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