);
4412         }
4413
4414         if (left < thresh) {
4415                 u64 flags = btrfs_system_alloc_profile(fs_info);
4416
4417                 /*
4418                  * Ignore failure to create system chunk. We might end up not
4419                  * needing it, as we might not need to COW all nodes/leafs from
4420                  * the paths we visit in the chunk tree (they were already COWed
4421                  * or created in the current transaction for example).
4422                  */
4423                 ret = btrfs_alloc_chunk(trans, flags);
4424         }
4425
4426         if (!ret) {
4427                 ret = btrfs_block_rsv_add(fs_info->chunk_root,
4428                                           &fs_info->chunk_block_rsv,
4429                                           thresh, BTRFS_RESERVE_NO_FLUSH);
4430                 if (!ret)
4431                         trans->chunk_bytes_reserved += thresh;
4432         }
4433 }
4434
4435 /*
4436  * If force is CHUNK_ALLOC_FORCE:
4437  *    - return 1 if it successfully allocates a chunk,
4438  *    - return errors including -ENOSPC otherwise.
4439  * If force is NOT CHUNK_ALLOC_FORCE:
4440  *    - return 0 if it doesn't need to allocate a new chunk,
4441  *    - return 1 if it successfully allocates a chunk,
4442  *    - return errors including -ENOSPC otherwise.
4443  */
4444 static int do_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags,
4445                           int force)
4446 {
4447         struct btrfs_fs_info *fs_info = trans->fs_info;
4448         struct btrfs_space_info *space_info;
4449         bool wait_for_alloc = false;
4450         bool should_alloc = false;
4451         int ret = 0;
4452
4453         /* Don't re-enter if we're already allocating a chunk */
4454         if (trans->allocating_chunk)
4455                 return -ENOSPC;
4456
4457         space_info = __find_space_info(fs_info, flags);
4458         ASSERT(space_info);
4459
4460         do {
4461                 spin_lock(&space_info->lock);
4462                 if (force < space_info->force_alloc)
4463                         force = space_info->force_alloc;
4464                 should_alloc = should_alloc_chunk(fs_info, space_info, force);
4465                 if (space_info->full) {
4466                         /* No more free physical space */
4467                         if (should_alloc)
4468                                 ret = -ENOSPC;
4469                         else
4470                                 ret = 0;
4471                         spin_unlock(&space_info->lock);
4472                         return ret;
4473                 } else if (!should_alloc) {
4474                         spin_unlock(&space_info->lock);
4475                         return 0;
4476                 } else if (space_info->chunk_alloc) {
4477                         /*
4478                          * Someone is already allocating, so we need to block
4479                          * until this someone is finished and then loop to
4480                          * recheck if we should continue with our allocation
4481                          * attempt.
4482                          */
4483                         wait_for_alloc = true;
4484                         spin_unlock(&space_info->lock);
4485                         mutex_lock(&fs_info->chunk_mutex);
4486                         mutex_unlock(&fs_info->chunk_mutex);
4487                 } else {
4488                         /* Proceed with allocation */
4489                         space_info->chunk_alloc = 1;
4490                         wait_for_alloc = false;
4491                         spin_unlock(&space_info->lock);
4492                 }
4493
4494                 cond_resched();
4495         } while (wait_for_alloc);
4496
4497         mutex_lock(&fs_info->chunk_mutex);
4498         trans->allocating_chunk = true;
4499
4500         /*
4501          * If we have mixed data/metadata chunks we want to make sure we keep
4502          * allocating mixed chunks instead of individual chunks.
4503          */
4504         if (btrfs_mixed_space_info(space_info))
4505                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4506
4507         /*
4508          * if we're doing a data chunk, go ahead and make sure that
4509          * we keep a reasonable number of metadata chunks allocated in the
4510          * FS as well.
4511          */
4512         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4513                 fs_info->data_chunk_allocations++;
4514                 if (!(fs_info->data_chunk_allocations %
4515                       fs_info->metadata_ratio))
4516                         force_metadata_allocation(fs_info);
4517         }
4518
4519         /*
4520          * Check if we have enough space in SYSTEM chunk because we may need
4521          * to update devices.
4522          */
4523         check_system_chunk(trans, flags);
4524
4525         ret = btrfs_alloc_chunk(trans, flags);
4526         trans->allocating_chunk = false;
4527
4528         spin_lock(&space_info->lock);
4529         if (ret < 0) {
4530                 if (ret == -ENOSPC)
4531                         space_info->full = 1;
4532                 else
4533                         goto out;
4534         } else {
4535                 ret = 1;
4536         }
4537
4538         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4539 out:
4540         space_info->chunk_alloc = 0;
4541         spin_unlock(&space_info->lock);
4542         mutex_unlock(&fs_info->chunk_mutex);
4543         /*
4544          * When we allocate a new chunk we reserve space in the chunk block
4545          * reserve to make sure we can COW nodes/leafs in the chunk tree or
4546          * add new nodes/leafs to it if we end up needing to do it when
4547          * inserting the chunk item and updating device items as part of the
4548          * second phase of chunk allocation, performed by
4549          * btrfs_finish_chunk_alloc(). So make sure we don't accumulate a
4550          * large number of new block groups to create in our transaction
4551          * handle's new_bgs list to avoid exhausting the chunk block reserve
4552          * in extreme cases - like having a single transaction create many new
4553          * block groups when starting to write out the free space caches of all
4554          * the block groups that were made dirty during the lifetime of the
4555          * transaction.
4556          */
4557         if (trans->can_flush_pending_bgs &&
4558             trans->chunk_bytes_reserved >= (u64)SZ_2M) {
4559                 btrfs_create_pending_block_groups(trans);
4560                 btrfs_trans_release_chunk_metadata(trans);
4561         }
4562         return ret;
4563 }
4564
4565 static int can_overcommit(struct btrfs_fs_info *fs_info,
4566                           struct btrfs_space_info *space_info, u64 bytes,
4567                           enum btrfs_reserve_flush_enum flush,
4568                           bool system_chunk)
4569 {
4570         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4571         u64 profile;
4572         u64 space_size;
4573         u64 avail;
4574         u64 used;
4575         int factor;
4576
4577         /* Don't overcommit when in mixed mode. */
4578         if (space_info->flags & BTRFS_BLOCK_GROUP_DATA)
4579                 return 0;
4580
4581         if (system_chunk)
4582                 profile = btrfs_system_alloc_profile(fs_info);
4583         else
4584                 profile = btrfs_metadata_alloc_profile(fs_info);
4585
4586         used = btrfs_space_info_used(space_info, false);
4587
4588         /*
4589          * We only want to allow over committing if we have lots of actual space
4590          * free, but if we don't have enough space to handle the global reserve
4591          * space then we could end up having a real enospc problem when trying
4592          * to allocate a chunk or some other such important allocation.
4593          */
4594         spin_lock(&global_rsv->lock);
4595         space_size = calc_global_rsv_need_space(global_rsv);
4596         spin_unlock(&global_rsv->lock);
4597         if (used + space_size >= space_info->total_bytes)
4598                 return 0;
4599
4600         used += space_info->bytes_may_use;
4601
4602         avail = atomic64_read(&fs_info->free_chunk_space);
4603
4604         /*
4605          * If we have dup, raid1 or raid10 then only half of the free
4606          * space is actually useable.  For raid56, the space info used
4607          * doesn't include the parity drive, so we don't have to
4608          * change the math
4609          */
4610         factor = btrfs_bg_type_to_factor(profile);
4611         avail = div_u64(avail, factor);
4612
4613         /*
4614          * If we aren't flushing all things, let us overcommit up to
4615          * 1/2th of the space. If we can flush, don't let us overcommit
4616          * too much, let it overcommit up to 1/8 of the space.
4617          */
4618         if (flush == BTRFS_RESERVE_FLUSH_ALL)
4619                 avail >>= 3;
4620         else
4621                 avail >>= 1;
4622
4623         if (used + bytes < space_info->total_bytes + avail)
4624                 return 1;
4625         return 0;
4626 }
4627
4628 static void btrfs_writeback_inodes_sb_nr(struct btrfs_fs_info *fs_info,
4629                                          unsigned long nr_pages, int nr_items)
4630 {
4631         struct super_block *sb = fs_info->sb;
4632
4633         if (down_read_trylock(&sb->s_umount)) {
4634                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
4635                 up_read(&sb->s_umount);
4636         } else {
4637                 /*
4638                  * We needn't worry the filesystem going from r/w to r/o though
4639                  * we don't acquire ->s_umount mutex, because the filesystem
4640                  * should guarantee the delalloc inodes list be empty after
4641                  * the filesystem is readonly(all dirty pages are written to
4642                  * the disk).
4643                  */
4644                 btrfs_start_delalloc_roots(fs_info, nr_items);
4645                 if (!current->journal_info)
4646                         btrfs_wait_ordered_roots(fs_info, nr_items, 0, (u64)-1);
4647         }
4648 }
4649
4650 static inline u64 calc_reclaim_items_nr(struct btrfs_fs_info *fs_info,
4651                                         u64 to_reclaim)
4652 {
4653         u64 bytes;
4654         u64 nr;
4655
4656         bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
4657         nr = div64_u64(to_reclaim, bytes);
4658         if (!nr)
4659                 nr = 1;
4660         return nr;
4661 }
4662
4663 #define EXTENT_SIZE_PER_ITEM    SZ_256K
4664
4665 /*
4666  * shrink metadata reservation for delalloc
4667  */
4668 static void shrink_delalloc(struct btrfs_fs_info *fs_info, u64 to_reclaim,
4669                             u64 orig, bool wait_ordered)
4670 {
4671         struct btrfs_space_info *space_info;
4672         struct btrfs_trans_handle *trans;
4673         u64 delalloc_bytes;
4674         u64 max_reclaim;
4675         u64 items;
4676         long time_left;
4677         unsigned long nr_pages;
4678         int loops;
4679
4680         /* Calc the number of the pages we need flush for space reservation */
4681         items = calc_reclaim_items_nr(fs_info, to_reclaim);
4682         to_reclaim = items * EXTENT_SIZE_PER_ITEM;
4683
4684         trans = (struct btrfs_trans_handle *)current->journal_info;
4685         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4686
4687         delalloc_bytes = percpu_counter_sum_positive(
4688                                                 &fs_info->delalloc_bytes);
4689         if (delalloc_bytes == 0) {
4690                 if (trans)
4691                         return;
4692                 if (wait_ordered)
4693                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
4694                 return;
4695         }
4696
4697         loops = 0;
4698         while (delalloc_bytes && loops < 3) {
4699                 max_reclaim = min(delalloc_bytes, to_reclaim);
4700                 nr_pages = max_reclaim >> PAGE_SHIFT;
4701                 btrfs_writeback_inodes_sb_nr(fs_info, nr_pages, items);
4702                 /*
4703                  * We need to wait for the async pages to actually start before
4704                  * we do anything.
4705                  */
4706                 max_reclaim = atomic_read(&fs_info->async_delalloc_pages);
4707                 if (!max_reclaim)
4708                         goto skip_async;
4709
4710                 if (max_reclaim <= nr_pages)
4711                         max_reclaim = 0;
4712                 else
4713                         max_reclaim -= nr_pages;
4714
4715                 wait_event(fs_info->async_submit_wait,
4716                            atomic_read(&fs_info->async_delalloc_pages) <=
4717                            (int)max_reclaim);
4718 skip_async:
4719                 spin_lock(&space_info->lock);
4720                 if (list_empty(&space_info->tickets) &&
4721                     list_empty(&space_info->priority_tickets)) {
4722                         spin_unlock(&space_info->lock);
4723                         break;
4724                 }
4725                 spin_unlock(&space_info->lock);
4726
4727                 loops++;
4728                 if (wait_ordered && !trans) {
4729                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
4730                 } else {
4731                         time_left = schedule_timeout_killable(1);
4732                         if (time_left)
4733                                 break;
4734                 }
4735                 delalloc_bytes = percpu_counter_sum_positive(
4736                                                 &fs_info->delalloc_bytes);
4737         }
4738 }
4739
4740 struct reserve_ticket {
4741         u64 bytes;
4742         int error;
4743         struct list_head list;
4744         wait_queue_head_t wait;
4745 };
4746
4747 /**
4748  * maybe_commit_transaction - possibly commit the transaction if its ok to
4749  * @root - the root we're allocating for
4750  * @bytes - the number of bytes we want to reserve
4751  * @force - force the commit
4752  *
4753  * This will check to make sure that committing the transaction will actually
4754  * get us somewhere and then commit the transaction if it does.  Otherwise it
4755  * will return -ENOSPC.
4756  */
4757 static int may_commit_transaction(struct btrfs_fs_info *fs_info,
4758                                   struct btrfs_space_info *space_info)
4759 {
4760         struct reserve_ticket *ticket = NULL;
4761         struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_block_rsv;
4762         struct btrfs_trans_handle *trans;
4763         u64 bytes;
4764
4765         trans = (struct btrfs_trans_handle *)current->journal_info;
4766         if (trans)
4767                 return -EAGAIN;
4768
4769         spin_lock(&space_info->lock);
4770         if (!list_empty(&space_info->priority_tickets))
4771                 ticket = list_first_entry(&space_info->priority_tickets,
4772                                           struct reserve_ticket, list);
4773         else if (!list_empty(&space_info->tickets))
4774                 ticket = list_first_entry(&space_info->tickets,
4775                                           struct reserve_ticket, list);
4776         bytes = (ticket) ? ticket->bytes : 0;
4777         spin_unlock(&space_info->lock);
4778
4779         if (!bytes)
4780                 return 0;
4781
4782         /* See if there is enough pinned space to make this reservation */
4783         if (__percpu_counter_compare(&space_info->total_bytes_pinned,
4784                                    bytes,
4785                                    BTRFS_TOTAL_BYTES_PINNED_BATCH) >= 0)
4786                 goto commit;
4787
4788         /*
4789          * See if there is some space in the delayed insertion reservation for
4790          * this reservation.
4791          */
4792         if (space_info != delayed_rsv->space_info)
4793                 return -ENOSPC;
4794
4795         spin_lock(&delayed_rsv->lock);
4796         if (delayed_rsv->size > bytes)
4797                 bytes = 0;
4798         else
4799                 bytes -= delayed_rsv->size;
4800         spin_unlock(&delayed_rsv->lock);
4801
4802         if (__percpu_counter_compare(&space_info->total_bytes_pinned,
4803                                    bytes,
4804                                    BTRFS_TOTAL_BYTES_PINNED_BATCH) < 0) {
4805                 return -ENOSPC;
4806         }
4807
4808 commit:
4809         trans = btrfs_join_transaction(fs_info->extent_root);
4810         if (IS_ERR(trans))
4811                 return -ENOSPC;
4812
4813         return btrfs_commit_transaction(trans);
4814 }
4815
4816 /*
4817  * Try to flush some data based on policy set by @state. This is only advisory
4818  * and may fail for various reasons. The caller is supposed to examine the
4819  * state of @space_info to detect the outcome.
4820  */
4821 static void flush_space(struct btrfs_fs_info *fs_info,
4822                        struct btrfs_space_info *space_info, u64 num_bytes,
4823                        int state)
4824 {
4825         struct btrfs_root *root = fs_info->extent_root;
4826         struct btrfs_trans_handle *trans;
4827         int nr;
4828         int ret = 0;
4829
4830         switch (state) {
4831         case FLUSH_DELAYED_ITEMS_NR:
4832         case FLUSH_DELAYED_ITEMS:
4833                 if (state == FLUSH_DELAYED_ITEMS_NR)
4834                         nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2;
4835                 else
4836                         nr = -1;
4837
4838                 trans = btrfs_join_transaction(root);
4839                 if (IS_ERR(trans)) {
4840                         ret = PTR_ERR(trans);
4841                         break;
4842                 }
4843                 ret = btrfs_run_delayed_items_nr(trans, nr);
4844                 btrfs_end_transaction(trans);
4845                 break;
4846         case FLUSH_DELALLOC:
4847         case FLUSH_DELALLOC_WAIT:
4848                 shrink_delalloc(fs_info, num_bytes * 2, num_bytes,
4849                                 state == FLUSH_DELALLOC_WAIT);
4850                 break;
4851         case ALLOC_CHUNK:
4852                 trans = btrfs_join_transaction(root);
4853                 if (IS_ERR(trans)) {
4854                         ret = PTR_ERR(trans);
4855                         break;
4856                 }
4857                 ret = do_chunk_alloc(trans,
4858                                      btrfs_metadata_alloc_profile(fs_info),
4859                                      CHUNK_ALLOC_NO_FORCE);
4860                 btrfs_end_transaction(trans);
4861                 if (ret > 0 || ret == -ENOSPC)
4862                         ret = 0;
4863                 break;
4864         case COMMIT_TRANS:
4865                 ret = may_commit_transaction(fs_info, space_info);
4866                 break;
4867         default:
4868                 ret = -ENOSPC;
4869                 break;
4870         }
4871
4872         trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state,
4873                                 ret);
4874         return;
4875 }
4876
4877 static inline u64
4878 btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info,
4879                                  struct btrfs_space_info *space_info,
4880                                  bool system_chunk)
4881 {
4882         struct reserve_ticket *ticket;
4883         u64 used;
4884         u64 expected;
4885         u64 to_reclaim = 0;
4886
4887         list_for_each_entry(ticket, &space_info->tickets, list)
4888                 to_reclaim += ticket->bytes;
4889         list_for_each_entry(ticket, &space_info->priority_tickets, list)
4890                 to_reclaim += ticket->bytes;
4891         if (to_reclaim)
4892                 return to_reclaim;
4893
4894         to_reclaim = min_t(u64, num_online_cpus() * SZ_1M, SZ_16M);
4895         if (can_overcommit(fs_info, space_info, to_reclaim,
4896                            BTRFS_RESERVE_FLUSH_ALL, system_chunk))
4897                 return 0;
4898
4899         used = btrfs_space_info_used(space_info, true);
4900
4901         if (can_overcommit(fs_info, space_info, SZ_1M,
4902                            BTRFS_RESERVE_FLUSH_ALL, system_chunk))
4903                 expected = div_factor_fine(space_info->total_bytes, 95);
4904         else
4905                 expected = div_factor_fine(space_info->total_bytes, 90);
4906
4907         if (used > expected)
4908                 to_reclaim = used - expected;
4909         else
4910                 to_reclaim = 0;
4911         to_reclaim = min(to_reclaim, space_info->bytes_may_use +
4912                                      space_info->bytes_reserved);
4913         return to_reclaim;
4914 }
4915
4916 static inline int need_do_async_reclaim(struct btrfs_fs_info *fs_info,
4917                                         struct btrfs_space_info *space_info,
4918                                         u64 used, bool system_chunk)
4919 {
4920         u64 thresh = div_factor_fine(space_info->total_bytes, 98);
4921
4922         /* If we're just plain full then async reclaim just slows us down. */
4923         if ((space_info->bytes_used + space_info->bytes_reserved) >= thresh)
4924                 return 0;
4925
4926         if (!btrfs_calc_reclaim_metadata_size(fs_info, space_info,
4927                                               system_chunk))
4928                 return 0;
4929
4930         return (used >= thresh && !btrfs_fs_closing(fs_info) &&
4931                 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
4932 }
4933
4934 static void wake_all_tickets(struct list_head *head)
4935 {
4936         struct reserve_ticket *ticket;
4937
4938         while (!list_empty(head)) {
4939                 ticket = list_first_entry(head, struct reserve_ticket, list);
4940                 list_del_init(&ticket->list);
4941                 ticket->error = -ENOSPC;
4942                 wake_up(&ticket->wait);
4943         }
4944 }
4945
4946 /*
4947  * This is for normal flushers, we can wait all goddamned day if we want to.  We
4948  * will loop and continuously try to flush as long as we are making progress.
4949  * We count progress as clearing off tickets each time we have to loop.
4950  */
4951 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
4952 {
4953         struct btrfs_fs_info *fs_info;
4954         struct btrfs_space_info *space_info;
4955         u64 to_reclaim;
4956         int flush_state;
4957         int commit_cycles = 0;
4958         u64 last_tickets_id;
4959
4960         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
4961         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4962
4963         spin_lock(&space_info->lock);
4964         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info,
4965                                                       false);
4966         if (!to_reclaim) {
4967                 space_info->flush = 0;
4968                 spin_unlock(&space_info->lock);
4969                 return;
4970         }
4971         last_tickets_id = space_info->tickets_id;
4972         spin_unlock(&space_info->lock);
4973
4974         flush_state = FLUSH_DELAYED_ITEMS_NR;
4975         do {
4976                 flush_space(fs_info, space_info, to_reclaim, flush_state);
4977                 spin_lock(&space_info->lock);
4978                 if (list_empty(&space_info->tickets)) {
4979                         space_info->flush = 0;
4980                         spin_unlock(&space_info->lock);
4981                         return;
4982                 }
4983                 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info,
4984                                                               space_info,
4985                                                               false);
4986                 if (last_tickets_id == space_info->tickets_id) {
4987                         flush_state++;
4988                 } else {
4989                         last_tickets_id = space_info->tickets_id;
4990                         flush_state = FLUSH_DELAYED_ITEMS_NR;
4991                         if (commit_cycles)
4992                                 commit_cycles--;
4993                 }
4994
4995                 if (flush_state > COMMIT_TRANS) {
4996                         commit_cycles++;
4997                         if (commit_cycles > 2) {
4998                                 wake_all_tickets(&space_info->tickets);
4999                                 space_info->flush = 0;
5000                         } else {
5001                                 flush_state = FLUSH_DELAYED_ITEMS_NR;
5002                         }
5003                 }
5004                 spin_unlock(&space_info->lock);
5005         } while (flush_state <= COMMIT_TRANS);
5006 }
5007
5008 void btrfs_init_async_reclaim_work(struct work_struct *work)
5009 {
5010         INIT_WORK(work, btrfs_async_reclaim_metadata_space);
5011 }
5012
5013 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info,
5014                                             struct btrfs_space_info *space_info,
5015                                             struct reserve_ticket *ticket)
5016 {
5017         u64 to_reclaim;
5018         int flush_state = FLUSH_DELAYED_ITEMS_NR;
5019
5020         spin_lock(&space_info->lock);
5021         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5022                                                       false);
5023         if (!to_reclaim) {
5024                 spin_unlock(&space_info->lock);
5025                 return;
5026         }
5027         spin_unlock(&space_info->lock);
5028
5029         do {
5030                 flush_space(fs_info, space_info, to_reclaim, flush_state);
5031                 flush_state++;
5032                 spin_lock(&space_info->lock);
5033                 if (ticket->bytes == 0) {
5034                         spin_unlock(&space_info->lock);
5035                         return;
5036                 }
5037                 spin_unlock(&space_info->lock);
5038
5039                 /*
5040                  * Priority flushers can't wait on delalloc without
5041                  * deadlocking.
5042                  */
5043                 if (flush_state == FLUSH_DELALLOC ||
5044                     flush_state == FLUSH_DELALLOC_WAIT)
5045                         flush_state = ALLOC_CHUNK;
5046         } while (flush_state < COMMIT_TRANS);
5047 }
5048
5049 static int wait_reserve_ticket(struct btrfs_fs_info *fs_info,
5050                                struct btrfs_space_info *space_info,
5051                                struct reserve_ticket *ticket, u64 orig_bytes)
5052
5053 {
5054         DEFINE_WAIT(wait);
5055         int ret = 0;
5056
5057         spin_lock(&space_info->lock);
5058         while (ticket->bytes > 0 && ticket->error == 0) {
5059                 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE);
5060                 if (ret) {
5061                         ret = -EINTR;
5062                         break;
5063                 }
5064                 spin_unlock(&space_info->lock);
5065
5066                 schedule();
5067
5068                 finish_wait(&ticket->wait, &wait);
5069                 spin_lock(&space_info->lock);
5070         }
5071         if (!ret)
5072                 ret = ticket->error;
5073         if (!list_empty(&ticket->list))
5074                 list_del_init(&ticket->list);
5075         if (ticket->bytes && ticket->bytes < orig_bytes) {
5076                 u64 num_bytes = orig_bytes - ticket->bytes;
5077                 space_info->bytes_may_use -= num_bytes;
5078                 trace_btrfs_space_reservation(fs_info, "space_info",
5079                                               space_info->flags, num_bytes, 0);
5080         }
5081         spin_unlock(&space_info->lock);
5082
5083         return ret;
5084 }
5085
5086 /**
5087  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5088  * @root - the root we're allocating for
5089  * @space_info - the space info we want to allocate from
5090  * @orig_bytes - the number of bytes we want
5091  * @flush - whether or not we can flush to make our reservation
5092  *
5093  * This will reserve orig_bytes number of bytes from the space info associated
5094  * with the block_rsv.  If there is not enough space it will make an attempt to
5095  * flush out space to make room.  It will do this by flushing delalloc if
5096  * possible or committing the transaction.  If flush is 0 then no attempts to
5097  * regain reservations will be made and this will fail if there is not enough
5098  * space already.
5099  */
5100 static int __reserve_metadata_bytes(struct btrfs_fs_info *fs_info,
5101                                     struct btrfs_space_info *space_info,
5102                                     u64 orig_bytes,
5103                                     enum btrfs_reserve_flush_enum flush,
5104                                     bool system_chunk)
5105 {
5106         struct reserve_ticket ticket;
5107         u64 used;
5108         int ret = 0;
5109
5110         ASSERT(orig_bytes);
5111         ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL);
5112
5113         spin_lock(&space_info->lock);
5114         ret = -ENOSPC;
5115         used = btrfs_space_info_used(space_info, true);
5116
5117         /*
5118          * If we have enough space then hooray, make our reservation and carry
5119          * on.  If not see if we can overcommit, and if we can, hooray carry on.
5120          * If not things get more complicated.
5121          */
5122         if (used + orig_bytes <= space_info->total_bytes) {
5123                 space_info->bytes_may_use += orig_bytes;
5124                 trace_btrfs_space_reservation(fs_info, "space_info",
5125                                               space_info->flags, orig_bytes, 1);
5126                 ret = 0;
5127         } else if (can_overcommit(fs_info, space_info, orig_bytes, flush,
5128                                   system_chunk)) {
5129                 space_info->bytes_may_use += orig_bytes;
5130                 trace_btrfs_space_reservation(fs_info, "space_info",
5131                                               space_info->flags, orig_bytes, 1);
5132                 ret = 0;
5133         }
5134
5135         /*
5136          * If we couldn't make a reservation then setup our reservation ticket
5137          * and kick the async worker if it's not already running.
5138          *
5139          * If we are a priority flusher then we just need to add our ticket to
5140          * the list and we will do our own flushing further down.
5141          */
5142         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
5143                 ticket.bytes = orig_bytes;
5144                 ticket.error = 0;
5145                 init_waitqueue_head(&ticket.wait);
5146                 if (flush == BTRFS_RESERVE_FLUSH_ALL) {
5147                         list_add_tail(&ticket.list, &space_info->tickets);
5148                         if (!space_info->flush) {
5149                                 space_info->flush = 1;
5150                                 trace_btrfs_trigger_flush(fs_info,
5151                                                           space_info->flags,
5152                                                           orig_bytes, flush,
5153                                                           "enospc");
5154                                 queue_work(system_unbound_wq,
5155                                            &fs_info->async_reclaim_work);
5156                         }
5157                 } else {
5158                         list_add_tail(&ticket.list,
5159                                       &space_info->priority_tickets);
5160                 }
5161         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
5162                 used += orig_bytes;
5163                 /*
5164                  * We will do the space reservation dance during log replay,
5165                  * which means we won't have fs_info->fs_root set, so don't do
5166                  * the async reclaim as we will panic.
5167                  */
5168                 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) &&
5169                     need_do_async_reclaim(fs_info, space_info,
5170                                           used, system_chunk) &&
5171                     !work_busy(&fs_info->async_reclaim_work)) {
5172                         trace_btrfs_trigger_flush(fs_info, space_info->flags,
5173                                                   orig_bytes, flush, "preempt");
5174                         queue_work(system_unbound_wq,
5175                                    &fs_info->async_reclaim_work);
5176                 }
5177         }
5178         spin_unlock(&space_info->lock);
5179         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
5180                 return ret;
5181
5182         if (flush == BTRFS_RESERVE_FLUSH_ALL)
5183                 return wait_reserve_ticket(fs_info, space_info, &ticket,
5184                                            orig_bytes);
5185
5186         ret = 0;
5187         priority_reclaim_metadata_space(fs_info, space_info, &ticket);
5188         spin_lock(&space_info->lock);
5189         if (ticket.bytes) {
5190                 if (ticket.bytes < orig_bytes) {
5191                         u64 num_bytes = orig_bytes - ticket.bytes;
5192                         space_info->bytes_may_use -= num_bytes;
5193                         trace_btrfs_space_reservation(fs_info, "space_info",
5194                                                       space_info->flags,
5195                                                       num_bytes, 0);
5196
5197                 }
5198                 list_del_init(&ticket.list);
5199                 ret = -ENOSPC;
5200         }
5201         spin_unlock(&space_info->lock);
5202         ASSERT(list_empty(&ticket.list));
5203         return ret;
5204 }
5205
5206 /**
5207  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5208  * @root - the root we're allocating for
5209  * @block_rsv - the block_rsv we're allocating for
5210  * @orig_bytes - the number of bytes we want
5211  * @flush - whether or not we can flush to make our reservation
5212  *
5213  * This will reserve orgi_bytes number of bytes from the space info associated
5214  * with the block_rsv.  If there is not enough space it will make an attempt to
5215  * flush out space to make room.  It will do this by flushing delalloc if
5216  * possible or committing the transaction.  If flush is 0 then no attempts to
5217  * regain reservations will be made and this will fail if there is not enough
5218  * space already.
5219  */
5220 static int reserve_metadata_bytes(struct btrfs_root *root,
5221                                   struct btrfs_block_rsv *block_rsv,
5222                                   u64 orig_bytes,
5223                                   enum btrfs_reserve_flush_enum flush)
5224 {
5225         struct btrfs_fs_info *fs_info = root->fs_info;
5226         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5227         int ret;
5228         bool system_chunk = (root == fs_info->chunk_root);
5229
5230         ret = __reserve_metadata_bytes(fs_info, block_rsv->space_info,
5231                                        orig_bytes, flush, system_chunk);
5232         if (ret == -ENOSPC &&
5233             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
5234                 if (block_rsv != global_rsv &&
5235                     !block_rsv_use_bytes(global_rsv, orig_bytes))
5236                         ret = 0;
5237         }
5238         if (ret == -ENOSPC) {
5239                 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
5240                                               block_rsv->space_info->flags,
5241                                               orig_bytes, 1);
5242
5243                 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
5244                         dump_space_info(fs_info, block_rsv->space_info,
5245                                         orig_bytes, 0);
5246         }
5247         return ret;
5248 }
5249
5250 static struct btrfs_block_rsv *get_block_rsv(
5251                                         const struct btrfs_trans_handle *trans,
5252                                         const struct btrfs_root *root)
5253 {
5254         struct btrfs_fs_info *fs_info = root->fs_info;
5255         struct btrfs_block_rsv *block_rsv = NULL;
5256
5257         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
5258             (root == fs_info->csum_root && trans->adding_csums) ||
5259             (root == fs_info->uuid_root))
5260                 block_rsv = trans->block_rsv;
5261
5262         if (!block_rsv)
5263                 block_rsv = root->block_rsv;
5264
5265         if (!block_rsv)
5266                 block_rsv = &fs_info->empty_block_rsv;
5267
5268         return block_rsv;
5269 }
5270
5271 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
5272                                u64 num_bytes)
5273 {
5274         int ret = -ENOSPC;
5275         spin_lock(&block_rsv->lock);
5276         if (block_rsv->reserved >= num_bytes) {
5277                 block_rsv->reserved -= num_bytes;
5278                 if (block_rsv->reserved < block_rsv->size)
5279                         block_rsv->full = 0;
5280                 ret = 0;
5281         }
5282         spin_unlock(&block_rsv->lock);
5283         return ret;
5284 }
5285
5286 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
5287                                 u64 num_bytes, int update_size)
5288 {
5289         spin_lock(&block_rsv->lock);
5290         block_rsv->reserved += num_bytes;
5291         if (update_size)
5292                 block_rsv->size += num_bytes;
5293         else if (block_rsv->reserved >= block_rsv->size)
5294                 block_rsv->full = 1;
5295         spin_unlock(&block_rsv->lock);
5296 }
5297
5298 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
5299                              struct btrfs_block_rsv *dest, u64 num_bytes,
5300                              int min_factor)
5301 {
5302         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5303         u64 min_bytes;
5304
5305         if (global_rsv->space_info != dest->space_info)
5306                 return -ENOSPC;
5307
5308         spin_lock(&global_rsv->lock);
5309         min_bytes = div_factor(global_rsv->size, min_factor);
5310         if (global_rsv->reserved < min_bytes + num_bytes) {
5311                 spin_unlock(&global_rsv->lock);
5312                 return -ENOSPC;
5313         }
5314         global_rsv->reserved -= num_bytes;
5315         if (global_rsv->reserved < global_rsv->size)
5316                 global_rsv->full = 0;
5317         spin_unlock(&global_rsv->lock);
5318
5319         block_rsv_add_bytes(dest, num_bytes, 1);
5320         return 0;
5321 }
5322
5323 /*
5324  * This is for space we already have accounted in space_info->bytes_may_use, so
5325  * basically when we're returning space from block_rsv's.
5326  */
5327 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
5328                                      struct btrfs_space_info *space_info,
5329                                      u64 num_bytes)
5330 {
5331         struct reserve_ticket *ticket;
5332         struct list_head *head;
5333         u64 used;
5334         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH;
5335         bool check_overcommit = false;
5336
5337         spin_lock(&space_info->lock);
5338         head = &space_info->priority_tickets;
5339
5340         /*
5341          * If we are over our limit then we need to check and see if we can
5342          * overcommit, and if we can't then we just need to free up our space
5343          * and not satisfy any requests.
5344          */
5345         used = btrfs_space_info_used(space_info, true);
5346         if (used - num_bytes >= space_info->total_bytes)
5347                 check_overcommit = true;
5348 again:
5349         while (!list_empty(head) && num_bytes) {
5350                 ticket = list_first_entry(head, struct reserve_ticket,
5351                                           list);
5352                 /*
5353                  * We use 0 bytes because this space is already reserved, so
5354                  * adding the ticket space would be a double count.
5355                  */
5356                 if (check_overcommit &&
5357                     !can_overcommit(fs_info, space_info, 0, flush, false))
5358                         break;
5359                 if (num_bytes >= ticket->bytes) {
5360                         list_del_init(&ticket->list);
5361                         num_bytes -= ticket->bytes;
5362                         ticket->bytes = 0;
5363                         space_info->tickets_id++;
5364                         wake_up(&ticket->wait);
5365                 } else {
5366                         ticket->bytes -= num_bytes;
5367                         num_bytes = 0;
5368                 }
5369         }
5370
5371         if (num_bytes && head == &space_info->priority_tickets) {
5372                 head = &space_info->tickets;
5373                 flush = BTRFS_RESERVE_FLUSH_ALL;
5374                 goto again;
5375         }
5376         space_info->bytes_may_use -= num_bytes;
5377         trace_btrfs_space_reservation(fs_info, "space_info",
5378                                       space_info->flags, num_bytes, 0);
5379         spin_unlock(&space_info->lock);
5380 }
5381
5382 /*
5383  * This is for newly allocated space that isn't accounted in
5384  * space_info->bytes_may_use yet.  So if we allocate a chunk or unpin an extent
5385  * we use this helper.
5386  */
5387 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
5388                                      struct btrfs_space_info *space_info,
5389                                      u64 num_bytes)
5390 {
5391         struct reserve_ticket *ticket;
5392         struct list_head *head = &space_info->priority_tickets;
5393
5394 again:
5395         while (!list_empty(head) && num_bytes) {
5396                 ticket = list_first_entry(head, struct reserve_ticket,
5397                                           list);
5398                 if (num_bytes >= ticket->bytes) {
5399                         trace_btrfs_space_reservation(fs_info, "space_info",
5400                                                       space_info->flags,
5401                                                       ticket->bytes, 1);
5402                         list_del_init(&ticket->list);
5403                         num_bytes -= ticket->bytes;
5404                         space_info->bytes_may_use += ticket->bytes;
5405                         ticket->bytes = 0;
5406                         space_info->tickets_id++;
5407                         wake_up(&ticket->wait);
5408                 } else {
5409                         trace_btrfs_space_reservation(fs_info, "space_info",
5410                                                       space_info->flags,
5411                                                       num_bytes, 1);
5412                         space_info->bytes_may_use += num_bytes;
5413                         ticket->bytes -= num_bytes;
5414                         num_bytes = 0;
5415                 }
5416         }
5417
5418         if (num_bytes && head == &space_info->priority_tickets) {
5419                 head = &space_info->tickets;
5420                 goto again;
5421         }
5422 }
5423
5424 static u64 block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
5425                                     struct btrfs_block_rsv *block_rsv,
5426                                     struct btrfs_block_rsv *dest, u64 num_bytes,
5427                                     u64 *qgroup_to_release_ret)
5428 {
5429         struct btrfs_space_info *space_info = block_rsv->space_info;
5430         u64 qgroup_to_release = 0;
5431         u64 ret;
5432
5433         spin_lock(&block_rsv->lock);
5434         if (num_bytes == (u64)-1) {
5435                 num_bytes = block_rsv->size;
5436                 qgroup_to_release = block_rsv->qgroup_rsv_size;
5437         }
5438         block_rsv->size -= num_bytes;
5439         if (block_rsv->reserved >= block_rsv->size) {
5440                 num_bytes = block_rsv->reserved - block_rsv->size;
5441                 block_rsv->reserved = block_rsv->size;
5442                 block_rsv->full = 1;
5443         } else {
5444                 num_bytes = 0;
5445         }
5446         if (block_rsv->qgroup_rsv_reserved >= block_rsv->qgroup_rsv_size) {
5447                 qgroup_to_release = block_rsv->qgroup_rsv_reserved -
5448                                     block_rsv->qgroup_rsv_size;
5449                 block_rsv->qgroup_rsv_reserved = block_rsv->qgroup_rsv_size;
5450         } else {
5451                 qgroup_to_release = 0;
5452         }
5453         spin_unlock(&block_rsv->lock);
5454
5455         ret = num_bytes;
5456         if (num_bytes > 0) {
5457                 if (dest) {
5458                         spin_lock(&dest->lock);
5459                         if (!dest->full) {
5460                                 u64 bytes_to_add;
5461
5462                                 bytes_to_add = dest->size - dest->reserved;
5463                                 bytes_to_add = min(num_bytes, bytes_to_add);
5464                                 dest->reserved += bytes_to_add;
5465                                 if (dest->reserved >= dest->size)
5466                                         dest->full = 1;
5467                                 num_bytes -= bytes_to_add;
5468                         }
5469                         spin_unlock(&dest->lock);
5470                 }
5471                 if (num_bytes)
5472                         space_info_add_old_bytes(fs_info, space_info,
5473                                                  num_bytes);
5474         }
5475         if (qgroup_to_release_ret)
5476                 *qgroup_to_release_ret = qgroup_to_release;
5477         return ret;
5478 }
5479
5480 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src,
5481                             struct btrfs_block_rsv *dst, u64 num_bytes,
5482                             int update_size)
5483 {
5484         int ret;
5485
5486         ret = block_rsv_use_bytes(src, num_bytes);
5487         if (ret)
5488                 return ret;
5489
5490         block_rsv_add_bytes(dst, num_bytes, update_size);
5491         return 0;
5492 }
5493
5494 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
5495 {
5496         memset(rsv, 0, sizeof(*rsv));
5497         spin_lock_init(&rsv->lock);
5498         rsv->type = type;
5499 }
5500
5501 void btrfs_init_metadata_block_rsv(struct btrfs_fs_info *fs_info,
5502                                    struct btrfs_block_rsv *rsv,
5503                                    unsigned short type)
5504 {
5505         btrfs_init_block_rsv(rsv, type);
5506         rsv->space_info = __find_space_info(fs_info,
5507                                             BTRFS_BLOCK_GROUP_METADATA);
5508 }
5509
5510 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_fs_info *fs_info,
5511                                               unsigned short type)
5512 {
5513         struct btrfs_block_rsv *block_rsv;
5514
5515         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
5516         if (!block_rsv)
5517                 return NULL;
5518
5519         btrfs_init_metadata_block_rsv(fs_info, block_rsv, type);
5520         return block_rsv;
5521 }
5522
5523 void btrfs_free_block_rsv(struct btrfs_fs_info *fs_info,
5524                           struct btrfs_block_rsv *rsv)
5525 {
5526         if (!rsv)
5527                 return;
5528         btrfs_block_rsv_release(fs_info, rsv, (u64)-1);
5529         kfree(rsv);
5530 }
5531
5532 int btrfs_block_rsv_add(struct btrfs_root *root,
5533                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
5534                         enum btrfs_reserve_flush_enum flush)
5535 {
5536         int ret;
5537
5538         if (num_bytes == 0)
5539                 return 0;
5540
5541         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5542         if (!ret) {
5543                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
5544                 return 0;
5545         }
5546
5547         return ret;
5548 }
5549
5550 int btrfs_block_rsv_check(struct btrfs_block_rsv *block_rsv, int min_factor)
5551 {
5552         u64 num_bytes = 0;
5553         int ret = -ENOSPC;
5554
5555         if (!block_rsv)
5556                 return 0;
5557
5558         spin_lock(&block_rsv->lock);
5559         num_bytes = div_factor(block_rsv->size, min_factor);
5560         if (block_rsv->reserved >= num_bytes)
5561                 ret = 0;
5562         spin_unlock(&block_rsv->lock);
5563
5564         return ret;
5565 }
5566
5567 int btrfs_block_rsv_refill(struct btrfs_root *root,
5568                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
5569                            enum btrfs_reserve_flush_enum flush)
5570 {
5571         u64 num_bytes = 0;
5572         int ret = -ENOSPC;
5573
5574         if (!block_rsv)
5575                 return 0;
5576
5577         spin_lock(&block_rsv->lock);
5578         num_bytes = min_reserved;
5579         if (block_rsv->reserved >= num_bytes)
5580                 ret = 0;
5581         else
5582                 num_bytes -= block_rsv->reserved;
5583         spin_unlock(&block_rsv->lock);
5584
5585         if (!ret)
5586                 return 0;
5587
5588         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5589         if (!ret) {
5590                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
5591                 return 0;
5592         }
5593
5594         return ret;
5595 }
5596
5597 /**
5598  * btrfs_inode_rsv_refill - refill the inode block rsv.
5599  * @inode - the inode we are refilling.
5600  * @flush - the flusing restriction.
5601  *
5602  * Essentially the same as btrfs_block_rsv_refill, except it uses the
5603  * block_rsv->size as the minimum size.  We'll either refill the missing amount
5604  * or return if we already have enough space.  This will also handle the resreve
5605  * tracepoint for the reserved amount.
5606  */
5607 static int btrfs_inode_rsv_refill(struct btrfs_inode *inode,
5608                                   enum btrfs_reserve_flush_enum flush)
5609 {
5610         struct btrfs_root *root = inode->root;
5611         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
5612         u64 num_bytes = 0;
5613         u64 qgroup_num_bytes = 0;
5614         int ret = -ENOSPC;
5615
5616         spin_lock(&block_rsv->lock);
5617         if (block_rsv->reserved < block_rsv->size)
5618                 num_bytes = block_rsv->size - block_rsv->reserved;
5619         if (block_rsv->qgroup_rsv_reserved < block_rsv->qgroup_rsv_size)
5620                 qgroup_num_bytes = block_rsv->qgroup_rsv_size -
5621                                    block_rsv->qgroup_rsv_reserved;
5622         spin_unlock(&block_rsv->lock);
5623
5624         if (num_bytes == 0)
5625                 return 0;
5626
5627         ret = btrfs_qgroup_reserve_meta_prealloc(root, qgroup_num_bytes, true);
5628         if (ret)
5629                 return ret;
5630         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5631         if (!ret) {
5632                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
5633                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5634                                               btrfs_ino(inode), num_bytes, 1);
5635
5636                 /* Don't forget to increase qgroup_rsv_reserved */
5637                 spin_lock(&block_rsv->lock);
5638                 block_rsv->qgroup_rsv_reserved += qgroup_num_bytes;
5639                 spin_unlock(&block_rsv->lock);
5640         } else
5641                 btrfs_qgroup_free_meta_prealloc(root, qgroup_num_bytes);
5642         return ret;
5643 }
5644
5645 /**
5646  * btrfs_inode_rsv_release - release any excessive reservation.
5647  * @inode - the inode we need to release from.
5648  * @qgroup_free - free or convert qgroup meta.
5649  *   Unlike normal operation, qgroup meta reservation needs to know if we are
5650  *   freeing qgroup reservation or just converting it into per-trans.  Normally
5651  *   @qgroup_free is true for error handling, and false for normal release.
5652  *
5653  * This is the same as btrfs_block_rsv_release, except that it handles the
5654  * tracepoint for the reservation.
5655  */
5656 static void btrfs_inode_rsv_release(struct btrfs_inode *inode, bool qgroup_free)
5657 {
5658         struct btrfs_fs_info *fs_info = inode->root->fs_info;
5659         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5660         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
5661         u64 released = 0;
5662         u64 qgroup_to_release = 0;
5663
5664         /*
5665          * Since we statically set the block_rsv->size we just want to say we
5666          * are releasing 0 bytes, and then we'll just get the reservation over
5667          * the size free'd.
5668          */
5669         released = block_rsv_release_bytes(fs_info, block_rsv, global_rsv, 0,
5670                                            &qgroup_to_release);
5671         if (released > 0)
5672                 trace_btrfs_space_reservation(fs_info, "delalloc",
5673                                               btrfs_ino(inode), released, 0);
5674         if (qgroup_free)
5675                 btrfs_qgroup_free_meta_prealloc(inode->root, qgroup_to_release);
5676         else
5677                 btrfs_qgroup_convert_reserved_meta(inode->root,
5678                                                    qgroup_to_release);
5679 }
5680
5681 void btrfs_block_rsv_release(struct btrfs_fs_info *fs_info,
5682                              struct btrfs_block_rsv *block_rsv,
5683                              u64 num_bytes)
5684 {
5685         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5686
5687         if (global_rsv == block_rsv ||
5688             block_rsv->space_info != global_rsv->space_info)
5689                 global_rsv = NULL;
5690         block_rsv_release_bytes(fs_info, block_rsv, global_rsv, num_bytes, NULL);
5691 }
5692
5693 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
5694 {
5695         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
5696         struct btrfs_space_info *sinfo = block_rsv->space_info;
5697         u64 num_bytes;
5698
5699         /*
5700          * The global block rsv is based on the size of the extent tree, the
5701          * checksum tree and the root tree.  If the fs is empty we want to set
5702          * it to a minimal amount for safety.
5703          */
5704         num_bytes = btrfs_root_used(&fs_info->extent_root->root_item) +
5705                 btrfs_root_used(&fs_info->csum_root->root_item) +
5706                 btrfs_root_used(&fs_info->tree_root->root_item);
5707         num_bytes = max_t(u64, num_bytes, SZ_16M);
5708
5709         spin_lock(&sinfo->lock);
5710         spin_lock(&block_rsv->lock);
5711
5712         block_rsv->size = min_t(u64, num_bytes, SZ_512M);
5713
5714         if (block_rsv->reserved < block_rsv->size) {
5715                 num_bytes = btrfs_space_info_used(sinfo, true);
5716                 if (sinfo->total_bytes > num_bytes) {
5717                         num_bytes = sinfo->total_bytes - num_bytes;
5718                         num_bytes = min(num_bytes,
5719                                         block_rsv->size - block_rsv->reserved);
5720                         block_rsv->reserved += num_bytes;
5721                         sinfo->bytes_may_use += num_bytes;
5722                         trace_btrfs_space_reservation(fs_info, "space_info",
5723                                                       sinfo->flags, num_bytes,
5724                                                       1);
5725                 }
5726         } else if (block_rsv->reserved > block_rsv->size) {
5727                 num_bytes = block_rsv->reserved - block_rsv->size;
5728                 sinfo->bytes_may_use -= num_bytes;
5729                 trace_btrfs_space_reservation(fs_info, "space_info",
5730                                       sinfo->flags, num_bytes, 0);
5731                 block_rsv->reserved = block_rsv->size;
5732         }
5733
5734         if (block_rsv->reserved == block_rsv->size)
5735                 block_rsv->full = 1;
5736         else
5737                 block_rsv->full = 0;
5738
5739         spin_unlock(&block_rsv->lock);
5740         spin_unlock(&sinfo->lock);
5741 }
5742
5743 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
5744 {
5745         struct btrfs_space_info *space_info;
5746
5747         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
5748         fs_info->chunk_block_rsv.space_info = space_info;
5749
5750         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5751         fs_info->global_block_rsv.space_info = space_info;
5752         fs_info->trans_block_rsv.space_info = space_info;
5753         fs_info->empty_block_rsv.space_info = space_info;
5754         fs_info->delayed_block_rsv.space_info = space_info;
5755
5756         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
5757         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
5758         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
5759         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
5760         if (fs_info->quota_root)
5761                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
5762         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
5763
5764         update_global_block_rsv(fs_info);
5765 }
5766
5767 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
5768 {
5769         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
5770                                 (u64)-1, NULL);
5771         WARN_ON(fs_info->trans_block_rsv.size > 0);
5772         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
5773         WARN_ON(fs_info->chunk_block_rsv.size > 0);
5774         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
5775         WARN_ON(fs_info->delayed_block_rsv.size > 0);
5776         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
5777 }
5778
5779
5780 /*
5781  * To be called after all the new block groups attached to the transaction
5782  * handle have been created (btrfs_create_pending_block_groups()).
5783  */
5784 void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans)
5785 {
5786         struct btrfs_fs_info *fs_info = trans->fs_info;
5787
5788         if (!trans->chunk_bytes_reserved)
5789                 return;
5790
5791         WARN_ON_ONCE(!list_empty(&trans->new_bgs));
5792
5793         block_rsv_release_bytes(fs_info, &fs_info->chunk_block_rsv, NULL,
5794                                 trans->chunk_bytes_reserved, NULL);
5795         trans->chunk_bytes_reserved = 0;
5796 }
5797
5798 /*
5799  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
5800  * root: the root of the parent directory
5801  * rsv: block reservation
5802  * items: the number of items that we need do reservation
5803  * use_global_rsv: allow fallback to the global block reservation
5804  *
5805  * This function is used to reserve the space for snapshot/subvolume
5806  * creation and deletion. Those operations are different with the
5807  * common file/directory operations, they change two fs/file trees
5808  * and root tree, the number of items that the qgroup reserves is
5809  * different with the free space reservation. So we can not use
5810  * the space reservation mechanism in start_transaction().
5811  */
5812 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
5813                                      struct btrfs_block_rsv *rsv, int items,
5814                                      bool use_global_rsv)
5815 {
5816         u64 qgroup_num_bytes = 0;
5817         u64 num_bytes;
5818         int ret;
5819         struct btrfs_fs_info *fs_info = root->fs_info;
5820         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5821
5822         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)) {
5823                 /* One for parent inode, two for dir entries */
5824                 qgroup_num_bytes = 3 * fs_info->nodesize;
5825                 ret = btrfs_qgroup_reserve_meta_prealloc(root,
5826                                 qgroup_num_bytes, true);
5827                 if (ret)
5828                         return ret;
5829         }
5830
5831         num_bytes = btrfs_calc_trans_metadata_size(fs_info, items);
5832         rsv->space_info = __find_space_info(fs_info,
5833                                             BTRFS_BLOCK_GROUP_METADATA);
5834         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
5835                                   BTRFS_RESERVE_FLUSH_ALL);
5836
5837         if (ret == -ENOSPC && use_global_rsv)
5838                 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes, 1);
5839
5840         if (ret && qgroup_num_bytes)
5841                 btrfs_qgroup_free_meta_prealloc(root, qgroup_num_bytes);
5842
5843         return ret;
5844 }
5845
5846 void btrfs_subvolume_release_metadata(struct btrfs_fs_info *fs_info,
5847                                       struct btrfs_block_rsv *rsv)
5848 {
5849         btrfs_block_rsv_release(fs_info, rsv, (u64)-1);
5850 }
5851
5852 static void btrfs_calculate_inode_block_rsv_size(struct btrfs_fs_info *fs_info,
5853                                                  struct btrfs_inode *inode)
5854 {
5855         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
5856         u64 reserve_size = 0;
5857         u64 qgroup_rsv_size = 0;
5858         u64 csum_leaves;
5859         unsigned outstanding_extents;
5860
5861         lockdep_assert_held(&inode->lock);
5862         outstanding_extents = inode->outstanding_extents;
5863         if (outstanding_extents)
5864                 reserve_size = btrfs_calc_trans_metadata_size(fs_info,
5865                                                 outstanding_extents + 1);
5866         csum_leaves = btrfs_csum_bytes_to_leaves(fs_info,
5867                                                  inode->csum_bytes);
5868         reserve_size += btrfs_calc_trans_metadata_size(fs_info,
5869                                                        csum_leaves);
5870         /*
5871          * For qgroup rsv, the calculation is very simple:
5872          * account one nodesize for each outstanding extent
5873          *
5874          * This is overestimating in most cases.
5875          */
5876         qgroup_rsv_size = outstanding_extents * fs_info->nodesize;
5877
5878         spin_lock(&block_rsv->lock);
5879         block_rsv->size = reserve_size;
5880         block_rsv->qgroup_rsv_size = qgroup_rsv_size;
5881         spin_unlock(&block_rsv->lock);
5882 }
5883
5884 int btrfs_delalloc_reserve_metadata(struct btrfs_inode *inode, u64 num_bytes)
5885 {
5886         struct btrfs_fs_info *fs_info = inode->root->fs_info;
5887         unsigned nr_extents;
5888         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
5889         int ret = 0;
5890         bool delalloc_lock = true;
5891
5892         /* If we are a free space inode we need to not flush since we will be in
5893          * the middle of a transaction commit.  We also don't need the delalloc
5894          * mutex since we won't race with anybody.  We need this mostly to make
5895          * lockdep shut its filthy mouth.
5896          *
5897          * If we have a transaction open (can happen if we call truncate_block
5898          * from truncate), then we need FLUSH_LIMIT so we don't deadlock.
5899          */
5900         if (btrfs_is_free_space_inode(inode)) {
5901                 flush = BTRFS_RESERVE_NO_FLUSH;
5902                 delalloc_lock = false;
5903         } else {
5904                 if (current->journal_info)
5905                         flush = BTRFS_RESERVE_FLUSH_LIMIT;
5906
5907                 if (btrfs_transaction_in_commit(fs_info))
5908                         schedule_timeout(1);
5909         }
5910
5911         if (delalloc_lock)
5912                 mutex_lock(&inode->delalloc_mutex);
5913
5914         num_bytes = ALIGN(num_bytes, fs_info->sectorsize);
5915
5916         /* Add our new extents and calculate the new rsv size. */
5917         spin_lock(&inode->lock);
5918         nr_extents = count_max_extents(num_bytes);
5919         btrfs_mod_outstanding_extents(inode, nr_extents);
5920         inode->csum_bytes += num_bytes;
5921         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
5922         spin_unlock(&inode->lock);
5923
5924         ret = btrfs_inode_rsv_refill(inode, flush);
5925         if (unlikely(ret))
5926                 goto out_fail;
5927
5928         if (delalloc_lock)
5929                 mutex_unlock(&inode->delalloc_mutex);
5930         return 0;
5931
5932 out_fail:
5933         spin_lock(&inode->lock);
5934         nr_extents = count_max_extents(num_bytes);
5935         btrfs_mod_outstanding_extents(inode, -nr_extents);
5936         inode->csum_bytes -= num_bytes;
5937         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
5938         spin_unlock(&inode->lock);
5939
5940         btrfs_inode_rsv_release(inode, true);
5941         if (delalloc_lock)
5942                 mutex_unlock(&inode->delalloc_mutex);
5943         return ret;
5944 }
5945
5946 /**
5947  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
5948  * @inode: the inode to release the reservation for.
5949  * @num_bytes: the number of bytes we are releasing.
5950  * @qgroup_free: free qgroup reservation or convert it to per-trans reservation
5951  *
5952  * This will release the metadata reservation for an inode.  This can be called
5953  * once we complete IO for a given set of bytes to release their metadata
5954  * reservations, or on error for the same reason.
5955  */
5956 void btrfs_delalloc_release_metadata(struct btrfs_inode *inode, u64 num_bytes,
5957                                      bool qgroup_free)
5958 {
5959         struct btrfs_fs_info *fs_info = inode->root->fs_info;
5960
5961         num_bytes = ALIGN(num_bytes, fs_info->sectorsize);
5962         spin_lock(&inode->lock);
5963         inode->csum_bytes -= num_bytes;
5964         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
5965         spin_unlock(&inode->lock);
5966
5967         if (btrfs_is_testing(fs_info))
5968                 return;
5969
5970         btrfs_inode_rsv_release(inode, qgroup_free);
5971 }
5972
5973 /**
5974  * btrfs_delalloc_release_extents - release our outstanding_extents
5975  * @inode: the inode to balance the reservation for.
5976  * @num_bytes: the number of bytes we originally reserved with
5977  * @qgroup_free: do we need to free qgroup meta reservation or convert them.
5978  *
5979  * When we reserve space we increase outstanding_extents for the extents we may
5980  * add.  Once we've set the range as delalloc or created our ordered extents we
5981  * have outstanding_extents to track the real usage, so we use this to free our
5982  * temporarily tracked outstanding_extents.  This _must_ be used in conjunction
5983  * with btrfs_delalloc_reserve_metadata.
5984  */
5985 void btrfs_delalloc_release_extents(struct btrfs_inode *inode, u64 num_bytes,
5986                                     bool qgroup_free)
5987 {
5988         struct btrfs_fs_info *fs_info = inode->root->fs_info;
5989         unsigned num_extents;
5990
5991         spin_lock(&inode->lock);
5992         num_extents = count_max_extents(num_bytes);
5993         btrfs_mod_outstanding_extents(inode, -num_extents);
5994         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
5995         spin_unlock(&inode->lock);
5996
5997         if (btrfs_is_testing(fs_info))
5998                 return;
5999
6000         btrfs_inode_rsv_release(inode, qgroup_free);
6001 }
6002
6003 /**
6004  * btrfs_delalloc_reserve_space - reserve data and metadata space for
6005  * delalloc
6006  * @inode: inode we're writing to
6007  * @start: start range we are writing to
6008  * @len: how long the range we are writing to
6009  * @reserved: mandatory parameter, record actually reserved qgroup ranges of
6010  *            current reservation.
6011  *
6012  * This will do the following things
6013  *
6014  * o reserve space in data space info for num bytes
6015  *   and reserve precious corresponding qgroup space
6016  *   (Done in check_data_free_space)
6017  *
6018  * o reserve space for metadata space, based on the number of outstanding
6019  *   extents and how much csums will be needed
6020  *   also reserve metadata space in a per root over-reserve method.
6021  * o add to the inodes->delalloc_bytes
6022  * o add it to the fs_info's delalloc inodes list.
6023  *   (Above 3 all done in delalloc_reserve_metadata)
6024  *
6025  * Return 0 for success
6026  * Return <0 for error(-ENOSPC or -EQUOT)
6027  */
6028 int btrfs_delalloc_reserve_space(struct inode *inode,
6029                         struct extent_changeset **reserved, u64 start, u64 len)
6030 {
6031         int ret;
6032
6033         ret = btrfs_check_data_free_space(inode, reserved, start, len);
6034         if (ret < 0)
6035                 return ret;
6036         ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode), len);
6037         if (ret < 0)
6038                 btrfs_free_reserved_data_space(inode, *reserved, start, len);
6039         return ret;
6040 }
6041
6042 /**
6043  * btrfs_delalloc_release_space - release data and metadata space for delalloc
6044  * @inode: inode we're releasing space for
6045  * @start: start position of the space already reserved
6046  * @len: the len of the space already reserved
6047  * @release_bytes: the len of the space we consumed or didn't use
6048  *
6049  * This function will release the metadata space that was not used and will
6050  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
6051  * list if there are no delalloc bytes left.
6052  * Also it will handle the qgroup reserved space.
6053  */
6054 void btrfs_delalloc_release_space(struct inode *inode,
6055                                   struct extent_changeset *reserved,
6056                                   u64 start, u64 len, bool qgroup_free)
6057 {
6058         btrfs_delalloc_release_metadata(BTRFS_I(inode), len, qgroup_free);
6059         btrfs_free_reserved_data_space(inode, reserved, start, len);
6060 }
6061
6062 static int update_block_group(struct btrfs_trans_handle *trans,
6063                               struct btrfs_fs_info *info, u64 bytenr,
6064                               u64 num_bytes, int alloc)
6065 {
6066         struct btrfs_block_group_cache *cache = NULL;
6067         u64 total = num_bytes;
6068         u64 old_val;
6069         u64 byte_in_group;
6070         int factor;
6071
6072         /* block accounting for super block */
6073         spin_lock(&info->delalloc_root_lock);
6074         old_val = btrfs_super_bytes_used(info->super_copy);
6075         if (alloc)
6076                 old_val += num_bytes;
6077         else
6078                 old_val -= num_bytes;
6079         btrfs_set_super_bytes_used(info->super_copy, old_val);
6080         spin_unlock(&info->delalloc_root_lock);
6081
6082         while (total) {
6083                 cache = btrfs_lookup_block_group(info, bytenr);
6084                 if (!cache)
6085                         return -ENOENT;
6086                 factor = btrfs_bg_type_to_factor(cache->flags);
6087
6088                 /*
6089                  * If this block group has free space cache written out, we
6090                  * need to make sure to load it if we are removing space.  This
6091                  * is because we need the unpinning stage to actually add the
6092                  * space back to the block group, otherwise we will leak space.
6093                  */
6094                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
6095                         cache_block_group(cache, 1);
6096
6097                 byte_in_group = bytenr - cache->key.objectid;
6098                 WARN_ON(byte_in_group > cache->key.offset);
6099
6100                 spin_lock(&cache->space_info->lock);
6101                 spin_lock(&cache->lock);
6102
6103                 if (btrfs_test_opt(info, SPACE_CACHE) &&
6104                     cache->disk_cache_state < BTRFS_DC_CLEAR)
6105                         cache->disk_cache_state = BTRFS_DC_CLEAR;
6106
6107                 old_val = btrfs_block_group_used(&cache->item);
6108                 num_bytes = min(total, cache->key.offset - byte_in_group);
6109                 if (alloc) {
6110                         old_val += num_bytes;
6111                         btrfs_set_block_group_used(&cache->item, old_val);
6112                         cache->reserved -= num_bytes;
6113                         cache->space_info->bytes_reserved -= num_bytes;
6114                         cache->space_info->bytes_used += num_bytes;
6115                         cache->space_info->disk_used += num_bytes * factor;
6116                         spin_unlock(&cache->lock);
6117                         spin_unlock(&cache->space_info->lock);
6118                 } else {
6119                         old_val -= num_bytes;
6120                         btrfs_set_block_group_used(&cache->item, old_val);
6121                         cache->pinned += num_bytes;
6122                         cache->space_info->bytes_pinned += num_bytes;
6123                         cache->space_info->bytes_used -= num_bytes;
6124                         cache->space_info->disk_used -= num_bytes * factor;
6125                         spin_unlock(&cache->lock);
6126                         spin_unlock(&cache->space_info->lock);
6127
6128                         trace_btrfs_space_reservation(info, "pinned",
6129                                                       cache->space_info->flags,
6130                                                       num_bytes, 1);
6131                         percpu_counter_add_batch(&cache->space_info->total_bytes_pinned,
6132                                            num_bytes,
6133                                            BTRFS_TOTAL_BYTES_PINNED_BATCH);
6134                         set_extent_dirty(info->pinned_extents,
6135                                          bytenr, bytenr + num_bytes - 1,
6136                                          GFP_NOFS | __GFP_NOFAIL);
6137                 }
6138
6139                 spin_lock(&trans->transaction->dirty_bgs_lock);
6140                 if (list_empty(&cache->dirty_list)) {
6141                         list_add_tail(&cache->dirty_list,
6142                                       &trans->transaction->dirty_bgs);
6143                         trans->transaction->num_dirty_bgs++;
6144                         btrfs_get_block_group(cache);
6145                 }
6146                 spin_unlock(&trans->transaction->dirty_bgs_lock);
6147
6148                 /*
6149                  * No longer have used bytes in this block group, queue it for
6150                  * deletion. We do this after adding the block group to the
6151                  * dirty list to avoid races between cleaner kthread and space
6152                  * cache writeout.
6153                  */
6154                 if (!alloc && old_val == 0)
6155                         btrfs_mark_bg_unused(cache);
6156
6157                 btrfs_put_block_group(cache);
6158                 total -= num_bytes;
6159                 bytenr += num_bytes;
6160         }
6161         return 0;
6162 }
6163
6164 static u64 first_logical_byte(struct btrfs_fs_info *fs_info, u64 search_start)
6165 {
6166         struct btrfs_block_group_cache *cache;
6167         u64 bytenr;
6168
6169         spin_lock(&fs_info->block_group_cache_lock);
6170         bytenr = fs_info->first_logical_byte;
6171         spin_unlock(&fs_info->block_group_cache_lock);
6172
6173         if (bytenr < (u64)-1)
6174                 return bytenr;
6175
6176         cache = btrfs_lookup_first_block_group(fs_info, search_start);
6177         if (!cache)
6178                 return 0;
6179
6180         bytenr = cache->key.objectid;
6181         btrfs_put_block_group(cache);
6182
6183         return bytenr;
6184 }
6185
6186 static int pin_down_extent(struct btrfs_fs_info *fs_info,
6187                            struct btrfs_block_group_cache *cache,
6188                            u64 bytenr, u64 num_bytes, int reserved)
6189 {
6190         spin_lock(&cache->space_info->lock);
6191         spin_lock(&cache->lock);
6192         cache->pinned += num_bytes;
6193         cache->space_info->bytes_pinned += num_bytes;
6194         if (reserved) {
6195                 cache->reserved -= num_bytes;
6196                 cache->space_info->bytes_reserved -= num_bytes;
6197         }
6198         spin_unlock(&cache->lock);
6199         spin_unlock(&cache->space_info->lock);
6200
6201         trace_btrfs_space_reservation(fs_info, "pinned",
6202                                       cache->space_info->flags, num_bytes, 1);
6203         percpu_counter_add_batch(&cache->space_info->total_bytes_pinned,
6204                     num_bytes, BTRFS_TOTAL_BYTES_PINNED_BATCH);
6205         set_extent_dirty(fs_info->pinned_extents, bytenr,
6206                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
6207         return 0;
6208 }
6209
6210 /*
6211  * this function must be called within transaction
6212  */
6213 int btrfs_pin_extent(struct btrfs_fs_info *fs_info,
6214                      u64 bytenr, u64 num_bytes, int reserved)
6215 {
6216         struct btrfs_block_group_cache *cache;
6217
6218         cache = btrfs_lookup_block_group(fs_info, bytenr);
6219         BUG_ON(!cache); /* Logic error */
6220
6221         pin_down_extent(fs_info, cache, bytenr, num_bytes, reserved);
6222
6223         btrfs_put_block_group(cache);
6224         return 0;
6225 }
6226
6227 /*
6228  * this function must be called within transaction
6229  */
6230 int btrfs_pin_extent_for_log_replay(struct btrfs_fs_info *fs_info,
6231                                     u64 bytenr, u64 num_bytes)
6232 {
6233         struct btrfs_block_group_cache *cache;
6234         int ret;
6235
6236         cache = btrfs_lookup_block_group(fs_info, bytenr);
6237         if (!cache)
6238                 return -EINVAL;
6239
6240         /*
6241          * pull in the free space cache (if any) so that our pin
6242          * removes the free space from the cache.  We have load_only set
6243          * to one because the slow code to read in the free extents does check
6244          * the pinned extents.
6245          */
6246         cache_block_group(cache, 1);
6247
6248         pin_down_extent(fs_info, cache, bytenr, num_bytes, 0);
6249
6250         /* remove us from the free space cache (if we're there at all) */
6251         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
6252         btrfs_put_block_group(cache);
6253         return ret;
6254 }
6255
6256 static int __exclude_logged_extent(struct btrfs_fs_info *fs_info,
6257                                    u64 start, u64 num_bytes)
6258 {
6259         int ret;
6260         struct btrfs_block_group_cache *block_group;
6261         struct btrfs_caching_control *caching_ctl;
6262
6263         block_group = btrfs_lookup_block_group(fs_info, start);
6264         if (!block_group)
6265                 return -EINVAL;
6266
6267         cache_block_group(block_group, 0);
6268         caching_ctl = get_caching_control(block_group);
6269
6270         if (!caching_ctl) {
6271                 /* Logic error */
6272                 BUG_ON(!block_group_cache_done(block_group));
6273                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6274         } else {
6275                 mutex_lock(&caching_ctl->mutex);
6276
6277                 if (start >= caching_ctl->progress) {
6278                         ret = add_excluded_extent(fs_info, start, num_bytes);
6279                 } else if (start + num_bytes <= caching_ctl->progress) {
6280                         ret = btrfs_remove_free_space(block_group,
6281                                                       start, num_bytes);
6282                 } else {
6283                         num_bytes = caching_ctl->progress - start;
6284                         ret = btrfs_remove_free_space(block_group,
6285                                                       start, num_bytes);
6286                         if (ret)
6287                                 goto out_lock;
6288
6289                         num_bytes = (start + num_bytes) -
6290                                 caching_ctl->progress;
6291                         start = caching_ctl->progress;
6292                         ret = add_excluded_extent(fs_info, start, num_bytes);
6293                 }
6294 out_lock:
6295                 mutex_unlock(&caching_ctl->mutex);
6296                 put_caching_control(caching_ctl);
6297         }
6298         btrfs_put_block_group(block_group);
6299         return ret;
6300 }
6301
6302 int btrfs_exclude_logged_extents(struct btrfs_fs_info *fs_info,
6303                                  struct extent_buffer *eb)
6304 {
6305         struct btrfs_file_extent_item *item;
6306         struct btrfs_key key;
6307         int found_type;
6308         int i;
6309         int ret = 0;
6310
6311         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS))
6312                 return 0;
6313
6314         for (i = 0; i < btrfs_header_nritems(eb); i++) {
6315                 btrfs_item_key_to_cpu(eb, &key, i);
6316                 if (key.type != BTRFS_EXTENT_DATA_KEY)
6317                         continue;
6318                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
6319                 found_type = btrfs_file_extent_type(eb, item);
6320                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
6321                         continue;
6322                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
6323                         continue;
6324                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
6325                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
6326                 ret = __exclude_logged_extent(fs_info, key.objectid, key.offset);
6327                 if (ret)
6328                         break;
6329         }
6330
6331         return ret;
6332 }
6333
6334 static void
6335 btrfs_inc_block_group_reservations(struct btrfs_block_group_cache *bg)
6336 {
6337         atomic_inc(&bg->reservations);
6338 }
6339
6340 void btrfs_dec_block_group_reservations(struct btrfs_fs_info *fs_info,
6341                                         const u64 start)
6342 {
6343         struct btrfs_block_group_cache *bg;
6344
6345         bg = btrfs_lookup_block_group(fs_info, start);
6346         ASSERT(bg);
6347         if (atomic_dec_and_test(&bg->reservations))
6348                 wake_up_var(&bg->reservations);
6349         btrfs_put_block_group(bg);
6350 }
6351
6352 void btrfs_wait_block_group_reservations(struct btrfs_block_group_cache *bg)
6353 {
6354         struct btrfs_space_info *space_info = bg->space_info;
6355
6356         ASSERT(bg->ro);
6357
6358         if (!(bg->flags & BTRFS_BLOCK_GROUP_DATA))
6359                 return;
6360
6361         /*
6362          * Our block group is read only but before we set it to read only,
6363          * some task might have had allocated an extent from it already, but it
6364          * has not yet created a respective ordered extent (and added it to a
6365          * root's list of ordered extents).
6366          * Therefore wait for any task currently allocating extents, since the
6367          * block group's reservations counter is incremented while a read lock
6368          * on the groups' semaphore is held and decremented after releasing
6369          * the read access on that semaphore and creating the ordered extent.
6370          */
6371         down_write(&space_info->groups_sem);
6372         up_write(&space_info->groups_sem);
6373
6374         wait_var_event(&bg->reservations, !atomic_read(&bg->reservations));
6375 }
6376
6377 /**
6378  * btrfs_add_reserved_bytes - update the block_group and space info counters
6379  * @cache:      The cache we are manipulating
6380  * @ram_bytes:  The number of bytes of file content, and will be same to
6381  *              @num_bytes except for the compress path.
6382  * @num_bytes:  The number of bytes in question
6383  * @delalloc:   The blocks are allocated for the delalloc write
6384  *
6385  * This is called by the allocator when it reserves space. If this is a
6386  * reservation and the block group has become read only we cannot make the
6387  * reservation and return -EAGAIN, otherwise this function always succeeds.
6388  */
6389 static int btrfs_add_reserved_bytes(struct btrfs_block_group_cache *cache,
6390                                     u64 ram_bytes, u64 num_bytes, int delalloc)
6391 {
6392         struct btrfs_space_info *space_info = cache->space_info;
6393         int ret = 0;
6394
6395         spin_lock(&space_info->lock);
6396         spin_lock(&cache->lock);
6397         if (cache->ro) {
6398                 ret = -EAGAIN;
6399         } else {
6400                 cache->reserved += num_bytes;
6401                 space_info->bytes_reserved += num_bytes;
6402
6403                 trace_btrfs_space_reservation(cache->fs_info,
6404                                 "space_info", space_info->flags,
6405                                 ram_bytes, 0);
6406                 space_info->bytes_may_use -= ram_bytes;
6407                 if (delalloc)
6408                         cache->delalloc_bytes += num_bytes;
6409         }
6410         spin_unlock(&cache->lock);
6411         spin_unlock(&space_info->lock);
6412         return ret;
6413 }
6414
6415 /**
6416  * btrfs_free_reserved_bytes - update the block_group and space info counters
6417  * @cache:      The cache we are manipulating
6418  * @num_bytes:  The number of bytes in question
6419  * @delalloc:   The blocks are allocated for the delalloc write
6420  *
6421  * This is called by somebody who is freeing space that was never actually used
6422  * on disk.  For example if you reserve some space for a new leaf in transaction
6423  * A and before transaction A commits you free that leaf, you call this with
6424  * reserve set to 0 in order to clear the reservation.
6425  */
6426
6427 static int btrfs_free_reserved_bytes(struct btrfs_block_group_cache *cache,
6428                                      u64 num_bytes, int delalloc)
6429 {
6430         struct btrfs_space_info *space_info = cache->space_info;
6431         int ret = 0;
6432
6433         spin_lock(&space_info->lock);
6434         spin_lock(&cache->lock);
6435         if (cache->ro)
6436                 space_info->bytes_readonly += num_bytes;
6437         cache->reserved -= num_bytes;
6438         space_info->bytes_reserved -= num_bytes;
6439
6440         if (delalloc)
6441                 cache->delalloc_bytes -= num_bytes;
6442         spin_unlock(&cache->lock);
6443         spin_unlock(&space_info->lock);
6444         return ret;
6445 }
6446 void btrfs_prepare_extent_commit(struct btrfs_fs_info *fs_info)
6447 {
6448         struct btrfs_caching_control *next;
6449         struct btrfs_caching_control *caching_ctl;
6450         struct btrfs_block_group_cache *cache;
6451
6452         down_write(&fs_info->commit_root_sem);
6453
6454         list_for_each_entry_safe(caching_ctl, next,
6455                                  &fs_info->caching_block_groups, list) {
6456                 cache = caching_ctl->block_group;
6457                 if (block_group_cache_done(cache)) {
6458                         cache->last_byte_to_unpin = (u64)-1;
6459                         list_del_init(&caching_ctl->list);
6460                         put_caching_control(caching_ctl);
6461                 } else {
6462                         cache->last_byte_to_unpin = caching_ctl->progress;
6463                 }
6464         }
6465
6466         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6467                 fs_info->pinned_extents = &fs_info->freed_extents[1];
6468         else
6469                 fs_info->pinned_extents = &fs_info->freed_extents[0];
6470
6471         up_write(&fs_info->commit_root_sem);
6472
6473         update_global_block_rsv(fs_info);
6474 }
6475
6476 /*
6477  * Returns the free cluster for the given space info and sets empty_cluster to
6478  * what it should be based on the mount options.
6479  */
6480 static struct btrfs_free_cluster *
6481 fetch_cluster_info(struct btrfs_fs_info *fs_info,
6482                    struct btrfs_space_info *space_info, u64 *empty_cluster)
6483 {
6484         struct btrfs_free_cluster *ret = NULL;
6485
6486         *empty_cluster = 0;
6487         if (btrfs_mixed_space_info(space_info))
6488                 return ret;
6489
6490         if (space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
6491                 ret = &fs_info->meta_alloc_cluster;
6492                 if (btrfs_test_opt(fs_info, SSD))
6493                         *empty_cluster = SZ_2M;
6494                 else
6495                         *empty_cluster = SZ_64K;
6496         } else if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) &&
6497                    btrfs_test_opt(fs_info, SSD_SPREAD)) {
6498                 *empty_cluster = SZ_2M;
6499                 ret = &fs_info->data_alloc_cluster;
6500         }
6501
6502         return ret;
6503 }
6504
6505 static int unpin_extent_range(struct btrfs_fs_info *fs_info,
6506                               u64 start, u64 end,
6507                               const bool return_free_space)
6508 {
6509         struct btrfs_block_group_cache *cache = NULL;
6510         struct btrfs_space_info *space_info;
6511         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
6512         struct btrfs_free_cluster *cluster = NULL;
6513         u64 len;
6514         u64 total_unpinned = 0;
6515         u64 empty_cluster = 0;
6516         bool readonly;
6517
6518         while (start <= end) {
6519                 readonly = false;
6520                 if (!cache ||
6521                     start >= cache->key.objectid + cache->key.offset) {
6522                         if (cache)
6523                                 btrfs_put_block_group(cache);
6524                         total_unpinned = 0;
6525                         cache = btrfs_lookup_block_group(fs_info, start);
6526                         BUG_ON(!cache); /* Logic error */
6527
6528                         cluster = fetch_cluster_info(fs_info,
6529                                                      cache->space_info,
6530                                                      &empty_cluster);
6531                         empty_cluster <<= 1;
6532                 }
6533
6534                 len = cache->key.objectid + cache->key.offset - start;
6535                 len = min(len, end + 1 - start);
6536
6537                 if (start < cache->last_byte_to_unpin) {
6538                         len = min(len, cache->last_byte_to_unpin - start);
6539                         if (return_free_space)
6540                                 btrfs_add_free_space(cache, start, len);
6541                 }
6542
6543                 start += len;
6544                 total_unpinned += len;
6545                 space_info = cache->space_info;
6546
6547                 /*
6548                  * If this space cluster has been marked as fragmented and we've
6549                  * unpinned enough in this block group to potentially allow a
6550                  * cluster to be created inside of it go ahead and clear the
6551                  * fragmented check.
6552                  */
6553                 if (cluster && cluster->fragmented &&
6554                     total_unpinned > empty_cluster) {
6555                         spin_lock(&cluster->lock);
6556                         cluster->fragmented = 0;
6557                         spin_unlock(&cluster->lock);
6558                 }
6559
6560                 spin_lock(&space_info->lock);
6561                 spin_lock(&cache->lock);
6562                 cache->pinned -= len;
6563                 space_info->bytes_pinned -= len;
6564
6565                 trace_btrfs_space_reservation(fs_info, "pinned",
6566                                               space_info->flags, len, 0);
6567                 space_info->max_extent_size = 0;
6568                 percpu_counter_add_batch(&space_info->total_bytes_pinned,
6569                             -len, BTRFS_TOTAL_BYTES_PINNED_BATCH);
6570                 if (cache->ro) {
6571                         space_info->bytes_readonly += len;
6572                         readonly = true;
6573                 }
6574                 spin_unlock(&cache->lock);
6575                 if (!readonly && return_free_space &&
6576                     global_rsv->space_info == space_info) {
6577                         u64 to_add = len;
6578
6579                         spin_lock(&global_rsv->lock);
6580                         if (!global_rsv->full) {
6581                                 to_add = min(len, global_rsv->size -
6582                                              global_rsv->reserved);
6583                                 global_rsv->reserved += to_add;
6584                                 space_info->bytes_may_use += to_add;
6585                                 if (global_rsv->reserved >= global_rsv->size)
6586                                         global_rsv->full = 1;
6587                                 trace_btrfs_space_reservation(fs_info,
6588                                                               "space_info",
6589                                                               space_info->flags,
6590                                                               to_add, 1);
6591                                 len -= to_add;
6592                         }
6593                         spin_unlock(&global_rsv->lock);
6594                         /* Add to any tickets we may have */
6595                         if (len)
6596                                 space_info_add_new_bytes(fs_info, space_info,
6597                                                          len);
6598                 }
6599                 spin_unlock(&space_info->lock);
6600         }
6601
6602         if (cache)
6603                 btrfs_put_block_group(cache);
6604         return 0;
6605 }
6606
6607 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans)
6608 {
6609         struct btrfs_fs_info *fs_info = trans->fs_info;
6610         struct btrfs_block_group_cache *block_group, *tmp;
6611         struct list_head *deleted_bgs;
6612         struct extent_io_tree *unpin;
6613         u64 start;
6614         u64 end;
6615         int ret;
6616
6617         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6618                 unpin = &fs_info->freed_extents[1];
6619         else
6620                 unpin = &fs_info->freed_extents[0];
6621
6622         while (!trans->aborted) {
6623                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
6624                 ret = find_first_extent_bit(unpin, 0, &start, &end,
6625                                             EXTENT_DIRTY, NULL);
6626                 if (ret) {
6627                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6628                         break;
6629                 }
6630
6631                 if (btrfs_test_opt(fs_info, DISCARD))
6632                         ret = btrfs_discard_extent(fs_info, start,
6633                                                    end + 1 - start, NULL);
6634
6635                 clear_extent_dirty(unpin, start, end);
6636                 unpin_extent_range(fs_info, start, end, true);
6637                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6638                 cond_resched();
6639         }
6640
6641         /*
6642          * Transaction is finished.  We don't need the lock anymore.  We
6643          * do need to clean up the block groups in case of a transaction
6644          * abort.
6645          */
6646         deleted_bgs = &trans->transaction->deleted_bgs;
6647         list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
6648                 u64 trimmed = 0;
6649
6650                 ret = -EROFS;
6651                 if (!trans->aborted)
6652                         ret = btrfs_discard_extent(fs_info,
6653                                                    block_group->key.objectid,
6654                                                    block_group->key.offset,
6655                                                    &trimmed);
6656
6657                 list_del_init(&block_group->bg_list);
6658                 btrfs_put_block_group_trimming(block_group);
6659                 btrfs_put_block_group(block_group);
6660
6661                 if (ret) {
6662                         const char *errstr = btrfs_decode_error(ret);
6663                         btrfs_warn(fs_info,
6664                            "discard failed while removing blockgroup: errno=%d %s",
6665                                    ret, errstr);
6666                 }
6667         }
6668
6669         return 0;
6670 }
6671
6672 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
6673                                struct btrfs_delayed_ref_node *node, u64 parent,
6674                                u64 root_objectid, u64 owner_objectid,
6675                                u64 owner_offset, int refs_to_drop,
6676                                struct btrfs_delayed_extent_op *extent_op)
6677 {
6678         struct btrfs_fs_info *info = trans->fs_info;
6679         struct btrfs_key key;
6680         struct btrfs_path *path;
6681         struct btrfs_root *extent_root = info->extent_root;
6682         struct extent_buffer *leaf;
6683         struct btrfs_extent_item *ei;
6684         struct btrfs_extent_inline_ref *iref;
6685         int ret;
6686         int is_data;
6687         int extent_slot = 0;
6688         int found_extent = 0;
6689         int num_to_del = 1;
6690         u32 item_size;
6691         u64 refs;
6692         u64 bytenr = node->bytenr;
6693         u64 num_bytes = node->num_bytes;
6694         int last_ref = 0;
6695         bool skinny_metadata = btrfs_fs_incompat(info, SKINNY_METADATA);
6696
6697         path = btrfs_alloc_path();
6698         if (!path)
6699                 return -ENOMEM;
6700
6701         path->reada = READA_FORWARD;
6702         path->leave_spinning = 1;
6703
6704         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
6705         BUG_ON(!is_data && refs_to_drop != 1);
6706
6707         if (is_data)
6708                 skinny_metadata = false;
6709
6710         ret = lookup_extent_backref(trans, path, &iref, bytenr, num_bytes,
6711                                     parent, root_objectid, owner_objectid,
6712                                     owner_offset);
6713         if (ret == 0) {
6714                 extent_slot = path->slots[0];
6715                 while (extent_slot >= 0) {
6716                         btrfs_item_key_to_cpu(path->nodes[0], &key,
6717                                               extent_slot);
6718                         if (key.objectid != bytenr)
6719                                 break;
6720                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
6721                             key.offset == num_bytes) {
6722                                 found_extent = 1;
6723                                 break;
6724                         }
6725                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
6726                             key.offset == owner_objectid) {
6727                                 found_extent = 1;
6728                                 break;
6729                         }
6730                         if (path->slots[0] - extent_slot > 5)
6731                                 break;
6732                         extent_slot--;
6733                 }
6734
6735                 if (!found_extent) {
6736                         BUG_ON(iref);
6737                         ret = remove_extent_backref(trans, path, NULL,
6738                                                     refs_to_drop,
6739                                                     is_data, &last_ref);
6740                         if (ret) {
6741                                 btrfs_abort_transaction(trans, ret);
6742                                 goto out;
6743                         }
6744                         btrfs_release_path(path);
6745                         path->leave_spinning = 1;
6746
6747                         key.objectid = bytenr;
6748                         key.type = BTRFS_EXTENT_ITEM_KEY;
6749                         key.offset = num_bytes;
6750
6751                         if (!is_data && skinny_metadata) {
6752                                 key.type = BTRFS_METADATA_ITEM_KEY;
6753                                 key.offset = owner_objectid;
6754                         }
6755
6756                         ret = btrfs_search_slot(trans, extent_root,
6757                                                 &key, path, -1, 1);
6758                         if (ret > 0 && skinny_metadata && path->slots[0]) {
6759                                 /*
6760                                  * Couldn't find our skinny metadata item,
6761                                  * see if we have ye olde extent item.
6762                                  */
6763                                 path->slots[0]--;
6764                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
6765                                                       path->slots[0]);
6766                                 if (key.objectid == bytenr &&
6767                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
6768                                     key.offset == num_bytes)
6769                                         ret = 0;
6770                         }
6771
6772                         if (ret > 0 && skinny_metadata) {
6773                                 skinny_metadata = false;
6774                                 key.objectid = bytenr;
6775                                 key.type = BTRFS_EXTENT_ITEM_KEY;
6776                                 key.offset = num_bytes;
6777                                 btrfs_release_path(path);
6778                                 ret = btrfs_search_slot(trans, extent_root,
6779                                                         &key, path, -1, 1);
6780                         }
6781
6782                         if (ret) {
6783                                 btrfs_err(info,
6784                                           "umm, got %d back from search, was looking for %llu",
6785                                           ret, bytenr);
6786                                 if (ret > 0)
6787                                         btrfs_print_leaf(path->nodes[0]);
6788                         }
6789                         if (ret < 0) {
6790                                 btrfs_abort_transaction(trans, ret);
6791                                 goto out;
6792                         }
6793                         extent_slot = path->slots[0];
6794                 }
6795         } else if (WARN_ON(ret == -ENOENT)) {
6796                 btrfs_print_leaf(path->nodes[0]);
6797                 btrfs_err(info,
6798                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
6799                         bytenr, parent, root_objectid, owner_objectid,
6800                         owner_offset);
6801                 btrfs_abort_transaction(trans, ret);
6802                 goto out;
6803         } else {
6804                 btrfs_abort_transaction(trans, ret);
6805                 goto out;
6806         }
6807
6808         leaf = path->nodes[0];
6809         item_size = btrfs_item_size_nr(leaf, extent_slot);
6810         if (unlikely(item_size < sizeof(*ei))) {
6811                 ret = -EINVAL;
6812                 btrfs_print_v0_err(info);
6813                 btrfs_abort_transaction(trans, ret);
6814                 goto out;
6815         }
6816         ei = btrfs_item_ptr(leaf, extent_slot,
6817                             struct btrfs_extent_item);
6818         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
6819             key.type == BTRFS_EXTENT_ITEM_KEY) {
6820                 struct btrfs_tree_block_info *bi;
6821                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
6822                 bi = (struct btrfs_tree_block_info *)(ei + 1);
6823                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
6824         }
6825
6826         refs = btrfs_extent_refs(leaf, ei);
6827         if (refs < refs_to_drop) {
6828                 btrfs_err(info,
6829                           "trying to drop %d refs but we only have %Lu for bytenr %Lu",
6830                           refs_to_drop, refs, bytenr);
6831                 ret = -EINVAL;
6832                 btrfs_abort_transaction(trans, ret);
6833                 goto out;
6834         }
6835         refs -= refs_to_drop;
6836
6837         if (refs > 0) {
6838                 if (extent_op)
6839                         __run_delayed_extent_op(extent_op, leaf, ei);
6840                 /*
6841                  * In the case of inline back ref, reference count will
6842                  * be updated by remove_extent_backref
6843                  */
6844                 if (iref) {
6845                         BUG_ON(!found_extent);
6846                 } else {
6847                         btrfs_set_extent_refs(leaf, ei, refs);
6848                         btrfs_mark_buffer_dirty(leaf);
6849                 }
6850                 if (found_extent) {
6851                         ret = remove_extent_backref(trans, path, iref,
6852                                                     refs_to_drop, is_data,
6853                                                     &last_ref);
6854                         if (ret) {
6855                                 btrfs_abort_transaction(trans, ret);
6856                                 goto out;
6857                         }
6858                 }
6859         } else {
6860                 if (found_extent) {
6861                         BUG_ON(is_data && refs_to_drop !=
6862                                extent_data_ref_count(path, iref));
6863                         if (iref) {
6864                                 BUG_ON(path->slots[0] != extent_slot);
6865                         } else {
6866                                 BUG_ON(path->slots[0] != extent_slot + 1);
6867                                 path->slots[0] = extent_slot;
6868                                 num_to_del = 2;
6869                         }
6870                 }
6871
6872                 last_ref = 1;
6873                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
6874                                       num_to_del);
6875                 if (ret) {
6876                         btrfs_abort_transaction(trans, ret);
6877                         goto out;
6878                 }
6879                 btrfs_release_path(path);
6880
6881                 if (is_data) {
6882                         ret = btrfs_del_csums(trans, info, bytenr, num_bytes);
6883                         if (ret) {
6884                                 btrfs_abort_transaction(trans, ret);
6885                                 goto out;
6886                         }
6887                 }
6888
6889                 ret = add_to_free_space_tree(trans, bytenr, num_bytes);
6890                 if (ret) {
6891                         btrfs_abort_transaction(trans, ret);
6892                         goto out;
6893                 }
6894
6895                 ret = update_block_group(trans, info, bytenr, num_bytes, 0);
6896                 if (ret) {
6897                         btrfs_abort_transaction(trans, ret);
6898                         goto out;
6899                 }
6900         }
6901         btrfs_release_path(path);
6902
6903 out:
6904         btrfs_free_path(path);
6905         return ret;
6906 }
6907
6908 /*
6909  * when we free an block, it is possible (and likely) that we free the last
6910  * delayed ref for that extent as well.  This searches the delayed ref tree for
6911  * a given extent, and if there are no other delayed refs to be processed, it
6912  * removes it from the tree.
6913  */
6914 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
6915                                       u64 bytenr)
6916 {
6917         struct btrfs_delayed_ref_head *head;
6918         struct btrfs_delayed_ref_root *delayed_refs;
6919         int ret = 0;
6920
6921         delayed_refs = &trans->transaction->delayed_refs;
6922         spin_lock(&delayed_refs->lock);
6923         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
6924         if (!head)
6925                 goto out_delayed_unlock;
6926
6927         spin_lock(&head->lock);
6928         if (!RB_EMPTY_ROOT(&head->ref_tree))
6929                 goto out;
6930
6931         if (head->extent_op) {
6932                 if (!head->must_insert_reserved)
6933                         goto out;
6934                 btrfs_free_delayed_extent_op(head->extent_op);
6935                 head->extent_op = NULL;
6936         }
6937
6938         /*
6939          * waiting for the lock here would deadlock.  If someone else has it
6940          * locked they are already in the process of dropping it anyway
6941          */
6942         if (!mutex_trylock(&head->mutex))
6943                 goto out;
6944
6945         /*
6946          * at this point we have a head with no other entries.  Go
6947          * ahead and process it.
6948          */
6949         rb_erase(&head->href_node, &delayed_refs->href_root);
6950         RB_CLEAR_NODE(&head->href_node);
6951         atomic_dec(&delayed_refs->num_entries);
6952
6953         /*
6954          * we don't take a ref on the node because we're removing it from the
6955          * tree, so we just steal the ref the tree was holding.
6956          */
6957         delayed_refs->num_heads--;
6958         if (head->processing == 0)
6959                 delayed_refs->num_heads_ready--;
6960         head->processing = 0;
6961         spin_unlock(&head->lock);
6962         spin_unlock(&delayed_refs->lock);
6963
6964         BUG_ON(head->extent_op);
6965         if (head->must_insert_reserved)
6966                 ret = 1;
6967
6968         mutex_unlock(&head->mutex);
6969         btrfs_put_delayed_ref_head(head);
6970         return ret;
6971 out:
6972         spin_unlock(&head->lock);
6973
6974 out_delayed_unlock:
6975         spin_unlock(&delayed_refs->lock);
6976         return 0;
6977 }
6978
6979 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
6980                            struct btrfs_root *root,
6981                            struct extent_buffer *buf,
6982                            u64 parent, int last_ref)
6983 {
6984         struct btrfs_fs_info *fs_info = root->fs_info;
6985         int pin = 1;
6986         int ret;
6987
6988         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
6989                 int old_ref_mod, new_ref_mod;
6990
6991                 btrfs_ref_tree_mod(root, buf->start, buf->len, parent,
6992                                    root->root_key.objectid,
6993                                    btrfs_header_level(buf), 0,
6994                                    BTRFS_DROP_DELAYED_REF);
6995                 ret = btrfs_add_delayed_tree_ref(trans, buf->start,
6996                                                  buf->len, parent,
6997                                                  root->root_key.objectid,
6998                                                  btrfs_header_level(buf),
6999                                                  BTRFS_DROP_DELAYED_REF, NULL,
7000                                                  &old_ref_mod, &new_ref_mod);
7001                 BUG_ON(ret); /* -ENOMEM */
7002                 pin = old_ref_mod >= 0 && new_ref_mod < 0;
7003         }
7004
7005         if (last_ref && btrfs_header_generation(buf) == trans->transid) {
7006                 struct btrfs_block_group_cache *cache;
7007
7008                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7009                         ret = check_ref_cleanup(trans, buf->start);
7010                         if (!ret)
7011                                 goto out;
7012                 }
7013
7014                 pin = 0;
7015                 cache = btrfs_lookup_block_group(fs_info, buf->start);
7016
7017                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
7018                         pin_down_extent(fs_info, cache, buf->start,
7019                                         buf->len, 1);
7020                         btrfs_put_block_group(cache);
7021                         goto out;
7022                 }
7023
7024                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
7025
7026                 btrfs_add_free_space(cache, buf->start, buf->len);
7027                 btrfs_free_reserved_bytes(cache, buf->len, 0);
7028                 btrfs_put_block_group(cache);
7029                 trace_btrfs_reserved_extent_free(fs_info, buf->start, buf->len);
7030         }
7031 out:
7032         if (pin)
7033                 add_pinned_bytes(fs_info, buf->len, true,
7034                                  root->root_key.objectid);
7035
7036         if (last_ref) {
7037                 /*
7038                  * Deleting the buffer, clear the corrupt flag since it doesn't
7039                  * matter anymore.
7040                  */
7041                 clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
7042         }
7043 }
7044
7045 /* Can return -ENOMEM */
7046 int btrfs_free_extent(struct btrfs_trans_handle *trans,
7047                       struct btrfs_root *root,
7048                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
7049                       u64 owner, u64 offset)
7050 {
7051         struct btrfs_fs_info *fs_info = root->fs_info;
7052         int old_ref_mod, new_ref_mod;
7053         int ret;
7054
7055         if (btrfs_is_testing(fs_info))
7056                 return 0;
7057
7058         if (root_objectid != BTRFS_TREE_LOG_OBJECTID)
7059                 btrfs_ref_tree_mod(root, bytenr, num_bytes, parent,
7060                                    root_objectid, owner, offset,
7061                                    BTRFS_DROP_DELAYED_REF);
7062
7063         /*
7064          * tree log blocks never actually go into the extent allocation
7065          * tree, just update pinning info and exit early.
7066          */
7067         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
7068                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
7069                 /* unlocks the pinned mutex */
7070                 btrfs_pin_extent(fs_info, bytenr, num_bytes, 1);
7071                 old_ref_mod = new_ref_mod = 0;
7072                 ret = 0;
7073         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
7074                 ret = btrfs_add_delayed_tree_ref(trans, bytenr,
7075                                                  num_bytes, parent,
7076                                                  root_objectid, (int)owner,
7077                                                  BTRFS_DROP_DELAYED_REF, NULL,
7078                                                  &old_ref_mod, &new_ref_mod);
7079         } else {
7080                 ret = btrfs_add_delayed_data_ref(trans, bytenr,
7081                                                  num_bytes, parent,
7082                                                  root_objectid, owner, offset,
7083                                                  0, BTRFS_DROP_DELAYED_REF,
7084                                                  &old_ref_mod, &new_ref_mod);
7085         }
7086
7087         if (ret == 0 && old_ref_mod >= 0 && new_ref_mod < 0) {
7088                 bool metadata = owner < BTRFS_FIRST_FREE_OBJECTID;
7089
7090                 add_pinned_bytes(fs_info, num_bytes, metadata, root_objectid);
7091         }
7092
7093         return ret;
7094 }
7095
7096 /*
7097  * when we wait for progress in the block group caching, its because
7098  * our allocation attempt failed at least once.  So, we must sleep
7099  * and let some progress happen before we try again.
7100  *
7101  * This function will sleep at least once waiting for new free space to
7102  * show up, and then it will check the block group free space numbers
7103  * for our min num_bytes.  Another option is to have it go ahead
7104  * and look in the rbtree for a free extent of a given size, but this
7105  * is a good start.
7106  *
7107  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
7108  * any of the information in this block group.
7109  */
7110 static noinline void
7111 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
7112                                 u64 num_bytes)
7113 {
7114         struct btrfs_caching_control *caching_ctl;
7115
7116         caching_ctl = get_caching_control(cache);
7117         if (!caching_ctl)
7118                 return;
7119
7120         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
7121                    (cache->free_space_ctl->free_space >= num_bytes));
7122
7123         put_caching_control(caching_ctl);
7124 }
7125
7126 static noinline int
7127 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
7128 {
7129         struct btrfs_caching_control *caching_ctl;
7130         int ret = 0;
7131
7132         caching_ctl = get_caching_control(cache);
7133         if (!caching_ctl)
7134                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
7135
7136         wait_event(caching_ctl->wait, block_group_cache_done(cache));
7137         if (cache->cached == BTRFS_CACHE_ERROR)
7138                 ret = -EIO;
7139         put_caching_control(caching_ctl);
7140         return ret;
7141 }
7142
7143 enum btrfs_loop_type {
7144         LOOP_CACHING_NOWAIT = 0,
7145         LOOP_CACHING_WAIT = 1,
7146         LOOP_ALLOC_CHUNK = 2,
7147         LOOP_NO_EMPTY_SIZE = 3,
7148 };
7149
7150 static inline void
7151 btrfs_lock_block_group(struct btrfs_block_group_cache *cache,
7152                        int delalloc)
7153 {
7154         if (delalloc)
7155                 down_read(&cache->data_rwsem);
7156 }
7157
7158 static inline void
7159 btrfs_grab_block_group(struct btrfs_block_group_cache *cache,
7160                        int delalloc)
7161 {
7162         btrfs_get_block_group(cache);
7163         if (delalloc)
7164                 down_read(&cache->data_rwsem);
7165 }
7166
7167 static struct btrfs_block_group_cache *
7168 btrfs_lock_cluster(struct btrfs_block_group_cache *block_group,
7169                    struct btrfs_free_cluster *cluster,
7170                    int delalloc)
7171 {
7172         struct btrfs_block_group_cache *used_bg = NULL;
7173
7174         spin_lock(&cluster->refill_lock);
7175         while (1) {
7176                 used_bg = cluster->block_group;
7177                 if (!used_bg)
7178                         return NULL;
7179
7180                 if (used_bg == block_group)
7181                         return used_bg;
7182
7183                 btrfs_get_block_group(used_bg);
7184
7185                 if (!delalloc)
7186                         return used_bg;
7187
7188                 if (down_read_trylock(&used_bg->data_rwsem))
7189                         return used_bg;
7190
7191                 spin_unlock(&cluster->refill_lock);
7192
7193                 /* We should only have one-level nested. */
7194                 down_read_nested(&used_bg->data_rwsem, SINGLE_DEPTH_NESTING);
7195
7196                 spin_lock(&cluster->refill_lock);
7197                 if (used_bg == cluster->block_group)
7198                         return used_bg;
7199
7200                 up_read(&used_bg->data_rwsem);
7201                 btrfs_put_block_group(used_bg);
7202         }
7203 }
7204
7205 static inline void
7206 btrfs_release_block_group(struct btrfs_block_group_cache *cache,
7207                          int delalloc)
7208 {
7209         if (delalloc)
7210                 up_read(&cache->data_rwsem);
7211         btrfs_put_block_group(cache);
7212 }
7213
7214 /*
7215  * walks the btree of allocated extents and find a hole of a given size.
7216  * The key ins is changed to record the hole:
7217  * ins->objectid == start position
7218  * ins->flags = BTRFS_EXTENT_ITEM_KEY
7219  * ins->offset == the size of the hole.
7220  * Any available blocks before search_start are skipped.
7221  *
7222  * If there is no suitable free space, we will record the max size of
7223  * the free space extent currently.
7224  */
7225 static noinline int find_free_extent(struct btrfs_fs_info *fs_info,
7226                                 u64 ram_bytes, u64 num_bytes, u64 empty_size,
7227                                 u64 hint_byte, struct btrfs_key *ins,
7228                                 u64 flags, int delalloc)
7229 {
7230         int ret = 0;
7231         struct btrfs_root *root = fs_info->extent_root;
7232         struct btrfs_free_cluster *last_ptr = NULL;
7233         struct btrfs_block_group_cache *block_group = NULL;
7234         u64 search_start = 0;
7235         u64 max_extent_size = 0;
7236         u64 empty_cluster = 0;
7237         struct btrfs_space_info *space_info;
7238         int loop = 0;
7239         int index = btrfs_bg_flags_to_raid_index(flags);
7240         bool failed_cluster_refill = false;
7241         bool failed_alloc = false;
7242         bool use_cluster = true;
7243         bool have_caching_bg = false;
7244         bool orig_have_caching_bg = false;
7245         bool full_search = false;
7246
7247         WARN_ON(num_bytes < fs_info->sectorsize);
7248         ins->type = BTRFS_EXTENT_ITEM_KEY;
7249         ins->objectid = 0;
7250         ins->offset = 0;
7251
7252         trace_find_free_extent(fs_info, num_bytes, empty_size, flags);
7253
7254         space_info = __find_space_info(fs_info, flags);
7255         if (!space_info) {
7256                 btrfs_err(fs_info, "No space info for %llu", flags);
7257                 return -ENOSPC;
7258         }
7259
7260         /*
7261          * If our free space is heavily fragmented we may not be able to make
7262          * big contiguous allocations, so instead of doing the expensive search
7263          * for free space, simply return ENOSPC with our max_extent_size so we
7264          * can go ahead and search for a more manageable chunk.
7265          *
7266          * If our max_extent_size is large enough for our allocation simply
7267          * disable clustering since we will likely not be able to find enough
7268          * space to create a cluster and induce latency trying.
7269          */
7270         if (unlikely(space_info->max_extent_size)) {
7271                 spin_lock(&space_info->lock);
7272                 if (space_info->max_extent_size &&
7273                     num_bytes > space_info->max_extent_size) {
7274                         ins->offset = space_info->max_extent_size;
7275                         spin_unlock(&space_info->lock);
7276                         return -ENOSPC;
7277                 } else if (space_info->max_extent_size) {
7278                         use_cluster = false;
7279                 }
7280                 spin_unlock(&space_info->lock);
7281         }
7282
7283         last_ptr = fetch_cluster_info(fs_info, space_info, &empty_cluster);
7284         if (last_ptr) {
7285                 spin_lock(&last_ptr->lock);
7286                 if (last_ptr->block_group)
7287                         hint_byte = last_ptr->window_start;
7288                 if (last_ptr->fragmented) {
7289                         /*
7290                          * We still set window_start so we can keep track of the
7291                          * last place we found an allocation to try and save
7292                          * some time.
7293                          */
7294                         hint_byte = last_ptr->window_start;
7295                         use_cluster = false;
7296                 }
7297                 spin_unlock(&last_ptr->lock);
7298         }
7299
7300         search_start = max(search_start, first_logical_byte(fs_info, 0));
7301         search_start = max(search_start, hint_byte);
7302         if (search_start == hint_byte) {
7303                 block_group = btrfs_lookup_block_group(fs_info, search_start);
7304                 /*
7305                  * we don't want to use the block group if it doesn't match our
7306                  * allocation bits, or if its not cached.
7307                  *
7308                  * However if we are re-searching with an ideal block group
7309                  * picked out then we don't care that the block group is cached.
7310                  */
7311                 if (block_group && block_group_bits(block_group, flags) &&
7312                     block_group->cached != BTRFS_CACHE_NO) {
7313                         down_read(&space_info->groups_sem);
7314                         if (list_empty(&block_group->list) ||
7315                             block_group->ro) {
7316                                 /*
7317                                  * someone is removing this block group,
7318                                  * we can't jump into the have_block_group
7319                                  * target because our list pointers are not
7320                                  * valid
7321                                  */
7322                                 btrfs_put_block_group(block_group);
7323                                 up_read(&space_info->groups_sem);
7324                         } else {
7325                                 index = btrfs_bg_flags_to_raid_index(
7326                                                 block_group->flags);
7327                                 btrfs_lock_block_group(block_group, delalloc);
7328                                 goto have_block_group;
7329                         }
7330                 } else if (block_group) {
7331                         btrfs_put_block_group(block_group);
7332                 }
7333         }
7334 search:
7335         have_caching_bg = false;
7336         if (index == 0 || index == btrfs_bg_flags_to_raid_index(flags))
7337                 full_search = true;
7338         down_read(&space_info->groups_sem);
7339         list_for_each_entry(block_group, &space_info->block_groups[index],
7340                             list) {
7341                 u64 offset;
7342                 int cached;
7343
7344                 /* If the block group is read-only, we can skip it entirely. */
7345                 if (unlikely(block_group->ro))
7346                         continue;
7347
7348                 btrfs_grab_block_group(block_group, delalloc);
7349                 search_start = block_group->key.objectid;
7350
7351                 /*
7352                  * this can happen if we end up cycling through all the
7353                  * raid types, but we want to make sure we only allocate
7354                  * for the proper type.
7355                  */
7356                 if (!block_group_bits(block_group, flags)) {
7357                         u64 extra = BTRFS_BLOCK_GROUP_DUP |
7358                                 BTRFS_BLOCK_GROUP_RAID1 |
7359                                 BTRFS_BLOCK_GROUP_RAID5 |
7360                                 BTRFS_BLOCK_GROUP_RAID6 |
7361                                 BTRFS_BLOCK_GROUP_RAID10;
7362
7363                         /*
7364                          * if they asked for extra copies and this block group
7365                          * doesn't provide them, bail.  This does allow us to
7366                          * fill raid0 from raid1.
7367                          */
7368                         if ((flags & extra) && !(block_group->flags & extra))
7369                                 goto loop;
7370                 }
7371
7372 have_block_group:
7373                 cached = block_group_cache_done(block_group);
7374                 if (unlikely(!cached)) {
7375                         have_caching_bg = true;
7376                         ret = cache_block_group(block_group, 0);
7377                         BUG_ON(ret < 0);
7378                         ret = 0;
7379                 }
7380
7381                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
7382                         goto loop;
7383
7384                 /*
7385                  * Ok we want to try and use the cluster allocator, so
7386                  * lets look there
7387                  */
7388                 if (last_ptr && use_cluster) {
7389                         struct btrfs_block_group_cache *used_block_group;
7390                         unsigned long aligned_cluster;
7391                         /*
7392                          * the refill lock keeps out other
7393                          * people trying to start a new cluster
7394                          */
7395                         used_block_group = btrfs_lock_cluster(block_group,
7396                                                               last_ptr,
7397                                                               delalloc);
7398                         if (!used_block_group)
7399                                 goto refill_cluster;
7400
7401                         if (used_block_group != block_group &&
7402                             (used_block_group->ro ||
7403                              !block_group_bits(used_block_group, flags)))
7404                                 goto release_cluster;
7405
7406                         offset = btrfs_alloc_from_cluster(used_block_group,
7407                                                 last_ptr,
7408                                                 num_bytes,
7409                                                 used_block_group->key.objectid,
7410                                                 &max_extent_size);
7411                         if (offset) {
7412                                 /* we have a block, we're done */
7413                                 spin_unlock(&last_ptr->refill_lock);
7414                                 trace_btrfs_reserve_extent_cluster(
7415                                                 used_block_group,
7416                                                 search_start, num_bytes);
7417                                 if (used_block_group != block_group) {
7418                                         btrfs_release_block_group(block_group,
7419                                                                   delalloc);
7420                                         block_group = used_block_group;
7421                                 }
7422                                 goto checks;
7423                         }
7424
7425                         WARN_ON(last_ptr->block_group != used_block_group);
7426 release_cluster:
7427                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
7428                          * set up a new clusters, so lets just skip it
7429                          * and let the allocator find whatever block
7430                          * it can find.  If we reach this point, we
7431                          * will have tried the cluster allocator
7432                          * plenty of times and not have found
7433                          * anything, so we are likely way too
7434                          * fragmented for the clustering stuff to find
7435                          * anything.
7436                          *
7437                          * However, if the cluster is taken from the
7438                          * current block group, release the cluster
7439                          * first, so that we stand a better chance of
7440                          * succeeding in the unclustered
7441                          * allocation.  */
7442                         if (loop >= LOOP_NO_EMPTY_SIZE &&
7443                             used_block_group != block_group) {
7444                                 spin_unlock(&last_ptr->refill_lock);
7445                                 btrfs_release_block_group(used_block_group,
7446                                                           delalloc);
7447                                 goto unclustered_alloc;
7448                         }
7449
7450                         /*
7451                          * this cluster didn't work out, free it and
7452                          * start over
7453                          */
7454                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7455
7456                         if (used_block_group != block_group)
7457                                 btrfs_release_block_group(used_block_group,
7458                                                           delalloc);
7459 refill_cluster:
7460                         if (loop >= LOOP_NO_EMPTY_SIZE) {
7461                                 spin_unlock(&last_ptr->refill_lock);
7462                                 goto unclustered_alloc;
7463                         }
7464
7465                         aligned_cluster = max_t(unsigned long,
7466                                                 empty_cluster + empty_size,
7467                                               block_group->full_stripe_len);
7468
7469                         /* allocate a cluster in this block group */
7470                         ret = btrfs_find_space_cluster(fs_info, block_group,
7471                                                        last_ptr, search_start,
7472                                                        num_bytes,
7473                                                        aligned_cluster);
7474                         if (ret == 0) {
7475                                 /*
7476                                  * now pull our allocation out of this
7477                                  * cluster
7478                                  */
7479                                 offset = btrfs_alloc_from_cluster(block_group,
7480                                                         last_ptr,
7481                                                         num_bytes,
7482                                                         search_start,
7483                                                         &max_extent_size);
7484                                 if (offset) {
7485                                         /* we found one, proceed */
7486                                         spin_unlock(&last_ptr->refill_lock);
7487                                         trace_btrfs_reserve_extent_cluster(
7488                                                 block_group, search_start,
7489                                                 num_bytes);
7490                                         goto checks;
7491                                 }
7492                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
7493                                    && !failed_cluster_refill) {
7494                                 spin_unlock(&last_ptr->refill_lock);
7495
7496                                 failed_cluster_refill = true;
7497                                 wait_block_group_cache_progress(block_group,
7498                                        num_bytes + empty_cluster + empty_size);
7499                                 goto have_block_group;
7500                         }
7501
7502                         /*
7503                          * at this point we either didn't find a cluster
7504                          * or we weren't able to allocate a block from our
7505                          * cluster.  Free the cluster we've been trying
7506                          * to use, and go to the next block group
7507                          */
7508                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7509                         spin_unlock(&last_ptr->refill_lock);
7510                         goto loop;
7511                 }
7512
7513 unclustered_alloc:
7514                 /*
7515                  * We are doing an unclustered alloc, set the fragmented flag so
7516                  * we don't bother trying to setup a cluster again until we get
7517                  * more space.
7518                  */
7519                 if (unlikely(last_ptr)) {
7520                         spin_lock(&last_ptr->lock);
7521                         last_ptr->fragmented = 1;
7522                         spin_unlock(&last_ptr->lock);
7523                 }
7524                 if (cached) {
7525                         struct btrfs_free_space_ctl *ctl =
7526                                 block_group->free_space_ctl;
7527
7528                         spin_lock(&ctl->tree_lock);
7529                         if (ctl->free_space <
7530                             num_bytes + empty_cluster + empty_size) {
7531                                 if (ctl->free_space > max_extent_size)
7532                                         max_extent_size = ctl->free_space;
7533                                 spin_unlock(&ctl->tree_lock);
7534                                 goto loop;
7535                         }
7536                         spin_unlock(&ctl->tree_lock);
7537                 }
7538
7539                 offset = btrfs_find_space_for_alloc(block_group, search_start,
7540                                                     num_bytes, empty_size,
7541                                                     &max_extent_size);
7542                 /*
7543                  * If we didn't find a chunk, and we haven't failed on this
7544                  * block group before, and this block group is in the middle of
7545                  * caching and we are ok with waiting, then go ahead and wait
7546                  * for progress to be made, and set failed_alloc to true.
7547                  *
7548                  * If failed_alloc is true then we've already waited on this
7549                  * block group once and should move on to the next block group.
7550                  */
7551                 if (!offset && !failed_alloc && !cached &&
7552                     loop > LOOP_CACHING_NOWAIT) {
7553                         wait_block_group_cache_progress(block_group,
7554                                                 num_bytes + empty_size);
7555                         failed_alloc = true;
7556                         goto have_block_group;
7557                 } else if (!offset) {
7558                         goto loop;
7559                 }
7560 checks:
7561                 search_start = round_up(offset, fs_info->stripesize);
7562
7563                 /* move on to the next group */
7564                 if (search_start + num_bytes >
7565                     block_group->key.objectid + block_group->key.offset) {
7566                         btrfs_add_free_space(block_group, offset, num_bytes);
7567                         goto loop;
7568                 }
7569
7570                 if (offset < search_start)
7571                         btrfs_add_free_space(block_group, offset,
7572                                              search_start - offset);
7573
7574                 ret = btrfs_add_reserved_bytes(block_group, ram_bytes,
7575                                 num_bytes, delalloc);
7576                 if (ret == -EAGAIN) {
7577                         btrfs_add_free_space(block_group, offset, num_bytes);
7578                         goto loop;
7579                 }
7580                 btrfs_inc_block_group_reservations(block_group);
7581
7582                 /* we are all good, lets return */
7583                 ins->objectid = search_start;
7584                 ins->offset = num_bytes;
7585
7586                 trace_btrfs_reserve_extent(block_group, search_start, num_bytes);
7587                 btrfs_release_block_group(block_group, delalloc);
7588                 break;
7589 loop:
7590                 failed_cluster_refill = false;
7591                 failed_alloc = false;
7592                 BUG_ON(btrfs_bg_flags_to_raid_index(block_group->flags) !=
7593                        index);
7594                 btrfs_release_block_group(block_group, delalloc);
7595                 cond_resched();
7596         }
7597         up_read(&space_info->groups_sem);
7598
7599         if ((loop == LOOP_CACHING_NOWAIT) && have_caching_bg
7600                 && !orig_have_caching_bg)
7601                 orig_have_caching_bg = true;
7602
7603         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
7604                 goto search;
7605
7606         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
7607                 goto search;
7608
7609         /*
7610          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
7611          *                      caching kthreads as we move along
7612          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
7613          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
7614          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
7615          *                      again
7616          */
7617         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
7618                 index = 0;
7619                 if (loop == LOOP_CACHING_NOWAIT) {
7620                         /*
7621                          * We want to skip the LOOP_CACHING_WAIT step if we
7622                          * don't have any uncached bgs and we've already done a
7623                          * full search through.
7624                          */
7625                         if (orig_have_caching_bg || !full_search)
7626                                 loop = LOOP_CACHING_WAIT;
7627                         else
7628                                 loop = LOOP_ALLOC_CHUNK;
7629                 } else {
7630                         loop++;
7631                 }
7632
7633                 if (loop == LOOP_ALLOC_CHUNK) {
7634                         struct btrfs_trans_handle *trans;
7635                         int exist = 0;
7636
7637                         trans = current->journal_info;
7638                         if (trans)
7639                                 exist = 1;
7640                         else
7641                                 trans = btrfs_join_transaction(root);
7642
7643                         if (IS_ERR(trans)) {
7644                                 ret = PTR_ERR(trans);
7645                                 goto out;
7646                         }
7647
7648                         ret = do_chunk_alloc(trans, flags, CHUNK_ALLOC_FORCE);
7649
7650                         /*
7651                          * If we can't allocate a new chunk we've already looped
7652                          * through at least once, move on to the NO_EMPTY_SIZE
7653                          * case.
7654                          */
7655                         if (ret == -ENOSPC)
7656                                 loop = LOOP_NO_EMPTY_SIZE;
7657
7658                         /*
7659                          * Do not bail out on ENOSPC since we
7660                          * can do more things.
7661                          */
7662                         if (ret < 0 && ret != -ENOSPC)
7663                                 btrfs_abort_transaction(trans, ret);
7664                         else
7665                                 ret = 0;
7666                         if (!exist)
7667                                 btrfs_end_transaction(trans);
7668                         if (ret)
7669                                 goto out;
7670                 }
7671
7672                 if (loop == LOOP_NO_EMPTY_SIZE) {
7673                         /*
7674                          * Don't loop again if we already have no empty_size and
7675                          * no empty_cluster.
7676                          */
7677                         if (empty_size == 0 &&
7678                             empty_cluster == 0) {
7679                                 ret = -ENOSPC;
7680                                 goto out;
7681                         }
7682                         empty_size = 0;
7683                         empty_cluster = 0;
7684                 }
7685
7686                 goto search;
7687         } else if (!ins->objectid) {
7688                 ret = -ENOSPC;
7689         } else if (ins->objectid) {
7690                 if (!use_cluster && last_ptr) {
7691                         spin_lock(&last_ptr->lock);
7692                         last_ptr->window_start = ins->objectid;
7693                         spin_unlock(&last_ptr->lock);
7694                 }
7695                 ret = 0;
7696         }
7697 out:
7698         if (ret == -ENOSPC) {
7699                 spin_lock(&space_info->lock);
7700                 space_info->max_extent_size = max_extent_size;
7701                 spin_unlock(&space_info->lock);
7702                 ins->offset = max_extent_size;
7703         }
7704         return ret;
7705 }
7706
7707 static void dump_space_info(struct btrfs_fs_info *fs_info,
7708                             struct btrfs_space_info *info, u64 bytes,
7709                             int dump_block_groups)
7710 {
7711         struct btrfs_block_group_cache *cache;
7712         int index = 0;
7713
7714         spin_lock(&info->lock);
7715         btrfs_info(fs_info, "space_info %llu has %llu free, is %sfull",
7716                    info->flags,
7717                    info->total_bytes - btrfs_space_info_used(info, true),
7718                    info->full ? "" : "not ");
7719         btrfs_info(fs_info,
7720                 "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu",
7721                 info->total_bytes, info->bytes_used, info->bytes_pinned,
7722                 info->bytes_reserved, info->bytes_may_use,
7723                 info->bytes_readonly);
7724         spin_unlock(&info->lock);
7725
7726         if (!dump_block_groups)
7727                 return;
7728
7729         down_read(&info->groups_sem);
7730 again:
7731         list_for_each_entry(cache, &info->block_groups[index], list) {
7732                 spin_lock(&cache->lock);
7733                 btrfs_info(fs_info,
7734                         "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s",
7735                         cache->key.objectid, cache->key.offset,
7736                         btrfs_block_group_used(&cache->item), cache->pinned,
7737                         cache->reserved, cache->ro ? "[readonly]" : "");
7738                 btrfs_dump_free_space(cache, bytes);
7739                 spin_unlock(&cache->lock);
7740         }
7741         if (++index < BTRFS_NR_RAID_TYPES)
7742                 goto again;
7743         up_read(&info->groups_sem);
7744 }
7745
7746 /*
7747  * btrfs_reserve_extent - entry point to the extent allocator. Tries to find a
7748  *                        hole that is at least as big as @num_bytes.
7749  *
7750  * @root           -    The root that will contain this extent
7751  *
7752  * @ram_bytes      -    The amount of space in ram that @num_bytes take. This
7753  *                      is used for accounting purposes. This value differs
7754  *                      from @num_bytes only in the case of compressed extents.
7755  *
7756  * @num_bytes      -    Number of bytes to allocate on-disk.
7757  *
7758  * @min_alloc_size -    Indicates the minimum amount of space that the
7759  *                      allocator should try to satisfy. In some cases
7760  *                      @num_bytes may be larger than what is required and if
7761  *                      the filesystem is fragmented then allocation fails.
7762  *                      However, the presence of @min_alloc_size gives a
7763  *                      chance to try and satisfy the smaller allocation.
7764  *
7765  * @empty_size     -    A hint that you plan on doing more COW. This is the
7766  *                      size in bytes the allocator should try to find free
7767  *                      next to the block it returns.  This is just a hint and
7768  *                      may be ignored by the allocator.
7769  *
7770  * @hint_byte      -    Hint to the allocator to start searching above the byte
7771  *                      address passed. It might be ignored.
7772  *
7773  * @ins            -    This key is modified to record the found hole. It will
7774  *                      have the following values:
7775  *                      ins->objectid == start position
7776  *                      ins->flags = BTRFS_EXTENT_ITEM_KEY
7777  *                      ins->offset == the size of the hole.
7778  *
7779  * @is_data        -    Boolean flag indicating whether an extent is
7780  *                      allocated for data (true) or metadata (false)
7781  *
7782  * @delalloc       -    Boolean flag indicating whether this allocation is for
7783  *                      delalloc or not. If 'true' data_rwsem of block groups
7784  *                      is going to be acquired.
7785  *
7786  *
7787  * Returns 0 when an allocation succeeded or < 0 when an error occurred. In
7788  * case -ENOSPC is returned then @ins->offset will contain the size of the
7789  * largest available hole the allocator managed to find.
7790  */
7791 int btrfs_reserve_extent(struct btrfs_root *root, u64 ram_bytes,
7792                          u64 num_bytes, u64 min_alloc_size,
7793                          u64 empty_size, u64 hint_byte,
7794                          struct btrfs_key *ins, int is_data, int delalloc)
7795 {
7796         struct btrfs_fs_info *fs_info = root->fs_info;
7797         bool final_tried = num_bytes == min_alloc_size;
7798         u64 flags;
7799         int ret;
7800
7801         flags = get_alloc_profile_by_root(root, is_data);
7802 again:
7803         WARN_ON(num_bytes < fs_info->sectorsize);
7804         ret = find_free_extent(fs_info, ram_bytes, num_bytes, empty_size,
7805                                hint_byte, ins, flags, delalloc);
7806         if (!ret && !is_data) {
7807                 btrfs_dec_block_group_reservations(fs_info, ins->objectid);
7808         } else if (ret == -ENOSPC) {
7809                 if (!final_tried && ins->offset) {
7810                         num_bytes = min(num_bytes >> 1, ins->offset);
7811                         num_bytes = round_down(num_bytes,
7812                                                fs_info->sectorsize);
7813                         num_bytes = max(num_bytes, min_alloc_size);
7814                         ram_bytes = num_bytes;
7815                         if (num_bytes == min_alloc_size)
7816                                 final_tried = true;
7817                         goto again;
7818                 } else if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
7819                         struct btrfs_space_info *sinfo;
7820
7821                         sinfo = __find_space_info(fs_info, flags);
7822                         btrfs_err(fs_info,
7823                                   "allocation failed flags %llu, wanted %llu",
7824                                   flags, num_bytes);
7825                         if (sinfo)
7826                                 dump_space_info(fs_info, sinfo, num_bytes, 1);
7827                 }
7828         }
7829
7830         return ret;
7831 }
7832
7833 static int __btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
7834                                         u64 start, u64 len,
7835                                         int pin, int delalloc)
7836 {
7837         struct btrfs_block_group_cache *cache;
7838         int ret = 0;
7839
7840         cache = btrfs_lookup_block_group(fs_info, start);
7841         if (!cache) {
7842                 btrfs_err(fs_info, "Unable to find block group for %llu",
7843                           start);
7844                 return -ENOSPC;
7845         }
7846
7847         if (pin)
7848                 pin_down_extent(fs_info, cache, start, len, 1);
7849         else {
7850                 if (btrfs_test_opt(fs_info, DISCARD))
7851                         ret = btrfs_discard_extent(fs_info, start, len, NULL);
7852                 btrfs_add_free_space(cache, start, len);
7853                 btrfs_free_reserved_bytes(cache, len, delalloc);
7854                 trace_btrfs_reserved_extent_free(fs_info, start, len);
7855         }
7856
7857         btrfs_put_block_group(cache);
7858         return ret;
7859 }
7860
7861 int btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
7862                                u64 start, u64 len, int delalloc)
7863 {
7864         return __btrfs_free_reserved_extent(fs_info, start, len, 0, delalloc);
7865 }
7866
7867 int btrfs_free_and_pin_reserved_extent(struct btrfs_fs_info *fs_info,
7868                                        u64 start, u64 len)
7869 {
7870         return __btrfs_free_reserved_extent(fs_info, start, len, 1, 0);
7871 }
7872
7873 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
7874                                       u64 parent, u64 root_objectid,
7875                                       u64 flags, u64 owner, u64 offset,
7876                                       struct btrfs_key *ins, int ref_mod)
7877 {
7878         struct btrfs_fs_info *fs_info = trans->fs_info;
7879         int ret;
7880         struct btrfs_extent_item *extent_item;
7881         struct btrfs_extent_inline_ref *iref;
7882         struct btrfs_path *path;
7883         struct extent_buffer *leaf;
7884         int type;
7885         u32 size;
7886
7887         if (parent > 0)
7888                 type = BTRFS_SHARED_DATA_REF_KEY;
7889         else
7890                 type = BTRFS_EXTENT_DATA_REF_KEY;
7891
7892         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
7893
7894         path = btrfs_alloc_path();
7895         if (!path)
7896                 return -ENOMEM;
7897
7898         path->leave_spinning = 1;
7899         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
7900                                       ins, size);
7901         if (ret) {
7902                 btrfs_free_path(path);
7903                 return ret;
7904         }
7905
7906         leaf = path->nodes[0];
7907         extent_item = btrfs_item_ptr(leaf, path->slots[0],
7908                                      struct btrfs_extent_item);
7909         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
7910         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
7911         btrfs_set_extent_flags(leaf, extent_item,
7912                                flags | BTRFS_EXTENT_FLAG_DATA);
7913
7914         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
7915         btrfs_set_extent_inline_ref_type(leaf, iref, type);
7916         if (parent > 0) {
7917                 struct btrfs_shared_data_ref *ref;
7918                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
7919                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
7920                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
7921         } else {
7922                 struct btrfs_extent_data_ref *ref;
7923                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
7924                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
7925                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
7926                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
7927                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
7928         }
7929
7930         btrfs_mark_buffer_dirty(path->nodes[0]);
7931         btrfs_free_path(path);
7932
7933         ret = remove_from_free_space_tree(trans, ins->objectid, ins->offset);
7934         if (ret)
7935                 return ret;
7936
7937         ret = update_block_group(trans, fs_info, ins->objectid, ins->offset, 1);
7938         if (ret) { /* -ENOENT, logic error */
7939                 btrfs_err(fs_info, "update block group failed for %llu %llu",
7940                         ins->objectid, ins->offset);
7941                 BUG();
7942         }
7943         trace_btrfs_reserved_extent_alloc(fs_info, ins->objectid, ins->offset);
7944         return ret;
7945 }
7946
7947 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
7948                                      struct btrfs_delayed_ref_node *node,
7949                                      struct btrfs_delayed_extent_op *extent_op)
7950 {
7951         struct btrfs_fs_info *fs_info = trans->fs_info;
7952         int ret;
7953         struct btrfs_extent_item *extent_item;
7954         struct btrfs_key extent_key;
7955         struct btrfs_tree_block_info *block_info;
7956         struct btrfs_extent_inline_ref *iref;
7957         struct btrfs_path *path;
7958         struct extent_buffer *leaf;
7959         struct btrfs_delayed_tree_ref *ref;
7960         u32 size = sizeof(*extent_item) + sizeof(*iref);
7961         u64 num_bytes;
7962         u64 flags = extent_op->flags_to_set;
7963         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
7964
7965         ref = btrfs_delayed_node_to_tree_ref(node);
7966
7967         extent_key.objectid = node->bytenr;
7968         if (skinny_metadata) {
7969                 extent_key.offset = ref->level;
7970                 extent_key.type = BTRFS_METADATA_ITEM_KEY;
7971                 num_bytes = fs_info->nodesize;
7972         } else {
7973                 extent_key.offset = node->num_bytes;
7974                 extent_key.type = BTRFS_EXTENT_ITEM_KEY;
7975                 size += sizeof(*block_info);
7976                 num_bytes = node->num_bytes;
7977         }
7978
7979         path = btrfs_alloc_path();
7980         if (!path) {
7981                 btrfs_free_and_pin_reserved_extent(fs_info,
7982                                                    extent_key.objectid,
7983                                                    fs_info->nodesize);
7984                 return -ENOMEM;
7985         }
7986
7987         path->leave_spinning = 1;
7988         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
7989                                       &extent_key, size);
7990         if (ret) {
7991                 btrfs_free_path(path);
7992                 btrfs_free_and_pin_reserved_extent(fs_info,
7993                                                    extent_key.objectid,
7994                                                    fs_info->nodesize);
7995                 return ret;
7996         }
7997
7998         leaf = path->nodes[0];
7999         extent_item = btrfs_item_ptr(leaf, path->slots[0],
8000                                      struct btrfs_extent_item);
8001         btrfs_set_extent_refs(leaf, extent_item, 1);
8002         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8003         btrfs_set_extent_flags(leaf, extent_item,
8004                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
8005
8006         if (skinny_metadata) {
8007                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8008         } else {
8009                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
8010                 btrfs_set_tree_block_key(leaf, block_info, &extent_op->key);
8011                 btrfs_set_tree_block_level(leaf, block_info, ref->level);
8012                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
8013         }
8014
8015         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY) {
8016                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
8017                 btrfs_set_extent_inline_ref_type(leaf, iref,
8018                                                  BTRFS_SHARED_BLOCK_REF_KEY);
8019                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->parent);
8020         } else {
8021                 btrfs_set_extent_inline_ref_type(leaf, iref,
8022                                                  BTRFS_TREE_BLOCK_REF_KEY);
8023                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->root);
8024         }
8025
8026         btrfs_mark_buffer_dirty(leaf);
8027         btrfs_free_path(path);
8028
8029         ret = remove_from_free_space_tree(trans, extent_key.objectid,
8030                                           num_bytes);
8031         if (ret)
8032                 return ret;
8033
8034         ret = update_block_group(trans, fs_info, extent_key.objectid,
8035                                  fs_info->nodesize, 1);
8036         if (ret) { /* -ENOENT, logic error */
8037                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8038                         extent_key.objectid, extent_key.offset);
8039                 BUG();
8040         }
8041
8042         trace_btrfs_reserved_extent_alloc(fs_info, extent_key.objectid,
8043                                           fs_info->nodesize);
8044         return ret;
8045 }
8046
8047 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8048                                      struct btrfs_root *root, u64 owner,
8049                                      u64 offset, u64 ram_bytes,
8050                                      struct btrfs_key *ins)
8051 {
8052         int ret;
8053
8054         BUG_ON(root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
8055
8056         btrfs_ref_tree_mod(root, ins->objectid, ins->offset, 0,
8057                            root->root_key.objectid, owner, offset,
8058                            BTRFS_ADD_DELAYED_EXTENT);
8059
8060         ret = btrfs_add_delayed_data_ref(trans, ins->objectid,
8061                                          ins->offset, 0,
8062                                          root->root_key.objectid, owner,
8063                                          offset, ram_bytes,
8064                                          BTRFS_ADD_DELAYED_EXTENT, NULL, NULL);
8065         return ret;
8066 }
8067
8068 /*
8069  * this is used by the tree logging recovery code.  It records that
8070  * an extent has been allocated and makes sure to clear the free
8071  * space cache bits as well
8072  */
8073 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
8074                                    u64 root_objectid, u64 owner, u64 offset,
8075                                    struct btrfs_key *ins)
8076 {
8077         struct btrfs_fs_info *fs_info = trans->fs_info;
8078         int ret;
8079         struct btrfs_block_group_cache *block_group;
8080         struct btrfs_space_info *space_info;
8081
8082         /*
8083          * Mixed block groups will exclude before processing the log so we only
8084          * need to do the exclude dance if this fs isn't mixed.
8085          */
8086         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS)) {
8087                 ret = __exclude_logged_extent(fs_info, ins->objectid,
8088                                               ins->offset);
8089                 if (ret)
8090                         return ret;
8091         }
8092
8093         block_group = btrfs_lookup_block_group(fs_info, ins->objectid);
8094         if (!block_group)
8095                 return -EINVAL;
8096
8097         space_info = block_group->space_info;
8098         spin_lock(&space_info->lock);
8099         spin_lock(&block_group->lock);
8100         space_info->bytes_reserved += ins->offset;
8101         block_group->reserved += ins->offset;
8102         spin_unlock(&block_group->lock);
8103         spin_unlock(&space_info->lock);
8104
8105         ret = alloc_reserved_file_extent(trans, 0, root_objectid, 0, owner,
8106                                          offset, ins, 1);
8107         btrfs_put_block_group(block_group);
8108         return ret;
8109 }
8110
8111 static struct extent_buffer *
8112 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
8113                       u64 bytenr, int level, u64 owner)
8114 {
8115         struct btrfs_fs_info *fs_info = root->fs_info;
8116         struct extent_buffer *buf;
8117
8118         buf = btrfs_find_create_tree_block(fs_info, bytenr);
8119         if (IS_ERR(buf))
8120                 return buf;
8121
8122         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
8123         btrfs_tree_lock(buf);
8124         clean_tree_block(fs_info, buf);
8125         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
8126
8127         btrfs_set_lock_blocking(buf);
8128         set_extent_buffer_uptodate(buf);
8129
8130         memzero_extent_buffer(buf, 0, sizeof(struct btrfs_header));
8131         btrfs_set_header_level(buf, level);
8132         btrfs_set_header_bytenr(buf, buf->start);
8133         btrfs_set_header_generation(buf, trans->transid);
8134         btrfs_set_header_backref_rev(buf, BTRFS_MIXED_BACKREF_REV);
8135         btrfs_set_header_owner(buf, owner);
8136         write_extent_buffer_fsid(buf, fs_info->fsid);
8137         write_extent_buffer_chunk_tree_uuid(buf, fs_info->chunk_tree_uuid);
8138         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
8139                 buf->log_index = root->log_transid % 2;
8140                 /*
8141                  * we allow two log transactions at a time, use different
8142                  * EXENT bit to differentiate dirty pages.
8143                  */
8144                 if (buf->log_index == 0)
8145                         set_extent_dirty(&root->dirty_log_pages, buf->start,
8146                                         buf->start + buf->len - 1, GFP_NOFS);
8147                 else
8148                         set_extent_new(&root->dirty_log_pages, buf->start,
8149                                         buf->start + buf->len - 1);
8150         } else {
8151                 buf->log_index = -1;
8152                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
8153                          buf->start + buf->len - 1, GFP_NOFS);
8154         }
8155         trans->dirty = true;
8156         /* this returns a buffer locked for blocking */
8157         return buf;
8158 }
8159
8160 static struct btrfs_block_rsv *
8161 use_block_rsv(struct btrfs_trans_handle *trans,
8162               struct btrfs_root *root, u32 blocksize)
8163 {
8164         struct btrfs_fs_info *fs_info = root->fs_info;
8165         struct btrfs_block_rsv *block_rsv;
8166         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
8167         int ret;
8168         bool global_updated = false;
8169
8170         block_rsv = get_block_rsv(trans, root);
8171
8172         if (unlikely(block_rsv->size == 0))
8173                 goto try_reserve;
8174 again:
8175         ret = block_rsv_use_bytes(block_rsv, blocksize);
8176         if (!ret)
8177                 return block_rsv;
8178
8179         if (block_rsv->failfast)
8180                 return ERR_PTR(ret);
8181
8182         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
8183                 global_updated = true;
8184                 update_global_block_rsv(fs_info);
8185                 goto again;
8186         }
8187
8188         if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
8189                 static DEFINE_RATELIMIT_STATE(_rs,
8190                                 DEFAULT_RATELIMIT_INTERVAL * 10,
8191                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
8192                 if (__ratelimit(&_rs))
8193                         WARN(1, KERN_DEBUG
8194                                 "BTRFS: block rsv returned %d\n", ret);
8195         }
8196 try_reserve:
8197         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
8198                                      BTRFS_RESERVE_NO_FLUSH);
8199         if (!ret)
8200                 return block_rsv;
8201         /*
8202          * If we couldn't reserve metadata bytes try and use some from
8203          * the global reserve if its space type is the same as the global
8204          * reservation.
8205          */
8206         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
8207             block_rsv->space_info == global_rsv->space_info) {
8208                 ret = block_rsv_use_bytes(global_rsv, blocksize);
8209                 if (!ret)
8210                         return global_rsv;
8211         }
8212         return ERR_PTR(ret);
8213 }
8214
8215 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
8216                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
8217 {
8218         block_rsv_add_bytes(block_rsv, blocksize, 0);
8219         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0, NULL);
8220 }
8221
8222 /*
8223  * finds a free extent and does all the dirty work required for allocation
8224  * returns the tree buffer or an ERR_PTR on error.
8225  */
8226 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
8227                                              struct btrfs_root *root,
8228                                              u64 parent, u64 root_objectid,
8229                                              const struct btrfs_disk_key *key,
8230                                              int level, u64 hint,
8231                                              u64 empty_size)
8232 {
8233         struct btrfs_fs_info *fs_info = root->fs_info;
8234         struct btrfs_key ins;
8235         struct btrfs_block_rsv *block_rsv;
8236         struct extent_buffer *buf;
8237         struct btrfs_delayed_extent_op *extent_op;
8238         u64 flags = 0;
8239         int ret;
8240         u32 blocksize = fs_info->nodesize;
8241         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
8242
8243 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
8244         if (btrfs_is_testing(fs_info)) {
8245                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
8246                                             level, root_objectid);
8247                 if (!IS_ERR(buf))
8248                         root->alloc_bytenr += blocksize;
8249                 return buf;
8250         }
8251 #endif
8252
8253         block_rsv = use_block_rsv(trans, root, blocksize);
8254         if (IS_ERR(block_rsv))
8255                 return ERR_CAST(block_rsv);
8256
8257         ret = btrfs_reserve_extent(root, blocksize, blocksize, blocksize,
8258                                    empty_size, hint, &ins, 0, 0);
8259         if (ret)
8260                 goto out_unuse;
8261
8262         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level,
8263                                     root_objectid);
8264         if (IS_ERR(buf)) {
8265                 ret = PTR_ERR(buf);
8266                 goto out_free_reserved;
8267         }
8268
8269         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
8270                 if (parent == 0)
8271                         parent = ins.objectid;
8272                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
8273         } else
8274                 BUG_ON(parent > 0);
8275
8276         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
8277                 extent_op = btrfs_alloc_delayed_extent_op();
8278                 if (!extent_op) {
8279                         ret = -ENOMEM;
8280                         goto out_free_buf;
8281                 }
8282                 if (key)
8283                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
8284                 else
8285                         memset(&extent_op->key, 0, sizeof(extent_op->key));
8286                 extent_op->flags_to_set = flags;
8287                 extent_op->update_key = skinny_metadata ? false : true;
8288                 extent_op->update_flags = true;
8289                 extent_op->is_data = false;
8290                 extent_op->level = level;
8291
8292                 btrfs_ref_tree_mod(root, ins.objectid, ins.offset, parent,
8293                                    root_objectid, level, 0,
8294                                    BTRFS_ADD_DELAYED_EXTENT);
8295                 ret = btrfs_add_delayed_tree_ref(trans, ins.objectid,
8296                                                  ins.offset, parent,
8297                                                  root_objectid, level,
8298                                                  BTRFS_ADD_DELAYED_EXTENT,
8299                                                  extent_op, NULL, NULL);
8300                 if (ret)
8301                         goto out_free_delayed;
8302         }
8303         return buf;
8304
8305 out_free_delayed:
8306         btrfs_free_delayed_extent_op(extent_op);
8307 out_free_buf:
8308         free_extent_buffer(buf);
8309 out_free_reserved:
8310         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 0);
8311 out_unuse:
8312         unuse_block_rsv(fs_info, block_rsv, blocksize);
8313         return ERR_PTR(ret);
8314 }
8315
8316 struct walk_control {
8317         u64 refs[BTRFS_MAX_LEVEL];
8318         u64 flags[BTRFS_MAX_LEVEL];
8319         struct btrfs_key update_progress;
8320         int stage;
8321         int level;
8322         int shared_level;
8323         int update_ref;
8324         int keep_locks;
8325         int reada_slot;
8326         int reada_count;
8327 };
8328
8329 #define DROP_REFERENCE  1
8330 #define UPDATE_BACKREF  2
8331
8332 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
8333                                      struct btrfs_root *root,
8334                                      struct walk_control *wc,
8335                                      struct btrfs_path *path)
8336 {
8337         struct btrfs_fs_info *fs_info = root->fs_info;
8338         u64 bytenr;
8339         u64 generation;
8340         u64 refs;
8341         u64 flags;
8342         u32 nritems;
8343         struct btrfs_key key;
8344         struct extent_buffer *eb;
8345         int ret;
8346         int slot;
8347         int nread = 0;
8348
8349         if (path->slots[wc->level] < wc->reada_slot) {
8350                 wc->reada_count = wc->reada_count * 2 / 3;
8351                 wc->reada_count = max(wc->reada_count, 2);
8352         } else {
8353                 wc->reada_count = wc->reada_count * 3 / 2;
8354                 wc->reada_count = min_t(int, wc->reada_count,
8355                                         BTRFS_NODEPTRS_PER_BLOCK(fs_info));
8356         }
8357
8358         eb = path->nodes[wc->level];
8359         nritems = btrfs_header_nritems(eb);
8360
8361         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
8362                 if (nread >= wc->reada_count)
8363                         break;
8364
8365                 cond_resched();
8366                 bytenr = btrfs_node_blockptr(eb, slot);
8367                 generation = btrfs_node_ptr_generation(eb, slot);
8368
8369                 if (slot == path->slots[wc->level])
8370                         goto reada;
8371
8372                 if (wc->stage == UPDATE_BACKREF &&
8373                     generation <= root->root_key.offset)
8374                         continue;
8375
8376                 /* We don't lock the tree block, it's OK to be racy here */
8377                 ret = btrfs_lookup_extent_info(trans, fs_info, bytenr,
8378                                                wc->level - 1, 1, &refs,
8379                                                &flags);
8380                 /* We don't care about errors in readahead. */
8381                 if (ret < 0)
8382                         continue;
8383                 BUG_ON(refs == 0);
8384
8385                 if (wc->stage == DROP_REFERENCE) {
8386                         if (refs == 1)
8387                                 goto reada;
8388
8389                         if (wc->level == 1 &&
8390                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8391                                 continue;
8392                         if (!wc->update_ref ||
8393                             generation <= root->root_key.offset)
8394                                 continue;
8395                         btrfs_node_key_to_cpu(eb, &key, slot);
8396                         ret = btrfs_comp_cpu_keys(&key,
8397                                                   &wc->update_progress);
8398                         if (ret < 0)
8399                                 continue;
8400                 } else {
8401                         if (wc->level == 1 &&
8402                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8403                                 continue;
8404                 }
8405 reada:
8406                 readahead_tree_block(fs_info, bytenr);
8407                 nread++;
8408         }
8409         wc->reada_slot = slot;
8410 }
8411
8412 /*
8413  * helper to process tree block while walking down the tree.
8414  *
8415  * when wc->stage == UPDATE_BACKREF, this function updates
8416  * back refs for pointers in the block.
8417  *
8418  * NOTE: return value 1 means we should stop walking down.
8419  */
8420 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
8421                                    struct btrfs_root *root,
8422                                    struct btrfs_path *path,
8423                                    struct walk_control *wc, int lookup_info)
8424 {
8425         struct btrfs_fs_info *fs_info = root->fs_info;
8426         int level = wc->level;
8427         struct extent_buffer *eb = path->nodes[level];
8428         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8429         int ret;
8430
8431         if (wc->stage == UPDATE_BACKREF &&
8432             btrfs_header_owner(eb) != root->root_key.objectid)
8433                 return 1;
8434
8435         /*
8436          * when reference count of tree block is 1, it won't increase
8437          * again. once full backref flag is set, we never clear it.
8438          */
8439         if (lookup_info &&
8440             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
8441              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
8442                 BUG_ON(!path->locks[level]);
8443                 ret = btrfs_lookup_extent_info(trans, fs_info,
8444                                                eb->start, level, 1,
8445                                                &wc->refs[level],
8446                                                &wc->flags[level]);
8447                 BUG_ON(ret == -ENOMEM);
8448                 if (ret)
8449                         return ret;
8450                 BUG_ON(wc->refs[level] == 0);
8451         }
8452
8453         if (wc->stage == DROP_REFERENCE) {
8454                 if (wc->refs[level] > 1)
8455                         return 1;
8456
8457                 if (path->locks[level] && !wc->keep_locks) {
8458                         btrfs_tree_unlock_rw(eb, path->locks[level]);
8459                         path->locks[level] = 0;
8460                 }
8461                 return 0;
8462         }
8463
8464         /* wc->stage == UPDATE_BACKREF */
8465         if (!(wc->flags[level] & flag)) {
8466                 BUG_ON(!path->locks[level]);
8467                 ret = btrfs_inc_ref(trans, root, eb, 1);
8468                 BUG_ON(ret); /* -ENOMEM */
8469                 ret = btrfs_dec_ref(trans, root, eb, 0);
8470                 BUG_ON(ret); /* -ENOMEM */
8471                 ret = btrfs_set_disk_extent_flags(trans, fs_info, eb->start,
8472                                                   eb->len, flag,
8473                                                   btrfs_header_level(eb), 0);
8474                 BUG_ON(ret); /* -ENOMEM */
8475                 wc->flags[level] |= flag;
8476         }
8477
8478         /*
8479          * the block is shared by multiple trees, so it's not good to
8480          * keep the tree lock
8481          */
8482         if (path->locks[level] && level > 0) {
8483                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8484                 path->locks[level] = 0;
8485         }
8486         return 0;
8487 }
8488
8489 /*
8490  * helper to process tree block pointer.
8491  *
8492  * when wc->stage == DROP_REFERENCE, this function checks
8493  * reference count of the block pointed to. if the block
8494  * is shared and we need update back refs for the subtree
8495  * rooted at the block, this function changes wc->stage to
8496  * UPDATE_BACKREF. if the block is shared and there is no
8497  * need to update back, this function drops the reference
8498  * to the block.
8499  *
8500  * NOTE: return value 1 means we should stop walking down.
8501  */
8502 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
8503                                  struct btrfs_root *root,
8504                                  struct btrfs_path *path,
8505                                  struct walk_control *wc, int *lookup_info)
8506 {
8507         struct btrfs_fs_info *fs_info = root->fs_info;
8508         u64 bytenr;
8509         u64 generation;
8510         u64 parent;
8511         u32 blocksize;
8512         struct btrfs_key key;
8513         struct btrfs_key first_key;
8514         struct extent_buffer *next;
8515         int level = wc->level;
8516         int reada = 0;
8517         int ret = 0;
8518         bool need_account = false;
8519
8520         generation = btrfs_node_ptr_generation(path->nodes[level],
8521                                                path->slots[level]);
8522         /*
8523          * if the lower level block was created before the snapshot
8524          * was created, we know there is no need to update back refs
8525          * for the subtree
8526          */
8527         if (wc->stage == UPDATE_BACKREF &&
8528             generation <= root->root_key.offset) {
8529                 *lookup_info = 1;
8530                 return 1;
8531         }
8532
8533         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
8534         btrfs_node_key_to_cpu(path->nodes[level], &first_key,
8535                               path->slots[level]);
8536         blocksize = fs_info->nodesize;
8537
8538         next = find_extent_buffer(fs_info, bytenr);
8539         if (!next) {
8540                 next = btrfs_find_create_tree_block(fs_info, bytenr);
8541                 if (IS_ERR(next))
8542                         return PTR_ERR(next);
8543
8544                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
8545                                                level - 1);
8546                 reada = 1;
8547         }
8548         btrfs_tree_lock(next);
8549         btrfs_set_lock_blocking(next);
8550
8551         ret = btrfs_lookup_extent_info(trans, fs_info, bytenr, level - 1, 1,
8552                                        &wc->refs[level - 1],
8553                                        &wc->flags[level - 1]);
8554         if (ret < 0)
8555                 goto out_unlock;
8556
8557         if (unlikely(wc->refs[level - 1] == 0)) {
8558                 btrfs_err(fs_info, "Missing references.");
8559                 ret = -EIO;
8560                 goto out_unlock;
8561         }
8562         *lookup_info = 0;
8563
8564         if (wc->stage == DROP_REFERENCE) {
8565                 if (wc->refs[level - 1] > 1) {
8566                         need_account = true;
8567                         if (level == 1 &&
8568                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8569                                 goto skip;
8570
8571                         if (!wc->update_ref ||
8572                             generation <= root->root_key.offset)
8573                                 goto skip;
8574
8575                         btrfs_node_key_to_cpu(path->nodes[level], &key,
8576                                               path->slots[level]);
8577                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
8578                         if (ret < 0)
8579                                 goto skip;
8580
8581                         wc->stage = UPDATE_BACKREF;
8582                         wc->shared_level = level - 1;
8583                 }
8584         } else {
8585                 if (level == 1 &&
8586                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8587                         goto skip;
8588         }
8589
8590         if (!btrfs_buffer_uptodate(next, generation, 0)) {
8591                 btrfs_tree_unlock(next);
8592                 free_extent_buffer(next);
8593                 next = NULL;
8594                 *lookup_info = 1;
8595         }
8596
8597         if (!next) {
8598                 if (reada && level == 1)
8599                         reada_walk_down(trans, root, wc, path);
8600                 next = read_tree_block(fs_info, bytenr, generation, level - 1,
8601                                        &first_key);
8602                 if (IS_ERR(next)) {
8603                         return PTR_ERR(next);
8604                 } else if (!extent_buffer_uptodate(next)) {
8605                         free_extent_buffer(next);
8606                         return -EIO;
8607                 }
8608                 btrfs_tree_lock(next);
8609                 btrfs_set_lock_blocking(next);
8610         }
8611
8612         level--;
8613         ASSERT(level == btrfs_header_level(next));
8614         if (level != btrfs_header_level(next)) {
8615                 btrfs_err(root->fs_info, "mismatched level");
8616                 ret = -EIO;
8617                 goto out_unlock;
8618         }
8619         path->nodes[level] = next;
8620         path->slots[level] = 0;
8621         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8622         wc->level = level;
8623         if (wc->level == 1)
8624                 wc->reada_slot = 0;
8625         return 0;
8626 skip:
8627         wc->refs[level - 1] = 0;
8628         wc->flags[level - 1] = 0;
8629         if (wc->stage == DROP_REFERENCE) {
8630                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
8631                         parent = path->nodes[level]->start;
8632                 } else {
8633                         ASSERT(root->root_key.objectid ==
8634                                btrfs_header_owner(path->nodes[level]));
8635                         if (root->root_key.objectid !=
8636                             btrfs_header_owner(path->nodes[level])) {
8637                                 btrfs_err(root->fs_info,
8638                                                 "mismatched block owner");
8639                                 ret = -EIO;
8640                                 goto out_unlock;
8641                         }
8642                         parent = 0;
8643                 }
8644
8645                 if (need_account) {
8646                         ret = btrfs_qgroup_trace_subtree(trans, next,
8647                                                          generation, level - 1);
8648                         if (ret) {
8649                                 btrfs_err_rl(fs_info,
8650                                              "Error %d accounting shared subtree. Quota is out of sync, rescan required.",
8651                                              ret);
8652                         }
8653                 }
8654                 ret = btrfs_free_extent(trans, root, bytenr, blocksize,
8655                                         parent, root->root_key.objectid,
8656                                         level - 1, 0);
8657                 if (ret)
8658                         goto out_unlock;
8659         }
8660
8661         *lookup_info = 1;
8662         ret = 1;
8663
8664 out_unlock:
8665         btrfs_tree_unlock(next);
8666         free_extent_buffer(next);
8667
8668         return ret;
8669 }
8670
8671 /*
8672  * helper to process tree block while walking up the tree.
8673  *
8674  * when wc->stage == DROP_REFERENCE, this function drops
8675  * reference count on the block.
8676  *
8677  * when wc->stage == UPDATE_BACKREF, this function changes
8678  * wc->stage back to DROP_REFERENCE if we changed wc->stage
8679  * to UPDATE_BACKREF previously while processing the block.
8680  *
8681  * NOTE: return value 1 means we should stop walking up.
8682  */
8683 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
8684                                  struct btrfs_root *root,
8685                                  struct btrfs_path *path,
8686                                  struct walk_control *wc)
8687 {
8688         struct btrfs_fs_info *fs_info = root->fs_info;
8689         int ret;
8690         int level = wc->level;
8691         struct extent_buffer *eb = path->nodes[level];
8692         u64 parent = 0;
8693
8694         if (wc->stage == UPDATE_BACKREF) {
8695                 BUG_ON(wc->shared_level < level);
8696                 if (level < wc->shared_level)
8697                         goto out;
8698
8699                 ret = find_next_key(path, level + 1, &wc->update_progress);
8700                 if (ret > 0)
8701                         wc->update_ref = 0;
8702
8703                 wc->stage = DROP_REFERENCE;
8704                 wc->shared_level = -1;
8705                 path->slots[level] = 0;
8706
8707                 /*
8708                  * check reference count again if the block isn't locked.
8709                  * we should start walking down the tree again if reference
8710                  * count is one.
8711                  */
8712                 if (!path->locks[level]) {
8713                         BUG_ON(level == 0);
8714                         btrfs_tree_lock(eb);
8715                         btrfs_set_lock_blocking(eb);
8716                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8717
8718                         ret = btrfs_lookup_extent_info(trans, fs_info,
8719                                                        eb->start, level, 1,
8720                                                        &wc->refs[level],
8721                                                        &wc->flags[level]);
8722                         if (ret < 0) {
8723                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8724                                 path->locks[level] = 0;
8725                                 return ret;
8726                         }
8727                         BUG_ON(wc->refs[level] == 0);
8728                         if (wc->refs[level] == 1) {
8729                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8730                                 path->locks[level] = 0;
8731                                 return 1;
8732                         }
8733                 }
8734         }
8735
8736         /* wc->stage == DROP_REFERENCE */
8737         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
8738
8739         if (wc->refs[level] == 1) {
8740                 if (level == 0) {
8741                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8742                                 ret = btrfs_dec_ref(trans, root, eb, 1);
8743                         else
8744                                 ret = btrfs_dec_ref(trans, root, eb, 0);
8745                         BUG_ON(ret); /* -ENOMEM */
8746                         ret = btrfs_qgroup_trace_leaf_items(trans, eb);
8747                         if (ret) {
8748                                 btrfs_err_rl(fs_info,
8749                                              "error %d accounting leaf items. Quota is out of sync, rescan required.",
8750                                              ret);
8751                         }
8752                 }
8753                 /* make block locked assertion in clean_tree_block happy */
8754                 if (!path->locks[level] &&
8755                     btrfs_header_generation(eb) == trans->transid) {
8756                         btrfs_tree_lock(eb);
8757                         btrfs_set_lock_blocking(eb);
8758                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8759                 }
8760                 clean_tree_block(fs_info, eb);
8761         }
8762
8763         if (eb == root->node) {
8764                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8765                         parent = eb->start;
8766                 else
8767                         BUG_ON(root->root_key.objectid !=
8768                                btrfs_header_owner(eb));
8769         } else {
8770                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8771                         parent = path->nodes[level + 1]->start;
8772                 else
8773                         BUG_ON(root->root_key.objectid !=
8774                                btrfs_header_owner(path->nodes[level + 1]));
8775         }
8776
8777         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
8778 out:
8779         wc->refs[level] = 0;
8780         wc->flags[level] = 0;
8781         return 0;
8782 }
8783
8784 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
8785                                    struct btrfs_root *root,
8786                                    struct btrfs_path *path,
8787                                    struct walk_control *wc)
8788 {
8789         int level = wc->level;
8790         int lookup_info = 1;
8791         int ret;
8792
8793         while (level >= 0) {
8794                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
8795                 if (ret > 0)
8796                         break;
8797
8798                 if (level == 0)
8799                         break;
8800
8801                 if (path->slots[level] >=
8802                     btrfs_header_nritems(path->nodes[level]))
8803                         break;
8804
8805                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
8806                 if (ret > 0) {
8807                         path->slots[level]++;
8808                         continue;
8809                 } else if (ret < 0)
8810                         return ret;
8811                 level = wc->level;
8812         }
8813         return 0;
8814 }
8815
8816 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
8817                                  struct btrfs_root *root,
8818                                  struct btrfs_path *path,
8819                                  struct walk_control *wc, int max_level)
8820 {
8821         int level = wc->level;
8822         int ret;
8823
8824         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
8825         while (level < max_level && path->nodes[level]) {
8826                 wc->level = level;
8827                 if (path->slots[level] + 1 <
8828                     btrfs_header_nritems(path->nodes[level])) {
8829                         path->slots[level]++;
8830                         return 0;
8831                 } else {
8832                         ret = walk_up_proc(trans, root, path, wc);
8833                         if (ret > 0)
8834                                 return 0;
8835
8836                         if (path->locks[level]) {
8837                                 btrfs_tree_unlock_rw(path->nodes[level],
8838                                                      path->locks[level]);
8839                                 path->locks[level] = 0;
8840                         }
8841                         free_extent_buffer(path->nodes[level]);
8842                         path->nodes[level] = NULL;
8843                         level++;
8844                 }
8845         }
8846         return 1;
8847 }
8848
8849 /*
8850  * drop a subvolume tree.
8851  *
8852  * this function traverses the tree freeing any blocks that only
8853  * referenced by the tree.
8854  *
8855  * when a shared tree block is found. this function decreases its
8856  * reference count by one. if update_ref is true, this function
8857  * also make sure backrefs for the shared block and all lower level
8858  * blocks are properly updated.
8859  *
8860  * If called with for_reloc == 0, may exit early with -EAGAIN
8861  */
8862 int btrfs_drop_snapshot(struct btrfs_root *root,
8863                          struct btrfs_block_rsv *block_rsv, int update_ref,
8864                          int for_reloc)
8865 {
8866         struct btrfs_fs_info *fs_info = root->fs_info;
8867         struct btrfs_path *path;
8868         struct btrfs_trans_handle *trans;
8869         struct btrfs_root *tree_root = fs_info->tree_root;
8870         struct btrfs_root_item *root_item = &root->root_item;
8871         struct walk_control *wc;
8872         struct btrfs_key key;
8873         int err = 0;
8874         int ret;
8875         int level;
8876         bool root_dropped = false;
8877
8878         btrfs_debug(fs_info, "Drop subvolume %llu", root->objectid);
8879
8880         path = btrfs_alloc_path();
8881         if (!path) {
8882                 err = -ENOMEM;
8883                 goto out;
8884         }
8885
8886         wc = kzalloc(sizeof(*wc), GFP_NOFS);
8887         if (!wc) {
8888                 btrfs_free_path(path);
8889                 err = -ENOMEM;
8890                 goto out;
8891         }
8892
8893         trans = btrfs_start_transaction(tree_root, 0);
8894         if (IS_ERR(trans)) {
8895                 err = PTR_ERR(trans);
8896                 goto out_free;
8897         }
8898
8899         if (block_rsv)
8900                 trans->block_rsv = block_rsv;
8901
8902         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
8903                 level = btrfs_header_level(root->node);
8904                 path->nodes[level] = btrfs_lock_root_node(root);
8905                 btrfs_set_lock_blocking(path->nodes[level]);
8906                 path->slots[level] = 0;
8907                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8908                 memset(&wc->update_progress, 0,
8909                        sizeof(wc->update_progress));
8910         } else {
8911                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
8912                 memcpy(&wc->update_progress, &key,
8913                        sizeof(wc->update_progress));
8914
8915                 level = root_item->drop_level;
8916                 BUG_ON(level == 0);
8917                 path->lowest_level = level;
8918                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
8919                 path->lowest_level = 0;
8920                 if (ret < 0) {
8921                         err = ret;
8922                         goto out_end_trans;
8923                 }
8924                 WARN_ON(ret > 0);
8925
8926                 /*
8927                  * unlock our path, this is safe because only this
8928                  * function is allowed to delete this snapshot
8929                  */
8930                 btrfs_unlock_up_safe(path, 0);
8931
8932                 level = btrfs_header_level(root->node);
8933                 while (1) {
8934                         btrfs_tree_lock(path->nodes[level]);
8935                         btrfs_set_lock_blocking(path->nodes[level]);
8936                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8937
8938                         ret = btrfs_lookup_extent_info(trans, fs_info,
8939                                                 path->nodes[level]->start,
8940                                                 level, 1, &wc->refs[level],
8941                                                 &wc->flags[level]);
8942                         if (ret < 0) {
8943                                 err = ret;
8944                                 goto out_end_trans;
8945                         }
8946                         BUG_ON(wc->refs[level] == 0);
8947
8948                         if (level == root_item->drop_level)
8949                                 break;
8950
8951                         btrfs_tree_unlock(path->nodes[level]);
8952                         path->locks[level] = 0;
8953                         WARN_ON(wc->refs[level] != 1);
8954                         level--;
8955                 }
8956         }
8957
8958         wc->level = level;
8959         wc->shared_level = -1;
8960         wc->stage = DROP_REFERENCE;
8961         wc->update_ref = update_ref;
8962         wc->keep_locks = 0;
8963         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
8964
8965         while (1) {
8966
8967                 ret = walk_down_tree(trans, root, path, wc);
8968                 if (ret < 0) {
8969                         err = ret;
8970                         break;
8971                 }
8972
8973                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
8974                 if (ret < 0) {
8975                         err = ret;
8976                         break;
8977                 }
8978
8979                 if (ret > 0) {
8980                         BUG_ON(wc->stage != DROP_REFERENCE);
8981                         break;
8982                 }
8983
8984                 if (wc->stage == DROP_REFERENCE) {
8985                         level = wc->level;
8986                         btrfs_node_key(path->nodes[level],
8987                                        &root_item->drop_progress,
8988                                        path->slots[level]);
8989                         root_item->drop_level = level;
8990                 }
8991
8992                 BUG_ON(wc->level == 0);
8993                 if (btrfs_should_end_transaction(trans) ||
8994                     (!for_reloc && btrfs_need_cleaner_sleep(fs_info))) {
8995                         ret = btrfs_update_root(trans, tree_root,
8996                                                 &root->root_key,
8997                                                 root_item);
8998                         if (ret) {
8999                                 btrfs_abort_transaction(trans, ret);
9000                                 err = ret;
9001                                 goto out_end_trans;
9002                         }
9003
9004                         btrfs_end_transaction_throttle(trans);
9005                         if (!for_reloc && btrfs_need_cleaner_sleep(fs_info)) {
9006                                 btrfs_debug(fs_info,
9007                                             "drop snapshot early exit");
9008                                 err = -EAGAIN;
9009                                 goto out_free;
9010                         }
9011
9012                         trans = btrfs_start_transaction(tree_root, 0);
9013                         if (IS_ERR(trans)) {
9014                                 err = PTR_ERR(trans);
9015                                 goto out_free;
9016                         }
9017                         if (block_rsv)
9018                                 trans->block_rsv = block_rsv;
9019                 }
9020         }
9021         btrfs_release_path(path);
9022         if (err)
9023                 goto out_end_trans;
9024
9025         ret = btrfs_del_root(trans, &root->root_key);
9026         if (ret) {
9027                 btrfs_abort_transaction(trans, ret);
9028                 err = ret;
9029                 goto out_end_trans;
9030         }
9031
9032         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
9033                 ret = btrfs_find_root(tree_root, &root->root_key, path,
9034                                       NULL, NULL);
9035                 if (ret < 0) {
9036                         btrfs_abort_transaction(trans, ret);
9037                         err = ret;
9038                         goto out_end_trans;
9039                 } else if (ret > 0) {
9040                         /* if we fail to delete the orphan item this time
9041                          * around, it'll get picked up the next time.
9042                          *
9043                          * The most common failure here is just -ENOENT.
9044                          */
9045                         btrfs_del_orphan_item(trans, tree_root,
9046                                               root->root_key.objectid);
9047                 }
9048         }
9049
9050         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
9051                 btrfs_add_dropped_root(trans, root);
9052         } else {
9053                 free_extent_buffer(root->node);
9054                 free_extent_buffer(root->commit_root);
9055                 btrfs_put_fs_root(root);
9056         }
9057         root_dropped = true;
9058 out_end_trans:
9059         btrfs_end_transaction_throttle(trans);
9060 out_free:
9061         kfree(wc);
9062         btrfs_free_path(path);
9063 out:
9064         /*
9065          * So if we need to stop dropping the snapshot for whatever reason we
9066          * need to make sure to add it back to the dead root list so that we
9067          * keep trying to do the work later.  This also cleans up roots if we
9068          * don't have it in the radix (like when we recover after a power fail
9069          * or unmount) so we don't leak memory.
9070          */
9071         if (!for_reloc && !root_dropped)
9072                 btrfs_add_dead_root(root);
9073         if (err && err != -EAGAIN)
9074                 btrfs_handle_fs_error(fs_info, err, NULL);
9075         return err;
9076 }
9077
9078 /*
9079  * drop subtree rooted at tree block 'node'.
9080  *
9081  * NOTE: this function will unlock and release tree block 'node'
9082  * only used by relocation code
9083  */
9084 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
9085                         struct btrfs_root *root,
9086                         struct extent_buffer *node,
9087                         struct extent_buffer *parent)
9088 {
9089         struct btrfs_fs_info *fs_info = root->fs_info;
9090         struct btrfs_path *path;
9091         struct walk_control *wc;
9092         int level;
9093         int parent_level;
9094         int ret = 0;
9095         int wret;
9096
9097         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
9098
9099         path = btrfs_alloc_path();
9100         if (!path)
9101                 return -ENOMEM;
9102
9103         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9104         if (!wc) {
9105                 btrfs_free_path(path);
9106                 return -ENOMEM;
9107         }
9108
9109         btrfs_assert_tree_locked(parent);
9110         parent_level = btrfs_header_level(parent);
9111         extent_buffer_get(parent);
9112         path->nodes[parent_level] = parent;
9113         path->slots[parent_level] = btrfs_header_nritems(parent);
9114
9115         btrfs_assert_tree_locked(node);
9116         level = btrfs_header_level(node);
9117         path->nodes[level] = node;
9118         path->slots[level] = 0;
9119         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9120
9121         wc->refs[parent_level] = 1;
9122         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
9123         wc->level = level;
9124         wc->shared_level = -1;
9125         wc->stage = DROP_REFERENCE;
9126         wc->update_ref = 0;
9127         wc->keep_locks = 1;
9128         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
9129
9130         while (1) {
9131                 wret = walk_down_tree(trans, root, path, wc);
9132                 if (wret < 0) {
9133                         ret = wret;
9134                         break;
9135                 }
9136
9137                 wret = walk_up_tree(trans, root, path, wc, parent_level);
9138                 if (wret < 0)
9139                         ret = wret;
9140                 if (wret != 0)
9141                         break;
9142         }
9143
9144         kfree(wc);
9145         btrfs_free_path(path);
9146         return ret;
9147 }
9148
9149 static u64 update_block_group_flags(struct btrfs_fs_info *fs_info, u64 flags)
9150 {
9151         u64 num_devices;
9152         u64 stripped;
9153
9154         /*
9155          * if restripe for this chunk_type is on pick target profile and
9156          * return, otherwise do the usual balance
9157          */
9158         stripped = get_restripe_target(fs_info, flags);
9159         if (stripped)
9160                 return extended_to_chunk(stripped);
9161
9162         num_devices = fs_info->fs_devices->rw_devices;
9163
9164         stripped = BTRFS_BLOCK_GROUP_RAID0 |
9165                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
9166                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
9167
9168         if (num_devices == 1) {
9169                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9170                 stripped = flags & ~stripped;
9171
9172                 /* turn raid0 into single device chunks */
9173                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
9174                         return stripped;
9175
9176                 /* turn mirroring into duplication */
9177                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
9178                              BTRFS_BLOCK_GROUP_RAID10))
9179                         return stripped | BTRFS_BLOCK_GROUP_DUP;
9180         } else {
9181                 /* they already had raid on here, just return */
9182                 if (flags & stripped)
9183                         return flags;
9184
9185                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9186                 stripped = flags & ~stripped;
9187
9188                 /* switch duplicated blocks with raid1 */
9189                 if (flags & BTRFS_BLOCK_GROUP_DUP)
9190                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
9191
9192                 /* this is drive concat, leave it alone */
9193         }
9194
9195         return flags;
9196 }
9197
9198 static int inc_block_group_ro(struct btrfs_block_group_cache *cache, int force)
9199 {
9200         struct btrfs_space_info *sinfo = cache->space_info;
9201         u64 num_bytes;
9202         u64 min_allocable_bytes;
9203         int ret = -ENOSPC;
9204
9205         /*
9206          * We need some metadata space and system metadata space for
9207          * allocating chunks in some corner cases until we force to set
9208          * it to be readonly.
9209          */
9210         if ((sinfo->flags &
9211              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
9212             !force)
9213                 min_allocable_bytes = SZ_1M;
9214         else
9215                 min_allocable_bytes = 0;
9216
9217         spin_lock(&sinfo->lock);
9218         spin_lock(&cache->lock);
9219
9220         if (cache->ro) {
9221                 cache->ro++;
9222                 ret = 0;
9223                 goto out;
9224         }
9225
9226         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
9227                     cache->bytes_super - btrfs_block_group_used(&cache->item);
9228
9229         if (btrfs_space_info_used(sinfo, true) + num_bytes +
9230             min_allocable_bytes <= sinfo->total_bytes) {
9231                 sinfo->bytes_readonly += num_bytes;
9232                 cache->ro++;
9233                 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
9234                 ret = 0;
9235         }
9236 out:
9237         spin_unlock(&cache->lock);
9238         spin_unlock(&sinfo->lock);
9239         return ret;
9240 }
9241
9242 int btrfs_inc_block_group_ro(struct btrfs_block_group_cache *cache)
9243
9244 {
9245         struct btrfs_fs_info *fs_info = cache->fs_info;
9246         struct btrfs_trans_handle *trans;
9247         u64 alloc_flags;
9248         int ret;
9249
9250 again:
9251         trans = btrfs_join_transaction(fs_info->extent_root);
9252         if (IS_ERR(trans))
9253                 return PTR_ERR(trans);
9254
9255         /*
9256          * we're not allowed to set block groups readonly after the dirty
9257          * block groups cache has started writing.  If it already started,
9258          * back off and let this transaction commit
9259          */
9260         mutex_lock(&fs_info->ro_block_group_mutex);
9261         if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
9262                 u64 transid = trans->transid;
9263
9264                 mutex_unlock(&fs_info->ro_block_group_mutex);
9265                 btrfs_end_transaction(trans);
9266
9267                 ret = btrfs_wait_for_commit(fs_info, transid);
9268                 if (ret)
9269                         return ret;
9270                 goto again;
9271         }
9272
9273         /*
9274          * if we are changing raid levels, try to allocate a corresponding
9275          * block group with the new raid level.
9276          */
9277         alloc_flags = update_block_group_flags(fs_info, cache->flags);
9278         if (alloc_flags != cache->flags) {
9279                 ret = do_chunk_alloc(trans, alloc_flags,
9280                                      CHUNK_ALLOC_FORCE);
9281                 /*
9282                  * ENOSPC is allowed here, we may have enough space
9283                  * already allocated at the new raid level to
9284                  * carry on
9285                  */
9286                 if (ret == -ENOSPC)
9287                         ret = 0;
9288                 if (ret < 0)
9289                         goto out;
9290         }
9291
9292         ret = inc_block_group_ro(cache, 0);
9293         if (!ret)
9294                 goto out;
9295         alloc_flags = get_alloc_profile(fs_info, cache->space_info->flags);
9296         ret = do_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
9297         if (ret < 0)
9298                 goto out;
9299         ret = inc_block_group_ro(cache, 0);
9300 out:
9301         if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
9302                 alloc_flags = update_block_group_flags(fs_info, cache->flags);
9303                 mutex_lock(&fs_info->chunk_mutex);
9304                 check_system_chunk(trans, alloc_flags);
9305                 mutex_unlock(&fs_info->chunk_mutex);
9306         }
9307         mutex_unlock(&fs_info->ro_block_group_mutex);
9308
9309         btrfs_end_transaction(trans);
9310         return ret;
9311 }
9312
9313 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans, u64 type)
9314 {
9315         u64 alloc_flags = get_alloc_profile(trans->fs_info, type);
9316
9317         return do_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
9318 }
9319
9320 /*
9321  * helper to account the unused space of all the readonly block group in the
9322  * space_info. takes mirrors into account.
9323  */
9324 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
9325 {
9326         struct btrfs_block_group_cache *block_group;
9327         u64 free_bytes = 0;
9328         int factor;
9329
9330         /* It's df, we don't care if it's racy */
9331         if (list_empty(&sinfo->ro_bgs))
9332                 return 0;
9333
9334         spin_lock(&sinfo->lock);
9335         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
9336                 spin_lock(&block_group->lock);
9337
9338                 if (!block_group->ro) {
9339                         spin_unlock(&block_group->lock);
9340                         continue;
9341                 }
9342
9343                 factor = btrfs_bg_type_to_factor(block_group->flags);
9344                 free_bytes += (block_group->key.offset -
9345                                btrfs_block_group_used(&block_group->item)) *
9346                                factor;
9347
9348                 spin_unlock(&block_group->lock);
9349         }
9350         spin_unlock(&sinfo->lock);
9351
9352         return free_bytes;
9353 }
9354
9355 void btrfs_dec_block_group_ro(struct btrfs_block_group_cache *cache)
9356 {
9357         struct btrfs_space_info *sinfo = cache->space_info;
9358         u64 num_bytes;
9359
9360         BUG_ON(!cache->ro);
9361
9362         spin_lock(&sinfo->lock);
9363         spin_lock(&cache->lock);
9364         if (!--cache->ro) {
9365                 num_bytes = cache->key.offset - cache->reserved -
9366                             cache->pinned - cache->bytes_super -
9367                             btrfs_block_group_used(&cache->item);
9368                 sinfo->bytes_readonly -= num_bytes;
9369                 list_del_init(&cache->ro_list);
9370         }
9371         spin_unlock(&cache->lock);
9372         spin_unlock(&sinfo->lock);
9373 }
9374
9375 /*
9376  * checks to see if its even possible to relocate this block group.
9377  *
9378  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
9379  * ok to go ahead and try.
9380  */
9381 int btrfs_can_relocate(struct btrfs_fs_info *fs_info, u64 bytenr)
9382 {
9383         struct btrfs_root *root = fs_info->extent_root;
9384         struct btrfs_block_group_cache *block_group;
9385         struct btrfs_space_info *space_info;
9386         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
9387         struct btrfs_device *device;
9388         struct btrfs_trans_handle *trans;
9389         u64 min_free;
9390         u64 dev_min = 1;
9391         u64 dev_nr = 0;
9392         u64 target;
9393         int debug;
9394         int index;
9395         int full = 0;
9396         int ret = 0;
9397
9398         debug = btrfs_test_opt(fs_info, ENOSPC_DEBUG);
9399
9400         block_group = btrfs_lookup_block_group(fs_info, bytenr);
9401
9402         /* odd, couldn't find the block group, leave it alone */
9403         if (!block_group) {
9404                 if (debug)
9405                         btrfs_warn(fs_info,
9406                                    "can't find block group for bytenr %llu",
9407                                    bytenr);
9408                 return -1;
9409         }
9410
9411         min_free = btrfs_block_group_used(&block_group->item);
9412
9413         /* no bytes used, we're good */
9414         if (!min_free)
9415                 goto out;
9416
9417         space_info = block_group->space_info;
9418         spin_lock(&space_info->lock);
9419
9420         full = space_info->full;
9421
9422         /*
9423          * if this is the last block group we have in this space, we can't
9424          * relocate it unless we're able to allocate a new chunk below.
9425          *
9426          * Otherwise, we need to make sure we have room in the space to handle
9427          * all of the extents from this block group.  If we can, we're good
9428          */
9429         if ((space_info->total_bytes != block_group->key.offset) &&
9430             (btrfs_space_info_used(space_info, false) + min_free <
9431              space_info->total_bytes)) {
9432                 spin_unlock(&space_info->lock);
9433                 goto out;
9434         }
9435         spin_unlock(&space_info->lock);
9436
9437         /*
9438          * ok we don't have enough space, but maybe we have free space on our
9439          * devices to allocate new chunks for relocation, so loop through our
9440          * alloc devices and guess if we have enough space.  if this block
9441          * group is going to be restriped, run checks against the target
9442          * profile instead of the current one.
9443          */
9444         ret = -1;
9445
9446         /*
9447          * index:
9448          *      0: raid10
9449          *      1: raid1
9450          *      2: dup
9451          *      3: raid0
9452          *      4: single
9453          */
9454         target = get_restripe_target(fs_info, block_group->flags);
9455         if (target) {
9456                 index = btrfs_bg_flags_to_raid_index(extended_to_chunk(target));
9457         } else {
9458                 /*
9459                  * this is just a balance, so if we were marked as full
9460                  * we know there is no space for a new chunk
9461                  */
9462                 if (full) {
9463                         if (debug)
9464                                 btrfs_warn(fs_info,
9465                                            "no space to alloc new chunk for block group %llu",
9466                                            block_group->key.objectid);
9467                         goto out;
9468                 }
9469
9470                 index = btrfs_bg_flags_to_raid_index(block_group->flags);
9471         }
9472
9473         if (index == BTRFS_RAID_RAID10) {
9474                 dev_min = 4;
9475                 /* Divide by 2 */
9476                 min_free >>= 1;
9477         } else if (index == BTRFS_RAID_RAID1) {
9478                 dev_min = 2;
9479         } else if (index == BTRFS_RAID_DUP) {
9480                 /* Multiply by 2 */
9481                 min_free <<= 1;
9482         } else if (index == BTRFS_RAID_RAID0) {
9483                 dev_min = fs_devices->rw_devices;
9484                 min_free = div64_u64(min_free, dev_min);
9485         }
9486
9487         /* We need to do this so that we can look at pending chunks */
9488         trans = btrfs_join_transaction(root);
9489         if (IS_ERR(trans)) {
9490                 ret = PTR_ERR(trans);
9491                 goto out;
9492         }
9493
9494         mutex_lock(&fs_info->chunk_mutex);
9495         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
9496                 u64 dev_offset;
9497
9498                 /*
9499                  * check to make sure we can actually find a chunk with enough
9500                  * space to fit our block group in.
9501                  */
9502                 if (device->total_bytes > device->bytes_used + min_free &&
9503                     !test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state)) {
9504                         ret = find_free_dev_extent(trans, device, min_free,
9505                                                    &dev_offset, NULL);
9506                         if (!ret)
9507                                 dev_nr++;
9508
9509                         if (dev_nr >= dev_min)
9510                                 break;
9511
9512                         ret = -1;
9513                 }
9514         }
9515         if (debug && ret == -1)
9516                 btrfs_warn(fs_info,
9517                            "no space to allocate a new chunk for block group %llu",
9518                            block_group->key.objectid);
9519         mutex_unlock(&fs_info->chunk_mutex);
9520         btrfs_end_transaction(trans);
9521 out:
9522         btrfs_put_block_group(block_group);
9523         return ret;
9524 }
9525
9526 static int find_first_block_group(struct btrfs_fs_info *fs_info,
9527                                   struct btrfs_path *path,
9528                                   struct btrfs_key *key)
9529 {
9530         struct btrfs_root *root = fs_info->extent_root;
9531         int ret = 0;
9532         struct btrfs_key found_key;
9533         struct extent_buffer *leaf;
9534         struct btrfs_block_group_item bg;
9535         u64 flags;
9536         int slot;
9537
9538         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
9539         if (ret < 0)
9540                 goto out;
9541
9542         while (1) {
9543                 slot = path->slots[0];
9544                 leaf = path->nodes[0];
9545                 if (slot >= btrfs_header_nritems(leaf)) {
9546                         ret = btrfs_next_leaf(root, path);
9547                         if (ret == 0)
9548                                 continue;
9549                         if (ret < 0)
9550                                 goto out;
9551                         break;
9552                 }
9553                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
9554
9555                 if (found_key.objectid >= key->objectid &&
9556                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
9557                         struct extent_map_tree *em_tree;
9558                         struct extent_map *em;
9559
9560                         em_tree = &root->fs_info->mapping_tree.map_tree;
9561                         read_lock(&em_tree->lock);
9562                         em = lookup_extent_mapping(em_tree, found_key.objectid,
9563                                                    found_key.offset);
9564                         read_unlock(&em_tree->lock);
9565                         if (!em) {
9566                                 btrfs_err(fs_info,
9567                         "logical %llu len %llu found bg but no related chunk",
9568                                           found_key.objectid, found_key.offset);
9569                                 ret = -ENOENT;
9570                         } else if (em->start != found_key.objectid ||
9571                                    em->len != found_key.offset) {
9572                                 btrfs_err(fs_info,
9573                 "block group %llu len %llu mismatch with chunk %llu len %llu",
9574                                           found_key.objectid, found_key.offset,
9575                                           em->start, em->len);
9576                                 ret = -EUCLEAN;
9577                         } else {
9578                                 read_extent_buffer(leaf, &bg,
9579                                         btrfs_item_ptr_offset(leaf, slot),
9580                                         sizeof(bg));
9581                                 flags = btrfs_block_group_flags(&bg) &
9582                                         BTRFS_BLOCK_GROUP_TYPE_MASK;
9583
9584                                 if (flags != (em->map_lookup->type &
9585                                               BTRFS_BLOCK_GROUP_TYPE_MASK)) {
9586                                         btrfs_err(fs_info,
9587 "block group %llu len %llu type flags 0x%llx mismatch with chunk type flags 0x%llx",
9588                                                 found_key.objectid,
9589                                                 found_key.offset, flags,
9590                                                 (BTRFS_BLOCK_GROUP_TYPE_MASK &
9591                                                  em->map_lookup->type));
9592                                         ret = -EUCLEAN;
9593                                 } else {
9594                                         ret = 0;
9595                                 }
9596                         }
9597                         free_extent_map(em);
9598                         goto out;
9599                 }
9600                 path->slots[0]++;
9601         }
9602 out:
9603         return ret;
9604 }
9605
9606 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
9607 {
9608         struct btrfs_block_group_cache *block_group;
9609         u64 last = 0;
9610
9611         while (1) {
9612                 struct inode *inode;
9613
9614                 block_group = btrfs_lookup_first_block_group(info, last);
9615                 while (block_group) {
9616                         spin_lock(&block_group->lock);
9617                         if (block_group->iref)
9618                                 break;
9619                         spin_unlock(&block_group->lock);
9620                         block_group = next_block_group(info, block_group);
9621                 }
9622                 if (!block_group) {
9623                         if (last == 0)
9624                                 break;
9625                         last = 0;
9626                         continue;
9627                 }
9628
9629                 inode = block_group->inode;
9630                 block_group->iref = 0;
9631                 block_group->inode = NULL;
9632                 spin_unlock(&block_group->lock);
9633                 ASSERT(block_group->io_ctl.inode == NULL);
9634                 iput(inode);
9635                 last = block_group->key.objectid + block_group->key.offset;
9636                 btrfs_put_block_group(block_group);
9637         }
9638 }
9639
9640 /*
9641  * Must be called only after stopping all workers, since we could have block
9642  * group caching kthreads running, and therefore they could race with us if we
9643  * freed the block groups before stopping them.
9644  */
9645 int btrfs_free_block_groups(struct btrfs_fs_info *info)
9646 {
9647         struct btrfs_block_group_cache *block_group;
9648         struct btrfs_space_info *space_info;
9649         struct btrfs_caching_control *caching_ctl;
9650         struct rb_node *n;
9651
9652         down_write(&info->commit_root_sem);
9653         while (!list_empty(&info->caching_block_groups)) {
9654                 caching_ctl = list_entry(info->caching_block_groups.next,
9655                                          struct btrfs_caching_control, list);
9656                 list_del(&caching_ctl->list);
9657                 put_caching_control(caching_ctl);
9658         }
9659         up_write(&info->commit_root_sem);
9660
9661         spin_lock(&info->unused_bgs_lock);
9662         while (!list_empty(&info->unused_bgs)) {
9663                 block_group = list_first_entry(&info->unused_bgs,
9664                                                struct btrfs_block_group_cache,
9665                                                bg_list);
9666                 list_del_init(&block_group->bg_list);
9667                 btrfs_put_block_group(block_group);
9668         }
9669         spin_unlock(&info->unused_bgs_lock);
9670
9671         spin_lock(&info->block_group_cache_lock);
9672         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
9673                 block_group = rb_entry(n, struct btrfs_block_group_cache,
9674                                        cache_node);
9675                 rb_erase(&block_group->cache_node,
9676                          &info->block_group_cache_tree);
9677                 RB_CLEAR_NODE(&block_group->cache_node);
9678                 spin_unlock(&info->block_group_cache_lock);
9679
9680                 down_write(&block_group->space_info->groups_sem);
9681                 list_del(&block_group->list);
9682                 up_write(&block_group->space_info->groups_sem);
9683
9684                 /*
9685                  * We haven't cached this block group, which means we could
9686                  * possibly have excluded extents on this block group.
9687                  */
9688                 if (block_group->cached == BTRFS_CACHE_NO ||
9689                     block_group->cached == BTRFS_CACHE_ERROR)
9690                         free_excluded_extents(block_group);
9691
9692                 btrfs_remove_free_space_cache(block_group);
9693                 ASSERT(block_group->cached != BTRFS_CACHE_STARTED);
9694                 ASSERT(list_empty(&block_group->dirty_list));
9695                 ASSERT(list_empty(&block_group->io_list));
9696                 ASSERT(list_empty(&block_group->bg_list));
9697                 ASSERT(atomic_read(&block_group->count) == 1);
9698                 btrfs_put_block_group(block_group);
9699
9700                 spin_lock(&info->block_group_cache_lock);
9701         }
9702         spin_unlock(&info->block_group_cache_lock);
9703
9704         /* now that all the block groups are freed, go through and
9705          * free all the space_info structs.  This is only called during
9706          * the final stages of unmount, and so we know nobody is
9707          * using them.  We call synchronize_rcu() once before we start,
9708          * just to be on the safe side.
9709          */
9710         synchronize_rcu();
9711
9712         release_global_block_rsv(info);
9713
9714         while (!list_empty(&info->space_info)) {
9715                 int i;
9716
9717                 space_info = list_entry(info->space_info.next,
9718                                         struct btrfs_space_info,
9719                                         list);
9720
9721                 /*
9722                  * Do not hide this behind enospc_debug, this is actually
9723                  * important and indicates a real bug if this happens.
9724                  */
9725                 if (WARN_ON(space_info->bytes_pinned > 0 ||
9726                             space_info->bytes_reserved > 0 ||
9727                             space_info->bytes_may_use > 0))
9728                         dump_space_info(info, space_info, 0, 0);
9729                 list_del(&space_info->list);
9730                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
9731                         struct kobject *kobj;
9732                         kobj = space_info->block_group_kobjs[i];
9733                         space_info->block_group_kobjs[i] = NULL;
9734                         if (kobj) {
9735                                 kobject_del(kobj);
9736                                 kobject_put(kobj);
9737                         }
9738                 }
9739                 kobject_del(&space_info->kobj);
9740                 kobject_put(&space_info->kobj);
9741         }
9742         return 0;
9743 }
9744
9745 /* link_block_group will queue up kobjects to add when we're reclaim-safe */
9746 void btrfs_add_raid_kobjects(struct btrfs_fs_info *fs_info)
9747 {
9748         struct btrfs_space_info *space_info;
9749         struct raid_kobject *rkobj;
9750         LIST_HEAD(list);
9751         int index;
9752         int ret = 0;
9753
9754         spin_lock(&fs_info->pending_raid_kobjs_lock);
9755         list_splice_init(&fs_info->pending_raid_kobjs, &list);
9756         spin_unlock(&fs_info->pending_raid_kobjs_lock);
9757
9758         list_for_each_entry(rkobj, &list, list) {
9759                 space_info = __find_space_info(fs_info, rkobj->flags);
9760                 index = btrfs_bg_flags_to_raid_index(rkobj->flags);
9761
9762                 ret = kobject_add(&rkobj->kobj, &space_info->kobj,
9763                                   "%s", get_raid_name(index));
9764                 if (ret) {
9765                         kobject_put(&rkobj->kobj);
9766                         break;
9767                 }
9768         }
9769         if (ret)
9770                 btrfs_warn(fs_info,
9771                            "failed to add kobject for block cache, ignoring");
9772 }
9773
9774 static void link_block_group(struct btrfs_block_group_cache *cache)
9775 {
9776         struct btrfs_space_info *space_info = cache->space_info;
9777         struct btrfs_fs_info *fs_info = cache->fs_info;
9778         int index = btrfs_bg_flags_to_raid_index(cache->flags);
9779         bool first = false;
9780
9781         down_write(&space_info->groups_sem);
9782         if (list_empty(&space_info->block_groups[index]))
9783                 first = true;
9784         list_add_tail(&cache->list, &space_info->block_groups[index]);
9785         up_write(&space_info->groups_sem);
9786
9787         if (first) {
9788                 struct raid_kobject *rkobj = kzalloc(sizeof(*rkobj), GFP_NOFS);
9789                 if (!rkobj) {
9790                         btrfs_warn(cache->fs_info,
9791                                 "couldn't alloc memory for raid level kobject");
9792                         return;
9793                 }
9794                 rkobj->flags = cache->flags;
9795                 kobject_init(&rkobj->kobj, &btrfs_raid_ktype);
9796
9797                 spin_lock(&fs_info->pending_raid_kobjs_lock);
9798                 list_add_tail(&rkobj->list, &fs_info->pending_raid_kobjs);
9799                 spin_unlock(&fs_info->pending_raid_kobjs_lock);
9800                 space_info->block_group_kobjs[index] = &rkobj->kobj;
9801         }
9802 }
9803
9804 static struct btrfs_block_group_cache *
9805 btrfs_create_block_group_cache(struct btrfs_fs_info *fs_info,
9806                                u64 start, u64 size)
9807 {
9808         struct btrfs_block_group_cache *cache;
9809
9810         cache = kzalloc(sizeof(*cache), GFP_NOFS);
9811         if (!cache)
9812                 return NULL;
9813
9814         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
9815                                         GFP_NOFS);
9816         if (!cache->free_space_ctl) {
9817                 kfree(cache);
9818                 return NULL;
9819         }
9820
9821         cache->key.objectid = start;
9822         cache->key.offset = size;
9823         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
9824
9825         cache->fs_info = fs_info;
9826         cache->full_stripe_len = btrfs_full_stripe_len(fs_info, start);
9827         set_free_space_tree_thresholds(cache);
9828
9829         atomic_set(&cache->count, 1);
9830         spin_lock_init(&cache->lock);
9831         init_rwsem(&cache->data_rwsem);
9832         INIT_LIST_HEAD(&cache->list);
9833         INIT_LIST_HEAD(&cache->cluster_list);
9834         INIT_LIST_HEAD(&cache->bg_list);
9835         INIT_LIST_HEAD(&cache->ro_list);
9836         INIT_LIST_HEAD(&cache->dirty_list);
9837         INIT_LIST_HEAD(&cache->io_list);
9838         btrfs_init_free_space_ctl(cache);
9839         atomic_set(&cache->trimming, 0);
9840         mutex_init(&cache->free_space_lock);
9841         btrfs_init_full_stripe_locks_tree(&cache->full_stripe_locks_root);
9842
9843         return cache;
9844 }
9845
9846
9847 /*
9848  * Iterate all chunks and verify that each of them has the corresponding block
9849  * group
9850  */
9851 static int check_chunk_block_group_mappings(struct btrfs_fs_info *fs_info)
9852 {
9853         struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
9854         struct extent_map *em;
9855         struct btrfs_block_group_cache *bg;
9856         u64 start = 0;
9857         int ret = 0;
9858
9859         while (1) {
9860                 read_lock(&map_tree->map_tree.lock);
9861                 /*
9862                  * lookup_extent_mapping will return the first extent map
9863                  * intersecting the range, so setting @len to 1 is enough to
9864                  * get the first chunk.
9865                  */
9866                 em = lookup_extent_mapping(&map_tree->map_tree, start, 1);
9867                 read_unlock(&map_tree->map_tree.lock);
9868                 if (!em)
9869                         break;
9870
9871                 bg = btrfs_lookup_block_group(fs_info, em->start);
9872                 if (!bg) {
9873                         btrfs_err(fs_info,
9874         "chunk start=%llu len=%llu doesn't have corresponding block group",
9875                                      em->start, em->len);
9876                         ret = -EUCLEAN;
9877                         free_extent_map(em);
9878                         break;
9879                 }
9880                 if (bg->key.objectid != em->start ||
9881                     bg->key.offset != em->len ||
9882                     (bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK) !=
9883                     (em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK)) {
9884                         btrfs_err(fs_info,
9885 "chunk start=%llu len=%llu flags=0x%llx doesn't match block group start=%llu len=%llu flags=0x%llx",
9886                                 em->start, em->len,
9887                                 em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK,
9888                                 bg->key.objectid, bg->key.offset,
9889                                 bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK);
9890                         ret = -EUCLEAN;
9891                         free_extent_map(em);
9892                         btrfs_put_block_group(bg);
9893                         break;
9894                 }
9895                 start = em->start + em->len;
9896                 free_extent_map(em);
9897                 btrfs_put_block_group(bg);
9898         }
9899         return ret;
9900 }
9901
9902 int btrfs_read_block_groups(struct btrfs_fs_info *info)
9903 {
9904         struct btrfs_path *path;
9905         int ret;
9906         struct btrfs_block_group_cache *cache;
9907         struct btrfs_space_info *space_info;
9908         struct btrfs_key key;
9909         struct btrfs_key found_key;
9910         struct extent_buffer *leaf;
9911         int need_clear = 0;
9912         u64 cache_gen;
9913         u64 feature;
9914         int mixed;
9915
9916         feature = btrfs_super_incompat_flags(info->super_copy);
9917         mixed = !!(feature & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS);
9918
9919         key.objectid = 0;
9920         key.offset = 0;
9921         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
9922         path = btrfs_alloc_path();
9923         if (!path)
9924                 return -ENOMEM;
9925         path->reada = READA_FORWARD;
9926
9927         cache_gen = btrfs_super_cache_generation(info->super_copy);
9928         if (btrfs_test_opt(info, SPACE_CACHE) &&
9929             btrfs_super_generation(info->super_copy) != cache_gen)
9930                 need_clear = 1;
9931         if (btrfs_test_opt(info, CLEAR_CACHE))
9932                 need_clear = 1;
9933
9934         while (1) {
9935                 ret = find_first_block_group(info, path, &key);
9936                 if (ret > 0)
9937                         break;
9938                 if (ret != 0)
9939                         goto error;
9940
9941                 leaf = path->nodes[0];
9942                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
9943
9944                 cache = btrfs_create_block_group_cache(info, found_key.objectid,
9945                                                        found_key.offset);
9946                 if (!cache) {
9947                         ret = -ENOMEM;
9948                         goto error;
9949                 }
9950
9951                 if (need_clear) {
9952                         /*
9953                          * When we mount with old space cache, we need to
9954                          * set BTRFS_DC_CLEAR and set dirty flag.
9955                          *
9956                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
9957                          *    truncate the old free space cache inode and
9958                          *    setup a new one.
9959                          * b) Setting 'dirty flag' makes sure that we flush
9960                          *    the new space cache info onto disk.
9961                          */
9962                         if (btrfs_test_opt(info, SPACE_CACHE))
9963                                 cache->disk_cache_state = BTRFS_DC_CLEAR;
9964                 }
9965
9966                 read_extent_buffer(leaf, &cache->item,
9967                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
9968                                    sizeof(cache->item));
9969                 cache->flags = btrfs_block_group_flags(&cache->item);
9970                 if (!mixed &&
9971                     ((cache->flags & BTRFS_BLOCK_GROUP_METADATA) &&
9972                     (cache->flags & BTRFS_BLOCK_GROUP_DATA))) {
9973                         btrfs_err(info,
9974 "bg %llu is a mixed block group but filesystem hasn't enabled mixed block groups",
9975                                   cache->key.objectid);
9976                         ret = -EINVAL;
9977                         goto error;
9978                 }
9979
9980                 key.objectid = found_key.objectid + found_key.offset;
9981                 btrfs_release_path(path);
9982
9983                 /*
9984                  * We need to exclude the super stripes now so that the space
9985                  * info has super bytes accounted for, otherwise we'll think
9986                  * we have more space than we actually do.
9987                  */
9988                 ret = exclude_super_stripes(cache);
9989                 if (ret) {
9990                         /*
9991                          * We may have excluded something, so call this just in
9992                          * case.
9993                          */
9994                         free_excluded_extents(cache);
9995                         btrfs_put_block_group(cache);
9996                         goto error;
9997                 }
9998
9999                 /*
10000                  * check for two cases, either we are full, and therefore
10001                  * don't need to bother with the caching work since we won't
10002                  * find any space, or we are empty, and we can just add all
10003                  * the space in and be done with it.  This saves us _alot_ of
10004                  * time, particularly in the full case.
10005                  */
10006                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
10007                         cache->last_byte_to_unpin = (u64)-1;
10008                         cache->cached = BTRFS_CACHE_FINISHED;
10009                         free_excluded_extents(cache);
10010                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10011                         cache->last_byte_to_unpin = (u64)-1;
10012                         cache->cached = BTRFS_CACHE_FINISHED;
10013                         add_new_free_space(cache, found_key.objectid,
10014                                            found_key.objectid +
10015                                            found_key.offset);
10016                         free_excluded_extents(cache);
10017                 }
10018
10019                 ret = btrfs_add_block_group_cache(info, cache);
10020                 if (ret) {
10021                         btrfs_remove_free_space_cache(cache);
10022                         btrfs_put_block_group(cache);
10023                         goto error;
10024                 }
10025
10026                 trace_btrfs_add_block_group(info, cache, 0);
10027                 update_space_info(info, cache->flags, found_key.offset,
10028                                   btrfs_block_group_used(&cache->item),
10029                                   cache->bytes_super, &space_info);
10030
10031                 cache->space_info = space_info;
10032
10033                 link_block_group(cache);
10034
10035                 set_avail_alloc_bits(info, cache->flags);
10036                 if (btrfs_chunk_readonly(info, cache->key.objectid)) {
10037                         inc_block_group_ro(cache, 1);
10038                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10039                         ASSERT(list_empty(&cache->bg_list));
10040                         btrfs_mark_bg_unused(cache);
10041                 }
10042         }
10043
10044         list_for_each_entry_rcu(space_info, &info->space_info, list) {
10045                 if (!(get_alloc_profile(info, space_info->flags) &
10046                       (BTRFS_BLOCK_GROUP_RAID10 |
10047                        BTRFS_BLOCK_GROUP_RAID1 |
10048                        BTRFS_BLOCK_GROUP_RAID5 |
10049                        BTRFS_BLOCK_GROUP_RAID6 |
10050                        BTRFS_BLOCK_GROUP_DUP)))
10051                         continue;
10052                 /*
10053                  * avoid allocating from un-mirrored block group if there are
10054                  * mirrored block groups.
10055                  */
10056                 list_for_each_entry(cache,
10057                                 &space_info->block_groups[BTRFS_RAID_RAID0],
10058                                 list)
10059                         inc_block_group_ro(cache, 1);
10060                 list_for_each_entry(cache,
10061                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
10062                                 list)
10063                         inc_block_group_ro(cache, 1);
10064         }
10065
10066         btrfs_add_raid_kobjects(info);
10067         init_global_block_rsv(info);
10068         ret = check_chunk_block_group_mappings(info);
10069 error:
10070         btrfs_free_path(path);
10071         return ret;
10072 }
10073
10074 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans)
10075 {
10076         struct btrfs_fs_info *fs_info = trans->fs_info;
10077         struct btrfs_block_group_cache *block_group, *tmp;
10078         struct btrfs_root *extent_root = fs_info->extent_root;
10079         struct btrfs_block_group_item item;
10080         struct btrfs_key key;
10081         int ret = 0;
10082         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
10083
10084         trans->can_flush_pending_bgs = false;
10085         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) {
10086                 if (ret)
10087                         goto next;
10088
10089                 spin_lock(&block_group->lock);
10090                 memcpy(&item, &block_group->item, sizeof(item));
10091                 memcpy(&key, &block_group->key, sizeof(key));
10092                 spin_unlock(&block_group->lock);
10093
10094                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
10095                                         sizeof(item));
10096                 if (ret)
10097                         btrfs_abort_transaction(trans, ret);
10098                 ret = btrfs_finish_chunk_alloc(trans, key.objectid, key.offset);
10099                 if (ret)
10100                         btrfs_abort_transaction(trans, ret);
10101                 add_block_group_free_space(trans, block_group);
10102                 /* already aborted the transaction if it failed. */
10103 next:
10104                 list_del_init(&block_group->bg_list);
10105         }
10106         trans->can_flush_pending_bgs = can_flush_pending_bgs;
10107 }
10108
10109 int btrfs_make_block_group(struct btrfs_trans_handle *trans, u64 bytes_used,
10110                            u64 type, u64 chunk_offset, u64 size)
10111 {
10112         struct btrfs_fs_info *fs_info = trans->fs_info;
10113         struct btrfs_block_group_cache *cache;
10114         int ret;
10115
10116         btrfs_set_log_full_commit(fs_info, trans);
10117
10118         cache = btrfs_create_block_group_cache(fs_info, chunk_offset, size);
10119         if (!cache)
10120                 return -ENOMEM;
10121
10122         btrfs_set_block_group_used(&cache->item, bytes_used);
10123         btrfs_set_block_group_chunk_objectid(&cache->item,
10124                                              BTRFS_FIRST_CHUNK_TREE_OBJECTID);
10125         btrfs_set_block_group_flags(&cache->item, type);
10126
10127         cache->flags = type;
10128         cache->last_byte_to_unpin = (u64)-1;
10129         cache->cached = BTRFS_CACHE_FINISHED;
10130         cache->needs_free_space = 1;
10131         ret = exclude_super_stripes(cache);
10132         if (ret) {
10133                 /*
10134                  * We may have excluded something, so call this just in
10135                  * case.
10136                  */
10137                 free_excluded_extents(cache);
10138                 btrfs_put_block_group(cache);
10139                 return ret;
10140         }
10141
10142         add_new_free_space(cache, chunk_offset, chunk_offset + size);
10143
10144         free_excluded_extents(cache);
10145
10146 #ifdef CONFIG_BTRFS_DEBUG
10147         if (btrfs_should_fragment_free_space(cache)) {
10148                 u64 new_bytes_used = size - bytes_used;
10149
10150                 bytes_used += new_bytes_used >> 1;
10151                 fragment_free_space(cache);
10152         }
10153 #endif
10154         /*
10155          * Ensure the corresponding space_info object is created and
10156          * assigned to our block group. We want our bg to be added to the rbtree
10157          * with its ->space_info set.
10158          */
10159         cache->space_info = __find_space_info(fs_info, cache->flags);
10160         ASSERT(cache->space_info);
10161
10162         ret = btrfs_add_block_group_cache(fs_info, cache);
10163         if (ret) {
10164                 btrfs_remove_free_space_cache(cache);
10165                 btrfs_put_block_group(cache);
10166                 return ret;
10167         }
10168
10169         /*
10170          * Now that our block group has its ->space_info set and is inserted in
10171          * the rbtree, update the space info's counters.
10172          */
10173         trace_btrfs_add_block_group(fs_info, cache, 1);
10174         update_space_info(fs_info, cache->flags, size, bytes_used,
10175                                 cache->bytes_super, &cache->space_info);
10176         update_global_block_rsv(fs_info);
10177
10178         link_block_group(cache);
10179
10180         list_add_tail(&cache->bg_list, &trans->new_bgs);
10181
10182         set_avail_alloc_bits(fs_info, type);
10183         return 0;
10184 }
10185
10186 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
10187 {
10188         u64 extra_flags = chunk_to_extended(flags) &
10189                                 BTRFS_EXTENDED_PROFILE_MASK;
10190
10191         write_seqlock(&fs_info->profiles_lock);
10192         if (flags & BTRFS_BLOCK_GROUP_DATA)
10193                 fs_info->avail_data_alloc_bits &= ~extra_flags;
10194         if (flags & BTRFS_BLOCK_GROUP_METADATA)
10195                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
10196         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
10197                 fs_info->avail_system_alloc_bits &= ~extra_flags;
10198         write_sequnlock(&fs_info->profiles_lock);
10199 }
10200
10201 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
10202                              u64 group_start, struct extent_map *em)
10203 {
10204         struct btrfs_fs_info *fs_info = trans->fs_info;
10205         struct btrfs_root *root = fs_info->extent_root;
10206         struct btrfs_path *path;
10207         struct btrfs_block_group_cache *block_group;
10208         struct btrfs_free_cluster *cluster;
10209         struct btrfs_root *tree_root = fs_info->tree_root;
10210         struct btrfs_key key;
10211         struct inode *inode;
10212         struct kobject *kobj = NULL;
10213         int ret;
10214         int index;
10215         int factor;
10216         struct btrfs_caching_control *caching_ctl = NULL;
10217         bool remove_em;
10218
10219         block_group = btrfs_lookup_block_group(fs_info, group_start);
10220         BUG_ON(!block_group);
10221         BUG_ON(!block_group->ro);
10222
10223         trace_btrfs_remove_block_group(block_group);
10224         /*
10225          * Free the reserved super bytes from this block group before
10226          * remove it.
10227          */
10228         free_excluded_extents(block_group);
10229         btrfs_free_ref_tree_range(fs_info, block_group->key.objectid,
10230                                   block_group->key.offset);
10231
10232         memcpy(&key, &block_group->key, sizeof(key));
10233         index = btrfs_bg_flags_to_raid_index(block_group->flags);
10234         factor = btrfs_bg_type_to_factor(block_group->flags);
10235
10236         /* make sure this block group isn't part of an allocation cluster */
10237         cluster = &fs_info->data_alloc_cluster;
10238         spin_lock(&cluster->refill_lock);
10239         btrfs_return_cluster_to_free_space(block_group, cluster);
10240         spin_unlock(&cluster->refill_lock);
10241
10242         /*
10243          * make sure this block group isn't part of a metadata
10244          * allocation cluster
10245          */
10246         cluster = &fs_info->meta_alloc_cluster;
10247         spin_lock(&cluster->refill_lock);
10248         btrfs_return_cluster_to_free_space(block_group, cluster);
10249         spin_unlock(&cluster->refill_lock);
10250
10251         path = btrfs_alloc_path();
10252         if (!path) {
10253                 ret = -ENOMEM;
10254                 goto out;
10255         }
10256
10257         /*
10258          * get the inode first so any iput calls done for the io_list
10259          * aren't the final iput (no unlinks allowed now)
10260          */
10261         inode = lookup_free_space_inode(fs_info, block_group, path);
10262
10263         mutex_lock(&trans->transaction->cache_write_mutex);
10264         /*
10265          * make sure our free spache cache IO is done before remove the
10266          * free space inode
10267          */
10268         spin_lock(&trans->transaction->dirty_bgs_lock);
10269         if (!list_empty(&block_group->io_list)) {
10270                 list_del_init(&block_group->io_list);
10271
10272                 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
10273
10274                 spin_unlock(&trans->transaction->dirty_bgs_lock);
10275                 btrfs_wait_cache_io(trans, block_group, path);
10276                 btrfs_put_block_group(block_group);
10277                 spin_lock(&trans->transaction->dirty_bgs_lock);
10278         }
10279
10280         if (!list_empty(&block_group->dirty_list)) {
10281                 list_del_init(&block_group->dirty_list);
10282                 btrfs_put_block_group(block_group);
10283         }
10284         spin_unlock(&trans->transaction->dirty_bgs_lock);
10285         mutex_unlock(&trans->transaction->cache_write_mutex);
10286
10287         if (!IS_ERR(inode)) {
10288                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
10289                 if (ret) {
10290                         btrfs_add_delayed_iput(inode);
10291                         goto out;
10292                 }
10293                 clear_nlink(inode);
10294                 /* One for the block groups ref */
10295                 spin_lock(&block_group->lock);
10296                 if (block_group->iref) {
10297                         block_group->iref = 0;
10298                         block_group->inode = NULL;
10299                         spin_unlock(&block_group->lock);
10300                         iput(inode);
10301                 } else {
10302                         spin_unlock(&block_group->lock);
10303                 }
10304                 /* One for our lookup ref */
10305                 btrfs_add_delayed_iput(inode);
10306         }
10307
10308         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
10309         key.offset = block_group->key.objectid;
10310         key.type = 0;
10311
10312         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
10313         if (ret < 0)
10314                 goto out;
10315         if (ret > 0)
10316                 btrfs_release_path(path);
10317         if (ret == 0) {
10318                 ret = btrfs_del_item(trans, tree_root, path);
10319                 if (ret)
10320                         goto out;
10321                 btrfs_release_path(path);
10322         }
10323
10324         spin_lock(&fs_info->block_group_cache_lock);
10325         rb_erase(&block_group->cache_node,
10326                  &fs_info->block_group_cache_tree);
10327         RB_CLEAR_NODE(&block_group->cache_node);
10328
10329         if (fs_info->first_logical_byte == block_group->key.objectid)
10330                 fs_info->first_logical_byte = (u64)-1;
10331         spin_unlock(&fs_info->block_group_cache_lock);
10332
10333         down_write(&block_group->space_info->groups_sem);
10334         /*
10335          * we must use list_del_init so people can check to see if they
10336          * are still on the list after taking the semaphore
10337          */
10338         list_del_init(&block_group->list);
10339         if (list_empty(&block_group->space_info->block_groups[index])) {
10340                 kobj = block_group->space_info->block_group_kobjs[index];
10341                 block_group->space_info->block_group_kobjs[index] = NULL;
10342                 clear_avail_alloc_bits(fs_info, block_group->flags);
10343         }
10344         up_write(&block_group->space_info->groups_sem);
10345         if (kobj) {
10346                 kobject_del(kobj);
10347                 kobject_put(kobj);
10348         }
10349
10350         if (block_group->has_caching_ctl)
10351                 caching_ctl = get_caching_control(block_group);
10352         if (block_group->cached == BTRFS_CACHE_STARTED)
10353                 wait_block_group_cache_done(block_group);
10354         if (block_group->has_caching_ctl) {
10355                 down_write(&fs_info->commit_root_sem);
10356                 if (!caching_ctl) {
10357                         struct btrfs_caching_control *ctl;
10358
10359                         list_for_each_entry(ctl,
10360                                     &fs_info->caching_block_groups, list)
10361                                 if (ctl->block_group == block_group) {
10362                                         caching_ctl = ctl;
10363                                         refcount_inc(&caching_ctl->count);
10364                                         break;
10365                                 }
10366                 }
10367                 if (caching_ctl)
10368                         list_del_init(&caching_ctl->list);
10369                 up_write(&fs_info->commit_root_sem);
10370                 if (caching_ctl) {
10371                         /* Once for the caching bgs list and once for us. */
10372                         put_caching_control(caching_ctl);
10373                         put_caching_control(caching_ctl);
10374                 }
10375         }
10376
10377         spin_lock(&trans->transaction->dirty_bgs_lock);
10378         if (!list_empty(&block_group->dirty_list)) {
10379                 WARN_ON(1);
10380         }
10381         if (!list_empty(&block_group->io_list)) {
10382                 WARN_ON(1);
10383         }
10384         spin_unlock(&trans->transaction->dirty_bgs_lock);
10385         btrfs_remove_free_space_cache(block_group);
10386
10387         spin_lock(&block_group->space_info->lock);
10388         list_del_init(&block_group->ro_list);
10389
10390         if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
10391                 WARN_ON(block_group->space_info->total_bytes
10392                         < block_group->key.offset);
10393                 WARN_ON(block_group->space_info->bytes_readonly
10394                         < block_group->key.offset);
10395                 WARN_ON(block_group->space_info->disk_total
10396                         < block_group->key.offset * factor);
10397         }
10398         block_group->space_info->total_bytes -= block_group->key.offset;
10399         block_group->space_info->bytes_readonly -= block_group->key.offset;
10400         block_group->space_info->disk_total -= block_group->key.offset * factor;
10401
10402         spin_unlock(&block_group->space_info->lock);
10403
10404         memcpy(&key, &block_group->key, sizeof(key));
10405
10406         mutex_lock(&fs_info->chunk_mutex);
10407         if (!list_empty(&em->list)) {
10408                 /* We're in the transaction->pending_chunks list. */
10409                 free_extent_map(em);
10410         }
10411         spin_lock(&block_group->lock);
10412         block_group->removed = 1;
10413         /*
10414          * At this point trimming can't start on this block group, because we
10415          * removed the block group from the tree fs_info->block_group_cache_tree
10416          * so no one can't find it anymore and even if someone already got this
10417          * block group before we removed it from the rbtree, they have already
10418          * incremented block_group->trimming - if they didn't, they won't find
10419          * any free space entries because we already removed them all when we
10420          * called btrfs_remove_free_space_cache().
10421          *
10422          * And we must not remove the extent map from the fs_info->mapping_tree
10423          * to prevent the same logical address range and physical device space
10424          * ranges from being reused for a new block group. This is because our
10425          * fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
10426          * completely transactionless, so while it is trimming a range the
10427          * currently running transaction might finish and a new one start,
10428          * allowing for new block groups to be created that can reuse the same
10429          * physical device locations unless we take this special care.
10430          *
10431          * There may also be an implicit trim operation if the file system
10432          * is mounted with -odiscard. The same protections must remain
10433          * in place until the extents have been discarded completely when
10434          * the transaction commit has completed.
10435          */
10436         remove_em = (atomic_read(&block_group->trimming) == 0);
10437         /*
10438          * Make sure a trimmer task always sees the em in the pinned_chunks list
10439          * if it sees block_group->removed == 1 (needs to lock block_group->lock
10440          * before checking block_group->removed).
10441          */
10442         if (!remove_em) {
10443                 /*
10444                  * Our em might be in trans->transaction->pending_chunks which
10445                  * is protected by fs_info->chunk_mutex ([lock|unlock]_chunks),
10446                  * and so is the fs_info->pinned_chunks list.
10447                  *
10448                  * So at this point we must be holding the chunk_mutex to avoid
10449                  * any races with chunk allocation (more specifically at
10450                  * volumes.c:contains_pending_extent()), to ensure it always
10451                  * sees the em, either in the pending_chunks list or in the
10452                  * pinned_chunks list.
10453                  */
10454                 list_move_tail(&em->list, &fs_info->pinned_chunks);
10455         }
10456         spin_unlock(&block_group->lock);
10457
10458         if (remove_em) {
10459                 struct extent_map_tree *em_tree;
10460
10461                 em_tree = &fs_info->mapping_tree.map_tree;
10462                 write_lock(&em_tree->lock);
10463                 /*
10464                  * The em might be in the pending_chunks list, so make sure the
10465                  * chunk mutex is locked, since remove_extent_mapping() will
10466                  * delete us from that list.
10467                  */
10468                 remove_extent_mapping(em_tree, em);
10469                 write_unlock(&em_tree->lock);
10470                 /* once for the tree */
10471                 free_extent_map(em);
10472         }
10473
10474         mutex_unlock(&fs_info->chunk_mutex);
10475
10476         ret = remove_block_group_free_space(trans, block_group);
10477         if (ret)
10478                 goto out;
10479
10480         btrfs_put_block_group(block_group);
10481         btrfs_put_block_group(block_group);
10482
10483         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
10484         if (ret > 0)
10485                 ret = -EIO;
10486         if (ret < 0)
10487                 goto out;
10488
10489         ret = btrfs_del_item(trans, root, path);
10490 out:
10491         btrfs_free_path(path);
10492         return ret;
10493 }
10494
10495 struct btrfs_trans_handle *
10496 btrfs_start_trans_remove_block_group(struct btrfs_fs_info *fs_info,
10497                                      const u64 chunk_offset)
10498 {
10499         struct extent_map_tree *em_tree = &fs_info->mapping_tree.map_tree;
10500         struct extent_map *em;
10501         struct map_lookup *map;
10502         unsigned int num_items;
10503
10504         read_lock(&em_tree->lock);
10505         em = lookup_extent_mapping(em_tree, chunk_offset, 1);
10506         read_unlock(&em_tree->lock);
10507         ASSERT(em && em->start == chunk_offset);
10508
10509         /*
10510          * We need to reserve 3 + N units from the metadata space info in order
10511          * to remove a block group (done at btrfs_remove_chunk() and at
10512          * btrfs_remove_block_group()), which are used for:
10513          *
10514          * 1 unit for adding the free space inode's orphan (located in the tree
10515          * of tree roots).
10516          * 1 unit for deleting the block group item (located in the extent
10517          * tree).
10518          * 1 unit for deleting the free space item (located in tree of tree
10519          * roots).
10520          * N units for deleting N device extent items corresponding to each
10521          * stripe (located in the device tree).
10522          *
10523          * In order to remove a block group we also need to reserve units in the
10524          * system space info in order to update the chunk tree (update one or
10525          * more device items and remove one chunk item), but this is done at
10526          * btrfs_remove_chunk() through a call to check_system_chunk().
10527          */
10528         map = em->map_lookup;
10529         num_items = 3 + map->num_stripes;
10530         free_extent_map(em);
10531
10532         return btrfs_start_transaction_fallback_global_rsv(fs_info->extent_root,
10533                                                            num_items, 1);
10534 }
10535
10536 /*
10537  * Process the unused_bgs list and remove any that don't have any allocated
10538  * space inside of them.
10539  */
10540 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
10541 {
10542         struct btrfs_block_group_cache *block_group;
10543         struct btrfs_space_info *space_info;
10544         struct btrfs_trans_handle *trans;
10545         int ret = 0;
10546
10547         if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
10548                 return;
10549
10550         spin_lock(&fs_info->unused_bgs_lock);
10551         while (!list_empty(&fs_info->unused_bgs)) {
10552                 u64 start, end;
10553                 int trimming;
10554
10555                 block_group = list_first_entry(&fs_info->unused_bgs,
10556                                                struct btrfs_block_group_cache,
10557                                                bg_list);
10558                 list_del_init(&block_group->bg_list);
10559
10560                 space_info = block_group->space_info;
10561
10562                 if (ret || btrfs_mixed_space_info(space_info)) {
10563                         btrfs_put_block_group(block_group);
10564                         continue;
10565                 }
10566                 spin_unlock(&fs_info->unused_bgs_lock);
10567
10568                 mutex_lock(&fs_info->delete_unused_bgs_mutex);
10569
10570                 /* Don't want to race with allocators so take the groups_sem */
10571                 down_write(&space_info->groups_sem);
10572                 spin_lock(&block_group->lock);
10573                 if (block_group->reserved || block_group->pinned ||
10574                     btrfs_block_group_used(&block_group->item) ||
10575                     block_group->ro ||
10576                     list_is_singular(&block_group->list)) {
10577                         /*
10578                          * We want to bail if we made new allocations or have
10579                          * outstanding allocations in this block group.  We do
10580                          * the ro check in case balance is currently acting on
10581                          * this block group.
10582                          */
10583                         trace_btrfs_skip_unused_block_group(block_group);
10584                         spin_unlock(&block_group->lock);
10585                         up_write(&space_info->groups_sem);
10586                         goto next;
10587                 }
10588                 spin_unlock(&block_group->lock);
10589
10590                 /* We don't want to force the issue, only flip if it's ok. */
10591                 ret = inc_block_group_ro(block_group, 0);
10592                 up_write(&space_info->groups_sem);
10593                 if (ret < 0) {
10594                         ret = 0;
10595                         goto next;
10596                 }
10597
10598                 /*
10599                  * Want to do this before we do anything else so we can recover
10600                  * properly if we fail to join the transaction.
10601                  */
10602                 trans = btrfs_start_trans_remove_block_group(fs_info,
10603                                                      block_group->key.objectid);
10604                 if (IS_ERR(trans)) {
10605                         btrfs_dec_block_group_ro(block_group);
10606                         ret = PTR_ERR(trans);
10607                         goto next;
10608                 }
10609
10610                 /*
10611                  * We could have pending pinned extents for this block group,
10612                  * just delete them, we don't care about them anymore.
10613                  */
10614                 start = block_group->key.objectid;
10615                 end = start + block_group->key.offset - 1;
10616                 /*
10617                  * Hold the unused_bg_unpin_mutex lock to avoid racing with
10618                  * btrfs_finish_extent_commit(). If we are at transaction N,
10619                  * another task might be running finish_extent_commit() for the
10620                  * previous transaction N - 1, and have seen a range belonging
10621                  * to the block group in freed_extents[] before we were able to
10622                  * clear the whole block group range from freed_extents[]. This
10623                  * means that task can lookup for the block group after we
10624                  * unpinned it from freed_extents[] and removed it, leading to
10625                  * a BUG_ON() at btrfs_unpin_extent_range().
10626                  */
10627                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
10628                 ret = clear_extent_bits(&fs_info->freed_extents[0], start, end,
10629                                   EXTENT_DIRTY);
10630                 if (ret) {
10631                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10632                         btrfs_dec_block_group_ro(block_group);
10633                         goto end_trans;
10634                 }
10635                 ret = clear_extent_bits(&fs_info->freed_extents[1], start, end,
10636                                   EXTENT_DIRTY);
10637                 if (ret) {
10638                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10639                         btrfs_dec_block_group_ro(block_group);
10640                         goto end_trans;
10641                 }
10642                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10643
10644                 /* Reset pinned so btrfs_put_block_group doesn't complain */
10645                 spin_lock(&space_info->lock);
10646                 spin_lock(&block_group->lock);
10647
10648                 space_info->bytes_pinned -= block_group->pinned;
10649                 space_info->bytes_readonly += block_group->pinned;
10650                 percpu_counter_add_batch(&space_info->total_bytes_pinned,
10651                                    -block_group->pinned,
10652                                    BTRFS_TOTAL_BYTES_PINNED_BATCH);
10653                 block_group->pinned = 0;
10654
10655                 spin_unlock(&block_group->lock);
10656                 spin_unlock(&space_info->lock);
10657
10658                 /* DISCARD can flip during remount */
10659                 trimming = btrfs_test_opt(fs_info, DISCARD);
10660
10661                 /* Implicit trim during transaction commit. */
10662                 if (trimming)
10663                         btrfs_get_block_group_trimming(block_group);
10664
10665                 /*
10666                  * Btrfs_remove_chunk will abort the transaction if things go
10667                  * horribly wrong.
10668                  */
10669                 ret = btrfs_remove_chunk(trans, block_group->key.objectid);
10670
10671                 if (ret) {
10672                         if (trimming)
10673                                 btrfs_put_block_group_trimming(block_group);
10674                         goto end_trans;
10675                 }
10676
10677                 /*
10678                  * If we're not mounted with -odiscard, we can just forget
10679                  * about this block group. Otherwise we'll need to wait
10680                  * until transaction commit to do the actual discard.
10681                  */
10682                 if (trimming) {
10683                         spin_lock(&fs_info->unused_bgs_lock);
10684                         /*
10685                          * A concurrent scrub might have added us to the list
10686                          * fs_info->unused_bgs, so use a list_move operation
10687                          * to add the block group to the deleted_bgs list.
10688                          */
10689                         list_move(&block_group->bg_list,
10690                                   &trans->transaction->deleted_bgs);
10691                         spin_unlock(&fs_info->unused_bgs_lock);
10692                         btrfs_get_block_group(block_group);
10693                 }
10694 end_trans:
10695                 btrfs_end_transaction(trans);
10696 next:
10697                 mutex_unlock(&fs_info->delete_unused_bgs_mutex);
10698                 btrfs_put_block_group(block_group);
10699                 spin_lock(&fs_info->unused_bgs_lock);
10700         }
10701         spin_unlock(&fs_info->unused_bgs_lock);
10702 }
10703
10704 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
10705 {
10706         struct btrfs_super_block *disk_super;
10707         u64 features;
10708         u64 flags;
10709         int mixed = 0;
10710         int ret;
10711
10712         disk_super = fs_info->super_copy;
10713         if (!btrfs_super_root(disk_super))
10714                 return -EINVAL;
10715
10716         features = btrfs_super_incompat_flags(disk_super);
10717         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
10718                 mixed = 1;
10719
10720         flags = BTRFS_BLOCK_GROUP_SYSTEM;
10721         ret = create_space_info(fs_info, flags);
10722         if (ret)
10723                 goto out;
10724
10725         if (mixed) {
10726                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
10727                 ret = create_space_info(fs_info, flags);
10728         } else {
10729                 flags = BTRFS_BLOCK_GROUP_METADATA;
10730                 ret = create_space_info(fs_info, flags);
10731                 if (ret)
10732                         goto out;
10733
10734                 flags = BTRFS_BLOCK_GROUP_DATA;
10735                 ret = create_space_info(fs_info, flags);
10736         }
10737 out:
10738         return ret;
10739 }
10740
10741 int btrfs_error_unpin_extent_range(struct btrfs_fs_info *fs_info,
10742                                    u64 start, u64 end)
10743 {
10744         return unpin_extent_range(fs_info, start, end, false);
10745 }
10746
10747 /*
10748  * It used to be that old block groups would be left around forever.
10749  * Iterating over them would be enough to trim unused space.  Since we
10750  * now automatically remove them, we also need to iterate over unallocated
10751  * space.
10752  *
10753  * We don't want a transaction for this since the discard may take a
10754  * substantial amount of time.  We don't require that a transaction be
10755  * running, but we do need to take a running transaction into account
10756  * to ensure that we're not discarding chunks that were released in
10757  * the current transaction.
10758  *
10759  * Holding the chunks lock will prevent other threads from allocating
10760  * or releasing chunks, but it won't prevent a running transaction
10761  * from committing and releasing the memory that the pending chunks
10762  * list head uses.  For that, we need to take a reference to the
10763  * transaction.
10764  */
10765 static int btrfs_trim_free_extents(struct btrfs_device *device,
10766                                    u64 minlen, u64 *trimmed)
10767 {
10768         u64 start = 0, len = 0;
10769         int ret;
10770
10771         *trimmed = 0;
10772
10773         /* Not writeable = nothing to do. */
10774         if (!test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state))
10775                 return 0;
10776
10777         /* No free space = nothing to do. */
10778         if (device->total_bytes <= device->bytes_used)
10779                 return 0;
10780
10781         ret = 0;
10782
10783         while (1) {
10784                 struct btrfs_fs_info *fs_info = device->fs_info;
10785                 struct btrfs_transaction *trans;
10786                 u64 bytes;
10787
10788                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
10789                 if (ret)
10790                         return ret;
10791
10792                 down_read(&fs_info->commit_root_sem);
10793
10794                 spin_lock(&fs_info->trans_lock);
10795                 trans = fs_info->running_transaction;
10796                 if (trans)
10797                         refcount_inc(&trans->use_count);
10798                 spin_unlock(&fs_info->trans_lock);
10799
10800                 ret = find_free_dev_extent_start(trans, device, minlen, start,
10801                                                  &start, &len);
10802                 if (trans)
10803                         btrfs_put_transaction(trans);
10804
10805                 if (ret) {
10806                         up_read(&fs_info->commit_root_sem);
10807                         mutex_unlock(&fs_info->chunk_mutex);
10808                         if (ret == -ENOSPC)
10809                                 ret = 0;
10810                         break;
10811                 }
10812
10813                 ret = btrfs_issue_discard(device->bdev, start, len, &bytes);
10814                 up_read(&fs_info->commit_root_sem);
10815                 mutex_unlock(&fs_info->chunk_mutex);
10816
10817                 if (ret)
10818                         break;
10819
10820                 start += len;
10821                 *trimmed += bytes;
10822
10823                 if (fatal_signal_pending(current)) {
10824                         ret = -ERESTARTSYS;
10825                         break;
10826                 }
10827
10828                 cond_resched();
10829         }
10830
10831         return ret;
10832 }
10833
10834 int btrfs_trim_fs(struct btrfs_fs_info *fs_info, struct fstrim_range *range)
10835 {
10836         struct btrfs_block_group_cache *cache = NULL;
10837         struct btrfs_device *device;
10838         struct list_head *devices;
10839         u64 group_trimmed;
10840         u64 start;
10841         u64 end;
10842         u64 trimmed = 0;
10843         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
10844         int ret = 0;
10845
10846         /*
10847          * try to trim all FS space, our block group may start from non-zero.
10848          */
10849         if (range->len == total_bytes)
10850                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
10851         else
10852                 cache = btrfs_lookup_block_group(fs_info, range->start);
10853
10854         while (cache) {
10855                 if (cache->key.objectid >= (range->start + range->len)) {
10856                         btrfs_put_block_group(cache);
10857                         break;
10858                 }
10859
10860                 start = max(range->start, cache->key.objectid);
10861                 end = min(range->start + range->len,
10862                                 cache->key.objectid + cache->key.offset);
10863
10864                 if (end - start >= range->minlen) {
10865                         if (!block_group_cache_done(cache)) {
10866                                 ret = cache_block_group(cache, 0);
10867                                 if (ret) {
10868                                         btrfs_put_block_group(cache);
10869                                         break;
10870                                 }
10871                                 ret = wait_block_group_cache_done(cache);
10872                                 if (ret) {
10873                                         btrfs_put_block_group(cache);
10874                                         break;
10875                                 }
10876                         }
10877                         ret = btrfs_trim_block_group(cache,
10878                                                      &group_trimmed,
10879                                                      start,
10880                                                      end,
10881                                                      range->minlen);
10882
10883                         trimmed += group_trimmed;
10884                         if (ret) {
10885                                 btrfs_put_block_group(cache);
10886                                 break;
10887                         }
10888                 }
10889
10890                 cache = next_block_group(fs_info, cache);
10891         }
10892
10893         mutex_lock(&fs_info->fs_devices->device_list_mutex);
10894         devices = &fs_info->fs_devices->alloc_list;
10895         list_for_each_entry(device, devices, dev_alloc_list) {
10896                 ret = btrfs_trim_free_extents(device, range->minlen,
10897                                               &group_trimmed);
10898                 if (ret)
10899                         break;
10900
10901                 trimmed += group_trimmed;
10902         }
10903         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
10904
10905         range->len = trimmed;
10906         return ret;
10907 }
10908
10909 /*
10910  * btrfs_{start,end}_write_no_snapshotting() are similar to
10911  * mnt_{want,drop}_write(), they are used to prevent some tasks from writing
10912  * data into the page cache through nocow before the subvolume is snapshoted,
10913  * but flush the data into disk after the snapshot creation, or to prevent
10914  * operations while snapshotting is ongoing and that cause the snapshot to be
10915  * inconsistent (writes followed by expanding truncates for example).
10916  */
10917 void btrfs_end_write_no_snapshotting(struct btrfs_root *root)
10918 {
10919         percpu_counter_dec(&root->subv_writers->counter);
10920         cond_wake_up(&root->subv_writers->wait);
10921 }
10922
10923 int btrfs_start_write_no_snapshotting(struct btrfs_root *root)
10924 {
10925         if (atomic_read(&root->will_be_snapshotted))
10926                 return 0;
10927
10928         percpu_counter_inc(&root->subv_writers->counter);
10929         /*
10930          * Make sure counter is updated before we check for snapshot creation.
10931          */
10932         smp_mb();
10933         if (atomic_read(&root->will_be_snapshotted)) {
10934                 btrfs_end_write_no_snapshotting(root);
10935                 return 0;
10936         }
10937         return 1;
10938 }
10939
10940 void btrfs_wait_for_snapshot_creation(struct btrfs_root *root)
10941 {
10942         while (true) {
10943                 int ret;
10944
10945                 ret = btrfs_start_write_no_snapshotting(root);
10946                 if (ret)
10947                         break;
10948                 wait_var_event(&root->will_be_snapshotted,
10949                                !atomic_read(&root->will_be_snapshotted));
10950         }
10951 }
10952
10953 void btrfs_mark_bg_unused(struct btrfs_block_group_cache *bg)
10954 {
10955         struct btrfs_fs_info *fs_info = bg->fs_info;
10956
10957         spin_lock(&fs_info->unused_bgs_lock);
10958         if (list_empty(&bg->bg_list)) {
10959                 btrfs_get_block_group(bg);
10960                 trace_btrfs_add_unused_block_group(bg);
10961                 list_add_tail(&bg->bg_list, &fs_info->unused_bgs);
10962         }
10963         spin_unlock(&fs_info->unused_bgs_lock);
10964 }