fs/btrfs/extent-tree.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2007 Oracle.  All rights reserved.
   4  */
   5
   6 #include <linux/sched.h>
   7 #include <linux/sched/signal.h>
   8 #include <linux/pagemap.h>
   9 #include <linux/writeback.h>
  10 #include <linux/blkdev.h>
  11 #include <linux/sort.h>
  12 #include <linux/rcupdate.h>
  13 #include <linux/kthread.h>
  14 #include <linux/slab.h>
  15 #include <linux/ratelimit.h>
  16 #include <linux/percpu_counter.h>
  17 #include <linux/lockdep.h>
  18 #include <linux/crc32c.h>
  19 #include "tree-log.h"
  20 #include "disk-io.h"
  21 #include "print-tree.h"
  22 #include "volumes.h"
  23 #include "raid56.h"
  24 #include "locking.h"
  25 #include "free-space-cache.h"
  26 #include "free-space-tree.h"
  27 #include "math.h"
  28 #include "sysfs.h"
  29 #include "qgroup.h"
  30 #include "ref-verify.h"
  31
  32 #undef SCRAMBLE_DELAYED_REFS
  33
  34 /*
  35  * control flags for do_chunk_alloc's force field
  36  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
  37  * if we really need one.
  38  *
  39  * CHUNK_ALLOC_LIMITED means to only try and allocate one
  40  * if we have very few chunks already allocated.  This is
  41  * used as part of the clustering code to help make sure
  42  * we have a good pool of storage to cluster in, without
  43  * filling the FS with empty chunks
  44  *
  45  * CHUNK_ALLOC_FORCE means it must try to allocate one
  46  *
  47  */
  48 enum {
  49         CHUNK_ALLOC_NO_FORCE = 0,
  50         CHUNK_ALLOC_LIMITED = 1,
  51         CHUNK_ALLOC_FORCE = 2,
  52 };
  53
  54 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  55                                struct btrfs_delayed_ref_node *node, u64 parent,
  56                                u64 root_objectid, u64 owner_objectid,
  57                                u64 owner_offset, int refs_to_drop,
  58                                struct btrfs_delayed_extent_op *extra_op);
  59 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  60                                     struct extent_buffer *leaf,
  61                                     struct btrfs_extent_item *ei);
  62 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  63                                       u64 parent, u64 root_objectid,
  64                                       u64 flags, u64 owner, u64 offset,
  65                                       struct btrfs_key *ins, int ref_mod);
  66 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  67                                      struct btrfs_delayed_ref_node *node,
  68                                      struct btrfs_delayed_extent_op *extent_op);
  69 static int do_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags,
  70                           int force);
  71 static int find_next_key(struct btrfs_path *path, int level,
  72                          struct btrfs_key *key);
  73 static void dump_space_info(struct btrfs_fs_info *fs_info,
  74                             struct btrfs_space_info *info, u64 bytes,
  75                             int dump_block_groups);
  76 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
  77                                u64 num_bytes);
  78 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
  79                                      struct btrfs_space_info *space_info,
  80                                      u64 num_bytes);
  81 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
  82                                      struct btrfs_space_info *space_info,
  83                                      u64 num_bytes);
  84
  85 static noinline int
  86 block_group_cache_done(struct btrfs_block_group_cache *cache)
  87 {
  88         smp_mb();
  89         return cache->cached == BTRFS_CACHE_FINISHED ||
  90                 cache->cached == BTRFS_CACHE_ERROR;
  91 }
  92
  93 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
  94 {
  95         return (cache->flags & bits) == bits;
  96 }
  97
  98 void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
  99 {
 100         atomic_inc(&cache->count);
 101 }
 102
 103 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 104 {
 105         if (atomic_dec_and_test(&cache->count)) {
 106                 WARN_ON(cache->pinned > 0);
 107                 WARN_ON(cache->reserved > 0);
 108
 109                 /*
 110                  * If not empty, someone is still holding mutex of
 111                  * full_stripe_lock, which can only be released by caller.
 112                  * And it will definitely cause use-after-free when caller
 113                  * tries to release full stripe lock.
 114                  *
 115                  * No better way to resolve, but only to warn.
 116                  */
 117                 WARN_ON(!RB_EMPTY_ROOT(&cache->full_stripe_locks_root.root));
 118                 kfree(cache->free_space_ctl);
 119                 kfree(cache);
 120         }
 121 }
 122
 123 /*
 124  * this adds the block group to the fs_info rb tree for the block group
 125  * cache
 126  */
 127 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 128                                 struct btrfs_block_group_cache *block_group)
 129 {
 130         struct rb_node **p;
 131         struct rb_node *parent = NULL;
 132         struct btrfs_block_group_cache *cache;
 133
 134         spin_lock(&info->block_group_cache_lock);
 135         p = &info->block_group_cache_tree.rb_node;
 136
 137         while (*p) {
 138                 parent = *p;
 139                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 140                                  cache_node);
 141                 if (block_group->key.objectid < cache->key.objectid) {
 142                         p = &(*p)->rb_left;
 143                 } else if (block_group->key.objectid > cache->key.objectid) {
 144                         p = &(*p)->rb_right;
 145                 } else {
 146                         spin_unlock(&info->block_group_cache_lock);
 147                         return -EEXIST;
 148                 }
 149         }
 150
 151         rb_link_node(&block_group->cache_node, parent, p);
 152         rb_insert_color(&block_group->cache_node,
 153                         &info->block_group_cache_tree);
 154
 155         if (info->first_logical_byte > block_group->key.objectid)
 156                 info->first_logical_byte = block_group->key.objectid;
 157
 158         spin_unlock(&info->block_group_cache_lock);
 159
 160         return 0;
 161 }
 162
 163 /*
 164  * This will return the block group at or after bytenr if contains is 0, else
 165  * it will return the block group that contains the bytenr
 166  */
 167 static struct btrfs_block_group_cache *
 168 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 169                               int contains)
 170 {
 171         struct btrfs_block_group_cache *cache, *ret = NULL;
 172         struct rb_node *n;
 173         u64 end, start;
 174
 175         spin_lock(&info->block_group_cache_lock);
 176         n = info->block_group_cache_tree.rb_node;
 177
 178         while (n) {
 179                 cache = rb_entry(n, struct btrfs_block_group_cache,
 180                                  cache_node);
 181                 end = cache->key.objectid + cache->key.offset - 1;
 182                 start = cache->key.objectid;
 183
 184                 if (bytenr < start) {
 185                         if (!contains && (!ret || start < ret->key.objectid))
 186                                 ret = cache;
 187                         n = n->rb_left;
 188                 } else if (bytenr > start) {
 189                         if (contains && bytenr <= end) {
 190                                 ret = cache;
 191                                 break;
 192                         }
 193                         n = n->rb_right;
 194                 } else {
 195                         ret = cache;
 196                         break;
 197                 }
 198         }
 199         if (ret) {
 200                 btrfs_get_block_group(ret);
 201                 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
 202                         info->first_logical_byte = ret->key.objectid;
 203         }
 204         spin_unlock(&info->block_group_cache_lock);
 205
 206         return ret;
 207 }
 208
 209 static int add_excluded_extent(struct btrfs_fs_info *fs_info,
 210                                u64 start, u64 num_bytes)
 211 {
 212         u64 end = start + num_bytes - 1;
 213         set_extent_bits(&fs_info->freed_extents[0],
 214                         start, end, EXTENT_UPTODATE);
 215         set_extent_bits(&fs_info->freed_extents[1],
 216                         start, end, EXTENT_UPTODATE);
 217         return 0;
 218 }
 219
 220 static void free_excluded_extents(struct btrfs_block_group_cache *cache)
 221 {
 222         struct btrfs_fs_info *fs_info = cache->fs_info;
 223         u64 start, end;
 224
 225         start = cache->key.objectid;
 226         end = start + cache->key.offset - 1;
 227
 228         clear_extent_bits(&fs_info->freed_extents[0],
 229                           start, end, EXTENT_UPTODATE);
 230         clear_extent_bits(&fs_info->freed_extents[1],
 231                           start, end, EXTENT_UPTODATE);
 232 }
 233
 234 static int exclude_super_stripes(struct btrfs_block_group_cache *cache)
 235 {
 236         struct btrfs_fs_info *fs_info = cache->fs_info;
 237         u64 bytenr;
 238         u64 *logical;
 239         int stripe_len;
 240         int i, nr, ret;
 241
 242         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 243                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 244                 cache->bytes_super += stripe_len;
 245                 ret = add_excluded_extent(fs_info, cache->key.objectid,
 246                                           stripe_len);
 247                 if (ret)
 248                         return ret;
 249         }
 250
 251         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 252                 bytenr = btrfs_sb_offset(i);
 253                 ret = btrfs_rmap_block(fs_info, cache->key.objectid,
 254                                        bytenr, &logical, &nr, &stripe_len);
 255                 if (ret)
 256                         return ret;
 257
 258                 while (nr--) {
 259                         u64 start, len;
 260
 261                         if (logical[nr] > cache->key.objectid +
 262                             cache->key.offset)
 263                                 continue;
 264
 265                         if (logical[nr] + stripe_len <= cache->key.objectid)
 266                                 continue;
 267
 268                         start = logical[nr];
 269                         if (start < cache->key.objectid) {
 270                                 start = cache->key.objectid;
 271                                 len = (logical[nr] + stripe_len) - start;
 272                         } else {
 273                                 len = min_t(u64, stripe_len,
 274                                             cache->key.objectid +
 275                                             cache->key.offset - start);
 276                         }
 277
 278                         cache->bytes_super += len;
 279                         ret = add_excluded_extent(fs_info, start, len);
 280                         if (ret) {
 281                                 kfree(logical);
 282                                 return ret;
 283                         }
 284                 }
 285
 286                 kfree(logical);
 287         }
 288         return 0;
 289 }
 290
 291 static struct btrfs_caching_control *
 292 get_caching_control(struct btrfs_block_group_cache *cache)
 293 {
 294         struct btrfs_caching_control *ctl;
 295
 296         spin_lock(&cache->lock);
 297         if (!cache->caching_ctl) {
 298                 spin_unlock(&cache->lock);
 299                 return NULL;
 300         }
 301
 302         ctl = cache->caching_ctl;
 303         refcount_inc(&ctl->count);
 304         spin_unlock(&cache->lock);
 305         return ctl;
 306 }
 307
 308 static void put_caching_control(struct btrfs_caching_control *ctl)
 309 {
 310         if (refcount_dec_and_test(&ctl->count))
 311                 kfree(ctl);
 312 }
 313
 314 #ifdef CONFIG_BTRFS_DEBUG
 315 static void fragment_free_space(struct btrfs_block_group_cache *block_group)
 316 {
 317         struct btrfs_fs_info *fs_info = block_group->fs_info;
 318         u64 start = block_group->key.objectid;
 319         u64 len = block_group->key.offset;
 320         u64 chunk = block_group->flags & BTRFS_BLOCK_GROUP_METADATA ?
 321                 fs_info->nodesize : fs_info->sectorsize;
 322         u64 step = chunk << 1;
 323
 324         while (len > chunk) {
 325                 btrfs_remove_free_space(block_group, start, chunk);
 326                 start += step;
 327                 if (len < step)
 328                         len = 0;
 329                 else
 330                         len -= step;
 331         }
 332 }
 333 #endif
 334
 335 /*
 336  * this is only called by cache_block_group, since we could have freed extents
 337  * we need to check the pinned_extents for any extents that can't be used yet
 338  * since their free space will be released as soon as the transaction commits.
 339  */
 340 u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 341                        u64 start, u64 end)
 342 {
 343         struct btrfs_fs_info *info = block_group->fs_info;
 344         u64 extent_start, extent_end, size, total_added = 0;
 345         int ret;
 346
 347         while (start < end) {
 348                 ret = find_first_extent_bit(info->pinned_extents, start,
 349                                             &extent_start, &extent_end,
 350                                             EXTENT_DIRTY | EXTENT_UPTODATE,
 351                                             NULL);
 352                 if (ret)
 353                         break;
 354
 355                 if (extent_start <= start) {
 356                         start = extent_end + 1;
 357                 } else if (extent_start > start && extent_start < end) {
 358                         size = extent_start - start;
 359                         total_added += size;
 360                         ret = btrfs_add_free_space(block_group, start,
 361                                                    size);
 362                         BUG_ON(ret); /* -ENOMEM or logic error */
 363                         start = extent_end + 1;
 364                 } else {
 365                         break;
 366                 }
 367         }
 368
 369         if (start < end) {
 370                 size = end - start;
 371                 total_added += size;
 372                 ret = btrfs_add_free_space(block_group, start, size);
 373                 BUG_ON(ret); /* -ENOMEM or logic error */
 374         }
 375
 376         return total_added;
 377 }
 378
 379 static int load_extent_tree_free(struct btrfs_caching_control *caching_ctl)
 380 {
 381         struct btrfs_block_group_cache *block_group = caching_ctl->block_group;
 382         struct btrfs_fs_info *fs_info = block_group->fs_info;
 383         struct btrfs_root *extent_root = fs_info->extent_root;
 384         struct btrfs_path *path;
 385         struct extent_buffer *leaf;
 386         struct btrfs_key key;
 387         u64 total_found = 0;
 388         u64 last = 0;
 389         u32 nritems;
 390         int ret;
 391         bool wakeup = true;
 392
 393         path = btrfs_alloc_path();
 394         if (!path)
 395                 return -ENOMEM;
 396
 397         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 398
 399 #ifdef CONFIG_BTRFS_DEBUG
 400         /*
 401          * If we're fragmenting we don't want to make anybody think we can
 402          * allocate from this block group until we've had a chance to fragment
 403          * the free space.
 404          */
 405         if (btrfs_should_fragment_free_space(block_group))
 406                 wakeup = false;
 407 #endif
 408         /*
 409          * We don't want to deadlock with somebody trying to allocate a new
 410          * extent for the extent root while also trying to search the extent
 411          * root to add free space.  So we skip locking and search the commit
 412          * root, since its read-only
 413          */
 414         path->skip_locking = 1;
 415         path->search_commit_root = 1;
 416         path->reada = READA_FORWARD;
 417
 418         key.objectid = last;
 419         key.offset = 0;
 420         key.type = BTRFS_EXTENT_ITEM_KEY;
 421
 422 next:
 423         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 424         if (ret < 0)
 425                 goto out;
 426
 427         leaf = path->nodes[0];
 428         nritems = btrfs_header_nritems(leaf);
 429
 430         while (1) {
 431                 if (btrfs_fs_closing(fs_info) > 1) {
 432                         last = (u64)-1;
 433                         break;
 434                 }
 435
 436                 if (path->slots[0] < nritems) {
 437                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 438                 } else {
 439                         ret = find_next_key(path, 0, &key);
 440                         if (ret)
 441                                 break;
 442
 443                         if (need_resched() ||
 444                             rwsem_is_contended(&fs_info->commit_root_sem)) {
 445                                 if (wakeup)
 446                                         caching_ctl->progress = last;
 447                                 btrfs_release_path(path);
 448                                 up_read(&fs_info->commit_root_sem);
 449                                 mutex_unlock(&caching_ctl->mutex);
 450                                 cond_resched();
 451                                 mutex_lock(&caching_ctl->mutex);
 452                                 down_read(&fs_info->commit_root_sem);
 453                                 goto next;
 454                         }
 455
 456                         ret = btrfs_next_leaf(extent_root, path);
 457                         if (ret < 0)
 458                                 goto out;
 459                         if (ret)
 460                                 break;
 461                         leaf = path->nodes[0];
 462                         nritems = btrfs_header_nritems(leaf);
 463                         continue;
 464                 }
 465
 466                 if (key.objectid < last) {
 467                         key.objectid = last;
 468                         key.offset = 0;
 469                         key.type = BTRFS_EXTENT_ITEM_KEY;
 470
 471                         if (wakeup)
 472                                 caching_ctl->progress = last;
 473                         btrfs_release_path(path);
 474                         goto next;
 475                 }
 476
 477                 if (key.objectid < block_group->key.objectid) {
 478                         path->slots[0]++;
 479                         continue;
 480                 }
 481
 482                 if (key.objectid >= block_group->key.objectid +
 483                     block_group->key.offset)
 484                         break;
 485
 486                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 487                     key.type == BTRFS_METADATA_ITEM_KEY) {
 488                         total_found += add_new_free_space(block_group, last,
 489                                                           key.objectid);
 490                         if (key.type == BTRFS_METADATA_ITEM_KEY)
 491                                 last = key.objectid +
 492                                         fs_info->nodesize;
 493                         else
 494                                 last = key.objectid + key.offset;
 495
 496                         if (total_found > CACHING_CTL_WAKE_UP) {
 497                                 total_found = 0;
 498                                 if (wakeup)
 499                                         wake_up(&caching_ctl->wait);
 500                         }
 501                 }
 502                 path->slots[0]++;
 503         }
 504         ret = 0;
 505
 506         total_found += add_new_free_space(block_group, last,
 507                                           block_group->key.objectid +
 508                                           block_group->key.offset);
 509         caching_ctl->progress = (u64)-1;
 510
 511 out:
 512         btrfs_free_path(path);
 513         return ret;
 514 }
 515
 516 static noinline void caching_thread(struct btrfs_work *work)
 517 {
 518         struct btrfs_block_group_cache *block_group;
 519         struct btrfs_fs_info *fs_info;
 520         struct btrfs_caching_control *caching_ctl;
 521         int ret;
 522
 523         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 524         block_group = caching_ctl->block_group;
 525         fs_info = block_group->fs_info;
 526
 527         mutex_lock(&caching_ctl->mutex);
 528         down_read(&fs_info->commit_root_sem);
 529
 530         if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
 531                 ret = load_free_space_tree(caching_ctl);
 532         else
 533                 ret = load_extent_tree_free(caching_ctl);
 534
 535         spin_lock(&block_group->lock);
 536         block_group->caching_ctl = NULL;
 537         block_group->cached = ret ? BTRFS_CACHE_ERROR : BTRFS_CACHE_FINISHED;
 538         spin_unlock(&block_group->lock);
 539
 540 #ifdef CONFIG_BTRFS_DEBUG
 541         if (btrfs_should_fragment_free_space(block_group)) {
 542                 u64 bytes_used;
 543
 544                 spin_lock(&block_group->space_info->lock);
 545                 spin_lock(&block_group->lock);
 546                 bytes_used = block_group->key.offset -
 547                         btrfs_block_group_used(&block_group->item);
 548                 block_group->space_info->bytes_used += bytes_used >> 1;
 549                 spin_unlock(&block_group->lock);
 550                 spin_unlock(&block_group->space_info->lock);
 551                 fragment_free_space(block_group);
 552         }
 553 #endif
 554
 555         caching_ctl->progress = (u64)-1;
 556
 557         up_read(&fs_info->commit_root_sem);
 558         free_excluded_extents(block_group);
 559         mutex_unlock(&caching_ctl->mutex);
 560
 561         wake_up(&caching_ctl->wait);
 562
 563         put_caching_control(caching_ctl);
 564         btrfs_put_block_group(block_group);
 565 }
 566
 567 static int cache_block_group(struct btrfs_block_group_cache *cache,
 568                              int load_cache_only)
 569 {
 570         DEFINE_WAIT(wait);
 571         struct btrfs_fs_info *fs_info = cache->fs_info;
 572         struct btrfs_caching_control *caching_ctl;
 573         int ret = 0;
 574
 575         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 576         if (!caching_ctl)
 577                 return -ENOMEM;
 578
 579         INIT_LIST_HEAD(&caching_ctl->list);
 580         mutex_init(&caching_ctl->mutex);
 581         init_waitqueue_head(&caching_ctl->wait);
 582         caching_ctl->block_group = cache;
 583         caching_ctl->progress = cache->key.objectid;
 584         refcount_set(&caching_ctl->count, 1);
 585         btrfs_init_work(&caching_ctl->work, btrfs_cache_helper,
 586                         caching_thread, NULL, NULL);
 587
 588         spin_lock(&cache->lock);
 589         /*
 590          * This should be a rare occasion, but this could happen I think in the
 591          * case where one thread starts to load the space cache info, and then
 592          * some other thread starts a transaction commit which tries to do an
 593          * allocation while the other thread is still loading the space cache
 594          * info.  The previous loop should have kept us from choosing this block
 595          * group, but if we've moved to the state where we will wait on caching
 596          * block groups we need to first check if we're doing a fast load here,
 597          * so we can wait for it to finish, otherwise we could end up allocating
 598          * from a block group who's cache gets evicted for one reason or
 599          * another.
 600          */
 601         while (cache->cached == BTRFS_CACHE_FAST) {
 602                 struct btrfs_caching_control *ctl;
 603
 604                 ctl = cache->caching_ctl;
 605                 refcount_inc(&ctl->count);
 606                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 607                 spin_unlock(&cache->lock);
 608
 609                 schedule();
 610
 611                 finish_wait(&ctl->wait, &wait);
 612                 put_caching_control(ctl);
 613                 spin_lock(&cache->lock);
 614         }
 615
 616         if (cache->cached != BTRFS_CACHE_NO) {
 617                 spin_unlock(&cache->lock);
 618                 kfree(caching_ctl);
 619                 return 0;
 620         }
 621         WARN_ON(cache->caching_ctl);
 622         cache->caching_ctl = caching_ctl;
 623         cache->cached = BTRFS_CACHE_FAST;
 624         spin_unlock(&cache->lock);
 625
 626         if (btrfs_test_opt(fs_info, SPACE_CACHE)) {
 627                 mutex_lock(&caching_ctl->mutex);
 628                 ret = load_free_space_cache(fs_info, cache);
 629
 630                 spin_lock(&cache->lock);
 631                 if (ret == 1) {
 632                         cache->caching_ctl = NULL;
 633                         cache->cached = BTRFS_CACHE_FINISHED;
 634                         cache->last_byte_to_unpin = (u64)-1;
 635                         caching_ctl->progress = (u64)-1;
 636                 } else {
 637                         if (load_cache_only) {
 638                                 cache->caching_ctl = NULL;
 639                                 cache->cached = BTRFS_CACHE_NO;
 640                         } else {
 641                                 cache->cached = BTRFS_CACHE_STARTED;
 642                                 cache->has_caching_ctl = 1;
 643                         }
 644                 }
 645                 spin_unlock(&cache->lock);
 646 #ifdef CONFIG_BTRFS_DEBUG
 647                 if (ret == 1 &&
 648                     btrfs_should_fragment_free_space(cache)) {
 649                         u64 bytes_used;
 650
 651                         spin_lock(&cache->space_info->lock);
 652                         spin_lock(&cache->lock);
 653                         bytes_used = cache->key.offset -
 654                                 btrfs_block_group_used(&cache->item);
 655                         cache->space_info->bytes_used += bytes_used >> 1;
 656                         spin_unlock(&cache->lock);
 657                         spin_unlock(&cache->space_info->lock);
 658                         fragment_free_space(cache);
 659                 }
 660 #endif
 661                 mutex_unlock(&caching_ctl->mutex);
 662
 663                 wake_up(&caching_ctl->wait);
 664                 if (ret == 1) {
 665                         put_caching_control(caching_ctl);
 666                         free_excluded_extents(cache);
 667                         return 0;
 668                 }
 669         } else {
 670                 /*
 671                  * We're either using the free space tree or no caching at all.
 672                  * Set cached to the appropriate value and wakeup any waiters.
 673                  */
 674                 spin_lock(&cache->lock);
 675                 if (load_cache_only) {
 676                         cache->caching_ctl = NULL;
 677                         cache->cached = BTRFS_CACHE_NO;
 678                 } else {
 679                         cache->cached = BTRFS_CACHE_STARTED;
 680                         cache->has_caching_ctl = 1;
 681                 }
 682                 spin_unlock(&cache->lock);
 683                 wake_up(&caching_ctl->wait);
 684         }
 685
 686         if (load_cache_only) {
 687                 put_caching_control(caching_ctl);
 688                 return 0;
 689         }
 690
 691         down_write(&fs_info->commit_root_sem);
 692         refcount_inc(&caching_ctl->count);
 693         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 694         up_write(&fs_info->commit_root_sem);
 695
 696         btrfs_get_block_group(cache);
 697
 698         btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
 699
 700         return ret;
 701 }
 702
 703 /*
 704  * return the block group that starts at or after bytenr
 705  */
 706 static struct btrfs_block_group_cache *
 707 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 708 {
 709         return block_group_cache_tree_search(info, bytenr, 0);
 710 }
 711
 712 /*
 713  * return the block group that contains the given bytenr
 714  */
 715 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 716                                                  struct btrfs_fs_info *info,
 717                                                  u64 bytenr)
 718 {
 719         return block_group_cache_tree_search(info, bytenr, 1);
 720 }
 721
 722 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 723                                                   u64 flags)
 724 {
 725         struct list_head *head = &info->space_info;
 726         struct btrfs_space_info *found;
 727
 728         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 729
 730         rcu_read_lock();
 731         list_for_each_entry_rcu(found, head, list) {
 732                 if (found->flags & flags) {
 733                         rcu_read_unlock();
 734                         return found;
 735                 }
 736         }
 737         rcu_read_unlock();
 738         return NULL;
 739 }
 740
 741 static void add_pinned_bytes(struct btrfs_fs_info *fs_info, s64 num_bytes,
 742                              bool metadata, u64 root_objectid)
 743 {
 744         struct btrfs_space_info *space_info;
 745         u64 flags;
 746
 747         if (metadata) {
 748                 if (root_objectid == BTRFS_CHUNK_TREE_OBJECTID)
 749                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
 750                 else
 751                         flags = BTRFS_BLOCK_GROUP_METADATA;
 752         } else {
 753                 flags = BTRFS_BLOCK_GROUP_DATA;
 754         }
 755
 756         space_info = __find_space_info(fs_info, flags);
 757         ASSERT(space_info);
 758         percpu_counter_add_batch(&space_info->total_bytes_pinned, num_bytes,
 759                     BTRFS_TOTAL_BYTES_PINNED_BATCH);
 760 }
 761
 762 /*
 763  * after adding space to the filesystem, we need to clear the full flags
 764  * on all the space infos.
 765  */
 766 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 767 {
 768         struct list_head *head = &info->space_info;
 769         struct btrfs_space_info *found;
 770
 771         rcu_read_lock();
 772         list_for_each_entry_rcu(found, head, list)
 773                 found->full = 0;
 774         rcu_read_unlock();
 775 }
 776
 777 /* simple helper to search for an existing data extent at a given offset */
 778 int btrfs_lookup_data_extent(struct btrfs_fs_info *fs_info, u64 start, u64 len)
 779 {
 780         int ret;
 781         struct btrfs_key key;
 782         struct btrfs_path *path;
 783
 784         path = btrfs_alloc_path();
 785         if (!path)
 786                 return -ENOMEM;
 787
 788         key.objectid = start;
 789         key.offset = len;
 790         key.type = BTRFS_EXTENT_ITEM_KEY;
 791         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
 792         btrfs_free_path(path);
 793         return ret;
 794 }
 795
 796 /*
 797  * helper function to lookup reference count and flags of a tree block.
 798  *
 799  * the head node for delayed ref is used to store the sum of all the
 800  * reference count modifications queued up in the rbtree. the head
 801  * node may also store the extent flags to set. This way you can check
 802  * to see what the reference count and extent flags would be if all of
 803  * the delayed refs are not processed.
 804  */
 805 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 806                              struct btrfs_fs_info *fs_info, u64 bytenr,
 807                              u64 offset, int metadata, u64 *refs, u64 *flags)
 808 {
 809         struct btrfs_delayed_ref_head *head;
 810         struct btrfs_delayed_ref_root *delayed_refs;
 811         struct btrfs_path *path;
 812         struct btrfs_extent_item *ei;
 813         struct extent_buffer *leaf;
 814         struct btrfs_key key;
 815         u32 item_size;
 816         u64 num_refs;
 817         u64 extent_flags;
 818         int ret;
 819
 820         /*
 821          * If we don't have skinny metadata, don't bother doing anything
 822          * different
 823          */
 824         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA)) {
 825                 offset = fs_info->nodesize;
 826                 metadata = 0;
 827         }
 828
 829         path = btrfs_alloc_path();
 830         if (!path)
 831                 return -ENOMEM;
 832
 833         if (!trans) {
 834                 path->skip_locking = 1;
 835                 path->search_commit_root = 1;
 836         }
 837
 838 search_again:
 839         key.objectid = bytenr;
 840         key.offset = offset;
 841         if (metadata)
 842                 key.type = BTRFS_METADATA_ITEM_KEY;
 843         else
 844                 key.type = BTRFS_EXTENT_ITEM_KEY;
 845
 846         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 0);
 847         if (ret < 0)
 848                 goto out_free;
 849
 850         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 851                 if (path->slots[0]) {
 852                         path->slots[0]--;
 853                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 854                                               path->slots[0]);
 855                         if (key.objectid == bytenr &&
 856                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 857                             key.offset == fs_info->nodesize)
 858                                 ret = 0;
 859                 }
 860         }
 861
 862         if (ret == 0) {
 863                 leaf = path->nodes[0];
 864                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 865                 if (item_size >= sizeof(*ei)) {
 866                         ei = btrfs_item_ptr(leaf, path->slots[0],
 867                                             struct btrfs_extent_item);
 868                         num_refs = btrfs_extent_refs(leaf, ei);
 869                         extent_flags = btrfs_extent_flags(leaf, ei);
 870                 } else {
 871                         ret = -EINVAL;
 872                         btrfs_print_v0_err(fs_info);
 873                         if (trans)
 874                                 btrfs_abort_transaction(trans, ret);
 875                         else
 876                                 btrfs_handle_fs_error(fs_info, ret, NULL);
 877
 878                         goto out_free;
 879                 }
 880
 881                 BUG_ON(num_refs == 0);
 882         } else {
 883                 num_refs = 0;
 884                 extent_flags = 0;
 885                 ret = 0;
 886         }
 887
 888         if (!trans)
 889                 goto out;
 890
 891         delayed_refs = &trans->transaction->delayed_refs;
 892         spin_lock(&delayed_refs->lock);
 893         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
 894         if (head) {
 895                 if (!mutex_trylock(&head->mutex)) {
 896                         refcount_inc(&head->refs);
 897                         spin_unlock(&delayed_refs->lock);
 898
 899                         btrfs_release_path(path);
 900
 901                         /*
 902                          * Mutex was contended, block until it's released and try
 903                          * again
 904                          */
 905                         mutex_lock(&head->mutex);
 906                         mutex_unlock(&head->mutex);
 907                         btrfs_put_delayed_ref_head(head);
 908                         goto search_again;
 909                 }
 910                 spin_lock(&head->lock);
 911                 if (head->extent_op && head->extent_op->update_flags)
 912                         extent_flags |= head->extent_op->flags_to_set;
 913                 else
 914                         BUG_ON(num_refs == 0);
 915
 916                 num_refs += head->ref_mod;
 917                 spin_unlock(&head->lock);
 918                 mutex_unlock(&head->mutex);
 919         }
 920         spin_unlock(&delayed_refs->lock);
 921 out:
 922         WARN_ON(num_refs == 0);
 923         if (refs)
 924                 *refs = num_refs;
 925         if (flags)
 926                 *flags = extent_flags;
 927 out_free:
 928         btrfs_free_path(path);
 929         return ret;
 930 }
 931
 932 /*
 933  * Back reference rules.  Back refs have three main goals:
 934  *
 935  * 1) differentiate between all holders of references to an extent so that
 936  *    when a reference is dropped we can make sure it was a valid reference
 937  *    before freeing the extent.
 938  *
 939  * 2) Provide enough information to quickly find the holders of an extent
 940  *    if we notice a given block is corrupted or bad.
 941  *
 942  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 943  *    maintenance.  This is actually the same as #2, but with a slightly
 944  *    different use case.
 945  *
 946  * There are two kinds of back refs. The implicit back refs is optimized
 947  * for pointers in non-shared tree blocks. For a given pointer in a block,
 948  * back refs of this kind provide information about the block's owner tree
 949  * and the pointer's key. These information allow us to find the block by
 950  * b-tree searching. The full back refs is for pointers in tree blocks not
 951  * referenced by their owner trees. The location of tree block is recorded
 952  * in the back refs. Actually the full back refs is generic, and can be
 953  * used in all cases the implicit back refs is used. The major shortcoming
 954  * of the full back refs is its overhead. Every time a tree block gets
 955  * COWed, we have to update back refs entry for all pointers in it.
 956  *
 957  * For a newly allocated tree block, we use implicit back refs for
 958  * pointers in it. This means most tree related operations only involve
 959  * implicit back refs. For a tree block created in old transaction, the
 960  * only way to drop a reference to it is COW it. So we can detect the
 961  * event that tree block loses its owner tree's reference and do the
 962  * back refs conversion.
 963  *
 964  * When a tree block is COWed through a tree, there are four cases:
 965  *
 966  * The reference count of the block is one and the tree is the block's
 967  * owner tree. Nothing to do in this case.
 968  *
 969  * The reference count of the block is one and the tree is not the
 970  * block's owner tree. In this case, full back refs is used for pointers
 971  * in the block. Remove these full back refs, add implicit back refs for
 972  * every pointers in the new block.
 973  *
 974  * The reference count of the block is greater than one and the tree is
 975  * the block's owner tree. In this case, implicit back refs is used for
 976  * pointers in the block. Add full back refs for every pointers in the
 977  * block, increase lower level extents' reference counts. The original
 978  * implicit back refs are entailed to the new block.
 979  *
 980  * The reference count of the block is greater than one and the tree is
 981  * not the block's owner tree. Add implicit back refs for every pointer in
 982  * the new block, increase lower level extents' reference count.
 983  *
 984  * Back Reference Key composing:
 985  *
 986  * The key objectid corresponds to the first byte in the extent,
 987  * The key type is used to differentiate between types of back refs.
 988  * There are different meanings of the key offset for different types
 989  * of back refs.
 990  *
 991  * File extents can be referenced by:
 992  *
 993  * - multiple snapshots, subvolumes, or different generations in one subvol
 994  * - different files inside a single subvolume
 995  * - different offsets inside a file (bookend extents in file.c)
 996  *
 997  * The extent ref structure for the implicit back refs has fields for:
 998  *
 999  * - Objectid of the subvolume root
1000  * - objectid of the file holding the reference
1001  * - original offset in the file
1002  * - how many bookend extents
1003  *
1004  * The key offset for the implicit back refs is hash of the first
1005  * three fields.
1006  *
1007  * The extent ref structure for the full back refs has field for:
1008  *
1009  * - number of pointers in the tree leaf
1010  *
1011  * The key offset for the implicit back refs is the first byte of
1012  * the tree leaf
1013  *
1014  * When a file extent is allocated, The implicit back refs is used.
1015  * the fields are filled in:
1016  *
1017  *     (root_key.objectid, inode objectid, offset in file, 1)
1018  *
1019  * When a file extent is removed file truncation, we find the
1020  * corresponding implicit back refs and check the following fields:
1021  *
1022  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
1023  *
1024  * Btree extents can be referenced by:
1025  *
1026  * - Different subvolumes
1027  *
1028  * Both the implicit back refs and the full back refs for tree blocks
1029  * only consist of key. The key offset for the implicit back refs is
1030  * objectid of block's owner tree. The key offset for the full back refs
1031  * is the first byte of parent block.
1032  *
1033  * When implicit back refs is used, information about the lowest key and
1034  * level of the tree block are required. These information are stored in
1035  * tree block info structure.
1036  */
1037
1038 /*
1039  * is_data == BTRFS_REF_TYPE_BLOCK, tree block type is required,
1040  * is_data == BTRFS_REF_TYPE_DATA, data type is requried,
1041  * is_data == BTRFS_REF_TYPE_ANY, either type is OK.
1042  */
1043 int btrfs_get_extent_inline_ref_type(const struct extent_buffer *eb,
1044                                      struct btrfs_extent_inline_ref *iref,
1045                                      enum btrfs_inline_ref_type is_data)
1046 {
1047         int type = btrfs_extent_inline_ref_type(eb, iref);
1048         u64 offset = btrfs_extent_inline_ref_offset(eb, iref);
1049
1050         if (type == BTRFS_TREE_BLOCK_REF_KEY ||
1051             type == BTRFS_SHARED_BLOCK_REF_KEY ||
1052             type == BTRFS_SHARED_DATA_REF_KEY ||
1053             type == BTRFS_EXTENT_DATA_REF_KEY) {
1054                 if (is_data == BTRFS_REF_TYPE_BLOCK) {
1055                         if (type == BTRFS_TREE_BLOCK_REF_KEY)
1056                                 return type;
1057                         if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1058                                 ASSERT(eb->fs_info);
1059                                 /*
1060                                  * Every shared one has parent tree
1061                                  * block, which must be aligned to
1062                                  * nodesize.
1063                                  */
1064                                 if (offset &&
1065                                     IS_ALIGNED(offset, eb->fs_info->nodesize))
1066                                         return type;
1067                         }
1068                 } else if (is_data == BTRFS_REF_TYPE_DATA) {
1069                         if (type == BTRFS_EXTENT_DATA_REF_KEY)
1070                                 return type;
1071                         if (type == BTRFS_SHARED_DATA_REF_KEY) {
1072                                 ASSERT(eb->fs_info);
1073                                 /*
1074                                  * Every shared one has parent tree
1075                                  * block, which must be aligned to
1076                                  * nodesize.
1077                                  */
1078                                 if (offset &&
1079                                     IS_ALIGNED(offset, eb->fs_info->nodesize))
1080                                         return type;
1081                         }
1082                 } else {
1083                         ASSERT(is_data == BTRFS_REF_TYPE_ANY);
1084                         return type;
1085                 }
1086         }
1087
1088         btrfs_print_leaf((struct extent_buffer *)eb);
1089         btrfs_err(eb->fs_info, "eb %llu invalid extent inline ref type %d",
1090                   eb->start, type);
1091         WARN_ON(1);
1092
1093         return BTRFS_REF_TYPE_INVALID;
1094 }
1095
1096 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1097 {
1098         u32 high_crc = ~(u32)0;
1099         u32 low_crc = ~(u32)0;
1100         __le64 lenum;
1101
1102         lenum = cpu_to_le64(root_objectid);
1103         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
1104         lenum = cpu_to_le64(owner);
1105         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1106         lenum = cpu_to_le64(offset);
1107         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1108
1109         return ((u64)high_crc << 31) ^ (u64)low_crc;
1110 }
1111
1112 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1113                                      struct btrfs_extent_data_ref *ref)
1114 {
1115         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1116                                     btrfs_extent_data_ref_objectid(leaf, ref),
1117                                     btrfs_extent_data_ref_offset(leaf, ref));
1118 }
1119
1120 static int match_extent_data_ref(struct extent_buffer *leaf,
1121                                  struct btrfs_extent_data_ref *ref,
1122                                  u64 root_objectid, u64 owner, u64 offset)
1123 {
1124         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1125             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1126             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1127                 return 0;
1128         return 1;
1129 }
1130
1131 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1132                                            struct btrfs_path *path,
1133                                            u64 bytenr, u64 parent,
1134                                            u64 root_objectid,
1135                                            u64 owner, u64 offset)
1136 {
1137         struct btrfs_root *root = trans->fs_info->extent_root;
1138         struct btrfs_key key;
1139         struct btrfs_extent_data_ref *ref;
1140         struct extent_buffer *leaf;
1141         u32 nritems;
1142         int ret;
1143         int recow;
1144         int err = -ENOENT;
1145
1146         key.objectid = bytenr;
1147         if (parent) {
1148                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1149                 key.offset = parent;
1150         } else {
1151                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1152                 key.offset = hash_extent_data_ref(root_objectid,
1153                                                   owner, offset);
1154         }
1155 again:
1156         recow = 0;
1157         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1158         if (ret < 0) {
1159                 err = ret;
1160                 goto fail;
1161         }
1162
1163         if (parent) {
1164                 if (!ret)
1165                         return 0;
1166                 goto fail;
1167         }
1168
1169         leaf = path->nodes[0];
1170         nritems = btrfs_header_nritems(leaf);
1171         while (1) {
1172                 if (path->slots[0] >= nritems) {
1173                         ret = btrfs_next_leaf(root, path);
1174                         if (ret < 0)
1175                                 err = ret;
1176                         if (ret)
1177                                 goto fail;
1178
1179                         leaf = path->nodes[0];
1180                         nritems = btrfs_header_nritems(leaf);
1181                         recow = 1;
1182                 }
1183
1184                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1185                 if (key.objectid != bytenr ||
1186                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1187                         goto fail;
1188
1189                 ref = btrfs_item_ptr(leaf, path->slots[0],
1190                                      struct btrfs_extent_data_ref);
1191
1192                 if (match_extent_data_ref(leaf, ref, root_objectid,
1193                                           owner, offset)) {
1194                         if (recow) {
1195                                 btrfs_release_path(path);
1196                                 goto again;
1197                         }
1198                         err = 0;
1199                         break;
1200                 }
1201                 path->slots[0]++;
1202         }
1203 fail:
1204         return err;
1205 }
1206
1207 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1208                                            struct btrfs_path *path,
1209                                            u64 bytenr, u64 parent,
1210                                            u64 root_objectid, u64 owner,
1211                                            u64 offset, int refs_to_add)
1212 {
1213         struct btrfs_root *root = trans->fs_info->extent_root;
1214         struct btrfs_key key;
1215         struct extent_buffer *leaf;
1216         u32 size;
1217         u32 num_refs;
1218         int ret;
1219
1220         key.objectid = bytenr;
1221         if (parent) {
1222                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1223                 key.offset = parent;
1224                 size = sizeof(struct btrfs_shared_data_ref);
1225         } else {
1226                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1227                 key.offset = hash_extent_data_ref(root_objectid,
1228                                                   owner, offset);
1229                 size = sizeof(struct btrfs_extent_data_ref);
1230         }
1231
1232         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1233         if (ret && ret != -EEXIST)
1234                 goto fail;
1235
1236         leaf = path->nodes[0];
1237         if (parent) {
1238                 struct btrfs_shared_data_ref *ref;
1239                 ref = btrfs_item_ptr(leaf, path->slots[0],
1240                                      struct btrfs_shared_data_ref);
1241                 if (ret == 0) {
1242                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1243                 } else {
1244                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1245                         num_refs += refs_to_add;
1246                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1247                 }
1248         } else {
1249                 struct btrfs_extent_data_ref *ref;
1250                 while (ret == -EEXIST) {
1251                         ref = btrfs_item_ptr(leaf, path->slots[0],
1252                                              struct btrfs_extent_data_ref);
1253                         if (match_extent_data_ref(leaf, ref, root_objectid,
1254                                                   owner, offset))
1255                                 break;
1256                         btrfs_release_path(path);
1257                         key.offset++;
1258                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1259                                                       size);
1260                         if (ret && ret != -EEXIST)
1261                                 goto fail;
1262
1263                         leaf = path->nodes[0];
1264                 }
1265                 ref = btrfs_item_ptr(leaf, path->slots[0],
1266                                      struct btrfs_extent_data_ref);
1267                 if (ret == 0) {
1268                         btrfs_set_extent_data_ref_root(leaf, ref,
1269                                                        root_objectid);
1270                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1271                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1272                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1273                 } else {
1274                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1275                         num_refs += refs_to_add;
1276                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1277                 }
1278         }
1279         btrfs_mark_buffer_dirty(leaf);
1280         ret = 0;
1281 fail:
1282         btrfs_release_path(path);
1283         return ret;
1284 }
1285
1286 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1287                                            struct btrfs_path *path,
1288                                            int refs_to_drop, int *last_ref)
1289 {
1290         struct btrfs_key key;
1291         struct btrfs_extent_data_ref *ref1 = NULL;
1292         struct btrfs_shared_data_ref *ref2 = NULL;
1293         struct extent_buffer *leaf;
1294         u32 num_refs = 0;
1295         int ret = 0;
1296
1297         leaf = path->nodes[0];
1298         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1299
1300         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1301                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1302                                       struct btrfs_extent_data_ref);
1303                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1304         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1305                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1306                                       struct btrfs_shared_data_ref);
1307                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1308         } else if (unlikely(key.type == BTRFS_EXTENT_REF_V0_KEY)) {
1309                 btrfs_print_v0_err(trans->fs_info);
1310                 btrfs_abort_transaction(trans, -EINVAL);
1311                 return -EINVAL;
1312         } else {
1313                 BUG();
1314         }
1315
1316         BUG_ON(num_refs < refs_to_drop);
1317         num_refs -= refs_to_drop;
1318
1319         if (num_refs == 0) {
1320                 ret = btrfs_del_item(trans, trans->fs_info->extent_root, path);
1321                 *last_ref = 1;
1322         } else {
1323                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1324                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1325                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1326                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1327                 btrfs_mark_buffer_dirty(leaf);
1328         }
1329         return ret;
1330 }
1331
1332 static noinline u32 extent_data_ref_count(struct btrfs_path *path,
1333                                           struct btrfs_extent_inline_ref *iref)
1334 {
1335         struct btrfs_key key;
1336         struct extent_buffer *leaf;
1337         struct btrfs_extent_data_ref *ref1;
1338         struct btrfs_shared_data_ref *ref2;
1339         u32 num_refs = 0;
1340         int type;
1341
1342         leaf = path->nodes[0];
1343         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1344
1345         BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
1346         if (iref) {
1347                 /*
1348                  * If type is invalid, we should have bailed out earlier than
1349                  * this call.
1350                  */
1351                 type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
1352                 ASSERT(type != BTRFS_REF_TYPE_INVALID);
1353                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1354                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1355                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1356                 } else {
1357                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1358                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1359                 }
1360         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1361                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1362                                       struct btrfs_extent_data_ref);
1363                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1364         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1365                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1366                                       struct btrfs_shared_data_ref);
1367                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1368         } else {
1369                 WARN_ON(1);
1370         }
1371         return num_refs;
1372 }
1373
1374 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1375                                           struct btrfs_path *path,
1376                                           u64 bytenr, u64 parent,
1377                                           u64 root_objectid)
1378 {
1379         struct btrfs_root *root = trans->fs_info->extent_root;
1380         struct btrfs_key key;
1381         int ret;
1382
1383         key.objectid = bytenr;
1384         if (parent) {
1385                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1386                 key.offset = parent;
1387         } else {
1388                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1389                 key.offset = root_objectid;
1390         }
1391
1392         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1393         if (ret > 0)
1394                 ret = -ENOENT;
1395         return ret;
1396 }
1397
1398 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1399                                           struct btrfs_path *path,
1400                                           u64 bytenr, u64 parent,
1401                                           u64 root_objectid)
1402 {
1403         struct btrfs_key key;
1404         int ret;
1405
1406         key.objectid = bytenr;
1407         if (parent) {
1408                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1409                 key.offset = parent;
1410         } else {
1411                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1412                 key.offset = root_objectid;
1413         }
1414
1415         ret = btrfs_insert_empty_item(trans, trans->fs_info->extent_root,
1416                                       path, &key, 0);
1417         btrfs_release_path(path);
1418         return ret;
1419 }
1420
1421 static inline int extent_ref_type(u64 parent, u64 owner)
1422 {
1423         int type;
1424         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1425                 if (parent > 0)
1426                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1427                 else
1428                         type = BTRFS_TREE_BLOCK_REF_KEY;
1429         } else {
1430                 if (parent > 0)
1431                         type = BTRFS_SHARED_DATA_REF_KEY;
1432                 else
1433                         type = BTRFS_EXTENT_DATA_REF_KEY;
1434         }
1435         return type;
1436 }
1437
1438 static int find_next_key(struct btrfs_path *path, int level,
1439                          struct btrfs_key *key)
1440
1441 {
1442         for (; level < BTRFS_MAX_LEVEL; level++) {
1443                 if (!path->nodes[level])
1444                         break;
1445                 if (path->slots[level] + 1 >=
1446                     btrfs_header_nritems(path->nodes[level]))
1447                         continue;
1448                 if (level == 0)
1449                         btrfs_item_key_to_cpu(path->nodes[level], key,
1450                                               path->slots[level] + 1);
1451                 else
1452                         btrfs_node_key_to_cpu(path->nodes[level], key,
1453                                               path->slots[level] + 1);
1454                 return 0;
1455         }
1456         return 1;
1457 }
1458
1459 /*
1460  * look for inline back ref. if back ref is found, *ref_ret is set
1461  * to the address of inline back ref, and 0 is returned.
1462  *
1463  * if back ref isn't found, *ref_ret is set to the address where it
1464  * should be inserted, and -ENOENT is returned.
1465  *
1466  * if insert is true and there are too many inline back refs, the path
1467  * points to the extent item, and -EAGAIN is returned.
1468  *
1469  * NOTE: inline back refs are ordered in the same way that back ref
1470  *       items in the tree are ordered.
1471  */
1472 static noinline_for_stack
1473 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1474                                  struct btrfs_path *path,
1475                                  struct btrfs_extent_inline_ref **ref_ret,
1476                                  u64 bytenr, u64 num_bytes,
1477                                  u64 parent, u64 root_objectid,
1478                                  u64 owner, u64 offset, int insert)
1479 {
1480         struct btrfs_fs_info *fs_info = trans->fs_info;
1481         struct btrfs_root *root = fs_info->extent_root;
1482         struct btrfs_key key;
1483         struct extent_buffer *leaf;
1484         struct btrfs_extent_item *ei;
1485         struct btrfs_extent_inline_ref *iref;
1486         u64 flags;
1487         u64 item_size;
1488         unsigned long ptr;
1489         unsigned long end;
1490         int extra_size;
1491         int type;
1492         int want;
1493         int ret;
1494         int err = 0;
1495         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
1496         int needed;
1497
1498         key.objectid = bytenr;
1499         key.type = BTRFS_EXTENT_ITEM_KEY;
1500         key.offset = num_bytes;
1501
1502         want = extent_ref_type(parent, owner);
1503         if (insert) {
1504                 extra_size = btrfs_extent_inline_ref_size(want);
1505                 path->keep_locks = 1;
1506         } else
1507                 extra_size = -1;
1508
1509         /*
1510          * Owner is our level, so we can just add one to get the level for the
1511          * block we are interested in.
1512          */
1513         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
1514                 key.type = BTRFS_METADATA_ITEM_KEY;
1515                 key.offset = owner;
1516         }
1517
1518 again:
1519         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1520         if (ret < 0) {
1521                 err = ret;
1522                 goto out;
1523         }
1524
1525         /*
1526          * We may be a newly converted file system which still has the old fat
1527          * extent entries for metadata, so try and see if we have one of those.
1528          */
1529         if (ret > 0 && skinny_metadata) {
1530                 skinny_metadata = false;
1531                 if (path->slots[0]) {
1532                         path->slots[0]--;
1533                         btrfs_item_key_to_cpu(path->nodes[0], &key,
1534                                               path->slots[0]);
1535                         if (key.objectid == bytenr &&
1536                             key.type == BTRFS_EXTENT_ITEM_KEY &&
1537                             key.offset == num_bytes)
1538                                 ret = 0;
1539                 }
1540                 if (ret) {
1541                         key.objectid = bytenr;
1542                         key.type = BTRFS_EXTENT_ITEM_KEY;
1543                         key.offset = num_bytes;
1544                         btrfs_release_path(path);
1545                         goto again;
1546                 }
1547         }
1548
1549         if (ret && !insert) {
1550                 err = -ENOENT;
1551                 goto out;
1552         } else if (WARN_ON(ret)) {
1553                 err = -EIO;
1554                 goto out;
1555         }
1556
1557         leaf = path->nodes[0];
1558         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1559         if (unlikely(item_size < sizeof(*ei))) {
1560                 err = -EINVAL;
1561                 btrfs_print_v0_err(fs_info);
1562                 btrfs_abort_transaction(trans, err);
1563                 goto out;
1564         }
1565
1566         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1567         flags = btrfs_extent_flags(leaf, ei);
1568
1569         ptr = (unsigned long)(ei + 1);
1570         end = (unsigned long)ei + item_size;
1571
1572         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
1573                 ptr += sizeof(struct btrfs_tree_block_info);
1574                 BUG_ON(ptr > end);
1575         }
1576
1577         if (owner >= BTRFS_FIRST_FREE_OBJECTID)
1578                 needed = BTRFS_REF_TYPE_DATA;
1579         else
1580                 needed = BTRFS_REF_TYPE_BLOCK;
1581
1582         err = -ENOENT;
1583         while (1) {
1584                 if (ptr >= end) {
1585                         WARN_ON(ptr > end);
1586                         break;
1587                 }
1588                 iref = (struct btrfs_extent_inline_ref *)ptr;
1589                 type = btrfs_get_extent_inline_ref_type(leaf, iref, needed);
1590                 if (type == BTRFS_REF_TYPE_INVALID) {
1591                         err = -EUCLEAN;
1592                         goto out;
1593                 }
1594
1595                 if (want < type)
1596                         break;
1597                 if (want > type) {
1598                         ptr += btrfs_extent_inline_ref_size(type);
1599                         continue;
1600                 }
1601
1602                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1603                         struct btrfs_extent_data_ref *dref;
1604                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1605                         if (match_extent_data_ref(leaf, dref, root_objectid,
1606                                                   owner, offset)) {
1607                                 err = 0;
1608                                 break;
1609                         }
1610                         if (hash_extent_data_ref_item(leaf, dref) <
1611                             hash_extent_data_ref(root_objectid, owner, offset))
1612                                 break;
1613                 } else {
1614                         u64 ref_offset;
1615                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1616                         if (parent > 0) {
1617                                 if (parent == ref_offset) {
1618                                         err = 0;
1619                                         break;
1620                                 }
1621                                 if (ref_offset < parent)
1622                                         break;
1623                         } else {
1624                                 if (root_objectid == ref_offset) {
1625                                         err = 0;
1626                                         break;
1627                                 }
1628                                 if (ref_offset < root_objectid)
1629                                         break;
1630                         }
1631                 }
1632                 ptr += btrfs_extent_inline_ref_size(type);
1633         }
1634         if (err == -ENOENT && insert) {
1635                 if (item_size + extra_size >=
1636                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1637                         err = -EAGAIN;
1638                         goto out;
1639                 }
1640                 /*
1641                  * To add new inline back ref, we have to make sure
1642                  * there is no corresponding back ref item.
1643                  * For simplicity, we just do not add new inline back
1644                  * ref if there is any kind of item for this block
1645                  */
1646                 if (find_next_key(path, 0, &key) == 0 &&
1647                     key.objectid == bytenr &&
1648                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1649                         err = -EAGAIN;
1650                         goto out;
1651                 }
1652         }
1653         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1654 out:
1655         if (insert) {
1656                 path->keep_locks = 0;
1657                 btrfs_unlock_up_safe(path, 1);
1658         }
1659         return err;
1660 }
1661
1662 /*
1663  * helper to add new inline back ref
1664  */
1665 static noinline_for_stack
1666 void setup_inline_extent_backref(struct btrfs_fs_info *fs_info,
1667                                  struct btrfs_path *path,
1668                                  struct btrfs_extent_inline_ref *iref,
1669                                  u64 parent, u64 root_objectid,
1670                                  u64 owner, u64 offset, int refs_to_add,
1671                                  struct btrfs_delayed_extent_op *extent_op)
1672 {
1673         struct extent_buffer *leaf;
1674         struct btrfs_extent_item *ei;
1675         unsigned long ptr;
1676         unsigned long end;
1677         unsigned long item_offset;
1678         u64 refs;
1679         int size;
1680         int type;
1681
1682         leaf = path->nodes[0];
1683         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1684         item_offset = (unsigned long)iref - (unsigned long)ei;
1685
1686         type = extent_ref_type(parent, owner);
1687         size = btrfs_extent_inline_ref_size(type);
1688
1689         btrfs_extend_item(fs_info, path, size);
1690
1691         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1692         refs = btrfs_extent_refs(leaf, ei);
1693         refs += refs_to_add;
1694         btrfs_set_extent_refs(leaf, ei, refs);
1695         if (extent_op)
1696                 __run_delayed_extent_op(extent_op, leaf, ei);
1697
1698         ptr = (unsigned long)ei + item_offset;
1699         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1700         if (ptr < end - size)
1701                 memmove_extent_buffer(leaf, ptr + size, ptr,
1702                                       end - size - ptr);
1703
1704         iref = (struct btrfs_extent_inline_ref *)ptr;
1705         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1706         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1707                 struct btrfs_extent_data_ref *dref;
1708                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1709                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1710                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1711                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1712                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1713         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1714                 struct btrfs_shared_data_ref *sref;
1715                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1716                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1717                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1718         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1719                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1720         } else {
1721                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1722         }
1723         btrfs_mark_buffer_dirty(leaf);
1724 }
1725
1726 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1727                                  struct btrfs_path *path,
1728                                  struct btrfs_extent_inline_ref **ref_ret,
1729                                  u64 bytenr, u64 num_bytes, u64 parent,
1730                                  u64 root_objectid, u64 owner, u64 offset)
1731 {
1732         int ret;
1733
1734         ret = lookup_inline_extent_backref(trans, path, ref_ret, bytenr,
1735                                            num_bytes, parent, root_objectid,
1736                                            owner, offset, 0);
1737         if (ret != -ENOENT)
1738                 return ret;
1739
1740         btrfs_release_path(path);
1741         *ref_ret = NULL;
1742
1743         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1744                 ret = lookup_tree_block_ref(trans, path, bytenr, parent,
1745                                             root_objectid);
1746         } else {
1747                 ret = lookup_extent_data_ref(trans, path, bytenr, parent,
1748                                              root_objectid, owner, offset);
1749         }
1750         return ret;
1751 }
1752
1753 /*
1754  * helper to update/remove inline back ref
1755  */
1756 static noinline_for_stack
1757 void update_inline_extent_backref(struct btrfs_path *path,
1758                                   struct btrfs_extent_inline_ref *iref,
1759                                   int refs_to_mod,
1760                                   struct btrfs_delayed_extent_op *extent_op,
1761                                   int *last_ref)
1762 {
1763         struct extent_buffer *leaf = path->nodes[0];
1764         struct btrfs_fs_info *fs_info = leaf->fs_info;
1765         struct btrfs_extent_item *ei;
1766         struct btrfs_extent_data_ref *dref = NULL;
1767         struct btrfs_shared_data_ref *sref = NULL;
1768         unsigned long ptr;
1769         unsigned long end;
1770         u32 item_size;
1771         int size;
1772         int type;
1773         u64 refs;
1774
1775         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1776         refs = btrfs_extent_refs(leaf, ei);
1777         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1778         refs += refs_to_mod;
1779         btrfs_set_extent_refs(leaf, ei, refs);
1780         if (extent_op)
1781                 __run_delayed_extent_op(extent_op, leaf, ei);
1782
1783         /*
1784          * If type is invalid, we should have bailed out after
1785          * lookup_inline_extent_backref().
1786          */
1787         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_ANY);
1788         ASSERT(type != BTRFS_REF_TYPE_INVALID);
1789
1790         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1791                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1792                 refs = btrfs_extent_data_ref_count(leaf, dref);
1793         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1794                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1795                 refs = btrfs_shared_data_ref_count(leaf, sref);
1796         } else {
1797                 refs = 1;
1798                 BUG_ON(refs_to_mod != -1);
1799         }
1800
1801         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1802         refs += refs_to_mod;
1803
1804         if (refs > 0) {
1805                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1806                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1807                 else
1808                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1809         } else {
1810                 *last_ref = 1;
1811                 size =  btrfs_extent_inline_ref_size(type);
1812                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1813                 ptr = (unsigned long)iref;
1814                 end = (unsigned long)ei + item_size;
1815                 if (ptr + size < end)
1816                         memmove_extent_buffer(leaf, ptr, ptr + size,
1817                                               end - ptr - size);
1818                 item_size -= size;
1819                 btrfs_truncate_item(fs_info, path, item_size, 1);
1820         }
1821         btrfs_mark_buffer_dirty(leaf);
1822 }
1823
1824 static noinline_for_stack
1825 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1826                                  struct btrfs_path *path,
1827                                  u64 bytenr, u64 num_bytes, u64 parent,
1828                                  u64 root_objectid, u64 owner,
1829                                  u64 offset, int refs_to_add,
1830                                  struct btrfs_delayed_extent_op *extent_op)
1831 {
1832         struct btrfs_extent_inline_ref *iref;
1833         int ret;
1834
1835         ret = lookup_inline_extent_backref(trans, path, &iref, bytenr,
1836                                            num_bytes, parent, root_objectid,
1837                                            owner, offset, 1);
1838         if (ret == 0) {
1839                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1840                 update_inline_extent_backref(path, iref, refs_to_add,
1841                                              extent_op, NULL);
1842         } else if (ret == -ENOENT) {
1843                 setup_inline_extent_backref(trans->fs_info, path, iref, parent,
1844                                             root_objectid, owner, offset,
1845                                             refs_to_add, extent_op);
1846                 ret = 0;
1847         }
1848         return ret;
1849 }
1850
1851 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1852                                  struct btrfs_path *path,
1853                                  u64 bytenr, u64 parent, u64 root_objectid,
1854                                  u64 owner, u64 offset, int refs_to_add)
1855 {
1856         int ret;
1857         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1858                 BUG_ON(refs_to_add != 1);
1859                 ret = insert_tree_block_ref(trans, path, bytenr, parent,
1860                                             root_objectid);
1861         } else {
1862                 ret = insert_extent_data_ref(trans, path, bytenr, parent,
1863                                              root_objectid, owner, offset,
1864                                              refs_to_add);
1865         }
1866         return ret;
1867 }
1868
1869 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1870                                  struct btrfs_path *path,
1871                                  struct btrfs_extent_inline_ref *iref,
1872                                  int refs_to_drop, int is_data, int *last_ref)
1873 {
1874         int ret = 0;
1875
1876         BUG_ON(!is_data && refs_to_drop != 1);
1877         if (iref) {
1878                 update_inline_extent_backref(path, iref, -refs_to_drop, NULL,
1879                                              last_ref);
1880         } else if (is_data) {
1881                 ret = remove_extent_data_ref(trans, path, refs_to_drop,
1882                                              last_ref);
1883         } else {
1884                 *last_ref = 1;
1885                 ret = btrfs_del_item(trans, trans->fs_info->extent_root, path);
1886         }
1887         return ret;
1888 }
1889
1890 #define in_range(b, first, len)        ((b) >= (first) && (b) < (first) + (len))
1891 static int btrfs_issue_discard(struct block_device *bdev, u64 start, u64 len,
1892                                u64 *discarded_bytes)
1893 {
1894         int j, ret = 0;
1895         u64 bytes_left, end;
1896         u64 aligned_start = ALIGN(start, 1 << 9);
1897
1898         if (WARN_ON(start != aligned_start)) {
1899                 len -= aligned_start - start;
1900                 len = round_down(len, 1 << 9);
1901                 start = aligned_start;
1902         }
1903
1904         *discarded_bytes = 0;
1905
1906         if (!len)
1907                 return 0;
1908
1909         end = start + len;
1910         bytes_left = len;
1911
1912         /* Skip any superblocks on this device. */
1913         for (j = 0; j < BTRFS_SUPER_MIRROR_MAX; j++) {
1914                 u64 sb_start = btrfs_sb_offset(j);
1915                 u64 sb_end = sb_start + BTRFS_SUPER_INFO_SIZE;
1916                 u64 size = sb_start - start;
1917
1918                 if (!in_range(sb_start, start, bytes_left) &&
1919                     !in_range(sb_end, start, bytes_left) &&
1920                     !in_range(start, sb_start, BTRFS_SUPER_INFO_SIZE))
1921                         continue;
1922
1923                 /*
1924                  * Superblock spans beginning of range.  Adjust start and
1925                  * try again.
1926                  */
1927                 if (sb_start <= start) {
1928                         start += sb_end - start;
1929                         if (start > end) {
1930                                 bytes_left = 0;
1931                                 break;
1932                         }
1933                         bytes_left = end - start;
1934                         continue;
1935                 }
1936
1937                 if (size) {
1938                         ret = blkdev_issue_discard(bdev, start >> 9, size >> 9,
1939                                                    GFP_NOFS, 0);
1940                         if (!ret)
1941                                 *discarded_bytes += size;
1942                         else if (ret != -EOPNOTSUPP)
1943                                 return ret;
1944                 }
1945
1946                 start = sb_end;
1947                 if (start > end) {
1948                         bytes_left = 0;
1949                         break;
1950                 }
1951                 bytes_left = end - start;
1952         }
1953
1954         if (bytes_left) {
1955                 ret = blkdev_issue_discard(bdev, start >> 9, bytes_left >> 9,
1956                                            GFP_NOFS, 0);
1957                 if (!ret)
1958                         *discarded_bytes += bytes_left;
1959         }
1960         return ret;
1961 }
1962
1963 int btrfs_discard_extent(struct btrfs_fs_info *fs_info, u64 bytenr,
1964                          u64 num_bytes, u64 *actual_bytes)
1965 {
1966         int ret;
1967         u64 discarded_bytes = 0;
1968         struct btrfs_bio *bbio = NULL;
1969
1970
1971         /*
1972          * Avoid races with device replace and make sure our bbio has devices
1973          * associated to its stripes that don't go away while we are discarding.
1974          */
1975         btrfs_bio_counter_inc_blocked(fs_info);
1976         /* Tell the block device(s) that the sectors can be discarded */
1977         ret = btrfs_map_block(fs_info, BTRFS_MAP_DISCARD, bytenr, &num_bytes,
1978                               &bbio, 0);
1979         /* Error condition is -ENOMEM */
1980         if (!ret) {
1981                 struct btrfs_bio_stripe *stripe = bbio->stripes;
1982                 int i;
1983
1984
1985                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
1986                         u64 bytes;
1987                         struct request_queue *req_q;
1988
1989                         if (!stripe->dev->bdev) {
1990                                 ASSERT(btrfs_test_opt(fs_info, DEGRADED));
1991                                 continue;
1992                         }
1993                         req_q = bdev_get_queue(stripe->dev->bdev);
1994                         if (!blk_queue_discard(req_q))
1995                                 continue;
1996
1997                         ret = btrfs_issue_discard(stripe->dev->bdev,
1998                                                   stripe->physical,
1999                                                   stripe->length,
2000                                                   &bytes);
2001                         if (!ret)
2002                                 discarded_bytes += bytes;
2003                         else if (ret != -EOPNOTSUPP)
2004                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
2005
2006                         /*
2007                          * Just in case we get back EOPNOTSUPP for some reason,
2008                          * just ignore the return value so we don't screw up
2009                          * people calling discard_extent.
2010                          */
2011                         ret = 0;
2012                 }
2013                 btrfs_put_bbio(bbio);
2014         }
2015         btrfs_bio_counter_dec(fs_info);
2016
2017         if (actual_bytes)
2018                 *actual_bytes = discarded_bytes;
2019
2020
2021         if (ret == -EOPNOTSUPP)
2022                 ret = 0;
2023         return ret;
2024 }
2025
2026 /* Can return -ENOMEM */
2027 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2028                          struct btrfs_root *root,
2029                          u64 bytenr, u64 num_bytes, u64 parent,
2030                          u64 root_objectid, u64 owner, u64 offset)
2031 {
2032         struct btrfs_fs_info *fs_info = root->fs_info;
2033         int old_ref_mod, new_ref_mod;
2034         int ret;
2035
2036         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
2037                root_objectid == BTRFS_TREE_LOG_OBJECTID);
2038
2039         btrfs_ref_tree_mod(root, bytenr, num_bytes, parent, root_objectid,
2040                            owner, offset, BTRFS_ADD_DELAYED_REF);
2041
2042         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
2043                 ret = btrfs_add_delayed_tree_ref(trans, bytenr,
2044                                                  num_bytes, parent,
2045                                                  root_objectid, (int)owner,
2046                                                  BTRFS_ADD_DELAYED_REF, NULL,
2047                                                  &old_ref_mod, &new_ref_mod);
2048         } else {
2049                 ret = btrfs_add_delayed_data_ref(trans, bytenr,
2050                                                  num_bytes, parent,
2051                                                  root_objectid, owner, offset,
2052                                                  0, BTRFS_ADD_DELAYED_REF,
2053                                                  &old_ref_mod, &new_ref_mod);
2054         }
2055
2056         if (ret == 0 && old_ref_mod < 0 && new_ref_mod >= 0) {
2057                 bool metadata = owner < BTRFS_FIRST_FREE_OBJECTID;
2058
2059                 add_pinned_bytes(fs_info, -num_bytes, metadata, root_objectid);
2060         }
2061
2062         return ret;
2063 }
2064
2065 /*
2066  * __btrfs_inc_extent_ref - insert backreference for a given extent
2067  *
2068  * @trans:          Handle of transaction
2069  *
2070  * @node:           The delayed ref node used to get the bytenr/length for
2071  *                  extent whose references are incremented.
2072  *
2073  * @parent:         If this is a shared extent (BTRFS_SHARED_DATA_REF_KEY/
2074  *                  BTRFS_SHARED_BLOCK_REF_KEY) then it holds the logical
2075  *                  bytenr of the parent block. Since new extents are always
2076  *                  created with indirect references, this will only be the case
2077  *                  when relocating a shared extent. In that case, root_objectid
2078  *                  will be BTRFS_TREE_RELOC_OBJECTID. Otheriwse, parent must
2079  *                  be 0
2080  *
2081  * @root_objectid:  The id of the root where this modification has originated,
2082  *                  this can be either one of the well-known metadata trees or
2083  *                  the subvolume id which references this extent.
2084  *
2085  * @owner:          For data extents it is the inode number of the owning file.
2086  *                  For metadata extents this parameter holds the level in the
2087  *                  tree of the extent.
2088  *
2089  * @offset:         For metadata extents the offset is ignored and is currently
2090  *                  always passed as 0. For data extents it is the fileoffset
2091  *                  this extent belongs to.
2092  *
2093  * @refs_to_add     Number of references to add
2094  *
2095  * @extent_op       Pointer to a structure, holding information necessary when
2096  *                  updating a tree block's flags
2097  *
2098  */
2099 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2100                                   struct btrfs_delayed_ref_node *node,
2101                                   u64 parent, u64 root_objectid,
2102                                   u64 owner, u64 offset, int refs_to_add,
2103                                   struct btrfs_delayed_extent_op *extent_op)
2104 {
2105         struct btrfs_path *path;
2106         struct extent_buffer *leaf;
2107         struct btrfs_extent_item *item;
2108         struct btrfs_key key;
2109         u64 bytenr = node->bytenr;
2110         u64 num_bytes = node->num_bytes;
2111         u64 refs;
2112         int ret;
2113
2114         path = btrfs_alloc_path();
2115         if (!path)
2116                 return -ENOMEM;
2117
2118         path->reada = READA_FORWARD;
2119         path->leave_spinning = 1;
2120         /* this will setup the path even if it fails to insert the back ref */
2121         ret = insert_inline_extent_backref(trans, path, bytenr, num_bytes,
2122                                            parent, root_objectid, owner,
2123                                            offset, refs_to_add, extent_op);
2124         if ((ret < 0 && ret != -EAGAIN) || !ret)
2125                 goto out;
2126
2127         /*
2128          * Ok we had -EAGAIN which means we didn't have space to insert and
2129          * inline extent ref, so just update the reference count and add a
2130          * normal backref.
2131          */
2132         leaf = path->nodes[0];
2133         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2134         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2135         refs = btrfs_extent_refs(leaf, item);
2136         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
2137         if (extent_op)
2138                 __run_delayed_extent_op(extent_op, leaf, item);
2139
2140         btrfs_mark_buffer_dirty(leaf);
2141         btrfs_release_path(path);
2142
2143         path->reada = READA_FORWARD;
2144         path->leave_spinning = 1;
2145         /* now insert the actual backref */
2146         ret = insert_extent_backref(trans, path, bytenr, parent, root_objectid,
2147                                     owner, offset, refs_to_add);
2148         if (ret)
2149                 btrfs_abort_transaction(trans, ret);
2150 out:
2151         btrfs_free_path(path);
2152         return ret;
2153 }
2154
2155 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
2156                                 struct btrfs_delayed_ref_node *node,
2157                                 struct btrfs_delayed_extent_op *extent_op,
2158                                 int insert_reserved)
2159 {
2160         int ret = 0;
2161         struct btrfs_delayed_data_ref *ref;
2162         struct btrfs_key ins;
2163         u64 parent = 0;
2164         u64 ref_root = 0;
2165         u64 flags = 0;
2166
2167         ins.objectid = node->bytenr;
2168         ins.offset = node->num_bytes;
2169         ins.type = BTRFS_EXTENT_ITEM_KEY;
2170
2171         ref = btrfs_delayed_node_to_data_ref(node);
2172         trace_run_delayed_data_ref(trans->fs_info, node, ref, node->action);
2173
2174         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2175                 parent = ref->parent;
2176         ref_root = ref->root;
2177
2178         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2179                 if (extent_op)
2180                         flags |= extent_op->flags_to_set;
2181                 ret = alloc_reserved_file_extent(trans, parent, ref_root,
2182                                                  flags, ref->objectid,
2183                                                  ref->offset, &ins,
2184                                                  node->ref_mod);
2185         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2186                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
2187                                              ref->objectid, ref->offset,
2188                                              node->ref_mod, extent_op);
2189         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2190                 ret = __btrfs_free_extent(trans, node, parent,
2191                                           ref_root, ref->objectid,
2192                                           ref->offset, node->ref_mod,
2193                                           extent_op);
2194         } else {
2195                 BUG();
2196         }
2197         return ret;
2198 }
2199
2200 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2201                                     struct extent_buffer *leaf,
2202                                     struct btrfs_extent_item *ei)
2203 {
2204         u64 flags = btrfs_extent_flags(leaf, ei);
2205         if (extent_op->update_flags) {
2206                 flags |= extent_op->flags_to_set;
2207                 btrfs_set_extent_flags(leaf, ei, flags);
2208         }
2209
2210         if (extent_op->update_key) {
2211                 struct btrfs_tree_block_info *bi;
2212                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2213                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2214                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2215         }
2216 }
2217
2218 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2219                                  struct btrfs_delayed_ref_head *head,
2220                                  struct btrfs_delayed_extent_op *extent_op)
2221 {
2222         struct btrfs_fs_info *fs_info = trans->fs_info;
2223         struct btrfs_key key;
2224         struct btrfs_path *path;
2225         struct btrfs_extent_item *ei;
2226         struct extent_buffer *leaf;
2227         u32 item_size;
2228         int ret;
2229         int err = 0;
2230         int metadata = !extent_op->is_data;
2231
2232         if (trans->aborted)
2233                 return 0;
2234
2235         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2236                 metadata = 0;
2237
2238         path = btrfs_alloc_path();
2239         if (!path)
2240                 return -ENOMEM;
2241
2242         key.objectid = head->bytenr;
2243
2244         if (metadata) {
2245                 key.type = BTRFS_METADATA_ITEM_KEY;
2246                 key.offset = extent_op->level;
2247         } else {
2248                 key.type = BTRFS_EXTENT_ITEM_KEY;
2249                 key.offset = head->num_bytes;
2250         }
2251
2252 again:
2253         path->reada = READA_FORWARD;
2254         path->leave_spinning = 1;
2255         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 1);
2256         if (ret < 0) {
2257                 err = ret;
2258                 goto out;
2259         }
2260         if (ret > 0) {
2261                 if (metadata) {
2262                         if (path->slots[0] > 0) {
2263                                 path->slots[0]--;
2264                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
2265                                                       path->slots[0]);
2266                                 if (key.objectid == head->bytenr &&
2267                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
2268                                     key.offset == head->num_bytes)
2269                                         ret = 0;
2270                         }
2271                         if (ret > 0) {
2272                                 btrfs_release_path(path);
2273                                 metadata = 0;
2274
2275                                 key.objectid = head->bytenr;
2276                                 key.offset = head->num_bytes;
2277                                 key.type = BTRFS_EXTENT_ITEM_KEY;
2278                                 goto again;
2279                         }
2280                 } else {
2281                         err = -EIO;
2282                         goto out;
2283                 }
2284         }
2285
2286         leaf = path->nodes[0];
2287         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2288
2289         if (unlikely(item_size < sizeof(*ei))) {
2290                 err = -EINVAL;
2291                 btrfs_print_v0_err(fs_info);
2292                 btrfs_abort_transaction(trans, err);
2293                 goto out;
2294         }
2295
2296         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2297         __run_delayed_extent_op(extent_op, leaf, ei);
2298
2299         btrfs_mark_buffer_dirty(leaf);
2300 out:
2301         btrfs_free_path(path);
2302         return err;
2303 }
2304
2305 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2306                                 struct btrfs_delayed_ref_node *node,
2307                                 struct btrfs_delayed_extent_op *extent_op,
2308                                 int insert_reserved)
2309 {
2310         int ret = 0;
2311         struct btrfs_delayed_tree_ref *ref;
2312         u64 parent = 0;
2313         u64 ref_root = 0;
2314
2315         ref = btrfs_delayed_node_to_tree_ref(node);
2316         trace_run_delayed_tree_ref(trans->fs_info, node, ref, node->action);
2317
2318         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2319                 parent = ref->parent;
2320         ref_root = ref->root;
2321
2322         if (node->ref_mod != 1) {
2323                 btrfs_err(trans->fs_info,
2324         "btree block(%llu) has %d references rather than 1: action %d ref_root %llu parent %llu",
2325                           node->bytenr, node->ref_mod, node->action, ref_root,
2326                           parent);
2327                 return -EIO;
2328         }
2329         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2330                 BUG_ON(!extent_op || !extent_op->update_flags);
2331                 ret = alloc_reserved_tree_block(trans, node, extent_op);
2332         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2333                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
2334                                              ref->level, 0, 1, extent_op);
2335         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2336                 ret = __btrfs_free_extent(trans, node, parent, ref_root,
2337                                           ref->level, 0, 1, extent_op);
2338         } else {
2339                 BUG();
2340         }
2341         return ret;
2342 }
2343
2344 /* helper function to actually process a single delayed ref entry */
2345 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2346                                struct btrfs_delayed_ref_node *node,
2347                                struct btrfs_delayed_extent_op *extent_op,
2348                                int insert_reserved)
2349 {
2350         int ret = 0;
2351
2352         if (trans->aborted) {
2353                 if (insert_reserved)
2354                         btrfs_pin_extent(trans->fs_info, node->bytenr,
2355                                          node->num_bytes, 1);
2356                 return 0;
2357         }
2358
2359         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2360             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2361                 ret = run_delayed_tree_ref(trans, node, extent_op,
2362                                            insert_reserved);
2363         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2364                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2365                 ret = run_delayed_data_ref(trans, node, extent_op,
2366                                            insert_reserved);
2367         else
2368                 BUG();
2369         return ret;
2370 }
2371
2372 static inline struct btrfs_delayed_ref_node *
2373 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2374 {
2375         struct btrfs_delayed_ref_node *ref;
2376
2377         if (RB_EMPTY_ROOT(&head->ref_tree))
2378                 return NULL;
2379
2380         /*
2381          * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
2382          * This is to prevent a ref count from going down to zero, which deletes
2383          * the extent item from the extent tree, when there still are references
2384          * to add, which would fail because they would not find the extent item.
2385          */
2386         if (!list_empty(&head->ref_add_list))
2387                 return list_first_entry(&head->ref_add_list,
2388                                 struct btrfs_delayed_ref_node, add_list);
2389
2390         ref = rb_entry(rb_first(&head->ref_tree),
2391                        struct btrfs_delayed_ref_node, ref_node);
2392         ASSERT(list_empty(&ref->add_list));
2393         return ref;
2394 }
2395
2396 static void unselect_delayed_ref_head(struct btrfs_delayed_ref_root *delayed_refs,
2397                                       struct btrfs_delayed_ref_head *head)
2398 {
2399         spin_lock(&delayed_refs->lock);
2400         head->processing = 0;
2401         delayed_refs->num_heads_ready++;
2402         spin_unlock(&delayed_refs->lock);
2403         btrfs_delayed_ref_unlock(head);
2404 }
2405
2406 static int cleanup_extent_op(struct btrfs_trans_handle *trans,
2407                              struct btrfs_delayed_ref_head *head)
2408 {
2409         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
2410         int ret;
2411
2412         if (!extent_op)
2413                 return 0;
2414         head->extent_op = NULL;
2415         if (head->must_insert_reserved) {
2416                 btrfs_free_delayed_extent_op(extent_op);
2417                 return 0;
2418         }
2419         spin_unlock(&head->lock);
2420         ret = run_delayed_extent_op(trans, head, extent_op);
2421         btrfs_free_delayed_extent_op(extent_op);
2422         return ret ? ret : 1;
2423 }
2424
2425 static int cleanup_ref_head(struct btrfs_trans_handle *trans,
2426                             struct btrfs_delayed_ref_head *head)
2427 {
2428
2429         struct btrfs_fs_info *fs_info = trans->fs_info;
2430         struct btrfs_delayed_ref_root *delayed_refs;
2431         int ret;
2432
2433         delayed_refs = &trans->transaction->delayed_refs;
2434
2435         ret = cleanup_extent_op(trans, head);
2436         if (ret < 0) {
2437                 unselect_delayed_ref_head(delayed_refs, head);
2438                 btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
2439                 return ret;
2440         } else if (ret) {
2441                 return ret;
2442         }
2443
2444         /*
2445          * Need to drop our head ref lock and re-acquire the delayed ref lock
2446          * and then re-check to make sure nobody got added.
2447          */
2448         spin_unlock(&head->lock);
2449         spin_lock(&delayed_refs->lock);
2450         spin_lock(&head->lock);
2451         if (!RB_EMPTY_ROOT(&head->ref_tree) || head->extent_op) {
2452                 spin_unlock(&head->lock);
2453                 spin_unlock(&delayed_refs->lock);
2454                 return 1;
2455         }
2456         delayed_refs->num_heads--;
2457         rb_erase(&head->href_node, &delayed_refs->href_root);
2458         RB_CLEAR_NODE(&head->href_node);
2459         spin_unlock(&head->lock);
2460         spin_unlock(&delayed_refs->lock);
2461         atomic_dec(&delayed_refs->num_entries);
2462
2463         trace_run_delayed_ref_head(fs_info, head, 0);
2464
2465         if (head->total_ref_mod < 0) {
2466                 struct btrfs_space_info *space_info;
2467                 u64 flags;
2468
2469                 if (head->is_data)
2470                         flags = BTRFS_BLOCK_GROUP_DATA;
2471                 else if (head->is_system)
2472                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
2473                 else
2474                         flags = BTRFS_BLOCK_GROUP_METADATA;
2475                 space_info = __find_space_info(fs_info, flags);
2476                 ASSERT(space_info);
2477                 percpu_counter_add_batch(&space_info->total_bytes_pinned,
2478                                    -head->num_bytes,
2479                                    BTRFS_TOTAL_BYTES_PINNED_BATCH);
2480
2481                 if (head->is_data) {
2482                         spin_lock(&delayed_refs->lock);
2483                         delayed_refs->pending_csums -= head->num_bytes;
2484                         spin_unlock(&delayed_refs->lock);
2485                 }
2486         }
2487
2488         if (head->must_insert_reserved) {
2489                 btrfs_pin_extent(fs_info, head->bytenr,
2490                                  head->num_bytes, 1);
2491                 if (head->is_data) {
2492                         ret = btrfs_del_csums(trans, fs_info, head->bytenr,
2493                                               head->num_bytes);
2494                 }
2495         }
2496
2497         /* Also free its reserved qgroup space */
2498         btrfs_qgroup_free_delayed_ref(fs_info, head->qgroup_ref_root,
2499                                       head->qgroup_reserved);
2500         btrfs_delayed_ref_unlock(head);
2501         btrfs_put_delayed_ref_head(head);
2502         return 0;
2503 }
2504
2505 /*
2506  * Returns 0 on success or if called with an already aborted transaction.
2507  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2508  */
2509 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2510                                              unsigned long nr)
2511 {
2512         struct btrfs_fs_info *fs_info = trans->fs_info;
2513         struct btrfs_delayed_ref_root *delayed_refs;
2514         struct btrfs_delayed_ref_node *ref;
2515         struct btrfs_delayed_ref_head *locked_ref = NULL;
2516         struct btrfs_delayed_extent_op *extent_op;
2517         ktime_t start = ktime_get();
2518         int ret;
2519         unsigned long count = 0;
2520         unsigned long actual_count = 0;
2521         int must_insert_reserved = 0;
2522
2523         delayed_refs = &trans->transaction->delayed_refs;
2524         while (1) {
2525                 if (!locked_ref) {
2526                         if (count >= nr)
2527                                 break;
2528
2529                         spin_lock(&delayed_refs->lock);
2530                         locked_ref = btrfs_select_ref_head(trans);
2531                         if (!locked_ref) {
2532                                 spin_unlock(&delayed_refs->lock);
2533                                 break;
2534                         }
2535
2536                         /* grab the lock that says we are going to process
2537                          * all the refs for this head */
2538                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2539                         spin_unlock(&delayed_refs->lock);
2540                         /*
2541                          * we may have dropped the spin lock to get the head
2542                          * mutex lock, and that might have given someone else
2543                          * time to free the head.  If that's true, it has been
2544                          * removed from our list and we can move on.
2545                          */
2546                         if (ret == -EAGAIN) {
2547                                 locked_ref = NULL;
2548                                 count++;
2549                                 continue;
2550                         }
2551                 }
2552
2553                 /*
2554                  * We need to try and merge add/drops of the same ref since we
2555                  * can run into issues with relocate dropping the implicit ref
2556                  * and then it being added back again before the drop can
2557                  * finish.  If we merged anything we need to re-loop so we can
2558                  * get a good ref.
2559                  * Or we can get node references of the same type that weren't
2560                  * merged when created due to bumps in the tree mod seq, and
2561                  * we need to merge them to prevent adding an inline extent
2562                  * backref before dropping it (triggering a BUG_ON at
2563                  * insert_inline_extent_backref()).
2564                  */
2565                 spin_lock(&locked_ref->lock);
2566                 btrfs_merge_delayed_refs(trans, delayed_refs, locked_ref);
2567
2568                 ref = select_delayed_ref(locked_ref);
2569
2570                 if (ref && ref->seq &&
2571                     btrfs_check_delayed_seq(fs_info, ref->seq)) {
2572                         spin_unlock(&locked_ref->lock);
2573                         unselect_delayed_ref_head(delayed_refs, locked_ref);
2574                         locked_ref = NULL;
2575                         cond_resched();
2576                         count++;
2577                         continue;
2578                 }
2579
2580                 /*
2581                  * We're done processing refs in this ref_head, clean everything
2582                  * up and move on to the next ref_head.
2583                  */
2584                 if (!ref) {
2585                         ret = cleanup_ref_head(trans, locked_ref);
2586                         if (ret > 0 ) {
2587                                 /* We dropped our lock, we need to loop. */
2588                                 ret = 0;
2589                                 continue;
2590                         } else if (ret) {
2591                                 return ret;
2592                         }
2593                         locked_ref = NULL;
2594                         count++;
2595                         continue;
2596                 }
2597
2598                 actual_count++;
2599                 ref->in_tree = 0;
2600                 rb_erase(&ref->ref_node, &locked_ref->ref_tree);
2601                 RB_CLEAR_NODE(&ref->ref_node);
2602                 if (!list_empty(&ref->add_list))
2603                         list_del(&ref->add_list);
2604                 /*
2605                  * When we play the delayed ref, also correct the ref_mod on
2606                  * head
2607                  */
2608                 switch (ref->action) {
2609                 case BTRFS_ADD_DELAYED_REF:
2610                 case BTRFS_ADD_DELAYED_EXTENT:
2611                         locked_ref->ref_mod -= ref->ref_mod;
2612                         break;
2613                 case BTRFS_DROP_DELAYED_REF:
2614                         locked_ref->ref_mod += ref->ref_mod;
2615                         break;
2616                 default:
2617                         WARN_ON(1);
2618                 }
2619                 atomic_dec(&delayed_refs->num_entries);
2620
2621                 /*
2622                  * Record the must-insert_reserved flag before we drop the spin
2623                  * lock.
2624                  */
2625                 must_insert_reserved = locked_ref->must_insert_reserved;
2626                 locked_ref->must_insert_reserved = 0;
2627
2628                 extent_op = locked_ref->extent_op;
2629                 locked_ref->extent_op = NULL;
2630                 spin_unlock(&locked_ref->lock);
2631
2632                 ret = run_one_delayed_ref(trans, ref, extent_op,
2633                                           must_insert_reserved);
2634
2635                 btrfs_free_delayed_extent_op(extent_op);
2636                 if (ret) {
2637                         unselect_delayed_ref_head(delayed_refs, locked_ref);
2638                         btrfs_put_delayed_ref(ref);
2639                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d",
2640                                     ret);
2641                         return ret;
2642                 }
2643
2644                 btrfs_put_delayed_ref(ref);
2645                 count++;
2646                 cond_resched();
2647         }
2648
2649         /*
2650          * We don't want to include ref heads since we can have empty ref heads
2651          * and those will drastically skew our runtime down since we just do
2652          * accounting, no actual extent tree updates.
2653          */
2654         if (actual_count > 0) {
2655                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2656                 u64 avg;
2657
2658                 /*
2659                  * We weigh the current average higher than our current runtime
2660                  * to avoid large swings in the average.
2661                  */
2662                 spin_lock(&delayed_refs->lock);
2663                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2664                 fs_info->avg_delayed_ref_runtime = avg >> 2;    /* div by 4 */
2665                 spin_unlock(&delayed_refs->lock);
2666         }
2667         return 0;
2668 }
2669
2670 #ifdef SCRAMBLE_DELAYED_REFS
2671 /*
2672  * Normally delayed refs get processed in ascending bytenr order. This
2673  * correlates in most cases to the order added. To expose dependencies on this
2674  * order, we start to process the tree in the middle instead of the beginning
2675  */
2676 static u64 find_middle(struct rb_root *root)
2677 {
2678         struct rb_node *n = root->rb_node;
2679         struct btrfs_delayed_ref_node *entry;
2680         int alt = 1;
2681         u64 middle;
2682         u64 first = 0, last = 0;
2683
2684         n = rb_first(root);
2685         if (n) {
2686                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2687                 first = entry->bytenr;
2688         }
2689         n = rb_last(root);
2690         if (n) {
2691                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2692                 last = entry->bytenr;
2693         }
2694         n = root->rb_node;
2695
2696         while (n) {
2697                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2698                 WARN_ON(!entry->in_tree);
2699
2700                 middle = entry->bytenr;
2701
2702                 if (alt)
2703                         n = n->rb_left;
2704                 else
2705                         n = n->rb_right;
2706
2707                 alt = 1 - alt;
2708         }
2709         return middle;
2710 }
2711 #endif
2712
2713 static inline u64 heads_to_leaves(struct btrfs_fs_info *fs_info, u64 heads)
2714 {
2715         u64 num_bytes;
2716
2717         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2718                              sizeof(struct btrfs_extent_inline_ref));
2719         if (!btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2720                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2721
2722         /*
2723          * We don't ever fill up leaves all the way so multiply by 2 just to be
2724          * closer to what we're really going to want to use.
2725          */
2726         return div_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(fs_info));
2727 }
2728
2729 /*
2730  * Takes the number of bytes to be csumm'ed and figures out how many leaves it
2731  * would require to store the csums for that many bytes.
2732  */
2733 u64 btrfs_csum_bytes_to_leaves(struct btrfs_fs_info *fs_info, u64 csum_bytes)
2734 {
2735         u64 csum_size;
2736         u64 num_csums_per_leaf;
2737         u64 num_csums;
2738
2739         csum_size = BTRFS_MAX_ITEM_SIZE(fs_info);
2740         num_csums_per_leaf = div64_u64(csum_size,
2741                         (u64)btrfs_super_csum_size(fs_info->super_copy));
2742         num_csums = div64_u64(csum_bytes, fs_info->sectorsize);
2743         num_csums += num_csums_per_leaf - 1;
2744         num_csums = div64_u64(num_csums, num_csums_per_leaf);
2745         return num_csums;
2746 }
2747
2748 int btrfs_check_space_for_delayed_refs(struct btrfs_trans_handle *trans,
2749                                        struct btrfs_fs_info *fs_info)
2750 {
2751         struct btrfs_block_rsv *global_rsv;
2752         u64 num_heads = trans->transaction->delayed_refs.num_heads_ready;
2753         u64 csum_bytes = trans->transaction->delayed_refs.pending_csums;
2754         unsigned int num_dirty_bgs = trans->transaction->num_dirty_bgs;
2755         u64 num_bytes, num_dirty_bgs_bytes;
2756         int ret = 0;
2757
2758         num_bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
2759         num_heads = heads_to_leaves(fs_info, num_heads);
2760         if (num_heads > 1)
2761                 num_bytes += (num_heads - 1) * fs_info->nodesize;
2762         num_bytes <<= 1;
2763         num_bytes += btrfs_csum_bytes_to_leaves(fs_info, csum_bytes) *
2764                                                         fs_info->nodesize;
2765         num_dirty_bgs_bytes = btrfs_calc_trans_metadata_size(fs_info,
2766                                                              num_dirty_bgs);
2767         global_rsv = &fs_info->global_block_rsv;
2768
2769         /*
2770          * If we can't allocate any more chunks lets make sure we have _lots_ of
2771          * wiggle room since running delayed refs can create more delayed refs.
2772          */
2773         if (global_rsv->space_info->full) {
2774                 num_dirty_bgs_bytes <<= 1;
2775                 num_bytes <<= 1;
2776         }
2777
2778         spin_lock(&global_rsv->lock);
2779         if (global_rsv->reserved <= num_bytes + num_dirty_bgs_bytes)
2780                 ret = 1;
2781         spin_unlock(&global_rsv->lock);
2782         return ret;
2783 }
2784
2785 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans,
2786                                        struct btrfs_fs_info *fs_info)
2787 {
2788         u64 num_entries =
2789                 atomic_read(&trans->transaction->delayed_refs.num_entries);
2790         u64 avg_runtime;
2791         u64 val;
2792
2793         smp_mb();
2794         avg_runtime = fs_info->avg_delayed_ref_runtime;
2795         val = num_entries * avg_runtime;
2796         if (val >= NSEC_PER_SEC)
2797                 return 1;
2798         if (val >= NSEC_PER_SEC / 2)
2799                 return 2;
2800
2801         return btrfs_check_space_for_delayed_refs(trans, fs_info);
2802 }
2803
2804 struct async_delayed_refs {
2805         struct btrfs_root *root;
2806         u64 transid;
2807         int count;
2808         int error;
2809         int sync;
2810         struct completion wait;
2811         struct btrfs_work work;
2812 };
2813
2814 static inline struct async_delayed_refs *
2815 to_async_delayed_refs(struct btrfs_work *work)
2816 {
2817         return container_of(work, struct async_delayed_refs, work);
2818 }
2819
2820 static void delayed_ref_async_start(struct btrfs_work *work)
2821 {
2822         struct async_delayed_refs *async = to_async_delayed_refs(work);
2823         struct btrfs_trans_handle *trans;
2824         struct btrfs_fs_info *fs_info = async->root->fs_info;
2825         int ret;
2826
2827         /* if the commit is already started, we don't need to wait here */
2828         if (btrfs_transaction_blocked(fs_info))
2829                 goto done;
2830
2831         trans = btrfs_join_transaction(async->root);
2832         if (IS_ERR(trans)) {
2833                 async->error = PTR_ERR(trans);
2834                 goto done;
2835         }
2836
2837         /*
2838          * trans->sync means that when we call end_transaction, we won't
2839          * wait on delayed refs
2840          */
2841         trans->sync = true;
2842
2843         /* Don't bother flushing if we got into a different transaction */
2844         if (trans->transid > async->transid)
2845                 goto end;
2846
2847         ret = btrfs_run_delayed_refs(trans, async->count);
2848         if (ret)
2849                 async->error = ret;
2850 end:
2851         ret = btrfs_end_transaction(trans);
2852         if (ret && !async->error)
2853                 async->error = ret;
2854 done:
2855         if (async->sync)
2856                 complete(&async->wait);
2857         else
2858                 kfree(async);
2859 }
2860
2861 int btrfs_async_run_delayed_refs(struct btrfs_fs_info *fs_info,
2862                                  unsigned long count, u64 transid, int wait)
2863 {
2864         struct async_delayed_refs *async;
2865         int ret;
2866
2867         async = kmalloc(sizeof(*async), GFP_NOFS);
2868         if (!async)
2869                 return -ENOMEM;
2870
2871         async->root = fs_info->tree_root;
2872         async->count = count;
2873         async->error = 0;
2874         async->transid = transid;
2875         if (wait)
2876                 async->sync = 1;
2877         else
2878                 async->sync = 0;
2879         init_completion(&async->wait);
2880
2881         btrfs_init_work(&async->work, btrfs_extent_refs_helper,
2882                         delayed_ref_async_start, NULL, NULL);
2883
2884         btrfs_queue_work(fs_info->extent_workers, &async->work);
2885
2886         if (wait) {
2887                 wait_for_completion(&async->wait);
2888                 ret = async->error;
2889                 kfree(async);
2890                 return ret;
2891         }
2892         return 0;
2893 }
2894
2895 /*
2896  * this starts processing the delayed reference count updates and
2897  * extent insertions we have queued up so far.  count can be
2898  * 0, which means to process everything in the tree at the start
2899  * of the run (but not newly added entries), or it can be some target
2900  * number you'd like to process.
2901  *
2902  * Returns 0 on success or if called with an aborted transaction
2903  * Returns <0 on error and aborts the transaction
2904  */
2905 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2906                            unsigned long count)
2907 {
2908         struct btrfs_fs_info *fs_info = trans->fs_info;
2909         struct rb_node *node;
2910         struct btrfs_delayed_ref_root *delayed_refs;
2911         struct btrfs_delayed_ref_head *head;
2912         int ret;
2913         int run_all = count == (unsigned long)-1;
2914         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
2915
2916         /* We'll clean this up in btrfs_cleanup_transaction */
2917         if (trans->aborted)
2918                 return 0;
2919
2920         if (test_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags))
2921                 return 0;
2922
2923         delayed_refs = &trans->transaction->delayed_refs;
2924         if (count == 0)
2925                 count = atomic_read(&delayed_refs->num_entries) * 2;
2926
2927 again:
2928 #ifdef SCRAMBLE_DELAYED_REFS
2929         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2930 #endif
2931         trans->can_flush_pending_bgs = false;
2932         ret = __btrfs_run_delayed_refs(trans, count);
2933         if (ret < 0) {
2934                 btrfs_abort_transaction(trans, ret);
2935                 return ret;
2936         }
2937
2938         if (run_all) {
2939                 if (!list_empty(&trans->new_bgs))
2940                         btrfs_create_pending_block_groups(trans);
2941
2942                 spin_lock(&delayed_refs->lock);
2943                 node = rb_first(&delayed_refs->href_root);
2944                 if (!node) {
2945                         spin_unlock(&delayed_refs->lock);
2946                         goto out;
2947                 }
2948                 head = rb_entry(node, struct btrfs_delayed_ref_head,
2949                                 href_node);
2950                 refcount_inc(&head->refs);
2951                 spin_unlock(&delayed_refs->lock);
2952
2953                 /* Mutex was contended, block until it's released and retry. */
2954                 mutex_lock(&head->mutex);
2955                 mutex_unlock(&head->mutex);
2956
2957                 btrfs_put_delayed_ref_head(head);
2958                 cond_resched();
2959                 goto again;
2960         }
2961 out:
2962         trans->can_flush_pending_bgs = can_flush_pending_bgs;
2963         return 0;
2964 }
2965
2966 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2967                                 struct btrfs_fs_info *fs_info,
2968                                 u64 bytenr, u64 num_bytes, u64 flags,
2969                                 int level, int is_data)
2970 {
2971         struct btrfs_delayed_extent_op *extent_op;
2972         int ret;
2973
2974         extent_op = btrfs_alloc_delayed_extent_op();
2975         if (!extent_op)
2976                 return -ENOMEM;
2977
2978         extent_op->flags_to_set = flags;
2979         extent_op->update_flags = true;
2980         extent_op->update_key = false;
2981         extent_op->is_data = is_data ? true : false;
2982         extent_op->level = level;
2983
2984         ret = btrfs_add_delayed_extent_op(fs_info, trans, bytenr,
2985                                           num_bytes, extent_op);
2986         if (ret)
2987                 btrfs_free_delayed_extent_op(extent_op);
2988         return ret;
2989 }
2990
2991 static noinline int check_delayed_ref(struct btrfs_root *root,
2992                                       struct btrfs_path *path,
2993                                       u64 objectid, u64 offset, u64 bytenr)
2994 {
2995         struct btrfs_delayed_ref_head *head;
2996         struct btrfs_delayed_ref_node *ref;
2997         struct btrfs_delayed_data_ref *data_ref;
2998         struct btrfs_delayed_ref_root *delayed_refs;
2999         struct btrfs_transaction *cur_trans;
3000         struct rb_node *node;
3001         int ret = 0;
3002
3003         spin_lock(&root->fs_info->trans_lock);
3004         cur_trans = root->fs_info->running_transaction;
3005         if (cur_trans)
3006                 refcount_inc(&cur_trans->use_count);
3007         spin_unlock(&root->fs_info->trans_lock);
3008         if (!cur_trans)
3009                 return 0;
3010
3011         delayed_refs = &cur_trans->delayed_refs;
3012         spin_lock(&delayed_refs->lock);
3013         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
3014         if (!head) {
3015                 spin_unlock(&delayed_refs->lock);
3016                 btrfs_put_transaction(cur_trans);
3017                 return 0;
3018         }
3019
3020         if (!mutex_trylock(&head->mutex)) {
3021                 refcount_inc(&head->refs);
3022                 spin_unlock(&delayed_refs->lock);
3023
3024                 btrfs_release_path(path);
3025
3026                 /*
3027                  * Mutex was contended, block until it's released and let
3028                  * caller try again
3029                  */
3030                 mutex_lock(&head->mutex);
3031                 mutex_unlock(&head->mutex);
3032                 btrfs_put_delayed_ref_head(head);
3033                 btrfs_put_transaction(cur_trans);
3034                 return -EAGAIN;
3035         }
3036         spin_unlock(&delayed_refs->lock);
3037
3038         spin_lock(&head->lock);
3039         /*
3040          * XXX: We should replace this with a proper search function in the
3041          * future.
3042          */
3043         for (node = rb_first(&head->ref_tree); node; node = rb_next(node)) {
3044                 ref = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
3045                 /* If it's a shared ref we know a cross reference exists */
3046                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
3047                         ret = 1;
3048                         break;
3049                 }
3050
3051                 data_ref = btrfs_delayed_node_to_data_ref(ref);
3052
3053                 /*
3054                  * If our ref doesn't match the one we're currently looking at
3055                  * then we have a cross reference.
3056                  */
3057                 if (data_ref->root != root->root_key.objectid ||
3058                     data_ref->objectid != objectid ||
3059                     data_ref->offset != offset) {
3060                         ret = 1;
3061                         break;
3062                 }
3063         }
3064         spin_unlock(&head->lock);
3065         mutex_unlock(&head->mutex);
3066         btrfs_put_transaction(cur_trans);
3067         return ret;
3068 }
3069
3070 static noinline int check_committed_ref(struct btrfs_root *root,
3071                                         struct btrfs_path *path,
3072                                         u64 objectid, u64 offset, u64 bytenr)
3073 {
3074         struct btrfs_fs_info *fs_info = root->fs_info;
3075         struct btrfs_root *extent_root = fs_info->extent_root;
3076         struct extent_buffer *leaf;
3077         struct btrfs_extent_data_ref *ref;
3078         struct btrfs_extent_inline_ref *iref;
3079         struct btrfs_extent_item *ei;
3080         struct btrfs_key key;
3081         u32 item_size;
3082         int type;
3083         int ret;
3084
3085         key.objectid = bytenr;
3086         key.offset = (u64)-1;
3087         key.type = BTRFS_EXTENT_ITEM_KEY;
3088
3089         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
3090         if (ret < 0)
3091                 goto out;
3092         BUG_ON(ret == 0); /* Corruption */
3093
3094         ret = -ENOENT;
3095         if (path->slots[0] == 0)
3096                 goto out;
3097
3098         path->slots[0]--;
3099         leaf = path->nodes[0];
3100         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3101
3102         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
3103                 goto out;
3104
3105         ret = 1;
3106         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
3107         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
3108
3109         if (item_size != sizeof(*ei) +
3110             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
3111                 goto out;
3112
3113         if (btrfs_extent_generation(leaf, ei) <=
3114             btrfs_root_last_snapshot(&root->root_item))
3115                 goto out;
3116
3117         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
3118
3119         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
3120         if (type != BTRFS_EXTENT_DATA_REF_KEY)
3121                 goto out;
3122
3123         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
3124         if (btrfs_extent_refs(leaf, ei) !=
3125             btrfs_extent_data_ref_count(leaf, ref) ||
3126             btrfs_extent_data_ref_root(leaf, ref) !=
3127             root->root_key.objectid ||
3128             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
3129             btrfs_extent_data_ref_offset(leaf, ref) != offset)
3130                 goto out;
3131
3132         ret = 0;
3133 out:
3134         return ret;
3135 }
3136
3137 int btrfs_cross_ref_exist(struct btrfs_root *root, u64 objectid, u64 offset,
3138                           u64 bytenr)
3139 {
3140         struct btrfs_path *path;
3141         int ret;
3142         int ret2;
3143
3144         path = btrfs_alloc_path();
3145         if (!path)
3146                 return -ENOMEM;
3147
3148         do {
3149                 ret = check_committed_ref(root, path, objectid,
3150                                           offset, bytenr);
3151                 if (ret && ret != -ENOENT)
3152                         goto out;
3153
3154                 ret2 = check_delayed_ref(root, path, objectid,
3155                                          offset, bytenr);
3156         } while (ret2 == -EAGAIN);
3157
3158         if (ret2 && ret2 != -ENOENT) {
3159                 ret = ret2;
3160                 goto out;
3161         }
3162
3163         if (ret != -ENOENT || ret2 != -ENOENT)
3164                 ret = 0;
3165 out:
3166         btrfs_free_path(path);
3167         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
3168                 WARN_ON(ret > 0);
3169         return ret;
3170 }
3171
3172 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
3173                            struct btrfs_root *root,
3174                            struct extent_buffer *buf,
3175                            int full_backref, int inc)
3176 {
3177         struct btrfs_fs_info *fs_info = root->fs_info;
3178         u64 bytenr;
3179         u64 num_bytes;
3180         u64 parent;
3181         u64 ref_root;
3182         u32 nritems;
3183         struct btrfs_key key;
3184         struct btrfs_file_extent_item *fi;
3185         int i;
3186         int level;
3187         int ret = 0;
3188         int (*process_func)(struct btrfs_trans_handle *,
3189                             struct btrfs_root *,
3190                             u64, u64, u64, u64, u64, u64);
3191
3192
3193         if (btrfs_is_testing(fs_info))
3194                 return 0;
3195
3196         ref_root = btrfs_header_owner(buf);
3197         nritems = btrfs_header_nritems(buf);
3198         level = btrfs_header_level(buf);
3199
3200         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state) && level == 0)
3201                 return 0;
3202
3203         if (inc)
3204                 process_func = btrfs_inc_extent_ref;
3205         else
3206                 process_func = btrfs_free_extent;
3207
3208         if (full_backref)
3209                 parent = buf->start;
3210         else
3211                 parent = 0;
3212
3213         for (i = 0; i < nritems; i++) {
3214                 if (level == 0) {
3215                         btrfs_item_key_to_cpu(buf, &key, i);
3216                         if (key.type != BTRFS_EXTENT_DATA_KEY)
3217                                 continue;
3218                         fi = btrfs_item_ptr(buf, i,
3219                                             struct btrfs_file_extent_item);
3220                         if (btrfs_file_extent_type(buf, fi) ==
3221                             BTRFS_FILE_EXTENT_INLINE)
3222                                 continue;
3223                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3224                         if (bytenr == 0)
3225                                 continue;
3226
3227                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3228                         key.offset -= btrfs_file_extent_offset(buf, fi);
3229                         ret = process_func(trans, root, bytenr, num_bytes,
3230                                            parent, ref_root, key.objectid,
3231                                            key.offset);
3232                         if (ret)
3233                                 goto fail;
3234                 } else {
3235                         bytenr = btrfs_node_blockptr(buf, i);
3236                         num_bytes = fs_info->nodesize;
3237                         ret = process_func(trans, root, bytenr, num_bytes,
3238                                            parent, ref_root, level - 1, 0);
3239                         if (ret)
3240                                 goto fail;
3241                 }
3242         }
3243         return 0;
3244 fail:
3245         return ret;
3246 }
3247
3248 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3249                   struct extent_buffer *buf, int full_backref)
3250 {
3251         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
3252 }
3253
3254 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3255                   struct extent_buffer *buf, int full_backref)
3256 {
3257         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
3258 }
3259
3260 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3261                                  struct btrfs_fs_info *fs_info,
3262                                  struct btrfs_path *path,
3263                                  struct btrfs_block_group_cache *cache)
3264 {
3265         int ret;
3266         struct btrfs_root *extent_root = fs_info->extent_root;
3267         unsigned long bi;
3268         struct extent_buffer *leaf;
3269
3270         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3271         if (ret) {
3272                 if (ret > 0)
3273                         ret = -ENOENT;
3274                 goto fail;
3275         }
3276
3277         leaf = path->nodes[0];
3278         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3279         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3280         btrfs_mark_buffer_dirty(leaf);
3281 fail:
3282         btrfs_release_path(path);
3283         return ret;
3284
3285 }
3286
3287 static struct btrfs_block_group_cache *
3288 next_block_group(struct btrfs_fs_info *fs_info,
3289                  struct btrfs_block_group_cache *cache)
3290 {
3291         struct rb_node *node;
3292
3293         spin_lock(&fs_info->block_group_cache_lock);
3294
3295         /* If our block group was removed, we need a full search. */
3296         if (RB_EMPTY_NODE(&cache->cache_node)) {
3297                 const u64 next_bytenr = cache->key.objectid + cache->key.offset;
3298
3299                 spin_unlock(&fs_info->block_group_cache_lock);
3300                 btrfs_put_block_group(cache);
3301                 cache = btrfs_lookup_first_block_group(fs_info, next_bytenr); return cache;
3302         }
3303         node = rb_next(&cache->cache_node);
3304         btrfs_put_block_group(cache);
3305         if (node) {
3306                 cache = rb_entry(node, struct btrfs_block_group_cache,
3307                                  cache_node);
3308                 btrfs_get_block_group(cache);
3309         } else
3310                 cache = NULL;
3311         spin_unlock(&fs_info->block_group_cache_lock);
3312         return cache;
3313 }
3314
3315 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3316                             struct btrfs_trans_handle *trans,
3317                             struct btrfs_path *path)
3318 {
3319         struct btrfs_fs_info *fs_info = block_group->fs_info;
3320         struct btrfs_root *root = fs_info->tree_root;
3321         struct inode *inode = NULL;
3322         struct extent_changeset *data_reserved = NULL;
3323         u64 alloc_hint = 0;
3324         int dcs = BTRFS_DC_ERROR;
3325         u64 num_pages = 0;
3326         int retries = 0;
3327         int ret = 0;
3328
3329         /*
3330          * If this block group is smaller than 100 megs don't bother caching the
3331          * block group.
3332          */
3333         if (block_group->key.offset < (100 * SZ_1M)) {
3334                 spin_lock(&block_group->lock);
3335                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3336                 spin_unlock(&block_group->lock);
3337                 return 0;
3338         }
3339
3340         if (trans->aborted)
3341                 return 0;
3342 again:
3343         inode = lookup_free_space_inode(fs_info, block_group, path);
3344         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3345                 ret = PTR_ERR(inode);
3346                 btrfs_release_path(path);
3347                 goto out;
3348         }
3349
3350         if (IS_ERR(inode)) {
3351                 BUG_ON(retries);
3352                 retries++;
3353
3354                 if (block_group->ro)
3355                         goto out_free;
3356
3357                 ret = create_free_space_inode(fs_info, trans, block_group,
3358                                               path);
3359                 if (ret)
3360                         goto out_free;
3361                 goto again;
3362         }
3363
3364         /*
3365          * We want to set the generation to 0, that way if anything goes wrong
3366          * from here on out we know not to trust this cache when we load up next
3367          * time.
3368          */
3369         BTRFS_I(inode)->generation = 0;
3370         ret = btrfs_update_inode(trans, root, inode);
3371         if (ret) {
3372                 /*
3373                  * So theoretically we could recover from this, simply set the
3374                  * super cache generation to 0 so we know to invalidate the
3375                  * cache, but then we'd have to keep track of the block groups
3376                  * that fail this way so we know we _have_ to reset this cache
3377                  * before the next commit or risk reading stale cache.  So to
3378                  * limit our exposure to horrible edge cases lets just abort the
3379                  * transaction, this only happens in really bad situations
3380                  * anyway.
3381                  */
3382                 btrfs_abort_transaction(trans, ret);
3383                 goto out_put;
3384         }
3385         WARN_ON(ret);
3386
3387         /* We've already setup this transaction, go ahead and exit */
3388         if (block_group->cache_generation == trans->transid &&
3389             i_size_read(inode)) {
3390                 dcs = BTRFS_DC_SETUP;
3391                 goto out_put;
3392         }
3393
3394         if (i_size_read(inode) > 0) {
3395                 ret = btrfs_check_trunc_cache_free_space(fs_info,
3396                                         &fs_info->global_block_rsv);
3397                 if (ret)
3398                         goto out_put;
3399
3400                 ret = btrfs_truncate_free_space_cache(trans, NULL, inode);
3401                 if (ret)
3402                         goto out_put;
3403         }
3404
3405         spin_lock(&block_group->lock);
3406         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3407             !btrfs_test_opt(fs_info, SPACE_CACHE)) {
3408                 /*
3409                  * don't bother trying to write stuff out _if_
3410                  * a) we're not cached,
3411                  * b) we're with nospace_cache mount option,
3412                  * c) we're with v2 space_cache (FREE_SPACE_TREE).
3413                  */
3414                 dcs = BTRFS_DC_WRITTEN;
3415                 spin_unlock(&block_group->lock);
3416                 goto out_put;
3417         }
3418         spin_unlock(&block_group->lock);
3419
3420         /*
3421          * We hit an ENOSPC when setting up the cache in this transaction, just
3422          * skip doing the setup, we've already cleared the cache so we're safe.
3423          */
3424         if (test_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags)) {
3425                 ret = -ENOSPC;
3426                 goto out_put;
3427         }
3428
3429         /*
3430          * Try to preallocate enough space based on how big the block group is.
3431          * Keep in mind this has to include any pinned space which could end up
3432          * taking up quite a bit since it's not folded into the other space
3433          * cache.
3434          */
3435         num_pages = div_u64(block_group->key.offset, SZ_256M);
3436         if (!num_pages)
3437                 num_pages = 1;
3438
3439         num_pages *= 16;
3440         num_pages *= PAGE_SIZE;
3441
3442         ret = btrfs_check_data_free_space(inode, &data_reserved, 0, num_pages);
3443         if (ret)
3444                 goto out_put;
3445
3446         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3447                                               num_pages, num_pages,
3448                                               &alloc_hint);
3449         /*
3450          * Our cache requires contiguous chunks so that we don't modify a bunch
3451          * of metadata or split extents when writing the cache out, which means
3452          * we can enospc if we are heavily fragmented in addition to just normal
3453          * out of space conditions.  So if we hit this just skip setting up any
3454          * other block groups for this transaction, maybe we'll unpin enough
3455          * space the next time around.
3456          */
3457         if (!ret)
3458                 dcs = BTRFS_DC_SETUP;
3459         else if (ret == -ENOSPC)
3460                 set_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags);
3461
3462 out_put:
3463         iput(inode);
3464 out_free:
3465         btrfs_release_path(path);
3466 out:
3467         spin_lock(&block_group->lock);
3468         if (!ret && dcs == BTRFS_DC_SETUP)
3469                 block_group->cache_generation = trans->transid;
3470         block_group->disk_cache_state = dcs;
3471         spin_unlock(&block_group->lock);
3472
3473         extent_changeset_free(data_reserved);
3474         return ret;
3475 }
3476
3477 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans,
3478                             struct btrfs_fs_info *fs_info)
3479 {
3480         struct btrfs_block_group_cache *cache, *tmp;
3481         struct btrfs_transaction *cur_trans = trans->transaction;
3482         struct btrfs_path *path;
3483
3484         if (list_empty(&cur_trans->dirty_bgs) ||
3485             !btrfs_test_opt(fs_info, SPACE_CACHE))
3486                 return 0;
3487
3488         path = btrfs_alloc_path();
3489         if (!path)
3490                 return -ENOMEM;
3491
3492         /* Could add new block groups, use _safe just in case */
3493         list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3494                                  dirty_list) {
3495                 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3496                         cache_save_setup(cache, trans, path);
3497         }
3498
3499         btrfs_free_path(path);
3500         return 0;
3501 }
3502
3503 /*
3504  * transaction commit does final block group cache writeback during a
3505  * critical section where nothing is allowed to change the FS.  This is
3506  * required in order for the cache to actually match the block group,
3507  * but can introduce a lot of latency into the commit.
3508  *
3509  * So, btrfs_start_dirty_block_groups is here to kick off block group
3510  * cache IO.  There's a chance we'll have to redo some of it if the
3511  * block group changes again during the commit, but it greatly reduces
3512  * the commit latency by getting rid of the easy block groups while
3513  * we're still allowing others to join the commit.
3514  */
3515 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans)
3516 {
3517         struct btrfs_fs_info *fs_info = trans->fs_info;
3518         struct btrfs_block_group_cache *cache;
3519         struct btrfs_transaction *cur_trans = trans->transaction;
3520         int ret = 0;
3521         int should_put;
3522         struct btrfs_path *path = NULL;
3523         LIST_HEAD(dirty);
3524         struct list_head *io = &cur_trans->io_bgs;
3525         int num_started = 0;
3526         int loops = 0;
3527
3528         spin_lock(&cur_trans->dirty_bgs_lock);
3529         if (list_empty(&cur_trans->dirty_bgs)) {
3530                 spin_unlock(&cur_trans->dirty_bgs_lock);
3531                 return 0;
3532         }
3533         list_splice_init(&cur_trans->dirty_bgs, &dirty);
3534         spin_unlock(&cur_trans->dirty_bgs_lock);
3535
3536 again:
3537         /*
3538          * make sure all the block groups on our dirty list actually
3539          * exist
3540          */
3541         btrfs_create_pending_block_groups(trans);
3542
3543         if (!path) {
3544                 path = btrfs_alloc_path();
3545                 if (!path)
3546                         return -ENOMEM;
3547         }
3548
3549         /*
3550          * cache_write_mutex is here only to save us from balance or automatic
3551          * removal of empty block groups deleting this block group while we are
3552          * writing out the cache
3553          */
3554         mutex_lock(&trans->transaction->cache_write_mutex);
3555         while (!list_empty(&dirty)) {
3556                 cache = list_first_entry(&dirty,
3557                                          struct btrfs_block_group_cache,
3558                                          dirty_list);
3559                 /*
3560                  * this can happen if something re-dirties a block
3561                  * group that is already under IO.  Just wait for it to
3562                  * finish and then do it all again
3563                  */
3564                 if (!list_empty(&cache->io_list)) {
3565                         list_del_init(&cache->io_list);
3566                         btrfs_wait_cache_io(trans, cache, path);
3567                         btrfs_put_block_group(cache);
3568                 }
3569
3570
3571                 /*
3572                  * btrfs_wait_cache_io uses the cache->dirty_list to decide
3573                  * if it should update the cache_state.  Don't delete
3574                  * until after we wait.
3575                  *
3576                  * Since we're not running in the commit critical section
3577                  * we need the dirty_bgs_lock to protect from update_block_group
3578                  */
3579                 spin_lock(&cur_trans->dirty_bgs_lock);
3580                 list_del_init(&cache->dirty_list);
3581                 spin_unlock(&cur_trans->dirty_bgs_lock);
3582
3583                 should_put = 1;
3584
3585                 cache_save_setup(cache, trans, path);
3586
3587                 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
3588                         cache->io_ctl.inode = NULL;
3589                         ret = btrfs_write_out_cache(fs_info, trans,
3590                                                     cache, path);
3591                         if (ret == 0 && cache->io_ctl.inode) {
3592                                 num_started++;
3593                                 should_put = 0;
3594
3595                                 /*
3596                                  * The cache_write_mutex is protecting the
3597                                  * io_list, also refer to the definition of
3598                                  * btrfs_transaction::io_bgs for more details
3599                                  */
3600                                 list_add_tail(&cache->io_list, io);
3601                         } else {
3602                                 /*
3603                                  * if we failed to write the cache, the
3604                                  * generation will be bad and life goes on
3605                                  */
3606                                 ret = 0;
3607                         }
3608                 }
3609                 if (!ret) {
3610                         ret = write_one_cache_group(trans, fs_info,
3611                                                     path, cache);
3612                         /*
3613                          * Our block group might still be attached to the list
3614                          * of new block groups in the transaction handle of some
3615                          * other task (struct btrfs_trans_handle->new_bgs). This
3616                          * means its block group item isn't yet in the extent
3617                          * tree. If this happens ignore the error, as we will
3618                          * try again later in the critical section of the
3619                          * transaction commit.
3620                          */
3621                         if (ret == -ENOENT) {
3622                                 ret = 0;
3623                                 spin_lock(&cur_trans->dirty_bgs_lock);
3624                                 if (list_empty(&cache->dirty_list)) {
3625                                         list_add_tail(&cache->dirty_list,
3626                                                       &cur_trans->dirty_bgs);
3627                                         btrfs_get_block_group(cache);
3628                                 }
3629                                 spin_unlock(&cur_trans->dirty_bgs_lock);
3630                         } else if (ret) {
3631                                 btrfs_abort_transaction(trans, ret);
3632                         }
3633                 }
3634
3635                 /* if its not on the io list, we need to put the block group */
3636                 if (should_put)
3637                         btrfs_put_block_group(cache);
3638
3639                 if (ret)
3640                         break;
3641
3642                 /*
3643                  * Avoid blocking other tasks for too long. It might even save
3644                  * us from writing caches for block groups that are going to be
3645                  * removed.
3646                  */
3647                 mutex_unlock(&trans->transaction->cache_write_mutex);
3648                 mutex_lock(&trans->transaction->cache_write_mutex);
3649         }
3650         mutex_unlock(&trans->transaction->cache_write_mutex);
3651
3652         /*
3653          * go through delayed refs for all the stuff we've just kicked off
3654          * and then loop back (just once)
3655          */
3656         ret = btrfs_run_delayed_refs(trans, 0);
3657         if (!ret && loops == 0) {
3658                 loops++;
3659                 spin_lock(&cur_trans->dirty_bgs_lock);
3660                 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3661                 /*
3662                  * dirty_bgs_lock protects us from concurrent block group
3663                  * deletes too (not just cache_write_mutex).
3664                  */
3665                 if (!list_empty(&dirty)) {
3666                         spin_unlock(&cur_trans->dirty_bgs_lock);
3667                         goto again;
3668                 }
3669                 spin_unlock(&cur_trans->dirty_bgs_lock);
3670         } else if (ret < 0) {
3671                 btrfs_cleanup_dirty_bgs(cur_trans, fs_info);
3672         }
3673
3674         btrfs_free_path(path);
3675         return ret;
3676 }
3677
3678 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3679                                    struct btrfs_fs_info *fs_info)
3680 {
3681         struct btrfs_block_group_cache *cache;
3682         struct btrfs_transaction *cur_trans = trans->transaction;
3683         int ret = 0;
3684         int should_put;
3685         struct btrfs_path *path;
3686         struct list_head *io = &cur_trans->io_bgs;
3687         int num_started = 0;
3688
3689         path = btrfs_alloc_path();
3690         if (!path)
3691                 return -ENOMEM;
3692
3693         /*
3694          * Even though we are in the critical section of the transaction commit,
3695          * we can still have concurrent tasks adding elements to this
3696          * transaction's list of dirty block groups. These tasks correspond to
3697          * endio free space workers started when writeback finishes for a
3698          * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
3699          * allocate new block groups as a result of COWing nodes of the root
3700          * tree when updating the free space inode. The writeback for the space
3701          * caches is triggered by an earlier call to
3702          * btrfs_start_dirty_block_groups() and iterations of the following
3703          * loop.
3704          * Also we want to do the cache_save_setup first and then run the
3705          * delayed refs to make sure we have the best chance at doing this all
3706          * in one shot.
3707          */
3708         spin_lock(&cur_trans->dirty_bgs_lock);
3709         while (!list_empty(&cur_trans->dirty_bgs)) {
3710                 cache = list_first_entry(&cur_trans->dirty_bgs,
3711                                          struct btrfs_block_group_cache,
3712                                          dirty_list);
3713
3714                 /*
3715                  * this can happen if cache_save_setup re-dirties a block
3716                  * group that is already under IO.  Just wait for it to
3717                  * finish and then do it all again
3718                  */
3719                 if (!list_empty(&cache->io_list)) {
3720                         spin_unlock(&cur_trans->dirty_bgs_lock);
3721                         list_del_init(&cache->io_list);
3722                         btrfs_wait_cache_io(trans, cache, path);
3723                         btrfs_put_block_group(cache);
3724                         spin_lock(&cur_trans->dirty_bgs_lock);
3725                 }
3726
3727                 /*
3728                  * don't remove from the dirty list until after we've waited
3729                  * on any pending IO
3730                  */
3731                 list_del_init(&cache->dirty_list);
3732                 spin_unlock(&cur_trans->dirty_bgs_lock);
3733                 should_put = 1;
3734
3735                 cache_save_setup(cache, trans, path);
3736
3737                 if (!ret)
3738                         ret = btrfs_run_delayed_refs(trans,
3739                                                      (unsigned long) -1);
3740
3741                 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3742                         cache->io_ctl.inode = NULL;
3743                         ret = btrfs_write_out_cache(fs_info, trans,
3744                                                     cache, path);
3745                         if (ret == 0 && cache->io_ctl.inode) {
3746                                 num_started++;
3747                                 should_put = 0;
3748                                 list_add_tail(&cache->io_list, io);
3749                         } else {
3750                                 /*
3751                                  * if we failed to write the cache, the
3752                                  * generation will be bad and life goes on
3753                                  */
3754                                 ret = 0;
3755                         }
3756                 }
3757                 if (!ret) {
3758                         ret = write_one_cache_group(trans, fs_info,
3759                                                     path, cache);
3760                         /*
3761                          * One of the free space endio workers might have
3762                          * created a new block group while updating a free space
3763                          * cache's inode (at inode.c:btrfs_finish_ordered_io())
3764                          * and hasn't released its transaction handle yet, in
3765                          * which case the new block group is still attached to
3766                          * its transaction handle and its creation has not
3767                          * finished yet (no block group item in the extent tree
3768                          * yet, etc). If this is the case, wait for all free
3769                          * space endio workers to finish and retry. This is a
3770                          * a very rare case so no need for a more efficient and
3771                          * complex approach.
3772                          */
3773                         if (ret == -ENOENT) {
3774                                 wait_event(cur_trans->writer_wait,
3775                                    atomic_read(&cur_trans->num_writers) == 1);
3776                                 ret = write_one_cache_group(trans, fs_info,
3777                                                             path, cache);
3778                         }
3779                         if (ret)
3780                                 btrfs_abort_transaction(trans, ret);
3781                 }
3782
3783                 /* if its not on the io list, we need to put the block group */
3784                 if (should_put)
3785                         btrfs_put_block_group(cache);
3786                 spin_lock(&cur_trans->dirty_bgs_lock);
3787         }
3788         spin_unlock(&cur_trans->dirty_bgs_lock);
3789
3790         /*
3791          * Refer to the definition of io_bgs member for details why it's safe
3792          * to use it without any locking
3793          */
3794         while (!list_empty(io)) {
3795                 cache = list_first_entry(io, struct btrfs_block_group_cache,
3796                                          io_list);
3797                 list_del_init(&cache->io_list);
3798                 btrfs_wait_cache_io(trans, cache, path);
3799                 btrfs_put_block_group(cache);
3800         }
3801
3802         btrfs_free_path(path);
3803         return ret;
3804 }
3805
3806 int btrfs_extent_readonly(struct btrfs_fs_info *fs_info, u64 bytenr)
3807 {
3808         struct btrfs_block_group_cache *block_group;
3809         int readonly = 0;
3810
3811         block_group = btrfs_lookup_block_group(fs_info, bytenr);
3812         if (!block_group || block_group->ro)
3813                 readonly = 1;
3814         if (block_group)
3815                 btrfs_put_block_group(block_group);
3816         return readonly;
3817 }
3818
3819 bool btrfs_inc_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3820 {
3821         struct btrfs_block_group_cache *bg;
3822         bool ret = true;
3823
3824         bg = btrfs_lookup_block_group(fs_info, bytenr);
3825         if (!bg)
3826                 return false;
3827
3828         spin_lock(&bg->lock);
3829         if (bg->ro)
3830                 ret = false;
3831         else
3832                 atomic_inc(&bg->nocow_writers);
3833         spin_unlock(&bg->lock);
3834
3835         /* no put on block group, done by btrfs_dec_nocow_writers */
3836         if (!ret)
3837                 btrfs_put_block_group(bg);
3838
3839         return ret;
3840
3841 }
3842
3843 void btrfs_dec_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3844 {
3845         struct btrfs_block_group_cache *bg;
3846
3847         bg = btrfs_lookup_block_group(fs_info, bytenr);
3848         ASSERT(bg);
3849         if (atomic_dec_and_test(&bg->nocow_writers))
3850                 wake_up_var(&bg->nocow_writers);
3851         /*
3852          * Once for our lookup and once for the lookup done by a previous call
3853          * to btrfs_inc_nocow_writers()
3854          */
3855         btrfs_put_block_group(bg);
3856         btrfs_put_block_group(bg);
3857 }
3858
3859 void btrfs_wait_nocow_writers(struct btrfs_block_group_cache *bg)
3860 {
3861         wait_var_event(&bg->nocow_writers, !atomic_read(&bg->nocow_writers));
3862 }
3863
3864 static const char *alloc_name(u64 flags)
3865 {
3866         switch (flags) {
3867         case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA:
3868                 return "mixed";
3869         case BTRFS_BLOCK_GROUP_METADATA:
3870                 return "metadata";
3871         case BTRFS_BLOCK_GROUP_DATA:
3872                 return "data";
3873         case BTRFS_BLOCK_GROUP_SYSTEM:
3874                 return "system";
3875         default:
3876                 WARN_ON(1);
3877                 return "invalid-combination";
3878         };
3879 }
3880
3881 static int create_space_info(struct btrfs_fs_info *info, u64 flags)
3882 {
3883
3884         struct btrfs_space_info *space_info;
3885         int i;
3886         int ret;
3887
3888         space_info = kzalloc(sizeof(*space_info), GFP_NOFS);
3889         if (!space_info)
3890                 return -ENOMEM;
3891
3892         ret = percpu_counter_init(&space_info->total_bytes_pinned, 0,
3893                                  GFP_KERNEL);
3894         if (ret) {
3895                 kfree(space_info);
3896                 return ret;
3897         }
3898
3899         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3900                 INIT_LIST_HEAD(&space_info->block_groups[i]);
3901         init_rwsem(&space_info->groups_sem);
3902         spin_lock_init(&space_info->lock);
3903         space_info->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3904         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3905         init_waitqueue_head(&space_info->wait);
3906         INIT_LIST_HEAD(&space_info->ro_bgs);
3907         INIT_LIST_HEAD(&space_info->tickets);
3908         INIT_LIST_HEAD(&space_info->priority_tickets);
3909
3910         ret = kobject_init_and_add(&space_info->kobj, &space_info_ktype,
3911                                     info->space_info_kobj, "%s",
3912                                     alloc_name(space_info->flags));
3913         if (ret) {
3914                 percpu_counter_destroy(&space_info->total_bytes_pinned);
3915                 kfree(space_info);
3916                 return ret;
3917         }
3918
3919         list_add_rcu(&space_info->list, &info->space_info);
3920         if (flags & BTRFS_BLOCK_GROUP_DATA)
3921                 info->data_sinfo = space_info;
3922
3923         return ret;
3924 }
3925
3926 static void update_space_info(struct btrfs_fs_info *info, u64 flags,
3927                              u64 total_bytes, u64 bytes_used,
3928                              u64 bytes_readonly,
3929                              struct btrfs_space_info **space_info)
3930 {
3931         struct btrfs_space_info *found;
3932         int factor;
3933
3934         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3935                      BTRFS_BLOCK_GROUP_RAID10))
3936                 factor = 2;
3937         else
3938                 factor = 1;
3939
3940         found = __find_space_info(info, flags);
3941         ASSERT(found);
3942         spin_lock(&found->lock);
3943         found->total_bytes += total_bytes;
3944         found->disk_total += total_bytes * factor;
3945         found->bytes_used += bytes_used;
3946         found->disk_used += bytes_used * factor;
3947         found->bytes_readonly += bytes_readonly;
3948         if (total_bytes > 0)
3949                 found->full = 0;
3950         space_info_add_new_bytes(info, found, total_bytes -
3951                                  bytes_used - bytes_readonly);
3952         spin_unlock(&found->lock);
3953         *space_info = found;
3954 }
3955
3956 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3957 {
3958         u64 extra_flags = chunk_to_extended(flags) &
3959                                 BTRFS_EXTENDED_PROFILE_MASK;
3960
3961         write_seqlock(&fs_info->profiles_lock);
3962         if (flags & BTRFS_BLOCK_GROUP_DATA)
3963                 fs_info->avail_data_alloc_bits |= extra_flags;
3964         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3965                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3966         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3967                 fs_info->avail_system_alloc_bits |= extra_flags;
3968         write_sequnlock(&fs_info->profiles_lock);
3969 }
3970
3971 /*
3972  * returns target flags in extended format or 0 if restripe for this
3973  * chunk_type is not in progress
3974  *
3975  * should be called with balance_lock held
3976  */
3977 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
3978 {
3979         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3980         u64 target = 0;
3981
3982         if (!bctl)
3983                 return 0;
3984
3985         if (flags & BTRFS_BLOCK_GROUP_DATA &&
3986             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3987                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
3988         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
3989                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3990                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
3991         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
3992                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3993                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
3994         }
3995
3996         return target;
3997 }
3998
3999 /*
4000  * @flags: available profiles in extended format (see ctree.h)
4001  *
4002  * Returns reduced profile in chunk format.  If profile changing is in
4003  * progress (either running or paused) picks the target profile (if it's
4004  * already available), otherwise falls back to plain reducing.
4005  */
4006 static u64 btrfs_reduce_alloc_profile(struct btrfs_fs_info *fs_info, u64 flags)
4007 {
4008         u64 num_devices = fs_info->fs_devices->rw_devices;
4009         u64 target;
4010         u64 raid_type;
4011         u64 allowed = 0;
4012
4013         /*
4014          * see if restripe for this chunk_type is in progress, if so
4015          * try to reduce to the target profile
4016          */
4017         spin_lock(&fs_info->balance_lock);
4018         target = get_restripe_target(fs_info, flags);
4019         if (target) {
4020                 /* pick target profile only if it's already available */
4021                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
4022                         spin_unlock(&fs_info->balance_lock);
4023                         return extended_to_chunk(target);
4024                 }
4025         }
4026         spin_unlock(&fs_info->balance_lock);
4027
4028         /* First, mask out the RAID levels which aren't possible */
4029         for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
4030                 if (num_devices >= btrfs_raid_array[raid_type].devs_min)
4031                         allowed |= btrfs_raid_array[raid_type].bg_flag;
4032         }
4033         allowed &= flags;
4034
4035         if (allowed & BTRFS_BLOCK_GROUP_RAID6)
4036                 allowed = BTRFS_BLOCK_GROUP_RAID6;
4037         else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
4038                 allowed = BTRFS_BLOCK_GROUP_RAID5;
4039         else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
4040                 allowed = BTRFS_BLOCK_GROUP_RAID10;
4041         else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
4042                 allowed = BTRFS_BLOCK_GROUP_RAID1;
4043         else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
4044                 allowed = BTRFS_BLOCK_GROUP_RAID0;
4045
4046         flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
4047
4048         return extended_to_chunk(flags | allowed);
4049 }
4050
4051 static u64 get_alloc_profile(struct btrfs_fs_info *fs_info, u64 orig_flags)
4052 {
4053         unsigned seq;
4054         u64 flags;
4055
4056         do {
4057                 flags = orig_flags;
4058                 seq = read_seqbegin(&fs_info->profiles_lock);
4059
4060                 if (flags & BTRFS_BLOCK_GROUP_DATA)
4061                         flags |= fs_info->avail_data_alloc_bits;
4062                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4063                         flags |= fs_info->avail_system_alloc_bits;
4064                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
4065                         flags |= fs_info->avail_metadata_alloc_bits;
4066         } while (read_seqretry(&fs_info->profiles_lock, seq));
4067
4068         return btrfs_reduce_alloc_profile(fs_info, flags);
4069 }
4070
4071 static u64 get_alloc_profile_by_root(struct btrfs_root *root, int data)
4072 {
4073         struct btrfs_fs_info *fs_info = root->fs_info;
4074         u64 flags;
4075         u64 ret;
4076
4077         if (data)
4078                 flags = BTRFS_BLOCK_GROUP_DATA;
4079         else if (root == fs_info->chunk_root)
4080                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
4081         else
4082                 flags = BTRFS_BLOCK_GROUP_METADATA;
4083
4084         ret = get_alloc_profile(fs_info, flags);
4085         return ret;
4086 }
4087
4088 u64 btrfs_data_alloc_profile(struct btrfs_fs_info *fs_info)
4089 {
4090         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_DATA);
4091 }
4092
4093 u64 btrfs_metadata_alloc_profile(struct btrfs_fs_info *fs_info)
4094 {
4095         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4096 }
4097
4098 u64 btrfs_system_alloc_profile(struct btrfs_fs_info *fs_info)
4099 {
4100         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4101 }
4102
4103 static u64 btrfs_space_info_used(struct btrfs_space_info *s_info,
4104                                  bool may_use_included)
4105 {
4106         ASSERT(s_info);
4107         return s_info->bytes_used + s_info->bytes_reserved +
4108                 s_info->bytes_pinned + s_info->bytes_readonly +
4109                 (may_use_included ? s_info->bytes_may_use : 0);
4110 }
4111
4112 int btrfs_alloc_data_chunk_ondemand(struct btrfs_inode *inode, u64 bytes)
4113 {
4114         struct btrfs_root *root = inode->root;
4115         struct btrfs_fs_info *fs_info = root->fs_info;
4116         struct btrfs_space_info *data_sinfo = fs_info->data_sinfo;
4117         u64 used;
4118         int ret = 0;
4119         int need_commit = 2;
4120         int have_pinned_space;
4121
4122         /* make sure bytes are sectorsize aligned */
4123         bytes = ALIGN(bytes, fs_info->sectorsize);
4124
4125         if (btrfs_is_free_space_inode(inode)) {
4126                 need_commit = 0;
4127                 ASSERT(current->journal_info);
4128         }
4129
4130 again:
4131         /* make sure we have enough space to handle the data first */
4132         spin_lock(&data_sinfo->lock);
4133         used = btrfs_space_info_used(data_sinfo, true);
4134
4135         if (used + bytes > data_sinfo->total_bytes) {
4136                 struct btrfs_trans_handle *trans;
4137
4138                 /*
4139                  * if we don't have enough free bytes in this space then we need
4140                  * to alloc a new chunk.
4141                  */
4142                 if (!data_sinfo->full) {
4143                         u64 alloc_target;
4144
4145                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
4146                         spin_unlock(&data_sinfo->lock);
4147
4148                         alloc_target = btrfs_data_alloc_profile(fs_info);
4149                         /*
4150                          * It is ugly that we don't call nolock join
4151                          * transaction for the free space inode case here.
4152                          * But it is safe because we only do the data space
4153                          * reservation for the free space cache in the
4154                          * transaction context, the common join transaction
4155                          * just increase the counter of the current transaction
4156                          * handler, doesn't try to acquire the trans_lock of
4157                          * the fs.
4158                          */
4159                         trans = btrfs_join_transaction(root);
4160                         if (IS_ERR(trans))
4161                                 return PTR_ERR(trans);
4162
4163                         ret = do_chunk_alloc(trans, alloc_target,
4164                                              CHUNK_ALLOC_NO_FORCE);
4165                         btrfs_end_transaction(trans);
4166                         if (ret < 0) {
4167                                 if (ret != -ENOSPC)
4168                                         return ret;
4169                                 else {
4170                                         have_pinned_space = 1;
4171                                         goto commit_trans;
4172                                 }
4173                         }
4174
4175                         goto again;
4176                 }
4177
4178                 /*
4179                  * If we don't have enough pinned space to deal with this
4180                  * allocation, and no removed chunk in current transaction,
4181                  * don't bother committing the transaction.
4182                  */
4183                 have_pinned_space = __percpu_counter_compare(
4184                         &data_sinfo->total_bytes_pinned,
4185                         used + bytes - data_sinfo->total_bytes,
4186                         BTRFS_TOTAL_BYTES_PINNED_BATCH);
4187                 spin_unlock(&data_sinfo->lock);
4188
4189                 /* commit the current transaction and try again */
4190 commit_trans:
4191                 if (need_commit) {
4192                         need_commit--;
4193
4194                         if (need_commit > 0) {
4195                                 btrfs_start_delalloc_roots(fs_info, -1);
4196                                 btrfs_wait_ordered_roots(fs_info, U64_MAX, 0,
4197                                                          (u64)-1);
4198                         }
4199
4200                         trans = btrfs_join_transaction(root);
4201                         if (IS_ERR(trans))
4202                                 return PTR_ERR(trans);
4203                         if (have_pinned_space >= 0 ||
4204                             test_bit(BTRFS_TRANS_HAVE_FREE_BGS,
4205                                      &trans->transaction->flags) ||
4206                             need_commit > 0) {
4207                                 ret = btrfs_commit_transaction(trans);
4208                                 if (ret)
4209                                         return ret;
4210                                 /*
4211                                  * The cleaner kthread might still be doing iput
4212                                  * operations. Wait for it to finish so that
4213                                  * more space is released.
4214                                  */
4215                                 mutex_lock(&fs_info->cleaner_delayed_iput_mutex);
4216                                 mutex_unlock(&fs_info->cleaner_delayed_iput_mutex);
4217                                 goto again;
4218                         } else {
4219                                 btrfs_end_transaction(trans);
4220                         }
4221                 }
4222
4223                 trace_btrfs_space_reservation(fs_info,
4224                                               "space_info:enospc",
4225                                               data_sinfo->flags, bytes, 1);
4226                 return -ENOSPC;
4227         }
4228         data_sinfo->bytes_may_use += bytes;
4229         trace_btrfs_space_reservation(fs_info, "space_info",
4230                                       data_sinfo->flags, bytes, 1);
4231         spin_unlock(&data_sinfo->lock);
4232
4233         return ret;
4234 }
4235
4236 int btrfs_check_data_free_space(struct inode *inode,
4237                         struct extent_changeset **reserved, u64 start, u64 len)
4238 {
4239         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4240         int ret;
4241
4242         /* align the range */
4243         len = round_up(start + len, fs_info->sectorsize) -
4244               round_down(start, fs_info->sectorsize);
4245         start = round_down(start, fs_info->sectorsize);
4246
4247         ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode), len);
4248         if (ret < 0)
4249                 return ret;
4250
4251         /* Use new btrfs_qgroup_reserve_data to reserve precious data space. */
4252         ret = btrfs_qgroup_reserve_data(inode, reserved, start, len);
4253         if (ret < 0)
4254                 btrfs_free_reserved_data_space_noquota(inode, start, len);
4255         else
4256                 ret = 0;
4257         return ret;
4258 }
4259
4260 /*
4261  * Called if we need to clear a data reservation for this inode
4262  * Normally in a error case.
4263  *
4264  * This one will *NOT* use accurate qgroup reserved space API, just for case
4265  * which we can't sleep and is sure it won't affect qgroup reserved space.
4266  * Like clear_bit_hook().
4267  */
4268 void btrfs_free_reserved_data_space_noquota(struct inode *inode, u64 start,
4269                                             u64 len)
4270 {
4271         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4272         struct btrfs_space_info *data_sinfo;
4273
4274         /* Make sure the range is aligned to sectorsize */
4275         len = round_up(start + len, fs_info->sectorsize) -
4276               round_down(start, fs_info->sectorsize);
4277         start = round_down(start, fs_info->sectorsize);
4278
4279         data_sinfo = fs_info->data_sinfo;
4280         spin_lock(&data_sinfo->lock);
4281         if (WARN_ON(data_sinfo->bytes_may_use < len))
4282                 data_sinfo->bytes_may_use = 0;
4283         else
4284                 data_sinfo->bytes_may_use -= len;
4285         trace_btrfs_space_reservation(fs_info, "space_info",
4286                                       data_sinfo->flags, len, 0);
4287         spin_unlock(&data_sinfo->lock);
4288 }
4289
4290 /*
4291  * Called if we need to clear a data reservation for this inode
4292  * Normally in a error case.
4293  *
4294  * This one will handle the per-inode data rsv map for accurate reserved
4295  * space framework.
4296  */
4297 void btrfs_free_reserved_data_space(struct inode *inode,
4298                         struct extent_changeset *reserved, u64 start, u64 len)
4299 {
4300         struct btrfs_root *root = BTRFS_I(inode)->root;
4301
4302         /* Make sure the range is aligned to sectorsize */
4303         len = round_up(start + len, root->fs_info->sectorsize) -
4304               round_down(start, root->fs_info->sectorsize);
4305         start = round_down(start, root->fs_info->sectorsize);
4306
4307         btrfs_free_reserved_data_space_noquota(inode, start, len);
4308         btrfs_qgroup_free_data(inode, reserved, start, len);
4309 }
4310
4311 static void force_metadata_allocation(struct btrfs_fs_info *info)
4312 {
4313         struct list_head *head = &info->space_info;
4314         struct btrfs_space_info *found;
4315
4316         rcu_read_lock();
4317         list_for_each_entry_rcu(found, head, list) {
4318                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
4319                         found->force_alloc = CHUNK_ALLOC_FORCE;
4320         }
4321         rcu_read_unlock();
4322 }
4323
4324 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
4325 {
4326         return (global->size << 1);
4327 }
4328
4329 static int should_alloc_chunk(struct btrfs_fs_info *fs_info,
4330                               struct btrfs_space_info *sinfo, int force)
4331 {
4332         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4333         u64 bytes_used = btrfs_space_info_used(sinfo, false);
4334         u64 thresh;
4335
4336         if (force == CHUNK_ALLOC_FORCE)
4337                 return 1;
4338
4339         /*
4340          * We need to take into account the global rsv because for all intents
4341          * and purposes it's used space.  Don't worry about locking the
4342          * global_rsv, it doesn't change except when the transaction commits.
4343          */
4344         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
4345                 bytes_used += calc_global_rsv_need_space(global_rsv);
4346
4347         /*
4348          * in limited mode, we want to have some free space up to
4349          * about 1% of the FS size.
4350          */
4351         if (force == CHUNK_ALLOC_LIMITED) {
4352                 thresh = btrfs_super_total_bytes(fs_info->super_copy);
4353                 thresh = max_t(u64, SZ_64M, div_factor_fine(thresh, 1));
4354
4355                 if (sinfo->total_bytes - bytes_used < thresh)
4356                         return 1;
4357         }
4358
4359         if (bytes_used + SZ_2M < div_factor(sinfo->total_bytes, 8))
4360                 return 0;
4361         return 1;
4362 }
4363
4364 static u64 get_profile_num_devs(struct btrfs_fs_info *fs_info, u64 type)
4365 {
4366         u64 num_dev;
4367
4368         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
4369                     BTRFS_BLOCK_GROUP_RAID0 |
4370                     BTRFS_BLOCK_GROUP_RAID5 |
4371                     BTRFS_BLOCK_GROUP_RAID6))
4372                 num_dev = fs_info->fs_devices->rw_devices;
4373         else if (type & BTRFS_BLOCK_GROUP_RAID1)
4374                 num_dev = 2;
4375         else
4376                 num_dev = 1;    /* DUP or single */
4377
4378         return num_dev;
4379 }
4380
4381 /*
4382  * If @is_allocation is true, reserve space in the system space info necessary
4383  * for allocating a chunk, otherwise if it's false, reserve space necessary for
4384  * removing a chunk.
4385  */
4386 void check_system_chunk(struct btrfs_trans_handle *trans, u64 type)
4387 {
4388         struct btrfs_fs_info *fs_info = trans->fs_info;
4389         struct btrfs_space_info *info;
4390         u64 left;
4391         u64 thresh;
4392         int ret = 0;
4393         u64 num_devs;
4394
4395         /*
4396          * Needed because we can end up allocating a system chunk and for an
4397          * atomic and race free space reservation in the chunk block reserve.
4398          */
4399         lockdep_assert_held(&fs_info->chunk_mutex);
4400
4401         info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4402         spin_lock(&info->lock);
4403         left = info->total_bytes - btrfs_space_info_used(info, true);
4404         spin_unlock(&info->lock);
4405
4406         num_devs = get_profile_num_devs(fs_info, type);
4407
4408         /* num_devs device items to update and 1 chunk item to add or remove */
4409         thresh = btrfs_calc_trunc_metadata_size(fs_info, num_devs) +
4410                 btrfs_calc_trans_metadata_size(fs_info, 1);
4411
4412         if (left < thresh && btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
4413                 btrfs_info(fs_info, "left=%llu, need=%llu, flags=%llu",
4414                            left, thresh, type);
4415                 dump_space_info(fs_info, info, 0, 0);
4416         }
4417
4418         if (left < thresh) {
4419                 u64 flags = btrfs_system_alloc_profile(fs_info);
4420
4421                 /*
4422                  * Ignore failure to create system chunk. We might end up not
4423                  * needing it, as we might not need to COW all nodes/leafs from
4424                  * the paths we visit in the chunk tree (they were already COWed
4425                  * or created in the current transaction for example).
4426                  */
4427                 ret = btrfs_alloc_chunk(trans, flags);
4428         }
4429
4430         if (!ret) {
4431                 ret = btrfs_block_rsv_add(fs_info->chunk_root,
4432                                           &fs_info->chunk_block_rsv,
4433                                           thresh, BTRFS_RESERVE_NO_FLUSH);
4434                 if (!ret)
4435                         trans->chunk_bytes_reserved += thresh;
4436         }
4437 }
4438
4439 /*
4440  * If force is CHUNK_ALLOC_FORCE:
4441  *    - return 1 if it successfully allocates a chunk,
4442  *    - return errors including -ENOSPC otherwise.
4443  * If force is NOT CHUNK_ALLOC_FORCE:
4444  *    - return 0 if it doesn't need to allocate a new chunk,
4445  *    - return 1 if it successfully allocates a chunk,
4446  *    - return errors including -ENOSPC otherwise.
4447  */
4448 static int do_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags,
4449                           int force)
4450 {
4451         struct btrfs_fs_info *fs_info = trans->fs_info;
4452         struct btrfs_space_info *space_info;
4453         bool wait_for_alloc = false;
4454         bool should_alloc = false;
4455         int ret = 0;
4456
4457         /* Don't re-enter if we're already allocating a chunk */
4458         if (trans->allocating_chunk)
4459                 return -ENOSPC;
4460
4461         space_info = __find_space_info(fs_info, flags);
4462         ASSERT(space_info);
4463
4464         do {
4465                 spin_lock(&space_info->lock);
4466                 if (force < space_info->force_alloc)
4467                         force = space_info->force_alloc;
4468                 should_alloc = should_alloc_chunk(fs_info, space_info, force);
4469                 if (space_info->full) {
4470                         /* No more free physical space */
4471                         if (should_alloc)
4472                                 ret = -ENOSPC;
4473                         else
4474                                 ret = 0;
4475                         spin_unlock(&space_info->lock);
4476                         return ret;
4477                 } else if (!should_alloc) {
4478                         spin_unlock(&space_info->lock);
4479                         return 0;
4480                 } else if (space_info->chunk_alloc) {
4481                         /*
4482                          * Someone is already allocating, so we need to block
4483                          * until this someone is finished and then loop to
4484                          * recheck if we should continue with our allocation
4485                          * attempt.
4486                          */
4487                         wait_for_alloc = true;
4488                         spin_unlock(&space_info->lock);
4489                         mutex_lock(&fs_info->chunk_mutex);
4490                         mutex_unlock(&fs_info->chunk_mutex);
4491                 } else {
4492                         /* Proceed with allocation */
4493                         space_info->chunk_alloc = 1;
4494                         wait_for_alloc = false;
4495                         spin_unlock(&space_info->lock);
4496                 }
4497
4498                 cond_resched();
4499         } while (wait_for_alloc);
4500
4501         mutex_lock(&fs_info->chunk_mutex);
4502         trans->allocating_chunk = true;
4503
4504         /*
4505          * If we have mixed data/metadata chunks we want to make sure we keep
4506          * allocating mixed chunks instead of individual chunks.
4507          */
4508         if (btrfs_mixed_space_info(space_info))
4509                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4510
4511         /*
4512          * if we're doing a data chunk, go ahead and make sure that
4513          * we keep a reasonable number of metadata chunks allocated in the
4514          * FS as well.
4515          */
4516         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4517                 fs_info->data_chunk_allocations++;
4518                 if (!(fs_info->data_chunk_allocations %
4519                       fs_info->metadata_ratio))
4520                         force_metadata_allocation(fs_info);
4521         }
4522
4523         /*
4524          * Check if we have enough space in SYSTEM chunk because we may need
4525          * to update devices.
4526          */
4527         check_system_chunk(trans, flags);
4528
4529         ret = btrfs_alloc_chunk(trans, flags);
4530         trans->allocating_chunk = false;
4531
4532         spin_lock(&space_info->lock);
4533         if (ret < 0) {
4534                 if (ret == -ENOSPC)
4535                         space_info->full = 1;
4536                 else
4537                         goto out;
4538         } else {
4539                 ret = 1;
4540         }
4541
4542         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4543 out:
4544         space_info->chunk_alloc = 0;
4545         spin_unlock(&space_info->lock);
4546         mutex_unlock(&fs_info->chunk_mutex);
4547         /*
4548          * When we allocate a new chunk we reserve space in the chunk block
4549          * reserve to make sure we can COW nodes/leafs in the chunk tree or
4550          * add new nodes/leafs to it if we end up needing to do it when
4551          * inserting the chunk item and updating device items as part of the
4552          * second phase of chunk allocation, performed by
4553          * btrfs_finish_chunk_alloc(). So make sure we don't accumulate a
4554          * large number of new block groups to create in our transaction
4555          * handle's new_bgs list to avoid exhausting the chunk block reserve
4556          * in extreme cases - like having a single transaction create many new
4557          * block groups when starting to write out the free space caches of all
4558          * the block groups that were made dirty during the lifetime of the
4559          * transaction.
4560          */
4561         if (trans->can_flush_pending_bgs &&
4562             trans->chunk_bytes_reserved >= (u64)SZ_2M) {
4563                 btrfs_create_pending_block_groups(trans);
4564                 btrfs_trans_release_chunk_metadata(trans);
4565         }
4566         return ret;
4567 }
4568
4569 static int can_overcommit(struct btrfs_fs_info *fs_info,
4570                           struct btrfs_space_info *space_info, u64 bytes,
4571                           enum btrfs_reserve_flush_enum flush,
4572                           bool system_chunk)
4573 {
4574         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4575         u64 profile;
4576         u64 space_size;
4577         u64 avail;
4578         u64 used;
4579
4580         /* Don't overcommit when in mixed mode. */
4581         if (space_info->flags & BTRFS_BLOCK_GROUP_DATA)
4582                 return 0;
4583
4584         if (system_chunk)
4585                 profile = btrfs_system_alloc_profile(fs_info);
4586         else
4587                 profile = btrfs_metadata_alloc_profile(fs_info);
4588
4589         used = btrfs_space_info_used(space_info, false);
4590
4591         /*
4592          * We only want to allow over committing if we have lots of actual space
4593          * free, but if we don't have enough space to handle the global reserve
4594          * space then we could end up having a real enospc problem when trying
4595          * to allocate a chunk or some other such important allocation.
4596          */
4597         spin_lock(&global_rsv->lock);
4598         space_size = calc_global_rsv_need_space(global_rsv);
4599         spin_unlock(&global_rsv->lock);
4600         if (used + space_size >= space_info->total_bytes)
4601                 return 0;
4602
4603         used += space_info->bytes_may_use;
4604
4605         avail = atomic64_read(&fs_info->free_chunk_space);
4606
4607         /*
4608          * If we have dup, raid1 or raid10 then only half of the free
4609          * space is actually useable.  For raid56, the space info used
4610          * doesn't include the parity drive, so we don't have to
4611          * change the math
4612          */
4613         if (profile & (BTRFS_BLOCK_GROUP_DUP |
4614                        BTRFS_BLOCK_GROUP_RAID1 |
4615                        BTRFS_BLOCK_GROUP_RAID10))
4616                 avail >>= 1;
4617
4618         /*
4619          * If we aren't flushing all things, let us overcommit up to
4620          * 1/2th of the space. If we can flush, don't let us overcommit
4621          * too much, let it overcommit up to 1/8 of the space.
4622          */
4623         if (flush == BTRFS_RESERVE_FLUSH_ALL)
4624                 avail >>= 3;
4625         else
4626                 avail >>= 1;
4627
4628         if (used + bytes < space_info->total_bytes + avail)
4629                 return 1;
4630         return 0;
4631 }
4632
4633 static void btrfs_writeback_inodes_sb_nr(struct btrfs_fs_info *fs_info,
4634                                          unsigned long nr_pages, int nr_items)
4635 {
4636         struct super_block *sb = fs_info->sb;
4637
4638         if (down_read_trylock(&sb->s_umount)) {
4639                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
4640                 up_read(&sb->s_umount);
4641         } else {
4642                 /*
4643                  * We needn't worry the filesystem going from r/w to r/o though
4644                  * we don't acquire ->s_umount mutex, because the filesystem
4645                  * should guarantee the delalloc inodes list be empty after
4646                  * the filesystem is readonly(all dirty pages are written to
4647                  * the disk).
4648                  */
4649                 btrfs_start_delalloc_roots(fs_info, nr_items);
4650                 if (!current->journal_info)
4651                         btrfs_wait_ordered_roots(fs_info, nr_items, 0, (u64)-1);
4652         }
4653 }
4654
4655 static inline u64 calc_reclaim_items_nr(struct btrfs_fs_info *fs_info,
4656                                         u64 to_reclaim)
4657 {
4658         u64 bytes;
4659         u64 nr;
4660
4661         bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
4662         nr = div64_u64(to_reclaim, bytes);
4663         if (!nr)
4664                 nr = 1;
4665         return nr;
4666 }
4667
4668 #define EXTENT_SIZE_PER_ITEM    SZ_256K
4669
4670 /*
4671  * shrink metadata reservation for delalloc
4672  */
4673 static void shrink_delalloc(struct btrfs_fs_info *fs_info, u64 to_reclaim,
4674                             u64 orig, bool wait_ordered)
4675 {
4676         struct btrfs_space_info *space_info;
4677         struct btrfs_trans_handle *trans;
4678         u64 delalloc_bytes;
4679         u64 max_reclaim;
4680         u64 items;
4681         long time_left;
4682         unsigned long nr_pages;
4683         int loops;
4684
4685         /* Calc the number of the pages we need flush for space reservation */
4686         items = calc_reclaim_items_nr(fs_info, to_reclaim);
4687         to_reclaim = items * EXTENT_SIZE_PER_ITEM;
4688
4689         trans = (struct btrfs_trans_handle *)current->journal_info;
4690         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4691
4692         delalloc_bytes = percpu_counter_sum_positive(
4693                                                 &fs_info->delalloc_bytes);
4694         if (delalloc_bytes == 0) {
4695                 if (trans)
4696                         return;
4697                 if (wait_ordered)
4698                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
4699                 return;
4700         }
4701
4702         loops = 0;
4703         while (delalloc_bytes && loops < 3) {
4704                 max_reclaim = min(delalloc_bytes, to_reclaim);
4705                 nr_pages = max_reclaim >> PAGE_SHIFT;
4706                 btrfs_writeback_inodes_sb_nr(fs_info, nr_pages, items);
4707                 /*
4708                  * We need to wait for the async pages to actually start before
4709                  * we do anything.
4710                  */
4711                 max_reclaim = atomic_read(&fs_info->async_delalloc_pages);
4712                 if (!max_reclaim)
4713                         goto skip_async;
4714
4715                 if (max_reclaim <= nr_pages)
4716                         max_reclaim = 0;
4717                 else
4718                         max_reclaim -= nr_pages;
4719
4720                 wait_event(fs_info->async_submit_wait,
4721                            atomic_read(&fs_info->async_delalloc_pages) <=
4722                            (int)max_reclaim);
4723 skip_async:
4724                 spin_lock(&space_info->lock);
4725                 if (list_empty(&space_info->tickets) &&
4726                     list_empty(&space_info->priority_tickets)) {
4727                         spin_unlock(&space_info->lock);
4728                         break;
4729                 }
4730                 spin_unlock(&space_info->lock);
4731
4732                 loops++;
4733                 if (wait_ordered && !trans) {
4734                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
4735                 } else {
4736                         time_left = schedule_timeout_killable(1);
4737                         if (time_left)
4738                                 break;
4739                 }
4740                 delalloc_bytes = percpu_counter_sum_positive(
4741                                                 &fs_info->delalloc_bytes);
4742         }
4743 }
4744
4745 struct reserve_ticket {
4746         u64 bytes;
4747         int error;
4748         struct list_head list;
4749         wait_queue_head_t wait;
4750 };
4751
4752 /**
4753  * maybe_commit_transaction - possibly commit the transaction if its ok to
4754  * @root - the root we're allocating for
4755  * @bytes - the number of bytes we want to reserve
4756  * @force - force the commit
4757  *
4758  * This will check to make sure that committing the transaction will actually
4759  * get us somewhere and then commit the transaction if it does.  Otherwise it
4760  * will return -ENOSPC.
4761  */
4762 static int may_commit_transaction(struct btrfs_fs_info *fs_info,
4763                                   struct btrfs_space_info *space_info)
4764 {
4765         struct reserve_ticket *ticket = NULL;
4766         struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_block_rsv;
4767         struct btrfs_trans_handle *trans;
4768         u64 bytes;
4769
4770         trans = (struct btrfs_trans_handle *)current->journal_info;
4771         if (trans)
4772                 return -EAGAIN;
4773
4774         spin_lock(&space_info->lock);
4775         if (!list_empty(&space_info->priority_tickets))
4776                 ticket = list_first_entry(&space_info->priority_tickets,
4777                                           struct reserve_ticket, list);
4778         else if (!list_empty(&space_info->tickets))
4779                 ticket = list_first_entry(&space_info->tickets,
4780                                           struct reserve_ticket, list);
4781         bytes = (ticket) ? ticket->bytes : 0;
4782         spin_unlock(&space_info->lock);
4783
4784         if (!bytes)
4785                 return 0;
4786
4787         /* See if there is enough pinned space to make this reservation */
4788         if (__percpu_counter_compare(&space_info->total_bytes_pinned,
4789                                    bytes,
4790                                    BTRFS_TOTAL_BYTES_PINNED_BATCH) >= 0)
4791                 goto commit;
4792
4793         /*
4794          * See if there is some space in the delayed insertion reservation for
4795          * this reservation.
4796          */
4797         if (space_info != delayed_rsv->space_info)
4798                 return -ENOSPC;
4799
4800         spin_lock(&delayed_rsv->lock);
4801         if (delayed_rsv->size > bytes)
4802                 bytes = 0;
4803         else
4804                 bytes -= delayed_rsv->size;
4805         spin_unlock(&delayed_rsv->lock);
4806
4807         if (__percpu_counter_compare(&space_info->total_bytes_pinned,
4808                                    bytes,
4809                                    BTRFS_TOTAL_BYTES_PINNED_BATCH) < 0) {
4810                 return -ENOSPC;
4811         }
4812
4813 commit:
4814         trans = btrfs_join_transaction(fs_info->extent_root);
4815         if (IS_ERR(trans))
4816                 return -ENOSPC;
4817
4818         return btrfs_commit_transaction(trans);
4819 }
4820
4821 /*
4822  * Try to flush some data based on policy set by @state. This is only advisory
4823  * and may fail for various reasons. The caller is supposed to examine the
4824  * state of @space_info to detect the outcome.
4825  */
4826 static void flush_space(struct btrfs_fs_info *fs_info,
4827                        struct btrfs_space_info *space_info, u64 num_bytes,
4828                        int state)
4829 {
4830         struct btrfs_root *root = fs_info->extent_root;
4831         struct btrfs_trans_handle *trans;
4832         int nr;
4833         int ret = 0;
4834
4835         switch (state) {
4836         case FLUSH_DELAYED_ITEMS_NR:
4837         case FLUSH_DELAYED_ITEMS:
4838                 if (state == FLUSH_DELAYED_ITEMS_NR)
4839                         nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2;
4840                 else
4841                         nr = -1;
4842
4843                 trans = btrfs_join_transaction(root);
4844                 if (IS_ERR(trans)) {
4845                         ret = PTR_ERR(trans);
4846                         break;
4847                 }
4848                 ret = btrfs_run_delayed_items_nr(trans, nr);
4849                 btrfs_end_transaction(trans);
4850                 break;
4851         case FLUSH_DELALLOC:
4852         case FLUSH_DELALLOC_WAIT:
4853                 shrink_delalloc(fs_info, num_bytes * 2, num_bytes,
4854                                 state == FLUSH_DELALLOC_WAIT);
4855                 break;
4856         case ALLOC_CHUNK:
4857                 trans = btrfs_join_transaction(root);
4858                 if (IS_ERR(trans)) {
4859                         ret = PTR_ERR(trans);
4860                         break;
4861                 }
4862                 ret = do_chunk_alloc(trans,
4863                                      btrfs_metadata_alloc_profile(fs_info),
4864                                      CHUNK_ALLOC_NO_FORCE);
4865                 btrfs_end_transaction(trans);
4866                 if (ret > 0 || ret == -ENOSPC)
4867                         ret = 0;
4868                 break;
4869         case COMMIT_TRANS:
4870                 ret = may_commit_transaction(fs_info, space_info);
4871                 break;
4872         default:
4873                 ret = -ENOSPC;
4874                 break;
4875         }
4876
4877         trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state,
4878                                 ret);
4879         return;
4880 }
4881
4882 static inline u64
4883 btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info,
4884                                  struct btrfs_space_info *space_info,
4885                                  bool system_chunk)
4886 {
4887         struct reserve_ticket *ticket;
4888         u64 used;
4889         u64 expected;
4890         u64 to_reclaim = 0;
4891
4892         list_for_each_entry(ticket, &space_info->tickets, list)
4893                 to_reclaim += ticket->bytes;
4894         list_for_each_entry(ticket, &space_info->priority_tickets, list)
4895                 to_reclaim += ticket->bytes;
4896         if (to_reclaim)
4897                 return to_reclaim;
4898
4899         to_reclaim = min_t(u64, num_online_cpus() * SZ_1M, SZ_16M);
4900         if (can_overcommit(fs_info, space_info, to_reclaim,
4901                            BTRFS_RESERVE_FLUSH_ALL, system_chunk))
4902                 return 0;
4903
4904         used = btrfs_space_info_used(space_info, true);
4905
4906         if (can_overcommit(fs_info, space_info, SZ_1M,
4907                            BTRFS_RESERVE_FLUSH_ALL, system_chunk))
4908                 expected = div_factor_fine(space_info->total_bytes, 95);
4909         else
4910                 expected = div_factor_fine(space_info->total_bytes, 90);
4911
4912         if (used > expected)
4913                 to_reclaim = used - expected;
4914         else
4915                 to_reclaim = 0;
4916         to_reclaim = min(to_reclaim, space_info->bytes_may_use +
4917                                      space_info->bytes_reserved);
4918         return to_reclaim;
4919 }
4920
4921 static inline int need_do_async_reclaim(struct btrfs_fs_info *fs_info,
4922                                         struct btrfs_space_info *space_info,
4923                                         u64 used, bool system_chunk)
4924 {
4925         u64 thresh = div_factor_fine(space_info->total_bytes, 98);
4926
4927         /* If we're just plain full then async reclaim just slows us down. */
4928         if ((space_info->bytes_used + space_info->bytes_reserved) >= thresh)
4929                 return 0;
4930
4931         if (!btrfs_calc_reclaim_metadata_size(fs_info, space_info,
4932                                               system_chunk))
4933                 return 0;
4934
4935         return (used >= thresh && !btrfs_fs_closing(fs_info) &&
4936                 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
4937 }
4938
4939 static void wake_all_tickets(struct list_head *head)
4940 {
4941         struct reserve_ticket *ticket;
4942
4943         while (!list_empty(head)) {
4944                 ticket = list_first_entry(head, struct reserve_ticket, list);
4945                 list_del_init(&ticket->list);
4946                 ticket->error = -ENOSPC;
4947                 wake_up(&ticket->wait);
4948         }
4949 }
4950
4951 /*
4952  * This is for normal flushers, we can wait all goddamned day if we want to.  We
4953  * will loop and continuously try to flush as long as we are making progress.
4954  * We count progress as clearing off tickets each time we have to loop.
4955  */
4956 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
4957 {
4958         struct btrfs_fs_info *fs_info;
4959         struct btrfs_space_info *space_info;
4960         u64 to_reclaim;
4961         int flush_state;
4962         int commit_cycles = 0;
4963         u64 last_tickets_id;
4964
4965         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
4966         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4967
4968         spin_lock(&space_info->lock);
4969         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info,
4970                                                       false);
4971         if (!to_reclaim) {
4972                 space_info->flush = 0;
4973                 spin_unlock(&space_info->lock);
4974                 return;
4975         }
4976         last_tickets_id = space_info->tickets_id;
4977         spin_unlock(&space_info->lock);
4978
4979         flush_state = FLUSH_DELAYED_ITEMS_NR;
4980         do {
4981                 flush_space(fs_info, space_info, to_reclaim, flush_state);
4982                 spin_lock(&space_info->lock);
4983                 if (list_empty(&space_info->tickets)) {
4984                         space_info->flush = 0;
4985                         spin_unlock(&space_info->lock);
4986                         return;
4987                 }
4988                 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info,
4989                                                               space_info,
4990                                                               false);
4991                 if (last_tickets_id == space_info->tickets_id) {
4992                         flush_state++;
4993                 } else {
4994                         last_tickets_id = space_info->tickets_id;
4995                         flush_state = FLUSH_DELAYED_ITEMS_NR;
4996                         if (commit_cycles)
4997                                 commit_cycles--;
4998                 }
4999
5000                 if (flush_state > COMMIT_TRANS) {
5001                         commit_cycles++;
5002                         if (commit_cycles > 2) {
5003                                 wake_all_tickets(&space_info->tickets);
5004                                 space_info->flush = 0;
5005                         } else {
5006                                 flush_state = FLUSH_DELAYED_ITEMS_NR;
5007                         }
5008                 }
5009                 spin_unlock(&space_info->lock);
5010         } while (flush_state <= COMMIT_TRANS);
5011 }
5012
5013 void btrfs_init_async_reclaim_work(struct work_struct *work)
5014 {
5015         INIT_WORK(work, btrfs_async_reclaim_metadata_space);
5016 }
5017
5018 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info,
5019                                             struct btrfs_space_info *space_info,
5020                                             struct reserve_ticket *ticket)
5021 {
5022         u64 to_reclaim;
5023         int flush_state = FLUSH_DELAYED_ITEMS_NR;
5024
5025         spin_lock(&space_info->lock);
5026         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5027                                                       false);
5028         if (!to_reclaim) {
5029                 spin_unlock(&space_info->lock);
5030                 return;
5031         }
5032         spin_unlock(&space_info->lock);
5033
5034         do {
5035                 flush_space(fs_info, space_info, to_reclaim, flush_state);
5036                 flush_state++;
5037                 spin_lock(&space_info->lock);
5038                 if (ticket->bytes == 0) {
5039                         spin_unlock(&space_info->lock);
5040                         return;
5041                 }
5042                 spin_unlock(&space_info->lock);
5043
5044                 /*
5045                  * Priority flushers can't wait on delalloc without
5046                  * deadlocking.
5047                  */
5048                 if (flush_state == FLUSH_DELALLOC ||
5049                     flush_state == FLUSH_DELALLOC_WAIT)
5050                         flush_state = ALLOC_CHUNK;
5051         } while (flush_state < COMMIT_TRANS);
5052 }
5053
5054 static int wait_reserve_ticket(struct btrfs_fs_info *fs_info,
5055                                struct btrfs_space_info *space_info,
5056                                struct reserve_ticket *ticket, u64 orig_bytes)
5057
5058 {
5059         DEFINE_WAIT(wait);
5060         int ret = 0;
5061
5062         spin_lock(&space_info->lock);
5063         while (ticket->bytes > 0 && ticket->error == 0) {
5064                 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE);
5065                 if (ret) {
5066                         ret = -EINTR;
5067                         break;
5068                 }
5069                 spin_unlock(&space_info->lock);
5070
5071                 schedule();
5072
5073                 finish_wait(&ticket->wait, &wait);
5074                 spin_lock(&space_info->lock);
5075         }
5076         if (!ret)
5077                 ret = ticket->error;
5078         if (!list_empty(&ticket->list))
5079                 list_del_init(&ticket->list);
5080         if (ticket->bytes && ticket->bytes < orig_bytes) {
5081                 u64 num_bytes = orig_bytes - ticket->bytes;
5082                 space_info->bytes_may_use -= num_bytes;
5083                 trace_btrfs_space_reservation(fs_info, "space_info",
5084                                               space_info->flags, num_bytes, 0);
5085         }
5086         spin_unlock(&space_info->lock);
5087
5088         return ret;
5089 }
5090
5091 /**
5092  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5093  * @root - the root we're allocating for
5094  * @space_info - the space info we want to allocate from
5095  * @orig_bytes - the number of bytes we want
5096  * @flush - whether or not we can flush to make our reservation
5097  *
5098  * This will reserve orig_bytes number of bytes from the space info associated
5099  * with the block_rsv.  If there is not enough space it will make an attempt to
5100  * flush out space to make room.  It will do this by flushing delalloc if
5101  * possible or committing the transaction.  If flush is 0 then no attempts to
5102  * regain reservations will be made and this will fail if there is not enough
5103  * space already.
5104  */
5105 static int __reserve_metadata_bytes(struct btrfs_fs_info *fs_info,
5106                                     struct btrfs_space_info *space_info,
5107                                     u64 orig_bytes,
5108                                     enum btrfs_reserve_flush_enum flush,
5109                                     bool system_chunk)
5110 {
5111         struct reserve_ticket ticket;
5112         u64 used;
5113         int ret = 0;
5114
5115         ASSERT(orig_bytes);
5116         ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL);
5117
5118         spin_lock(&space_info->lock);
5119         ret = -ENOSPC;
5120         used = btrfs_space_info_used(space_info, true);
5121
5122         /*
5123          * If we have enough space then hooray, make our reservation and carry
5124          * on.  If not see if we can overcommit, and if we can, hooray carry on.
5125          * If not things get more complicated.
5126          */
5127         if (used + orig_bytes <= space_info->total_bytes) {
5128                 space_info->bytes_may_use += orig_bytes;
5129                 trace_btrfs_space_reservation(fs_info, "space_info",
5130                                               space_info->flags, orig_bytes, 1);
5131                 ret = 0;
5132         } else if (can_overcommit(fs_info, space_info, orig_bytes, flush,
5133                                   system_chunk)) {
5134                 space_info->bytes_may_use += orig_bytes;
5135                 trace_btrfs_space_reservation(fs_info, "space_info",
5136                                               space_info->flags, orig_bytes, 1);
5137                 ret = 0;
5138         }
5139
5140         /*
5141          * If we couldn't make a reservation then setup our reservation ticket
5142          * and kick the async worker if it's not already running.
5143          *
5144          * If we are a priority flusher then we just need to add our ticket to
5145          * the list and we will do our own flushing further down.
5146          */
5147         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
5148                 ticket.bytes = orig_bytes;
5149                 ticket.error = 0;
5150                 init_waitqueue_head(&ticket.wait);
5151                 if (flush == BTRFS_RESERVE_FLUSH_ALL) {
5152                         list_add_tail(&ticket.list, &space_info->tickets);
5153                         if (!space_info->flush) {
5154                                 space_info->flush = 1;
5155                                 trace_btrfs_trigger_flush(fs_info,
5156                                                           space_info->flags,
5157                                                           orig_bytes, flush,
5158                                                           "enospc");
5159                                 queue_work(system_unbound_wq,
5160                                            &fs_info->async_reclaim_work);
5161                         }
5162                 } else {
5163                         list_add_tail(&ticket.list,
5164                                       &space_info->priority_tickets);
5165                 }
5166         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
5167                 used += orig_bytes;
5168                 /*
5169                  * We will do the space reservation dance during log replay,
5170                  * which means we won't have fs_info->fs_root set, so don't do
5171                  * the async reclaim as we will panic.
5172                  */
5173                 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) &&
5174                     need_do_async_reclaim(fs_info, space_info,
5175                                           used, system_chunk) &&
5176                     !work_busy(&fs_info->async_reclaim_work)) {
5177                         trace_btrfs_trigger_flush(fs_info, space_info->flags,
5178                                                   orig_bytes, flush, "preempt");
5179                         queue_work(system_unbound_wq,
5180                                    &fs_info->async_reclaim_work);
5181                 }
5182         }
5183         spin_unlock(&space_info->lock);
5184         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
5185                 return ret;
5186
5187         if (flush == BTRFS_RESERVE_FLUSH_ALL)
5188                 return wait_reserve_ticket(fs_info, space_info, &ticket,
5189                                            orig_bytes);
5190
5191         ret = 0;
5192         priority_reclaim_metadata_space(fs_info, space_info, &ticket);
5193         spin_lock(&space_info->lock);
5194         if (ticket.bytes) {
5195                 if (ticket.bytes < orig_bytes) {
5196                         u64 num_bytes = orig_bytes - ticket.bytes;
5197                         space_info->bytes_may_use -= num_bytes;
5198                         trace_btrfs_space_reservation(fs_info, "space_info",
5199                                                       space_info->flags,
5200                                                       num_bytes, 0);
5201
5202                 }
5203                 list_del_init(&ticket.list);
5204                 ret = -ENOSPC;
5205         }
5206         spin_unlock(&space_info->lock);
5207         ASSERT(list_empty(&ticket.list));
5208         return ret;
5209 }
5210
5211 /**
5212  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5213  * @root - the root we're allocating for
5214  * @block_rsv - the block_rsv we're allocating for
5215  * @orig_bytes - the number of bytes we want
5216  * @flush - whether or not we can flush to make our reservation
5217  *
5218  * This will reserve orgi_bytes number of bytes from the space info associated
5219  * with the block_rsv.  If there is not enough space it will make an attempt to
5220  * flush out space to make room.  It will do this by flushing delalloc if
5221  * possible or committing the transaction.  If flush is 0 then no attempts to
5222  * regain reservations will be made and this will fail if there is not enough
5223  * space already.
5224  */
5225 static int reserve_metadata_bytes(struct btrfs_root *root,
5226                                   struct btrfs_block_rsv *block_rsv,
5227                                   u64 orig_bytes,
5228                                   enum btrfs_reserve_flush_enum flush)
5229 {
5230         struct btrfs_fs_info *fs_info = root->fs_info;
5231         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5232         int ret;
5233         bool system_chunk = (root == fs_info->chunk_root);
5234
5235         ret = __reserve_metadata_bytes(fs_info, block_rsv->space_info,
5236                                        orig_bytes, flush, system_chunk);
5237         if (ret == -ENOSPC &&
5238             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
5239                 if (block_rsv != global_rsv &&
5240                     !block_rsv_use_bytes(global_rsv, orig_bytes))
5241                         ret = 0;
5242         }
5243         if (ret == -ENOSPC) {
5244                 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
5245                                               block_rsv->space_info->flags,
5246                                               orig_bytes, 1);
5247
5248                 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
5249                         dump_space_info(fs_info, block_rsv->space_info,
5250                                         orig_bytes, 0);
5251         }
5252         return ret;
5253 }
5254
5255 static struct btrfs_block_rsv *get_block_rsv(
5256                                         const struct btrfs_trans_handle *trans,
5257                                         const struct btrfs_root *root)
5258 {
5259         struct btrfs_fs_info *fs_info = root->fs_info;
5260         struct btrfs_block_rsv *block_rsv = NULL;
5261
5262         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
5263             (root == fs_info->csum_root && trans->adding_csums) ||
5264             (root == fs_info->uuid_root))
5265                 block_rsv = trans->block_rsv;
5266
5267         if (!block_rsv)
5268                 block_rsv = root->block_rsv;
5269
5270         if (!block_rsv)
5271                 block_rsv = &fs_info->empty_block_rsv;
5272
5273         return block_rsv;
5274 }
5275
5276 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
5277                                u64 num_bytes)
5278 {
5279         int ret = -ENOSPC;
5280         spin_lock(&block_rsv->lock);
5281         if (block_rsv->reserved >= num_bytes) {
5282                 block_rsv->reserved -= num_bytes;
5283                 if (block_rsv->reserved < block_rsv->size)
5284                         block_rsv->full = 0;
5285                 ret = 0;
5286         }
5287         spin_unlock(&block_rsv->lock);
5288         return ret;
5289 }
5290
5291 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
5292                                 u64 num_bytes, int update_size)
5293 {
5294         spin_lock(&block_rsv->lock);
5295         block_rsv->reserved += num_bytes;
5296         if (update_size)
5297                 block_rsv->size += num_bytes;
5298         else if (block_rsv->reserved >= block_rsv->size)
5299                 block_rsv->full = 1;
5300         spin_unlock(&block_rsv->lock);
5301 }
5302
5303 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
5304                              struct btrfs_block_rsv *dest, u64 num_bytes,
5305                              int min_factor)
5306 {
5307         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5308         u64 min_bytes;
5309
5310         if (global_rsv->space_info != dest->space_info)
5311                 return -ENOSPC;
5312
5313         spin_lock(&global_rsv->lock);
5314         min_bytes = div_factor(global_rsv->size, min_factor);
5315         if (global_rsv->reserved < min_bytes + num_bytes) {
5316                 spin_unlock(&global_rsv->lock);
5317                 return -ENOSPC;
5318         }
5319         global_rsv->reserved -= num_bytes;
5320         if (global_rsv->reserved < global_rsv->size)
5321                 global_rsv->full = 0;
5322         spin_unlock(&global_rsv->lock);
5323
5324         block_rsv_add_bytes(dest, num_bytes, 1);
5325         return 0;
5326 }
5327
5328 /*
5329  * This is for space we already have accounted in space_info->bytes_may_use, so
5330  * basically when we're returning space from block_rsv's.
5331  */
5332 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
5333                                      struct btrfs_space_info *space_info,
5334                                      u64 num_bytes)
5335 {
5336         struct reserve_ticket *ticket;
5337         struct list_head *head;
5338         u64 used;
5339         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH;
5340         bool check_overcommit = false;
5341
5342         spin_lock(&space_info->lock);
5343         head = &space_info->priority_tickets;
5344
5345         /*
5346          * If we are over our limit then we need to check and see if we can
5347          * overcommit, and if we can't then we just need to free up our space
5348          * and not satisfy any requests.
5349          */
5350         used = btrfs_space_info_used(space_info, true);
5351         if (used - num_bytes >= space_info->total_bytes)
5352                 check_overcommit = true;
5353 again:
5354         while (!list_empty(head) && num_bytes) {
5355                 ticket = list_first_entry(head, struct reserve_ticket,
5356                                           list);
5357                 /*
5358                  * We use 0 bytes because this space is already reserved, so
5359                  * adding the ticket space would be a double count.
5360                  */
5361                 if (check_overcommit &&
5362                     !can_overcommit(fs_info, space_info, 0, flush, false))
5363                         break;
5364                 if (num_bytes >= ticket->bytes) {
5365                         list_del_init(&ticket->list);
5366                         num_bytes -= ticket->bytes;
5367                         ticket->bytes = 0;
5368                         space_info->tickets_id++;
5369                         wake_up(&ticket->wait);
5370                 } else {
5371                         ticket->bytes -= num_bytes;
5372                         num_bytes = 0;
5373                 }
5374         }
5375
5376         if (num_bytes && head == &space_info->priority_tickets) {
5377                 head = &space_info->tickets;
5378                 flush = BTRFS_RESERVE_FLUSH_ALL;
5379                 goto again;
5380         }
5381         space_info->bytes_may_use -= num_bytes;
5382         trace_btrfs_space_reservation(fs_info, "space_info",
5383                                       space_info->flags, num_bytes, 0);
5384         spin_unlock(&space_info->lock);
5385 }
5386
5387 /*
5388  * This is for newly allocated space that isn't accounted in
5389  * space_info->bytes_may_use yet.  So if we allocate a chunk or unpin an extent
5390  * we use this helper.
5391  */
5392 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
5393                                      struct btrfs_space_info *space_info,
5394                                      u64 num_bytes)
5395 {
5396         struct reserve_ticket *ticket;
5397         struct list_head *head = &space_info->priority_tickets;
5398
5399 again:
5400         while (!list_empty(head) && num_bytes) {
5401                 ticket = list_first_entry(head, struct reserve_ticket,
5402                                           list);
5403                 if (num_bytes >= ticket->bytes) {
5404                         trace_btrfs_space_reservation(fs_info, "space_info",
5405                                                       space_info->flags,
5406                                                       ticket->bytes, 1);
5407                         list_del_init(&ticket->list);
5408                         num_bytes -= ticket->bytes;
5409                         space_info->bytes_may_use += ticket->bytes;
5410                         ticket->bytes = 0;
5411                         space_info->tickets_id++;
5412                         wake_up(&ticket->wait);
5413                 } else {
5414                         trace_btrfs_space_reservation(fs_info, "space_info",
5415                                                       space_info->flags,
5416                                                       num_bytes, 1);
5417                         space_info->bytes_may_use += num_bytes;
5418                         ticket->bytes -= num_bytes;
5419                         num_bytes = 0;
5420                 }
5421         }
5422
5423         if (num_bytes && head == &space_info->priority_tickets) {
5424                 head = &space_info->tickets;
5425                 goto again;
5426         }
5427 }
5428
5429 static u64 block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
5430                                     struct btrfs_block_rsv *block_rsv,
5431                                     struct btrfs_block_rsv *dest, u64 num_bytes,
5432                                     u64 *qgroup_to_release_ret)
5433 {
5434         struct btrfs_space_info *space_info = block_rsv->space_info;
5435         u64 qgroup_to_release = 0;
5436         u64 ret;
5437
5438         spin_lock(&block_rsv->lock);
5439         if (num_bytes == (u64)-1) {
5440                 num_bytes = block_rsv->size;
5441                 qgroup_to_release = block_rsv->qgroup_rsv_size;
5442         }
5443         block_rsv->size -= num_bytes;
5444         if (block_rsv->reserved >= block_rsv->size) {
5445                 num_bytes = block_rsv->reserved - block_rsv->size;
5446                 block_rsv->reserved = block_rsv->size;
5447                 block_rsv->full = 1;
5448         } else {
5449                 num_bytes = 0;
5450         }
5451         if (block_rsv->qgroup_rsv_reserved >= block_rsv->qgroup_rsv_size) {
5452                 qgroup_to_release = block_rsv->qgroup_rsv_reserved -
5453                                     block_rsv->qgroup_rsv_size;
5454                 block_rsv->qgroup_rsv_reserved = block_rsv->qgroup_rsv_size;
5455         } else {
5456                 qgroup_to_release = 0;
5457         }
5458         spin_unlock(&block_rsv->lock);
5459
5460         ret = num_bytes;
5461         if (num_bytes > 0) {
5462                 if (dest) {
5463                         spin_lock(&dest->lock);
5464                         if (!dest->full) {
5465                                 u64 bytes_to_add;
5466
5467                                 bytes_to_add = dest->size - dest->reserved;
5468                                 bytes_to_add = min(num_bytes, bytes_to_add);
5469                                 dest->reserved += bytes_to_add;
5470                                 if (dest->reserved >= dest->size)
5471                                         dest->full = 1;
5472                                 num_bytes -= bytes_to_add;
5473                         }
5474                         spin_unlock(&dest->lock);
5475                 }
5476                 if (num_bytes)
5477                         space_info_add_old_bytes(fs_info, space_info,
5478                                                  num_bytes);
5479         }
5480         if (qgroup_to_release_ret)
5481                 *qgroup_to_release_ret = qgroup_to_release;
5482         return ret;
5483 }
5484
5485 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src,
5486                             struct btrfs_block_rsv *dst, u64 num_bytes,
5487                             int update_size)
5488 {
5489         int ret;
5490
5491         ret = block_rsv_use_bytes(src, num_bytes);
5492         if (ret)
5493                 return ret;
5494
5495         block_rsv_add_bytes(dst, num_bytes, update_size);
5496         return 0;
5497 }
5498
5499 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
5500 {
5501         memset(rsv, 0, sizeof(*rsv));
5502         spin_lock_init(&rsv->lock);
5503         rsv->type = type;
5504 }
5505
5506 void btrfs_init_metadata_block_rsv(struct btrfs_fs_info *fs_info,
5507                                    struct btrfs_block_rsv *rsv,
5508                                    unsigned short type)
5509 {
5510         btrfs_init_block_rsv(rsv, type);
5511         rsv->space_info = __find_space_info(fs_info,
5512                                             BTRFS_BLOCK_GROUP_METADATA);
5513 }
5514
5515 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_fs_info *fs_info,
5516                                               unsigned short type)
5517 {
5518         struct btrfs_block_rsv *block_rsv;
5519
5520         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
5521         if (!block_rsv)
5522                 return NULL;
5523
5524         btrfs_init_metadata_block_rsv(fs_info, block_rsv, type);
5525         return block_rsv;
5526 }
5527
5528 void btrfs_free_block_rsv(struct btrfs_fs_info *fs_info,
5529                           struct btrfs_block_rsv *rsv)
5530 {
5531         if (!rsv)
5532                 return;
5533         btrfs_block_rsv_release(fs_info, rsv, (u64)-1);
5534         kfree(rsv);
5535 }
5536
5537 void __btrfs_free_block_rsv(struct btrfs_block_rsv *rsv)
5538 {
5539         kfree(rsv);
5540 }
5541
5542 int btrfs_block_rsv_add(struct btrfs_root *root,
5543                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
5544                         enum btrfs_reserve_flush_enum flush)
5545 {
5546         int ret;
5547
5548         if (num_bytes == 0)
5549                 return 0;
5550
5551         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5552         if (!ret) {
5553                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
5554                 return 0;
5555         }
5556
5557         return ret;
5558 }
5559
5560 int btrfs_block_rsv_check(struct btrfs_block_rsv *block_rsv, int min_factor)
5561 {
5562         u64 num_bytes = 0;
5563         int ret = -ENOSPC;
5564
5565         if (!block_rsv)
5566                 return 0;
5567
5568         spin_lock(&block_rsv->lock);
5569         num_bytes = div_factor(block_rsv->size, min_factor);
5570         if (block_rsv->reserved >= num_bytes)
5571                 ret = 0;
5572         spin_unlock(&block_rsv->lock);
5573
5574         return ret;
5575 }
5576
5577 int btrfs_block_rsv_refill(struct btrfs_root *root,
5578                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
5579                            enum btrfs_reserve_flush_enum flush)
5580 {
5581         u64 num_bytes = 0;
5582         int ret = -ENOSPC;
5583
5584         if (!block_rsv)
5585                 return 0;
5586
5587         spin_lock(&block_rsv->lock);
5588         num_bytes = min_reserved;
5589         if (block_rsv->reserved >= num_bytes)
5590                 ret = 0;
5591         else
5592                 num_bytes -= block_rsv->reserved;
5593         spin_unlock(&block_rsv->lock);
5594
5595         if (!ret)
5596                 return 0;
5597
5598         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5599         if (!ret) {
5600                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
5601                 return 0;
5602         }
5603
5604         return ret;
5605 }
5606
5607 /**
5608  * btrfs_inode_rsv_refill - refill the inode block rsv.
5609  * @inode - the inode we are refilling.
5610  * @flush - the flusing restriction.
5611  *
5612  * Essentially the same as btrfs_block_rsv_refill, except it uses the
5613  * block_rsv->size as the minimum size.  We'll either refill the missing amount
5614  * or return if we already have enough space.  This will also handle the resreve
5615  * tracepoint for the reserved amount.
5616  */
5617 static int btrfs_inode_rsv_refill(struct btrfs_inode *inode,
5618                                   enum btrfs_reserve_flush_enum flush)
5619 {
5620         struct btrfs_root *root = inode->root;
5621         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
5622         u64 num_bytes = 0;
5623         u64 qgroup_num_bytes = 0;
5624         int ret = -ENOSPC;
5625
5626         spin_lock(&block_rsv->lock);
5627         if (block_rsv->reserved < block_rsv->size)
5628                 num_bytes = block_rsv->size - block_rsv->reserved;
5629         if (block_rsv->qgroup_rsv_reserved < block_rsv->qgroup_rsv_size)
5630                 qgroup_num_bytes = block_rsv->qgroup_rsv_size -
5631                                    block_rsv->qgroup_rsv_reserved;
5632         spin_unlock(&block_rsv->lock);
5633
5634         if (num_bytes == 0)
5635                 return 0;
5636
5637         ret = btrfs_qgroup_reserve_meta_prealloc(root, qgroup_num_bytes, true);
5638         if (ret)
5639                 return ret;
5640         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5641         if (!ret) {
5642                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
5643                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5644                                               btrfs_ino(inode), num_bytes, 1);
5645
5646                 /* Don't forget to increase qgroup_rsv_reserved */
5647                 spin_lock(&block_rsv->lock);
5648                 block_rsv->qgroup_rsv_reserved += qgroup_num_bytes;
5649                 spin_unlock(&block_rsv->lock);
5650         } else
5651                 btrfs_qgroup_free_meta_prealloc(root, qgroup_num_bytes);
5652         return ret;
5653 }
5654
5655 /**
5656  * btrfs_inode_rsv_release - release any excessive reservation.
5657  * @inode - the inode we need to release from.
5658  * @qgroup_free - free or convert qgroup meta.
5659  *   Unlike normal operation, qgroup meta reservation needs to know if we are
5660  *   freeing qgroup reservation or just converting it into per-trans.  Normally
5661  *   @qgroup_free is true for error handling, and false for normal release.
5662  *
5663  * This is the same as btrfs_block_rsv_release, except that it handles the
5664  * tracepoint for the reservation.
5665  */
5666 static void btrfs_inode_rsv_release(struct btrfs_inode *inode, bool qgroup_free)
5667 {
5668         struct btrfs_fs_info *fs_info = inode->root->fs_info;
5669         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5670         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
5671         u64 released = 0;
5672         u64 qgroup_to_release = 0;
5673
5674         /*
5675          * Since we statically set the block_rsv->size we just want to say we
5676          * are releasing 0 bytes, and then we'll just get the reservation over
5677          * the size free'd.
5678          */
5679         released = block_rsv_release_bytes(fs_info, block_rsv, global_rsv, 0,
5680                                            &qgroup_to_release);
5681         if (released > 0)
5682                 trace_btrfs_space_reservation(fs_info, "delalloc",
5683                                               btrfs_ino(inode), released, 0);
5684         if (qgroup_free)
5685                 btrfs_qgroup_free_meta_prealloc(inode->root, qgroup_to_release);
5686         else
5687                 btrfs_qgroup_convert_reserved_meta(inode->root,
5688                                                    qgroup_to_release);
5689 }
5690
5691 void btrfs_block_rsv_release(struct btrfs_fs_info *fs_info,
5692                              struct btrfs_block_rsv *block_rsv,
5693                              u64 num_bytes)
5694 {
5695         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5696
5697         if (global_rsv == block_rsv ||
5698             block_rsv->space_info != global_rsv->space_info)
5699                 global_rsv = NULL;
5700         block_rsv_release_bytes(fs_info, block_rsv, global_rsv, num_bytes, NULL);
5701 }
5702
5703 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
5704 {
5705         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
5706         struct btrfs_space_info *sinfo = block_rsv->space_info;
5707         u64 num_bytes;
5708
5709         /*
5710          * The global block rsv is based on the size of the extent tree, the
5711          * checksum tree and the root tree.  If the fs is empty we want to set
5712          * it to a minimal amount for safety.
5713          */
5714         num_bytes = btrfs_root_used(&fs_info->extent_root->root_item) +
5715                 btrfs_root_used(&fs_info->csum_root->root_item) +
5716                 btrfs_root_used(&fs_info->tree_root->root_item);
5717         num_bytes = max_t(u64, num_bytes, SZ_16M);
5718
5719         spin_lock(&sinfo->lock);
5720         spin_lock(&block_rsv->lock);
5721
5722         block_rsv->size = min_t(u64, num_bytes, SZ_512M);
5723
5724         if (block_rsv->reserved < block_rsv->size) {
5725                 num_bytes = btrfs_space_info_used(sinfo, true);
5726                 if (sinfo->total_bytes > num_bytes) {
5727                         num_bytes = sinfo->total_bytes - num_bytes;
5728                         num_bytes = min(num_bytes,
5729                                         block_rsv->size - block_rsv->reserved);
5730                         block_rsv->reserved += num_bytes;
5731                         sinfo->bytes_may_use += num_bytes;
5732                         trace_btrfs_space_reservation(fs_info, "space_info",
5733                                                       sinfo->flags, num_bytes,
5734                                                       1);
5735                 }
5736         } else if (block_rsv->reserved > block_rsv->size) {
5737                 num_bytes = block_rsv->reserved - block_rsv->size;
5738                 sinfo->bytes_may_use -= num_bytes;
5739                 trace_btrfs_space_reservation(fs_info, "space_info",
5740                                       sinfo->flags, num_bytes, 0);
5741                 block_rsv->reserved = block_rsv->size;
5742         }
5743
5744         if (block_rsv->reserved == block_rsv->size)
5745                 block_rsv->full = 1;
5746         else
5747                 block_rsv->full = 0;
5748
5749         spin_unlock(&block_rsv->lock);
5750         spin_unlock(&sinfo->lock);
5751 }
5752
5753 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
5754 {
5755         struct btrfs_space_info *space_info;
5756
5757         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
5758         fs_info->chunk_block_rsv.space_info = space_info;
5759
5760         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5761         fs_info->global_block_rsv.space_info = space_info;
5762         fs_info->trans_block_rsv.space_info = space_info;
5763         fs_info->empty_block_rsv.space_info = space_info;
5764         fs_info->delayed_block_rsv.space_info = space_info;
5765
5766         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
5767         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
5768         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
5769         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
5770         if (fs_info->quota_root)
5771                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
5772         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
5773
5774         update_global_block_rsv(fs_info);
5775 }
5776
5777 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
5778 {
5779         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
5780                                 (u64)-1, NULL);
5781         WARN_ON(fs_info->trans_block_rsv.size > 0);
5782         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
5783         WARN_ON(fs_info->chunk_block_rsv.size > 0);
5784         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
5785         WARN_ON(fs_info->delayed_block_rsv.size > 0);
5786         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
5787 }
5788
5789
5790 /*
5791  * To be called after all the new block groups attached to the transaction
5792  * handle have been created (btrfs_create_pending_block_groups()).
5793  */
5794 void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans)
5795 {
5796         struct btrfs_fs_info *fs_info = trans->fs_info;
5797
5798         if (!trans->chunk_bytes_reserved)
5799                 return;
5800
5801         WARN_ON_ONCE(!list_empty(&trans->new_bgs));
5802
5803         block_rsv_release_bytes(fs_info, &fs_info->chunk_block_rsv, NULL,
5804                                 trans->chunk_bytes_reserved, NULL);
5805         trans->chunk_bytes_reserved = 0;
5806 }
5807
5808 /*
5809  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
5810  * root: the root of the parent directory
5811  * rsv: block reservation
5812  * items: the number of items that we need do reservation
5813  * qgroup_reserved: used to return the reserved size in qgroup
5814  *
5815  * This function is used to reserve the space for snapshot/subvolume
5816  * creation and deletion. Those operations are different with the
5817  * common file/directory operations, they change two fs/file trees
5818  * and root tree, the number of items that the qgroup reserves is
5819  * different with the free space reservation. So we can not use
5820  * the space reservation mechanism in start_transaction().
5821  */
5822 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
5823                                      struct btrfs_block_rsv *rsv,
5824                                      int items,
5825                                      bool use_global_rsv)
5826 {
5827         u64 num_bytes;
5828         int ret;
5829         struct btrfs_fs_info *fs_info = root->fs_info;
5830         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5831
5832         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)) {
5833                 /* One for parent inode, two for dir entries */
5834                 num_bytes = 3 * fs_info->nodesize;
5835                 ret = btrfs_qgroup_reserve_meta_prealloc(root, num_bytes, true);
5836                 if (ret)
5837                         return ret;
5838         } else {
5839                 num_bytes = 0;
5840         }
5841
5842         num_bytes = btrfs_calc_trans_metadata_size(fs_info, items);
5843         rsv->space_info = __find_space_info(fs_info,
5844                                             BTRFS_BLOCK_GROUP_METADATA);
5845         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
5846                                   BTRFS_RESERVE_FLUSH_ALL);
5847
5848         if (ret == -ENOSPC && use_global_rsv)
5849                 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes, 1);
5850
5851         if (ret && num_bytes)
5852                 btrfs_qgroup_free_meta_prealloc(root, num_bytes);
5853
5854         return ret;
5855 }
5856
5857 void btrfs_subvolume_release_metadata(struct btrfs_fs_info *fs_info,
5858                                       struct btrfs_block_rsv *rsv)
5859 {
5860         btrfs_block_rsv_release(fs_info, rsv, (u64)-1);
5861 }
5862
5863 static void btrfs_calculate_inode_block_rsv_size(struct btrfs_fs_info *fs_info,
5864                                                  struct btrfs_inode *inode)
5865 {
5866         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
5867         u64 reserve_size = 0;
5868         u64 qgroup_rsv_size = 0;
5869         u64 csum_leaves;
5870         unsigned outstanding_extents;
5871
5872         lockdep_assert_held(&inode->lock);
5873         outstanding_extents = inode->outstanding_extents;
5874         if (outstanding_extents)
5875                 reserve_size = btrfs_calc_trans_metadata_size(fs_info,
5876                                                 outstanding_extents + 1);
5877         csum_leaves = btrfs_csum_bytes_to_leaves(fs_info,
5878                                                  inode->csum_bytes);
5879         reserve_size += btrfs_calc_trans_metadata_size(fs_info,
5880                                                        csum_leaves);
5881         /*
5882          * For qgroup rsv, the calculation is very simple:
5883          * account one nodesize for each outstanding extent
5884          *
5885          * This is overestimating in most cases.
5886          */
5887         qgroup_rsv_size = outstanding_extents * fs_info->nodesize;
5888
5889         spin_lock(&block_rsv->lock);
5890         block_rsv->size = reserve_size;
5891         block_rsv->qgroup_rsv_size = qgroup_rsv_size;
5892         spin_unlock(&block_rsv->lock);
5893 }
5894
5895 int btrfs_delalloc_reserve_metadata(struct btrfs_inode *inode, u64 num_bytes)
5896 {
5897         struct btrfs_fs_info *fs_info = inode->root->fs_info;
5898         unsigned nr_extents;
5899         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
5900         int ret = 0;
5901         bool delalloc_lock = true;
5902
5903         /* If we are a free space inode we need to not flush since we will be in
5904          * the middle of a transaction commit.  We also don't need the delalloc
5905          * mutex since we won't race with anybody.  We need this mostly to make
5906          * lockdep shut its filthy mouth.
5907          *
5908          * If we have a transaction open (can happen if we call truncate_block
5909          * from truncate), then we need FLUSH_LIMIT so we don't deadlock.
5910          */
5911         if (btrfs_is_free_space_inode(inode)) {
5912                 flush = BTRFS_RESERVE_NO_FLUSH;
5913                 delalloc_lock = false;
5914         } else {
5915                 if (current->journal_info)
5916                         flush = BTRFS_RESERVE_FLUSH_LIMIT;
5917
5918                 if (btrfs_transaction_in_commit(fs_info))
5919                         schedule_timeout(1);
5920         }
5921
5922         if (delalloc_lock)
5923                 mutex_lock(&inode->delalloc_mutex);
5924
5925         num_bytes = ALIGN(num_bytes, fs_info->sectorsize);
5926
5927         /* Add our new extents and calculate the new rsv size. */
5928         spin_lock(&inode->lock);
5929         nr_extents = count_max_extents(num_bytes);
5930         btrfs_mod_outstanding_extents(inode, nr_extents);
5931         inode->csum_bytes += num_bytes;
5932         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
5933         spin_unlock(&inode->lock);
5934
5935         ret = btrfs_inode_rsv_refill(inode, flush);
5936         if (unlikely(ret))
5937                 goto out_fail;
5938
5939         if (delalloc_lock)
5940                 mutex_unlock(&inode->delalloc_mutex);
5941         return 0;
5942
5943 out_fail:
5944         spin_lock(&inode->lock);
5945         nr_extents = count_max_extents(num_bytes);
5946         btrfs_mod_outstanding_extents(inode, -nr_extents);
5947         inode->csum_bytes -= num_bytes;
5948         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
5949         spin_unlock(&inode->lock);
5950
5951         btrfs_inode_rsv_release(inode, true);
5952         if (delalloc_lock)
5953                 mutex_unlock(&inode->delalloc_mutex);
5954         return ret;
5955 }
5956
5957 /**
5958  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
5959  * @inode: the inode to release the reservation for.
5960  * @num_bytes: the number of bytes we are releasing.
5961  * @qgroup_free: free qgroup reservation or convert it to per-trans reservation
5962  *
5963  * This will release the metadata reservation for an inode.  This can be called
5964  * once we complete IO for a given set of bytes to release their metadata
5965  * reservations, or on error for the same reason.
5966  */
5967 void btrfs_delalloc_release_metadata(struct btrfs_inode *inode, u64 num_bytes,
5968                                      bool qgroup_free)
5969 {
5970         struct btrfs_fs_info *fs_info = inode->root->fs_info;
5971
5972         num_bytes = ALIGN(num_bytes, fs_info->sectorsize);
5973         spin_lock(&inode->lock);
5974         inode->csum_bytes -= num_bytes;
5975         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
5976         spin_unlock(&inode->lock);
5977
5978         if (btrfs_is_testing(fs_info))
5979                 return;
5980
5981         btrfs_inode_rsv_release(inode, qgroup_free);
5982 }
5983
5984 /**
5985  * btrfs_delalloc_release_extents - release our outstanding_extents
5986  * @inode: the inode to balance the reservation for.
5987  * @num_bytes: the number of bytes we originally reserved with
5988  * @qgroup_free: do we need to free qgroup meta reservation or convert them.
5989  *
5990  * When we reserve space we increase outstanding_extents for the extents we may
5991  * add.  Once we've set the range as delalloc or created our ordered extents we
5992  * have outstanding_extents to track the real usage, so we use this to free our
5993  * temporarily tracked outstanding_extents.  This _must_ be used in conjunction
5994  * with btrfs_delalloc_reserve_metadata.
5995  */
5996 void btrfs_delalloc_release_extents(struct btrfs_inode *inode, u64 num_bytes,
5997                                     bool qgroup_free)
5998 {
5999         struct btrfs_fs_info *fs_info = inode->root->fs_info;
6000         unsigned num_extents;
6001
6002         spin_lock(&inode->lock);
6003         num_extents = count_max_extents(num_bytes);
6004         btrfs_mod_outstanding_extents(inode, -num_extents);
6005         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6006         spin_unlock(&inode->lock);
6007
6008         if (btrfs_is_testing(fs_info))
6009                 return;
6010
6011         btrfs_inode_rsv_release(inode, qgroup_free);
6012 }
6013
6014 /**
6015  * btrfs_delalloc_reserve_space - reserve data and metadata space for
6016  * delalloc
6017  * @inode: inode we're writing to
6018  * @start: start range we are writing to
6019  * @len: how long the range we are writing to
6020  * @reserved: mandatory parameter, record actually reserved qgroup ranges of
6021  *            current reservation.
6022  *
6023  * This will do the following things
6024  *
6025  * o reserve space in data space info for num bytes
6026  *   and reserve precious corresponding qgroup space
6027  *   (Done in check_data_free_space)
6028  *
6029  * o reserve space for metadata space, based on the number of outstanding
6030  *   extents and how much csums will be needed
6031  *   also reserve metadata space in a per root over-reserve method.
6032  * o add to the inodes->delalloc_bytes
6033  * o add it to the fs_info's delalloc inodes list.
6034  *   (Above 3 all done in delalloc_reserve_metadata)
6035  *
6036  * Return 0 for success
6037  * Return <0 for error(-ENOSPC or -EQUOT)
6038  */
6039 int btrfs_delalloc_reserve_space(struct inode *inode,
6040                         struct extent_changeset **reserved, u64 start, u64 len)
6041 {
6042         int ret;
6043
6044         ret = btrfs_check_data_free_space(inode, reserved, start, len);
6045         if (ret < 0)
6046                 return ret;
6047         ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode), len);
6048         if (ret < 0)
6049                 btrfs_free_reserved_data_space(inode, *reserved, start, len);
6050         return ret;
6051 }
6052
6053 /**
6054  * btrfs_delalloc_release_space - release data and metadata space for delalloc
6055  * @inode: inode we're releasing space for
6056  * @start: start position of the space already reserved
6057  * @len: the len of the space already reserved
6058  * @release_bytes: the len of the space we consumed or didn't use
6059  *
6060  * This function will release the metadata space that was not used and will
6061  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
6062  * list if there are no delalloc bytes left.
6063  * Also it will handle the qgroup reserved space.
6064  */
6065 void btrfs_delalloc_release_space(struct inode *inode,
6066                                   struct extent_changeset *reserved,
6067                                   u64 start, u64 len, bool qgroup_free)
6068 {
6069         btrfs_delalloc_release_metadata(BTRFS_I(inode), len, qgroup_free);
6070         btrfs_free_reserved_data_space(inode, reserved, start, len);
6071 }
6072
6073 static int update_block_group(struct btrfs_trans_handle *trans,
6074                               struct btrfs_fs_info *info, u64 bytenr,
6075                               u64 num_bytes, int alloc)
6076 {
6077         struct btrfs_block_group_cache *cache = NULL;
6078         u64 total = num_bytes;
6079         u64 old_val;
6080         u64 byte_in_group;
6081         int factor;
6082
6083         /* block accounting for super block */
6084         spin_lock(&info->delalloc_root_lock);
6085         old_val = btrfs_super_bytes_used(info->super_copy);
6086         if (alloc)
6087                 old_val += num_bytes;
6088         else
6089                 old_val -= num_bytes;
6090         btrfs_set_super_bytes_used(info->super_copy, old_val);
6091         spin_unlock(&info->delalloc_root_lock);
6092
6093         while (total) {
6094                 cache = btrfs_lookup_block_group(info, bytenr);
6095                 if (!cache)
6096                         return -ENOENT;
6097                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
6098                                     BTRFS_BLOCK_GROUP_RAID1 |
6099                                     BTRFS_BLOCK_GROUP_RAID10))
6100                         factor = 2;
6101                 else
6102                         factor = 1;
6103                 /*
6104                  * If this block group has free space cache written out, we
6105                  * need to make sure to load it if we are removing space.  This
6106                  * is because we need the unpinning stage to actually add the
6107                  * space back to the block group, otherwise we will leak space.
6108                  */
6109                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
6110                         cache_block_group(cache, 1);
6111
6112                 byte_in_group = bytenr - cache->key.objectid;
6113                 WARN_ON(byte_in_group > cache->key.offset);
6114
6115                 spin_lock(&cache->space_info->lock);
6116                 spin_lock(&cache->lock);
6117
6118                 if (btrfs_test_opt(info, SPACE_CACHE) &&
6119                     cache->disk_cache_state < BTRFS_DC_CLEAR)
6120                         cache->disk_cache_state = BTRFS_DC_CLEAR;
6121
6122                 old_val = btrfs_block_group_used(&cache->item);
6123                 num_bytes = min(total, cache->key.offset - byte_in_group);
6124                 if (alloc) {
6125                         old_val += num_bytes;
6126                         btrfs_set_block_group_used(&cache->item, old_val);
6127                         cache->reserved -= num_bytes;
6128                         cache->space_info->bytes_reserved -= num_bytes;
6129                         cache->space_info->bytes_used += num_bytes;
6130                         cache->space_info->disk_used += num_bytes * factor;
6131                         spin_unlock(&cache->lock);
6132                         spin_unlock(&cache->space_info->lock);
6133                 } else {
6134                         old_val -= num_bytes;
6135                         btrfs_set_block_group_used(&cache->item, old_val);
6136                         cache->pinned += num_bytes;
6137                         cache->space_info->bytes_pinned += num_bytes;
6138                         cache->space_info->bytes_used -= num_bytes;
6139                         cache->space_info->disk_used -= num_bytes * factor;
6140                         spin_unlock(&cache->lock);
6141                         spin_unlock(&cache->space_info->lock);
6142
6143                         trace_btrfs_space_reservation(info, "pinned",
6144                                                       cache->space_info->flags,
6145                                                       num_bytes, 1);
6146                         percpu_counter_add_batch(&cache->space_info->total_bytes_pinned,
6147                                            num_bytes,
6148                                            BTRFS_TOTAL_BYTES_PINNED_BATCH);
6149                         set_extent_dirty(info->pinned_extents,
6150                                          bytenr, bytenr + num_bytes - 1,
6151                                          GFP_NOFS | __GFP_NOFAIL);
6152                 }
6153
6154                 spin_lock(&trans->transaction->dirty_bgs_lock);
6155                 if (list_empty(&cache->dirty_list)) {
6156                         list_add_tail(&cache->dirty_list,
6157                                       &trans->transaction->dirty_bgs);
6158                         trans->transaction->num_dirty_bgs++;
6159                         btrfs_get_block_group(cache);
6160                 }
6161                 spin_unlock(&trans->transaction->dirty_bgs_lock);
6162
6163                 /*
6164                  * No longer have used bytes in this block group, queue it for
6165                  * deletion. We do this after adding the block group to the
6166                  * dirty list to avoid races between cleaner kthread and space
6167                  * cache writeout.
6168                  */
6169                 if (!alloc && old_val == 0)
6170                         btrfs_mark_bg_unused(cache);
6171
6172                 btrfs_put_block_group(cache);
6173                 total -= num_bytes;
6174                 bytenr += num_bytes;
6175         }
6176         return 0;
6177 }
6178
6179 static u64 first_logical_byte(struct btrfs_fs_info *fs_info, u64 search_start)
6180 {
6181         struct btrfs_block_group_cache *cache;
6182         u64 bytenr;
6183
6184         spin_lock(&fs_info->block_group_cache_lock);
6185         bytenr = fs_info->first_logical_byte;
6186         spin_unlock(&fs_info->block_group_cache_lock);
6187
6188         if (bytenr < (u64)-1)
6189                 return bytenr;
6190
6191         cache = btrfs_lookup_first_block_group(fs_info, search_start);
6192         if (!cache)
6193                 return 0;
6194
6195         bytenr = cache->key.objectid;
6196         btrfs_put_block_group(cache);
6197
6198         return bytenr;
6199 }
6200
6201 static int pin_down_extent(struct btrfs_fs_info *fs_info,
6202                            struct btrfs_block_group_cache *cache,
6203                            u64 bytenr, u64 num_bytes, int reserved)
6204 {
6205         spin_lock(&cache->space_info->lock);
6206         spin_lock(&cache->lock);
6207         cache->pinned += num_bytes;
6208         cache->space_info->bytes_pinned += num_bytes;
6209         if (reserved) {
6210                 cache->reserved -= num_bytes;
6211                 cache->space_info->bytes_reserved -= num_bytes;
6212         }
6213         spin_unlock(&cache->lock);
6214         spin_unlock(&cache->space_info->lock);
6215
6216         trace_btrfs_space_reservation(fs_info, "pinned",
6217                                       cache->space_info->flags, num_bytes, 1);
6218         percpu_counter_add_batch(&cache->space_info->total_bytes_pinned,
6219                     num_bytes, BTRFS_TOTAL_BYTES_PINNED_BATCH);
6220         set_extent_dirty(fs_info->pinned_extents, bytenr,
6221                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
6222         return 0;
6223 }
6224
6225 /*
6226  * this function must be called within transaction
6227  */
6228 int btrfs_pin_extent(struct btrfs_fs_info *fs_info,
6229                      u64 bytenr, u64 num_bytes, int reserved)
6230 {
6231         struct btrfs_block_group_cache *cache;
6232
6233         cache = btrfs_lookup_block_group(fs_info, bytenr);
6234         BUG_ON(!cache); /* Logic error */
6235
6236         pin_down_extent(fs_info, cache, bytenr, num_bytes, reserved);
6237
6238         btrfs_put_block_group(cache);
6239         return 0;
6240 }
6241
6242 /*
6243  * this function must be called within transaction
6244  */
6245 int btrfs_pin_extent_for_log_replay(struct btrfs_fs_info *fs_info,
6246                                     u64 bytenr, u64 num_bytes)
6247 {
6248         struct btrfs_block_group_cache *cache;
6249         int ret;
6250
6251         cache = btrfs_lookup_block_group(fs_info, bytenr);
6252         if (!cache)
6253                 return -EINVAL;
6254
6255         /*
6256          * pull in the free space cache (if any) so that our pin
6257          * removes the free space from the cache.  We have load_only set
6258          * to one because the slow code to read in the free extents does check
6259          * the pinned extents.
6260          */
6261         cache_block_group(cache, 1);
6262
6263         pin_down_extent(fs_info, cache, bytenr, num_bytes, 0);
6264
6265         /* remove us from the free space cache (if we're there at all) */
6266         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
6267         btrfs_put_block_group(cache);
6268         return ret;
6269 }
6270
6271 static int __exclude_logged_extent(struct btrfs_fs_info *fs_info,
6272                                    u64 start, u64 num_bytes)
6273 {
6274         int ret;
6275         struct btrfs_block_group_cache *block_group;
6276         struct btrfs_caching_control *caching_ctl;
6277
6278         block_group = btrfs_lookup_block_group(fs_info, start);
6279         if (!block_group)
6280                 return -EINVAL;
6281
6282         cache_block_group(block_group, 0);
6283         caching_ctl = get_caching_control(block_group);
6284
6285         if (!caching_ctl) {
6286                 /* Logic error */
6287                 BUG_ON(!block_group_cache_done(block_group));
6288                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6289         } else {
6290                 mutex_lock(&caching_ctl->mutex);
6291
6292                 if (start >= caching_ctl->progress) {
6293                         ret = add_excluded_extent(fs_info, start, num_bytes);
6294                 } else if (start + num_bytes <= caching_ctl->progress) {
6295                         ret = btrfs_remove_free_space(block_group,
6296                                                       start, num_bytes);
6297                 } else {
6298                         num_bytes = caching_ctl->progress - start;
6299                         ret = btrfs_remove_free_space(block_group,
6300                                                       start, num_bytes);
6301                         if (ret)
6302                                 goto out_lock;
6303
6304                         num_bytes = (start + num_bytes) -
6305                                 caching_ctl->progress;
6306                         start = caching_ctl->progress;
6307                         ret = add_excluded_extent(fs_info, start, num_bytes);
6308                 }
6309 out_lock:
6310                 mutex_unlock(&caching_ctl->mutex);
6311                 put_caching_control(caching_ctl);
6312         }
6313         btrfs_put_block_group(block_group);
6314         return ret;
6315 }
6316
6317 int btrfs_exclude_logged_extents(struct btrfs_fs_info *fs_info,
6318                                  struct extent_buffer *eb)
6319 {
6320         struct btrfs_file_extent_item *item;
6321         struct btrfs_key key;
6322         int found_type;
6323         int i;
6324         int ret = 0;
6325
6326         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS))
6327                 return 0;
6328
6329         for (i = 0; i < btrfs_header_nritems(eb); i++) {
6330                 btrfs_item_key_to_cpu(eb, &key, i);
6331                 if (key.type != BTRFS_EXTENT_DATA_KEY)
6332                         continue;
6333                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
6334                 found_type = btrfs_file_extent_type(eb, item);
6335                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
6336                         continue;
6337                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
6338                         continue;
6339                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
6340                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
6341                 ret = __exclude_logged_extent(fs_info, key.objectid, key.offset);
6342                 if (ret)
6343                         break;
6344         }
6345
6346         return ret;
6347 }
6348
6349 static void
6350 btrfs_inc_block_group_reservations(struct btrfs_block_group_cache *bg)
6351 {
6352         atomic_inc(&bg->reservations);
6353 }
6354
6355 void btrfs_dec_block_group_reservations(struct btrfs_fs_info *fs_info,
6356                                         const u64 start)
6357 {
6358         struct btrfs_block_group_cache *bg;
6359
6360         bg = btrfs_lookup_block_group(fs_info, start);
6361         ASSERT(bg);
6362         if (atomic_dec_and_test(&bg->reservations))
6363                 wake_up_var(&bg->reservations);
6364         btrfs_put_block_group(bg);
6365 }
6366
6367 void btrfs_wait_block_group_reservations(struct btrfs_block_group_cache *bg)
6368 {
6369         struct btrfs_space_info *space_info = bg->space_info;
6370
6371         ASSERT(bg->ro);
6372
6373         if (!(bg->flags & BTRFS_BLOCK_GROUP_DATA))
6374                 return;
6375
6376         /*
6377          * Our block group is read only but before we set it to read only,
6378          * some task might have had allocated an extent from it already, but it
6379          * has not yet created a respective ordered extent (and added it to a
6380          * root's list of ordered extents).
6381          * Therefore wait for any task currently allocating extents, since the
6382          * block group's reservations counter is incremented while a read lock
6383          * on the groups' semaphore is held and decremented after releasing
6384          * the read access on that semaphore and creating the ordered extent.
6385          */
6386         down_write(&space_info->groups_sem);
6387         up_write(&space_info->groups_sem);
6388
6389         wait_var_event(&bg->reservations, !atomic_read(&bg->reservations));
6390 }
6391
6392 /**
6393  * btrfs_add_reserved_bytes - update the block_group and space info counters
6394  * @cache:      The cache we are manipulating
6395  * @ram_bytes:  The number of bytes of file content, and will be same to
6396  *              @num_bytes except for the compress path.
6397  * @num_bytes:  The number of bytes in question
6398  * @delalloc:   The blocks are allocated for the delalloc write
6399  *
6400  * This is called by the allocator when it reserves space. If this is a
6401  * reservation and the block group has become read only we cannot make the
6402  * reservation and return -EAGAIN, otherwise this function always succeeds.
6403  */
6404 static int btrfs_add_reserved_bytes(struct btrfs_block_group_cache *cache,
6405                                     u64 ram_bytes, u64 num_bytes, int delalloc)
6406 {
6407         struct btrfs_space_info *space_info = cache->space_info;
6408         int ret = 0;
6409
6410         spin_lock(&space_info->lock);
6411         spin_lock(&cache->lock);
6412         if (cache->ro) {
6413                 ret = -EAGAIN;
6414         } else {
6415                 cache->reserved += num_bytes;
6416                 space_info->bytes_reserved += num_bytes;
6417
6418                 trace_btrfs_space_reservation(cache->fs_info,
6419                                 "space_info", space_info->flags,
6420                                 ram_bytes, 0);
6421                 space_info->bytes_may_use -= ram_bytes;
6422                 if (delalloc)
6423                         cache->delalloc_bytes += num_bytes;
6424         }
6425         spin_unlock(&cache->lock);
6426         spin_unlock(&space_info->lock);
6427         return ret;
6428 }
6429
6430 /**
6431  * btrfs_free_reserved_bytes - update the block_group and space info counters
6432  * @cache:      The cache we are manipulating
6433  * @num_bytes:  The number of bytes in question
6434  * @delalloc:   The blocks are allocated for the delalloc write
6435  *
6436  * This is called by somebody who is freeing space that was never actually used
6437  * on disk.  For example if you reserve some space for a new leaf in transaction
6438  * A and before transaction A commits you free that leaf, you call this with
6439  * reserve set to 0 in order to clear the reservation.
6440  */
6441
6442 static int btrfs_free_reserved_bytes(struct btrfs_block_group_cache *cache,
6443                                      u64 num_bytes, int delalloc)
6444 {
6445         struct btrfs_space_info *space_info = cache->space_info;
6446         int ret = 0;
6447
6448         spin_lock(&space_info->lock);
6449         spin_lock(&cache->lock);
6450         if (cache->ro)
6451                 space_info->bytes_readonly += num_bytes;
6452         cache->reserved -= num_bytes;
6453         space_info->bytes_reserved -= num_bytes;
6454
6455         if (delalloc)
6456                 cache->delalloc_bytes -= num_bytes;
6457         spin_unlock(&cache->lock);
6458         spin_unlock(&space_info->lock);
6459         return ret;
6460 }
6461 void btrfs_prepare_extent_commit(struct btrfs_fs_info *fs_info)
6462 {
6463         struct btrfs_caching_control *next;
6464         struct btrfs_caching_control *caching_ctl;
6465         struct btrfs_block_group_cache *cache;
6466
6467         down_write(&fs_info->commit_root_sem);
6468
6469         list_for_each_entry_safe(caching_ctl, next,
6470                                  &fs_info->caching_block_groups, list) {
6471                 cache = caching_ctl->block_group;
6472                 if (block_group_cache_done(cache)) {
6473                         cache->last_byte_to_unpin = (u64)-1;
6474                         list_del_init(&caching_ctl->list);
6475                         put_caching_control(caching_ctl);
6476                 } else {
6477                         cache->last_byte_to_unpin = caching_ctl->progress;
6478                 }
6479         }
6480
6481         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6482                 fs_info->pinned_extents = &fs_info->freed_extents[1];
6483         else
6484                 fs_info->pinned_extents = &fs_info->freed_extents[0];
6485
6486         up_write(&fs_info->commit_root_sem);
6487
6488         update_global_block_rsv(fs_info);
6489 }
6490
6491 /*
6492  * Returns the free cluster for the given space info and sets empty_cluster to
6493  * what it should be based on the mount options.
6494  */
6495 static struct btrfs_free_cluster *
6496 fetch_cluster_info(struct btrfs_fs_info *fs_info,
6497                    struct btrfs_space_info *space_info, u64 *empty_cluster)
6498 {
6499         struct btrfs_free_cluster *ret = NULL;
6500
6501         *empty_cluster = 0;
6502         if (btrfs_mixed_space_info(space_info))
6503                 return ret;
6504
6505         if (space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
6506                 ret = &fs_info->meta_alloc_cluster;
6507                 if (btrfs_test_opt(fs_info, SSD))
6508                         *empty_cluster = SZ_2M;
6509                 else
6510                         *empty_cluster = SZ_64K;
6511         } else if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) &&
6512                    btrfs_test_opt(fs_info, SSD_SPREAD)) {
6513                 *empty_cluster = SZ_2M;
6514                 ret = &fs_info->data_alloc_cluster;
6515         }
6516
6517         return ret;
6518 }
6519
6520 static int unpin_extent_range(struct btrfs_fs_info *fs_info,
6521                               u64 start, u64 end,
6522                               const bool return_free_space)
6523 {
6524         struct btrfs_block_group_cache *cache = NULL;
6525         struct btrfs_space_info *space_info;
6526         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
6527         struct btrfs_free_cluster *cluster = NULL;
6528         u64 len;
6529         u64 total_unpinned = 0;
6530         u64 empty_cluster = 0;
6531         bool readonly;
6532
6533         while (start <= end) {
6534                 readonly = false;
6535                 if (!cache ||
6536                     start >= cache->key.objectid + cache->key.offset) {
6537                         if (cache)
6538                                 btrfs_put_block_group(cache);
6539                         total_unpinned = 0;
6540                         cache = btrfs_lookup_block_group(fs_info, start);
6541                         BUG_ON(!cache); /* Logic error */
6542
6543                         cluster = fetch_cluster_info(fs_info,
6544                                                      cache->space_info,
6545                                                      &empty_cluster);
6546                         empty_cluster <<= 1;
6547                 }
6548
6549                 len = cache->key.objectid + cache->key.offset - start;
6550                 len = min(len, end + 1 - start);
6551
6552                 if (start < cache->last_byte_to_unpin) {
6553                         len = min(len, cache->last_byte_to_unpin - start);
6554                         if (return_free_space)
6555                                 btrfs_add_free_space(cache, start, len);
6556                 }
6557
6558                 start += len;
6559                 total_unpinned += len;
6560                 space_info = cache->space_info;
6561
6562                 /*
6563                  * If this space cluster has been marked as fragmented and we've
6564                  * unpinned enough in this block group to potentially allow a
6565                  * cluster to be created inside of it go ahead and clear the
6566                  * fragmented check.
6567                  */
6568                 if (cluster && cluster->fragmented &&
6569                     total_unpinned > empty_cluster) {
6570                         spin_lock(&cluster->lock);
6571                         cluster->fragmented = 0;
6572                         spin_unlock(&cluster->lock);
6573                 }
6574
6575                 spin_lock(&space_info->lock);
6576                 spin_lock(&cache->lock);
6577                 cache->pinned -= len;
6578                 space_info->bytes_pinned -= len;
6579
6580                 trace_btrfs_space_reservation(fs_info, "pinned",
6581                                               space_info->flags, len, 0);
6582                 space_info->max_extent_size = 0;
6583                 percpu_counter_add_batch(&space_info->total_bytes_pinned,
6584                             -len, BTRFS_TOTAL_BYTES_PINNED_BATCH);
6585                 if (cache->ro) {
6586                         space_info->bytes_readonly += len;
6587                         readonly = true;
6588                 }
6589                 spin_unlock(&cache->lock);
6590                 if (!readonly && return_free_space &&
6591                     global_rsv->space_info == space_info) {
6592                         u64 to_add = len;
6593
6594                         spin_lock(&global_rsv->lock);
6595                         if (!global_rsv->full) {
6596                                 to_add = min(len, global_rsv->size -
6597                                              global_rsv->reserved);
6598                                 global_rsv->reserved += to_add;
6599                                 space_info->bytes_may_use += to_add;
6600                                 if (global_rsv->reserved >= global_rsv->size)
6601                                         global_rsv->full = 1;
6602                                 trace_btrfs_space_reservation(fs_info,
6603                                                               "space_info",
6604                                                               space_info->flags,
6605                                                               to_add, 1);
6606                                 len -= to_add;
6607                         }
6608                         spin_unlock(&global_rsv->lock);
6609                         /* Add to any tickets we may have */
6610                         if (len)
6611                                 space_info_add_new_bytes(fs_info, space_info,
6612                                                          len);
6613                 }
6614                 spin_unlock(&space_info->lock);
6615         }
6616
6617         if (cache)
6618                 btrfs_put_block_group(cache);
6619         return 0;
6620 }
6621
6622 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans)
6623 {
6624         struct btrfs_fs_info *fs_info = trans->fs_info;
6625         struct btrfs_block_group_cache *block_group, *tmp;
6626         struct list_head *deleted_bgs;
6627         struct extent_io_tree *unpin;
6628         u64 start;
6629         u64 end;
6630         int ret;
6631
6632         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6633                 unpin = &fs_info->freed_extents[1];
6634         else
6635                 unpin = &fs_info->freed_extents[0];
6636
6637         while (!trans->aborted) {
6638                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
6639                 ret = find_first_extent_bit(unpin, 0, &start, &end,
6640                                             EXTENT_DIRTY, NULL);
6641                 if (ret) {
6642                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6643                         break;
6644                 }
6645
6646                 if (btrfs_test_opt(fs_info, DISCARD))
6647                         ret = btrfs_discard_extent(fs_info, start,
6648                                                    end + 1 - start, NULL);
6649
6650                 clear_extent_dirty(unpin, start, end);
6651                 unpin_extent_range(fs_info, start, end, true);
6652                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6653                 cond_resched();
6654         }
6655
6656         /*
6657          * Transaction is finished.  We don't need the lock anymore.  We
6658          * do need to clean up the block groups in case of a transaction
6659          * abort.
6660          */
6661         deleted_bgs = &trans->transaction->deleted_bgs;
6662         list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
6663                 u64 trimmed = 0;
6664
6665                 ret = -EROFS;
6666                 if (!trans->aborted)
6667                         ret = btrfs_discard_extent(fs_info,
6668                                                    block_group->key.objectid,
6669                                                    block_group->key.offset,
6670                                                    &trimmed);
6671
6672                 list_del_init(&block_group->bg_list);
6673                 btrfs_put_block_group_trimming(block_group);
6674                 btrfs_put_block_group(block_group);
6675
6676                 if (ret) {
6677                         const char *errstr = btrfs_decode_error(ret);
6678                         btrfs_warn(fs_info,
6679                            "discard failed while removing blockgroup: errno=%d %s",
6680                                    ret, errstr);
6681                 }
6682         }
6683
6684         return 0;
6685 }
6686
6687 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
6688                                struct btrfs_delayed_ref_node *node, u64 parent,
6689                                u64 root_objectid, u64 owner_objectid,
6690                                u64 owner_offset, int refs_to_drop,
6691                                struct btrfs_delayed_extent_op *extent_op)
6692 {
6693         struct btrfs_fs_info *info = trans->fs_info;
6694         struct btrfs_key key;
6695         struct btrfs_path *path;
6696         struct btrfs_root *extent_root = info->extent_root;
6697         struct extent_buffer *leaf;
6698         struct btrfs_extent_item *ei;
6699         struct btrfs_extent_inline_ref *iref;
6700         int ret;
6701         int is_data;
6702         int extent_slot = 0;
6703         int found_extent = 0;
6704         int num_to_del = 1;
6705         u32 item_size;
6706         u64 refs;
6707         u64 bytenr = node->bytenr;
6708         u64 num_bytes = node->num_bytes;
6709         int last_ref = 0;
6710         bool skinny_metadata = btrfs_fs_incompat(info, SKINNY_METADATA);
6711
6712         path = btrfs_alloc_path();
6713         if (!path)
6714                 return -ENOMEM;
6715
6716         path->reada = READA_FORWARD;
6717         path->leave_spinning = 1;
6718
6719         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
6720         BUG_ON(!is_data && refs_to_drop != 1);
6721
6722         if (is_data)
6723                 skinny_metadata = false;
6724
6725         ret = lookup_extent_backref(trans, path, &iref, bytenr, num_bytes,
6726                                     parent, root_objectid, owner_objectid,
6727                                     owner_offset);
6728         if (ret == 0) {
6729                 extent_slot = path->slots[0];
6730                 while (extent_slot >= 0) {
6731                         btrfs_item_key_to_cpu(path->nodes[0], &key,
6732                                               extent_slot);
6733                         if (key.objectid != bytenr)
6734                                 break;
6735                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
6736                             key.offset == num_bytes) {
6737                                 found_extent = 1;
6738                                 break;
6739                         }
6740                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
6741                             key.offset == owner_objectid) {
6742                                 found_extent = 1;
6743                                 break;
6744                         }
6745                         if (path->slots[0] - extent_slot > 5)
6746                                 break;
6747                         extent_slot--;
6748                 }
6749
6750                 if (!found_extent) {
6751                         BUG_ON(iref);
6752                         ret = remove_extent_backref(trans, path, NULL,
6753                                                     refs_to_drop,
6754                                                     is_data, &last_ref);
6755                         if (ret) {
6756                                 btrfs_abort_transaction(trans, ret);
6757                                 goto out;
6758                         }
6759                         btrfs_release_path(path);
6760                         path->leave_spinning = 1;
6761
6762                         key.objectid = bytenr;
6763                         key.type = BTRFS_EXTENT_ITEM_KEY;
6764                         key.offset = num_bytes;
6765
6766                         if (!is_data && skinny_metadata) {
6767                                 key.type = BTRFS_METADATA_ITEM_KEY;
6768                                 key.offset = owner_objectid;
6769                         }
6770
6771                         ret = btrfs_search_slot(trans, extent_root,
6772                                                 &key, path, -1, 1);
6773                         if (ret > 0 && skinny_metadata && path->slots[0]) {
6774                                 /*
6775                                  * Couldn't find our skinny metadata item,
6776                                  * see if we have ye olde extent item.
6777                                  */
6778                                 path->slots[0]--;
6779                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
6780                                                       path->slots[0]);
6781                                 if (key.objectid == bytenr &&
6782                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
6783                                     key.offset == num_bytes)
6784                                         ret = 0;
6785                         }
6786
6787                         if (ret > 0 && skinny_metadata) {
6788                                 skinny_metadata = false;
6789                                 key.objectid = bytenr;
6790                                 key.type = BTRFS_EXTENT_ITEM_KEY;
6791                                 key.offset = num_bytes;
6792                                 btrfs_release_path(path);
6793                                 ret = btrfs_search_slot(trans, extent_root,
6794                                                         &key, path, -1, 1);
6795                         }
6796
6797                         if (ret) {
6798                                 btrfs_err(info,
6799                                           "umm, got %d back from search, was looking for %llu",
6800                                           ret, bytenr);
6801                                 if (ret > 0)
6802                                         btrfs_print_leaf(path->nodes[0]);
6803                         }
6804                         if (ret < 0) {
6805                                 btrfs_abort_transaction(trans, ret);
6806                                 goto out;
6807                         }
6808                         extent_slot = path->slots[0];
6809                 }
6810         } else if (WARN_ON(ret == -ENOENT)) {
6811                 btrfs_print_leaf(path->nodes[0]);
6812                 btrfs_err(info,
6813                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
6814                         bytenr, parent, root_objectid, owner_objectid,
6815                         owner_offset);
6816                 btrfs_abort_transaction(trans, ret);
6817                 goto out;
6818         } else {
6819                 btrfs_abort_transaction(trans, ret);
6820                 goto out;
6821         }
6822
6823         leaf = path->nodes[0];
6824         item_size = btrfs_item_size_nr(leaf, extent_slot);
6825         if (unlikely(item_size < sizeof(*ei))) {
6826                 ret = -EINVAL;
6827                 btrfs_print_v0_err(info);
6828                 btrfs_abort_transaction(trans, ret);
6829                 goto out;
6830         }
6831         ei = btrfs_item_ptr(leaf, extent_slot,
6832                             struct btrfs_extent_item);
6833         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
6834             key.type == BTRFS_EXTENT_ITEM_KEY) {
6835                 struct btrfs_tree_block_info *bi;
6836                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
6837                 bi = (struct btrfs_tree_block_info *)(ei + 1);
6838                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
6839         }
6840
6841         refs = btrfs_extent_refs(leaf, ei);
6842         if (refs < refs_to_drop) {
6843                 btrfs_err(info,
6844                           "trying to drop %d refs but we only have %Lu for bytenr %Lu",
6845                           refs_to_drop, refs, bytenr);
6846                 ret = -EINVAL;
6847                 btrfs_abort_transaction(trans, ret);
6848                 goto out;
6849         }
6850         refs -= refs_to_drop;
6851
6852         if (refs > 0) {
6853                 if (extent_op)
6854                         __run_delayed_extent_op(extent_op, leaf, ei);
6855                 /*
6856                  * In the case of inline back ref, reference count will
6857                  * be updated by remove_extent_backref
6858                  */
6859                 if (iref) {
6860                         BUG_ON(!found_extent);
6861                 } else {
6862                         btrfs_set_extent_refs(leaf, ei, refs);
6863                         btrfs_mark_buffer_dirty(leaf);
6864                 }
6865                 if (found_extent) {
6866                         ret = remove_extent_backref(trans, path, iref,
6867                                                     refs_to_drop, is_data,
6868                                                     &last_ref);
6869                         if (ret) {
6870                                 btrfs_abort_transaction(trans, ret);
6871                                 goto out;
6872                         }
6873                 }
6874         } else {
6875                 if (found_extent) {
6876                         BUG_ON(is_data && refs_to_drop !=
6877                                extent_data_ref_count(path, iref));
6878                         if (iref) {
6879                                 BUG_ON(path->slots[0] != extent_slot);
6880                         } else {
6881                                 BUG_ON(path->slots[0] != extent_slot + 1);
6882                                 path->slots[0] = extent_slot;
6883                                 num_to_del = 2;
6884                         }
6885                 }
6886
6887                 last_ref = 1;
6888                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
6889                                       num_to_del);
6890                 if (ret) {
6891                         btrfs_abort_transaction(trans, ret);
6892                         goto out;
6893                 }
6894                 btrfs_release_path(path);
6895
6896                 if (is_data) {
6897                         ret = btrfs_del_csums(trans, info, bytenr, num_bytes);
6898                         if (ret) {
6899                                 btrfs_abort_transaction(trans, ret);
6900                                 goto out;
6901                         }
6902                 }
6903
6904                 ret = add_to_free_space_tree(trans, bytenr, num_bytes);
6905                 if (ret) {
6906                         btrfs_abort_transaction(trans, ret);
6907                         goto out;
6908                 }
6909
6910                 ret = update_block_group(trans, info, bytenr, num_bytes, 0);
6911                 if (ret) {
6912                         btrfs_abort_transaction(trans, ret);
6913                         goto out;
6914                 }
6915         }
6916         btrfs_release_path(path);
6917
6918 out:
6919         btrfs_free_path(path);
6920         return ret;
6921 }
6922
6923 /*
6924  * when we free an block, it is possible (and likely) that we free the last
6925  * delayed ref for that extent as well.  This searches the delayed ref tree for
6926  * a given extent, and if there are no other delayed refs to be processed, it
6927  * removes it from the tree.
6928  */
6929 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
6930                                       u64 bytenr)
6931 {
6932         struct btrfs_delayed_ref_head *head;
6933         struct btrfs_delayed_ref_root *delayed_refs;
6934         int ret = 0;
6935
6936         delayed_refs = &trans->transaction->delayed_refs;
6937         spin_lock(&delayed_refs->lock);
6938         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
6939         if (!head)
6940                 goto out_delayed_unlock;
6941
6942         spin_lock(&head->lock);
6943         if (!RB_EMPTY_ROOT(&head->ref_tree))
6944                 goto out;
6945
6946         if (head->extent_op) {
6947                 if (!head->must_insert_reserved)
6948                         goto out;
6949                 btrfs_free_delayed_extent_op(head->extent_op);
6950                 head->extent_op = NULL;
6951         }
6952
6953         /*
6954          * waiting for the lock here would deadlock.  If someone else has it
6955          * locked they are already in the process of dropping it anyway
6956          */
6957         if (!mutex_trylock(&head->mutex))
6958                 goto out;
6959
6960         /*
6961          * at this point we have a head with no other entries.  Go
6962          * ahead and process it.
6963          */
6964         rb_erase(&head->href_node, &delayed_refs->href_root);
6965         RB_CLEAR_NODE(&head->href_node);
6966         atomic_dec(&delayed_refs->num_entries);
6967
6968         /*
6969          * we don't take a ref on the node because we're removing it from the
6970          * tree, so we just steal the ref the tree was holding.
6971          */
6972         delayed_refs->num_heads--;
6973         if (head->processing == 0)
6974                 delayed_refs->num_heads_ready--;
6975         head->processing = 0;
6976         spin_unlock(&head->lock);
6977         spin_unlock(&delayed_refs->lock);
6978
6979         BUG_ON(head->extent_op);
6980         if (head->must_insert_reserved)
6981                 ret = 1;
6982
6983         mutex_unlock(&head->mutex);
6984         btrfs_put_delayed_ref_head(head);
6985         return ret;
6986 out:
6987         spin_unlock(&head->lock);
6988
6989 out_delayed_unlock:
6990         spin_unlock(&delayed_refs->lock);
6991         return 0;
6992 }
6993
6994 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
6995                            struct btrfs_root *root,
6996                            struct extent_buffer *buf,
6997                            u64 parent, int last_ref)
6998 {
6999         struct btrfs_fs_info *fs_info = root->fs_info;
7000         int pin = 1;
7001         int ret;
7002
7003         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7004                 int old_ref_mod, new_ref_mod;
7005
7006                 btrfs_ref_tree_mod(root, buf->start, buf->len, parent,
7007                                    root->root_key.objectid,
7008                                    btrfs_header_level(buf), 0,
7009                                    BTRFS_DROP_DELAYED_REF);
7010                 ret = btrfs_add_delayed_tree_ref(trans, buf->start,
7011                                                  buf->len, parent,
7012                                                  root->root_key.objectid,
7013                                                  btrfs_header_level(buf),
7014                                                  BTRFS_DROP_DELAYED_REF, NULL,
7015                                                  &old_ref_mod, &new_ref_mod);
7016                 BUG_ON(ret); /* -ENOMEM */
7017                 pin = old_ref_mod >= 0 && new_ref_mod < 0;
7018         }
7019
7020         if (last_ref && btrfs_header_generation(buf) == trans->transid) {
7021                 struct btrfs_block_group_cache *cache;
7022
7023                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7024                         ret = check_ref_cleanup(trans, buf->start);
7025                         if (!ret)
7026                                 goto out;
7027                 }
7028
7029                 pin = 0;
7030                 cache = btrfs_lookup_block_group(fs_info, buf->start);
7031
7032                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
7033                         pin_down_extent(fs_info, cache, buf->start,
7034                                         buf->len, 1);
7035                         btrfs_put_block_group(cache);
7036                         goto out;
7037                 }
7038
7039                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
7040
7041                 btrfs_add_free_space(cache, buf->start, buf->len);
7042                 btrfs_free_reserved_bytes(cache, buf->len, 0);
7043                 btrfs_put_block_group(cache);
7044                 trace_btrfs_reserved_extent_free(fs_info, buf->start, buf->len);
7045         }
7046 out:
7047         if (pin)
7048                 add_pinned_bytes(fs_info, buf->len, true,
7049                                  root->root_key.objectid);
7050
7051         if (last_ref) {
7052                 /*
7053                  * Deleting the buffer, clear the corrupt flag since it doesn't
7054                  * matter anymore.
7055                  */
7056                 clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
7057         }
7058 }
7059
7060 /* Can return -ENOMEM */
7061 int btrfs_free_extent(struct btrfs_trans_handle *trans,
7062                       struct btrfs_root *root,
7063                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
7064                       u64 owner, u64 offset)
7065 {
7066         struct btrfs_fs_info *fs_info = root->fs_info;
7067         int old_ref_mod, new_ref_mod;
7068         int ret;
7069
7070         if (btrfs_is_testing(fs_info))
7071                 return 0;
7072
7073         if (root_objectid != BTRFS_TREE_LOG_OBJECTID)
7074                 btrfs_ref_tree_mod(root, bytenr, num_bytes, parent,
7075                                    root_objectid, owner, offset,
7076                                    BTRFS_DROP_DELAYED_REF);
7077
7078         /*
7079          * tree log blocks never actually go into the extent allocation
7080          * tree, just update pinning info and exit early.
7081          */
7082         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
7083                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
7084                 /* unlocks the pinned mutex */
7085                 btrfs_pin_extent(fs_info, bytenr, num_bytes, 1);
7086                 old_ref_mod = new_ref_mod = 0;
7087                 ret = 0;
7088         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
7089                 ret = btrfs_add_delayed_tree_ref(trans, bytenr,
7090                                                  num_bytes, parent,
7091                                                  root_objectid, (int)owner,
7092                                                  BTRFS_DROP_DELAYED_REF, NULL,
7093                                                  &old_ref_mod, &new_ref_mod);
7094         } else {
7095                 ret = btrfs_add_delayed_data_ref(trans, bytenr,
7096                                                  num_bytes, parent,
7097                                                  root_objectid, owner, offset,
7098                                                  0, BTRFS_DROP_DELAYED_REF,
7099                                                  &old_ref_mod, &new_ref_mod);
7100         }
7101
7102         if (ret == 0 && old_ref_mod >= 0 && new_ref_mod < 0) {
7103                 bool metadata = owner < BTRFS_FIRST_FREE_OBJECTID;
7104
7105                 add_pinned_bytes(fs_info, num_bytes, metadata, root_objectid);
7106         }
7107
7108         return ret;
7109 }
7110
7111 /*
7112  * when we wait for progress in the block group caching, its because
7113  * our allocation attempt failed at least once.  So, we must sleep
7114  * and let some progress happen before we try again.
7115  *
7116  * This function will sleep at least once waiting for new free space to
7117  * show up, and then it will check the block group free space numbers
7118  * for our min num_bytes.  Another option is to have it go ahead
7119  * and look in the rbtree for a free extent of a given size, but this
7120  * is a good start.
7121  *
7122  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
7123  * any of the information in this block group.
7124  */
7125 static noinline void
7126 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
7127                                 u64 num_bytes)
7128 {
7129         struct btrfs_caching_control *caching_ctl;
7130
7131         caching_ctl = get_caching_control(cache);
7132         if (!caching_ctl)
7133                 return;
7134
7135         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
7136                    (cache->free_space_ctl->free_space >= num_bytes));
7137
7138         put_caching_control(caching_ctl);
7139 }
7140
7141 static noinline int
7142 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
7143 {
7144         struct btrfs_caching_control *caching_ctl;
7145         int ret = 0;
7146
7147         caching_ctl = get_caching_control(cache);
7148         if (!caching_ctl)
7149                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
7150
7151         wait_event(caching_ctl->wait, block_group_cache_done(cache));
7152         if (cache->cached == BTRFS_CACHE_ERROR)
7153                 ret = -EIO;
7154         put_caching_control(caching_ctl);
7155         return ret;
7156 }
7157
7158 enum btrfs_loop_type {
7159         LOOP_CACHING_NOWAIT = 0,
7160         LOOP_CACHING_WAIT = 1,
7161         LOOP_ALLOC_CHUNK = 2,
7162         LOOP_NO_EMPTY_SIZE = 3,
7163 };
7164
7165 static inline void
7166 btrfs_lock_block_group(struct btrfs_block_group_cache *cache,
7167                        int delalloc)
7168 {
7169         if (delalloc)
7170                 down_read(&cache->data_rwsem);
7171 }
7172
7173 static inline void
7174 btrfs_grab_block_group(struct btrfs_block_group_cache *cache,
7175                        int delalloc)
7176 {
7177         btrfs_get_block_group(cache);
7178         if (delalloc)
7179                 down_read(&cache->data_rwsem);
7180 }
7181
7182 static struct btrfs_block_group_cache *
7183 btrfs_lock_cluster(struct btrfs_block_group_cache *block_group,
7184                    struct btrfs_free_cluster *cluster,
7185                    int delalloc)
7186 {
7187         struct btrfs_block_group_cache *used_bg = NULL;
7188
7189         spin_lock(&cluster->refill_lock);
7190         while (1) {
7191                 used_bg = cluster->block_group;
7192                 if (!used_bg)
7193                         return NULL;
7194
7195                 if (used_bg == block_group)
7196                         return used_bg;
7197
7198                 btrfs_get_block_group(used_bg);
7199
7200                 if (!delalloc)
7201                         return used_bg;
7202
7203                 if (down_read_trylock(&used_bg->data_rwsem))
7204                         return used_bg;
7205
7206                 spin_unlock(&cluster->refill_lock);
7207
7208                 /* We should only have one-level nested. */
7209                 down_read_nested(&used_bg->data_rwsem, SINGLE_DEPTH_NESTING);
7210
7211                 spin_lock(&cluster->refill_lock);
7212                 if (used_bg == cluster->block_group)
7213                         return used_bg;
7214
7215                 up_read(&used_bg->data_rwsem);
7216                 btrfs_put_block_group(used_bg);
7217         }
7218 }
7219
7220 static inline void
7221 btrfs_release_block_group(struct btrfs_block_group_cache *cache,
7222                          int delalloc)
7223 {
7224         if (delalloc)
7225                 up_read(&cache->data_rwsem);
7226         btrfs_put_block_group(cache);
7227 }
7228
7229 /*
7230  * walks the btree of allocated extents and find a hole of a given size.
7231  * The key ins is changed to record the hole:
7232  * ins->objectid == start position
7233  * ins->flags = BTRFS_EXTENT_ITEM_KEY
7234  * ins->offset == the size of the hole.
7235  * Any available blocks before search_start are skipped.
7236  *
7237  * If there is no suitable free space, we will record the max size of
7238  * the free space extent currently.
7239  */
7240 static noinline int find_free_extent(struct btrfs_fs_info *fs_info,
7241                                 u64 ram_bytes, u64 num_bytes, u64 empty_size,
7242                                 u64 hint_byte, struct btrfs_key *ins,
7243                                 u64 flags, int delalloc)
7244 {
7245         int ret = 0;
7246         struct btrfs_root *root = fs_info->extent_root;
7247         struct btrfs_free_cluster *last_ptr = NULL;
7248         struct btrfs_block_group_cache *block_group = NULL;
7249         u64 search_start = 0;
7250         u64 max_extent_size = 0;
7251         u64 empty_cluster = 0;
7252         struct btrfs_space_info *space_info;
7253         int loop = 0;
7254         int index = btrfs_bg_flags_to_raid_index(flags);
7255         bool failed_cluster_refill = false;
7256         bool failed_alloc = false;
7257         bool use_cluster = true;
7258         bool have_caching_bg = false;
7259         bool orig_have_caching_bg = false;
7260         bool full_search = false;
7261
7262         WARN_ON(num_bytes < fs_info->sectorsize);
7263         ins->type = BTRFS_EXTENT_ITEM_KEY;
7264         ins->objectid = 0;
7265         ins->offset = 0;
7266
7267         trace_find_free_extent(fs_info, num_bytes, empty_size, flags);
7268
7269         space_info = __find_space_info(fs_info, flags);
7270         if (!space_info) {
7271                 btrfs_err(fs_info, "No space info for %llu", flags);
7272                 return -ENOSPC;
7273         }
7274
7275         /*
7276          * If our free space is heavily fragmented we may not be able to make
7277          * big contiguous allocations, so instead of doing the expensive search
7278          * for free space, simply return ENOSPC with our max_extent_size so we
7279          * can go ahead and search for a more manageable chunk.
7280          *
7281          * If our max_extent_size is large enough for our allocation simply
7282          * disable clustering since we will likely not be able to find enough
7283          * space to create a cluster and induce latency trying.
7284          */
7285         if (unlikely(space_info->max_extent_size)) {
7286                 spin_lock(&space_info->lock);
7287                 if (space_info->max_extent_size &&
7288                     num_bytes > space_info->max_extent_size) {
7289                         ins->offset = space_info->max_extent_size;
7290                         spin_unlock(&space_info->lock);
7291                         return -ENOSPC;
7292                 } else if (space_info->max_extent_size) {
7293                         use_cluster = false;
7294                 }
7295                 spin_unlock(&space_info->lock);
7296         }
7297
7298         last_ptr = fetch_cluster_info(fs_info, space_info, &empty_cluster);
7299         if (last_ptr) {
7300                 spin_lock(&last_ptr->lock);
7301                 if (last_ptr->block_group)
7302                         hint_byte = last_ptr->window_start;
7303                 if (last_ptr->fragmented) {
7304                         /*
7305                          * We still set window_start so we can keep track of the
7306                          * last place we found an allocation to try and save
7307                          * some time.
7308                          */
7309                         hint_byte = last_ptr->window_start;
7310                         use_cluster = false;
7311                 }
7312                 spin_unlock(&last_ptr->lock);
7313         }
7314
7315         search_start = max(search_start, first_logical_byte(fs_info, 0));
7316         search_start = max(search_start, hint_byte);
7317         if (search_start == hint_byte) {
7318                 block_group = btrfs_lookup_block_group(fs_info, search_start);
7319                 /*
7320                  * we don't want to use the block group if it doesn't match our
7321                  * allocation bits, or if its not cached.
7322                  *
7323                  * However if we are re-searching with an ideal block group
7324                  * picked out then we don't care that the block group is cached.
7325                  */
7326                 if (block_group && block_group_bits(block_group, flags) &&
7327                     block_group->cached != BTRFS_CACHE_NO) {
7328                         down_read(&space_info->groups_sem);
7329                         if (list_empty(&block_group->list) ||
7330                             block_group->ro) {
7331                                 /*
7332                                  * someone is removing this block group,
7333                                  * we can't jump into the have_block_group
7334                                  * target because our list pointers are not
7335                                  * valid
7336                                  */
7337                                 btrfs_put_block_group(block_group);
7338                                 up_read(&space_info->groups_sem);
7339                         } else {
7340                                 index = btrfs_bg_flags_to_raid_index(
7341                                                 block_group->flags);
7342                                 btrfs_lock_block_group(block_group, delalloc);
7343                                 goto have_block_group;
7344                         }
7345                 } else if (block_group) {
7346                         btrfs_put_block_group(block_group);
7347                 }
7348         }
7349 search:
7350         have_caching_bg = false;
7351         if (index == 0 || index == btrfs_bg_flags_to_raid_index(flags))
7352                 full_search = true;
7353         down_read(&space_info->groups_sem);
7354         list_for_each_entry(block_group, &space_info->block_groups[index],
7355                             list) {
7356                 u64 offset;
7357                 int cached;
7358
7359                 /* If the block group is read-only, we can skip it entirely. */
7360                 if (unlikely(block_group->ro))
7361                         continue;
7362
7363                 btrfs_grab_block_group(block_group, delalloc);
7364                 search_start = block_group->key.objectid;
7365
7366                 /*
7367                  * this can happen if we end up cycling through all the
7368                  * raid types, but we want to make sure we only allocate
7369                  * for the proper type.
7370                  */
7371                 if (!block_group_bits(block_group, flags)) {
7372                         u64 extra = BTRFS_BLOCK_GROUP_DUP |
7373                                 BTRFS_BLOCK_GROUP_RAID1 |
7374                                 BTRFS_BLOCK_GROUP_RAID5 |
7375                                 BTRFS_BLOCK_GROUP_RAID6 |
7376                                 BTRFS_BLOCK_GROUP_RAID10;
7377
7378                         /*
7379                          * if they asked for extra copies and this block group
7380                          * doesn't provide them, bail.  This does allow us to
7381                          * fill raid0 from raid1.
7382                          */
7383                         if ((flags & extra) && !(block_group->flags & extra))
7384                                 goto loop;
7385                 }
7386
7387 have_block_group:
7388                 cached = block_group_cache_done(block_group);
7389                 if (unlikely(!cached)) {
7390                         have_caching_bg = true;
7391                         ret = cache_block_group(block_group, 0);
7392                         BUG_ON(ret < 0);
7393                         ret = 0;
7394                 }
7395
7396                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
7397                         goto loop;
7398
7399                 /*
7400                  * Ok we want to try and use the cluster allocator, so
7401                  * lets look there
7402                  */
7403                 if (last_ptr && use_cluster) {
7404                         struct btrfs_block_group_cache *used_block_group;
7405                         unsigned long aligned_cluster;
7406                         /*
7407                          * the refill lock keeps out other
7408                          * people trying to start a new cluster
7409                          */
7410                         used_block_group = btrfs_lock_cluster(block_group,
7411                                                               last_ptr,
7412                                                               delalloc);
7413                         if (!used_block_group)
7414                                 goto refill_cluster;
7415
7416                         if (used_block_group != block_group &&
7417                             (used_block_group->ro ||
7418                              !block_group_bits(used_block_group, flags)))
7419                                 goto release_cluster;
7420
7421                         offset = btrfs_alloc_from_cluster(used_block_group,
7422                                                 last_ptr,
7423                                                 num_bytes,
7424                                                 used_block_group->key.objectid,
7425                                                 &max_extent_size);
7426                         if (offset) {
7427                                 /* we have a block, we're done */
7428                                 spin_unlock(&last_ptr->refill_lock);
7429                                 trace_btrfs_reserve_extent_cluster(
7430                                                 used_block_group,
7431                                                 search_start, num_bytes);
7432                                 if (used_block_group != block_group) {
7433                                         btrfs_release_block_group(block_group,
7434                                                                   delalloc);
7435                                         block_group = used_block_group;
7436                                 }
7437                                 goto checks;
7438                         }
7439
7440                         WARN_ON(last_ptr->block_group != used_block_group);
7441 release_cluster:
7442                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
7443                          * set up a new clusters, so lets just skip it
7444                          * and let the allocator find whatever block
7445                          * it can find.  If we reach this point, we
7446                          * will have tried the cluster allocator
7447                          * plenty of times and not have found
7448                          * anything, so we are likely way too
7449                          * fragmented for the clustering stuff to find
7450                          * anything.
7451                          *
7452                          * However, if the cluster is taken from the
7453                          * current block group, release the cluster
7454                          * first, so that we stand a better chance of
7455                          * succeeding in the unclustered
7456                          * allocation.  */
7457                         if (loop >= LOOP_NO_EMPTY_SIZE &&
7458                             used_block_group != block_group) {
7459                                 spin_unlock(&last_ptr->refill_lock);
7460                                 btrfs_release_block_group(used_block_group,
7461                                                           delalloc);
7462                                 goto unclustered_alloc;
7463                         }
7464
7465                         /*
7466                          * this cluster didn't work out, free it and
7467                          * start over
7468                          */
7469                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7470
7471                         if (used_block_group != block_group)
7472                                 btrfs_release_block_group(used_block_group,
7473                                                           delalloc);
7474 refill_cluster:
7475                         if (loop >= LOOP_NO_EMPTY_SIZE) {
7476                                 spin_unlock(&last_ptr->refill_lock);
7477                                 goto unclustered_alloc;
7478                         }
7479
7480                         aligned_cluster = max_t(unsigned long,
7481                                                 empty_cluster + empty_size,
7482                                               block_group->full_stripe_len);
7483
7484                         /* allocate a cluster in this block group */
7485                         ret = btrfs_find_space_cluster(fs_info, block_group,
7486                                                        last_ptr, search_start,
7487                                                        num_bytes,
7488                                                        aligned_cluster);
7489                         if (ret == 0) {
7490                                 /*
7491                                  * now pull our allocation out of this
7492                                  * cluster
7493                                  */
7494                                 offset = btrfs_alloc_from_cluster(block_group,
7495                                                         last_ptr,
7496                                                         num_bytes,
7497                                                         search_start,
7498                                                         &max_extent_size);
7499                                 if (offset) {
7500                                         /* we found one, proceed */
7501                                         spin_unlock(&last_ptr->refill_lock);
7502                                         trace_btrfs_reserve_extent_cluster(
7503                                                 block_group, search_start,
7504                                                 num_bytes);
7505                                         goto checks;
7506                                 }
7507                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
7508                                    && !failed_cluster_refill) {
7509                                 spin_unlock(&last_ptr->refill_lock);
7510
7511                                 failed_cluster_refill = true;
7512                                 wait_block_group_cache_progress(block_group,
7513                                        num_bytes + empty_cluster + empty_size);
7514                                 goto have_block_group;
7515                         }
7516
7517                         /*
7518                          * at this point we either didn't find a cluster
7519                          * or we weren't able to allocate a block from our
7520                          * cluster.  Free the cluster we've been trying
7521                          * to use, and go to the next block group
7522                          */
7523                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7524                         spin_unlock(&last_ptr->refill_lock);
7525                         goto loop;
7526                 }
7527
7528 unclustered_alloc:
7529                 /*
7530                  * We are doing an unclustered alloc, set the fragmented flag so
7531                  * we don't bother trying to setup a cluster again until we get
7532                  * more space.
7533                  */
7534                 if (unlikely(last_ptr)) {
7535                         spin_lock(&last_ptr->lock);
7536                         last_ptr->fragmented = 1;
7537                         spin_unlock(&last_ptr->lock);
7538                 }
7539                 if (cached) {
7540                         struct btrfs_free_space_ctl *ctl =
7541                                 block_group->free_space_ctl;
7542
7543                         spin_lock(&ctl->tree_lock);
7544                         if (ctl->free_space <
7545                             num_bytes + empty_cluster + empty_size) {
7546                                 if (ctl->free_space > max_extent_size)
7547                                         max_extent_size = ctl->free_space;
7548                                 spin_unlock(&ctl->tree_lock);
7549                                 goto loop;
7550                         }
7551                         spin_unlock(&ctl->tree_lock);
7552                 }
7553
7554                 offset = btrfs_find_space_for_alloc(block_group, search_start,
7555                                                     num_bytes, empty_size,
7556                                                     &max_extent_size);
7557                 /*
7558                  * If we didn't find a chunk, and we haven't failed on this
7559                  * block group before, and this block group is in the middle of
7560                  * caching and we are ok with waiting, then go ahead and wait
7561                  * for progress to be made, and set failed_alloc to true.
7562                  *
7563                  * If failed_alloc is true then we've already waited on this
7564                  * block group once and should move on to the next block group.
7565                  */
7566                 if (!offset && !failed_alloc && !cached &&
7567                     loop > LOOP_CACHING_NOWAIT) {
7568                         wait_block_group_cache_progress(block_group,
7569                                                 num_bytes + empty_size);
7570                         failed_alloc = true;
7571                         goto have_block_group;
7572                 } else if (!offset) {
7573                         goto loop;
7574                 }
7575 checks:
7576                 search_start = ALIGN(offset, fs_info->stripesize);
7577
7578                 /* move on to the next group */
7579                 if (search_start + num_bytes >
7580                     block_group->key.objectid + block_group->key.offset) {
7581                         btrfs_add_free_space(block_group, offset, num_bytes);
7582                         goto loop;
7583                 }
7584
7585                 if (offset < search_start)
7586                         btrfs_add_free_space(block_group, offset,
7587                                              search_start - offset);
7588                 BUG_ON(offset > search_start);
7589
7590                 ret = btrfs_add_reserved_bytes(block_group, ram_bytes,
7591                                 num_bytes, delalloc);
7592                 if (ret == -EAGAIN) {
7593                         btrfs_add_free_space(block_group, offset, num_bytes);
7594                         goto loop;
7595                 }
7596                 btrfs_inc_block_group_reservations(block_group);
7597
7598                 /* we are all good, lets return */
7599                 ins->objectid = search_start;
7600                 ins->offset = num_bytes;
7601
7602                 trace_btrfs_reserve_extent(block_group, search_start, num_bytes);
7603                 btrfs_release_block_group(block_group, delalloc);
7604                 break;
7605 loop:
7606                 failed_cluster_refill = false;
7607                 failed_alloc = false;
7608                 BUG_ON(btrfs_bg_flags_to_raid_index(block_group->flags) !=
7609                        index);
7610                 btrfs_release_block_group(block_group, delalloc);
7611                 cond_resched();
7612         }
7613         up_read(&space_info->groups_sem);
7614
7615         if ((loop == LOOP_CACHING_NOWAIT) && have_caching_bg
7616                 && !orig_have_caching_bg)
7617                 orig_have_caching_bg = true;
7618
7619         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
7620                 goto search;
7621
7622         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
7623                 goto search;
7624
7625         /*
7626          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
7627          *                      caching kthreads as we move along
7628          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
7629          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
7630          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
7631          *                      again
7632          */
7633         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
7634                 index = 0;
7635                 if (loop == LOOP_CACHING_NOWAIT) {
7636                         /*
7637                          * We want to skip the LOOP_CACHING_WAIT step if we
7638                          * don't have any uncached bgs and we've already done a
7639                          * full search through.
7640                          */
7641                         if (orig_have_caching_bg || !full_search)
7642                                 loop = LOOP_CACHING_WAIT;
7643                         else
7644                                 loop = LOOP_ALLOC_CHUNK;
7645                 } else {
7646                         loop++;
7647                 }
7648
7649                 if (loop == LOOP_ALLOC_CHUNK) {
7650                         struct btrfs_trans_handle *trans;
7651                         int exist = 0;
7652
7653                         trans = current->journal_info;
7654                         if (trans)
7655                                 exist = 1;
7656                         else
7657                                 trans = btrfs_join_transaction(root);
7658
7659                         if (IS_ERR(trans)) {
7660                                 ret = PTR_ERR(trans);
7661                                 goto out;
7662                         }
7663
7664                         ret = do_chunk_alloc(trans, flags, CHUNK_ALLOC_FORCE);
7665
7666                         /*
7667                          * If we can't allocate a new chunk we've already looped
7668                          * through at least once, move on to the NO_EMPTY_SIZE
7669                          * case.
7670                          */
7671                         if (ret == -ENOSPC)
7672                                 loop = LOOP_NO_EMPTY_SIZE;
7673
7674                         /*
7675                          * Do not bail out on ENOSPC since we
7676                          * can do more things.
7677                          */
7678                         if (ret < 0 && ret != -ENOSPC)
7679                                 btrfs_abort_transaction(trans, ret);
7680                         else
7681                                 ret = 0;
7682                         if (!exist)
7683                                 btrfs_end_transaction(trans);
7684                         if (ret)
7685                                 goto out;
7686                 }
7687
7688                 if (loop == LOOP_NO_EMPTY_SIZE) {
7689                         /*
7690                          * Don't loop again if we already have no empty_size and
7691                          * no empty_cluster.
7692                          */
7693                         if (empty_size == 0 &&
7694                             empty_cluster == 0) {
7695                                 ret = -ENOSPC;
7696                                 goto out;
7697                         }
7698                         empty_size = 0;
7699                         empty_cluster = 0;
7700                 }
7701
7702                 goto search;
7703         } else if (!ins->objectid) {
7704                 ret = -ENOSPC;
7705         } else if (ins->objectid) {
7706                 if (!use_cluster && last_ptr) {
7707                         spin_lock(&last_ptr->lock);
7708                         last_ptr->window_start = ins->objectid;
7709                         spin_unlock(&last_ptr->lock);
7710                 }
7711                 ret = 0;
7712         }
7713 out:
7714         if (ret == -ENOSPC) {
7715                 spin_lock(&space_info->lock);
7716                 space_info->max_extent_size = max_extent_size;
7717                 spin_unlock(&space_info->lock);
7718                 ins->offset = max_extent_size;
7719         }
7720         return ret;
7721 }
7722
7723 static void dump_space_info(struct btrfs_fs_info *fs_info,
7724                             struct btrfs_space_info *info, u64 bytes,
7725                             int dump_block_groups)
7726 {
7727         struct btrfs_block_group_cache *cache;
7728         int index = 0;
7729
7730         spin_lock(&info->lock);
7731         btrfs_info(fs_info, "space_info %llu has %llu free, is %sfull",
7732                    info->flags,
7733                    info->total_bytes - btrfs_space_info_used(info, true),
7734                    info->full ? "" : "not ");
7735         btrfs_info(fs_info,
7736                 "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu",
7737                 info->total_bytes, info->bytes_used, info->bytes_pinned,
7738                 info->bytes_reserved, info->bytes_may_use,
7739                 info->bytes_readonly);
7740         spin_unlock(&info->lock);
7741
7742         if (!dump_block_groups)
7743                 return;
7744
7745         down_read(&info->groups_sem);
7746 again:
7747         list_for_each_entry(cache, &info->block_groups[index], list) {
7748                 spin_lock(&cache->lock);
7749                 btrfs_info(fs_info,
7750                         "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s",
7751                         cache->key.objectid, cache->key.offset,
7752                         btrfs_block_group_used(&cache->item), cache->pinned,
7753                         cache->reserved, cache->ro ? "[readonly]" : "");
7754                 btrfs_dump_free_space(cache, bytes);
7755                 spin_unlock(&cache->lock);
7756         }
7757         if (++index < BTRFS_NR_RAID_TYPES)
7758                 goto again;
7759         up_read(&info->groups_sem);
7760 }
7761
7762 /*
7763  * btrfs_reserve_extent - entry point to the extent allocator. Tries to find a
7764  *                        hole that is at least as big as @num_bytes.
7765  *
7766  * @root           -    The root that will contain this extent
7767  *
7768  * @ram_bytes      -    The amount of space in ram that @num_bytes take. This
7769  *                      is used for accounting purposes. This value differs
7770  *                      from @num_bytes only in the case of compressed extents.
7771  *
7772  * @num_bytes      -    Number of bytes to allocate on-disk.
7773  *
7774  * @min_alloc_size -    Indicates the minimum amount of space that the
7775  *                      allocator should try to satisfy. In some cases
7776  *                      @num_bytes may be larger than what is required and if
7777  *                      the filesystem is fragmented then allocation fails.
7778  *                      However, the presence of @min_alloc_size gives a
7779  *                      chance to try and satisfy the smaller allocation.
7780  *
7781  * @empty_size     -    A hint that you plan on doing more COW. This is the
7782  *                      size in bytes the allocator should try to find free
7783  *                      next to the block it returns.  This is just a hint and
7784  *                      may be ignored by the allocator.
7785  *
7786  * @hint_byte      -    Hint to the allocator to start searching above the byte
7787  *                      address passed. It might be ignored.
7788  *
7789  * @ins            -    This key is modified to record the found hole. It will
7790  *                      have the following values:
7791  *                      ins->objectid == start position
7792  *                      ins->flags = BTRFS_EXTENT_ITEM_KEY
7793  *                      ins->offset == the size of the hole.
7794  *
7795  * @is_data        -    Boolean flag indicating whether an extent is
7796  *                      allocated for data (true) or metadata (false)
7797  *
7798  * @delalloc       -    Boolean flag indicating whether this allocation is for
7799  *                      delalloc or not. If 'true' data_rwsem of block groups
7800  *                      is going to be acquired.
7801  *
7802  *
7803  * Returns 0 when an allocation succeeded or < 0 when an error occurred. In
7804  * case -ENOSPC is returned then @ins->offset will contain the size of the
7805  * largest available hole the allocator managed to find.
7806  */
7807 int btrfs_reserve_extent(struct btrfs_root *root, u64 ram_bytes,
7808                          u64 num_bytes, u64 min_alloc_size,
7809                          u64 empty_size, u64 hint_byte,
7810                          struct btrfs_key *ins, int is_data, int delalloc)
7811 {
7812         struct btrfs_fs_info *fs_info = root->fs_info;
7813         bool final_tried = num_bytes == min_alloc_size;
7814         u64 flags;
7815         int ret;
7816
7817         flags = get_alloc_profile_by_root(root, is_data);
7818 again:
7819         WARN_ON(num_bytes < fs_info->sectorsize);
7820         ret = find_free_extent(fs_info, ram_bytes, num_bytes, empty_size,
7821                                hint_byte, ins, flags, delalloc);
7822         if (!ret && !is_data) {
7823                 btrfs_dec_block_group_reservations(fs_info, ins->objectid);
7824         } else if (ret == -ENOSPC) {
7825                 if (!final_tried && ins->offset) {
7826                         num_bytes = min(num_bytes >> 1, ins->offset);
7827                         num_bytes = round_down(num_bytes,
7828                                                fs_info->sectorsize);
7829                         num_bytes = max(num_bytes, min_alloc_size);
7830                         ram_bytes = num_bytes;
7831                         if (num_bytes == min_alloc_size)
7832                                 final_tried = true;
7833                         goto again;
7834                 } else if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
7835                         struct btrfs_space_info *sinfo;
7836
7837                         sinfo = __find_space_info(fs_info, flags);
7838                         btrfs_err(fs_info,
7839                                   "allocation failed flags %llu, wanted %llu",
7840                                   flags, num_bytes);
7841                         if (sinfo)
7842                                 dump_space_info(fs_info, sinfo, num_bytes, 1);
7843                 }
7844         }
7845
7846         return ret;
7847 }
7848
7849 static int __btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
7850                                         u64 start, u64 len,
7851                                         int pin, int delalloc)
7852 {
7853         struct btrfs_block_group_cache *cache;
7854         int ret = 0;
7855
7856         cache = btrfs_lookup_block_group(fs_info, start);
7857         if (!cache) {
7858                 btrfs_err(fs_info, "Unable to find block group for %llu",
7859                           start);
7860                 return -ENOSPC;
7861         }
7862
7863         if (pin)
7864                 pin_down_extent(fs_info, cache, start, len, 1);
7865         else {
7866                 if (btrfs_test_opt(fs_info, DISCARD))
7867                         ret = btrfs_discard_extent(fs_info, start, len, NULL);
7868                 btrfs_add_free_space(cache, start, len);
7869                 btrfs_free_reserved_bytes(cache, len, delalloc);
7870                 trace_btrfs_reserved_extent_free(fs_info, start, len);
7871         }
7872
7873         btrfs_put_block_group(cache);
7874         return ret;
7875 }
7876
7877 int btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
7878                                u64 start, u64 len, int delalloc)
7879 {
7880         return __btrfs_free_reserved_extent(fs_info, start, len, 0, delalloc);
7881 }
7882
7883 int btrfs_free_and_pin_reserved_extent(struct btrfs_fs_info *fs_info,
7884                                        u64 start, u64 len)
7885 {
7886         return __btrfs_free_reserved_extent(fs_info, start, len, 1, 0);
7887 }
7888
7889 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
7890                                       u64 parent, u64 root_objectid,
7891                                       u64 flags, u64 owner, u64 offset,
7892                                       struct btrfs_key *ins, int ref_mod)
7893 {
7894         struct btrfs_fs_info *fs_info = trans->fs_info;
7895         int ret;
7896         struct btrfs_extent_item *extent_item;
7897         struct btrfs_extent_inline_ref *iref;
7898         struct btrfs_path *path;
7899         struct extent_buffer *leaf;
7900         int type;
7901         u32 size;
7902
7903         if (parent > 0)
7904                 type = BTRFS_SHARED_DATA_REF_KEY;
7905         else
7906                 type = BTRFS_EXTENT_DATA_REF_KEY;
7907
7908         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
7909
7910         path = btrfs_alloc_path();
7911         if (!path)
7912                 return -ENOMEM;
7913
7914         path->leave_spinning = 1;
7915         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
7916                                       ins, size);
7917         if (ret) {
7918                 btrfs_free_path(path);
7919                 return ret;
7920         }
7921
7922         leaf = path->nodes[0];
7923         extent_item = btrfs_item_ptr(leaf, path->slots[0],
7924                                      struct btrfs_extent_item);
7925         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
7926         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
7927         btrfs_set_extent_flags(leaf, extent_item,
7928                                flags | BTRFS_EXTENT_FLAG_DATA);
7929
7930         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
7931         btrfs_set_extent_inline_ref_type(leaf, iref, type);
7932         if (parent > 0) {
7933                 struct btrfs_shared_data_ref *ref;
7934                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
7935                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
7936                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
7937         } else {
7938                 struct btrfs_extent_data_ref *ref;
7939                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
7940                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
7941                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
7942                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
7943                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
7944         }
7945
7946         btrfs_mark_buffer_dirty(path->nodes[0]);
7947         btrfs_free_path(path);
7948
7949         ret = remove_from_free_space_tree(trans, ins->objectid, ins->offset);
7950         if (ret)
7951                 return ret;
7952
7953         ret = update_block_group(trans, fs_info, ins->objectid, ins->offset, 1);
7954         if (ret) { /* -ENOENT, logic error */
7955                 btrfs_err(fs_info, "update block group failed for %llu %llu",
7956                         ins->objectid, ins->offset);
7957                 BUG();
7958         }
7959         trace_btrfs_reserved_extent_alloc(fs_info, ins->objectid, ins->offset);
7960         return ret;
7961 }
7962
7963 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
7964                                      struct btrfs_delayed_ref_node *node,
7965                                      struct btrfs_delayed_extent_op *extent_op)
7966 {
7967         struct btrfs_fs_info *fs_info = trans->fs_info;
7968         int ret;
7969         struct btrfs_extent_item *extent_item;
7970         struct btrfs_key extent_key;
7971         struct btrfs_tree_block_info *block_info;
7972         struct btrfs_extent_inline_ref *iref;
7973         struct btrfs_path *path;
7974         struct extent_buffer *leaf;
7975         struct btrfs_delayed_tree_ref *ref;
7976         u32 size = sizeof(*extent_item) + sizeof(*iref);
7977         u64 num_bytes;
7978         u64 flags = extent_op->flags_to_set;
7979         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
7980
7981         ref = btrfs_delayed_node_to_tree_ref(node);
7982
7983         extent_key.objectid = node->bytenr;
7984         if (skinny_metadata) {
7985                 extent_key.offset = ref->level;
7986                 extent_key.type = BTRFS_METADATA_ITEM_KEY;
7987                 num_bytes = fs_info->nodesize;
7988         } else {
7989                 extent_key.offset = node->num_bytes;
7990                 extent_key.type = BTRFS_EXTENT_ITEM_KEY;
7991                 size += sizeof(*block_info);
7992                 num_bytes = node->num_bytes;
7993         }
7994
7995         path = btrfs_alloc_path();
7996         if (!path) {
7997                 btrfs_free_and_pin_reserved_extent(fs_info,
7998                                                    extent_key.objectid,
7999                                                    fs_info->nodesize);
8000                 return -ENOMEM;
8001         }
8002
8003         path->leave_spinning = 1;
8004         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
8005                                       &extent_key, size);
8006         if (ret) {
8007                 btrfs_free_path(path);
8008                 btrfs_free_and_pin_reserved_extent(fs_info,
8009                                                    extent_key.objectid,
8010                                                    fs_info->nodesize);
8011                 return ret;
8012         }
8013
8014         leaf = path->nodes[0];
8015         extent_item = btrfs_item_ptr(leaf, path->slots[0],
8016                                      struct btrfs_extent_item);
8017         btrfs_set_extent_refs(leaf, extent_item, 1);
8018         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8019         btrfs_set_extent_flags(leaf, extent_item,
8020                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
8021
8022         if (skinny_metadata) {
8023                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8024         } else {
8025                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
8026                 btrfs_set_tree_block_key(leaf, block_info, &extent_op->key);
8027                 btrfs_set_tree_block_level(leaf, block_info, ref->level);
8028                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
8029         }
8030
8031         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY) {
8032                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
8033                 btrfs_set_extent_inline_ref_type(leaf, iref,
8034                                                  BTRFS_SHARED_BLOCK_REF_KEY);
8035                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->parent);
8036         } else {
8037                 btrfs_set_extent_inline_ref_type(leaf, iref,
8038                                                  BTRFS_TREE_BLOCK_REF_KEY);
8039                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->root);
8040         }
8041
8042         btrfs_mark_buffer_dirty(leaf);
8043         btrfs_free_path(path);
8044
8045         ret = remove_from_free_space_tree(trans, extent_key.objectid,
8046                                           num_bytes);
8047         if (ret)
8048                 return ret;
8049
8050         ret = update_block_group(trans, fs_info, extent_key.objectid,
8051                                  fs_info->nodesize, 1);
8052         if (ret) { /* -ENOENT, logic error */
8053                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8054                         extent_key.objectid, extent_key.offset);
8055                 BUG();
8056         }
8057
8058         trace_btrfs_reserved_extent_alloc(fs_info, extent_key.objectid,
8059                                           fs_info->nodesize);
8060         return ret;
8061 }
8062
8063 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8064                                      struct btrfs_root *root, u64 owner,
8065                                      u64 offset, u64 ram_bytes,
8066                                      struct btrfs_key *ins)
8067 {
8068         int ret;
8069
8070         BUG_ON(root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
8071
8072         btrfs_ref_tree_mod(root, ins->objectid, ins->offset, 0,
8073                            root->root_key.objectid, owner, offset,
8074                            BTRFS_ADD_DELAYED_EXTENT);
8075
8076         ret = btrfs_add_delayed_data_ref(trans, ins->objectid,
8077                                          ins->offset, 0,
8078                                          root->root_key.objectid, owner,
8079                                          offset, ram_bytes,
8080                                          BTRFS_ADD_DELAYED_EXTENT, NULL, NULL);
8081         return ret;
8082 }
8083
8084 /*
8085  * this is used by the tree logging recovery code.  It records that
8086  * an extent has been allocated and makes sure to clear the free
8087  * space cache bits as well
8088  */
8089 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
8090                                    u64 root_objectid, u64 owner, u64 offset,
8091                                    struct btrfs_key *ins)
8092 {
8093         struct btrfs_fs_info *fs_info = trans->fs_info;
8094         int ret;
8095         struct btrfs_block_group_cache *block_group;
8096         struct btrfs_space_info *space_info;
8097
8098         /*
8099          * Mixed block groups will exclude before processing the log so we only
8100          * need to do the exclude dance if this fs isn't mixed.
8101          */
8102         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS)) {
8103                 ret = __exclude_logged_extent(fs_info, ins->objectid,
8104                                               ins->offset);
8105                 if (ret)
8106                         return ret;
8107         }
8108
8109         block_group = btrfs_lookup_block_group(fs_info, ins->objectid);
8110         if (!block_group)
8111                 return -EINVAL;
8112
8113         space_info = block_group->space_info;
8114         spin_lock(&space_info->lock);
8115         spin_lock(&block_group->lock);
8116         space_info->bytes_reserved += ins->offset;
8117         block_group->reserved += ins->offset;
8118         spin_unlock(&block_group->lock);
8119         spin_unlock(&space_info->lock);
8120
8121         ret = alloc_reserved_file_extent(trans, 0, root_objectid, 0, owner,
8122                                          offset, ins, 1);
8123         btrfs_put_block_group(block_group);
8124         return ret;
8125 }
8126
8127 static struct extent_buffer *
8128 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
8129                       u64 bytenr, int level, u64 owner)
8130 {
8131         struct btrfs_fs_info *fs_info = root->fs_info;
8132         struct extent_buffer *buf;
8133
8134         buf = btrfs_find_create_tree_block(fs_info, bytenr);
8135         if (IS_ERR(buf))
8136                 return buf;
8137
8138         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
8139         btrfs_tree_lock(buf);
8140         clean_tree_block(fs_info, buf);
8141         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
8142
8143         btrfs_set_lock_blocking(buf);
8144         set_extent_buffer_uptodate(buf);
8145
8146         memzero_extent_buffer(buf, 0, sizeof(struct btrfs_header));
8147         btrfs_set_header_level(buf, level);
8148         btrfs_set_header_bytenr(buf, buf->start);
8149         btrfs_set_header_generation(buf, trans->transid);
8150         btrfs_set_header_backref_rev(buf, BTRFS_MIXED_BACKREF_REV);
8151         btrfs_set_header_owner(buf, owner);
8152         write_extent_buffer_fsid(buf, fs_info->fsid);
8153         write_extent_buffer_chunk_tree_uuid(buf, fs_info->chunk_tree_uuid);
8154         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
8155                 buf->log_index = root->log_transid % 2;
8156                 /*
8157                  * we allow two log transactions at a time, use different
8158                  * EXENT bit to differentiate dirty pages.
8159                  */
8160                 if (buf->log_index == 0)
8161                         set_extent_dirty(&root->dirty_log_pages, buf->start,
8162                                         buf->start + buf->len - 1, GFP_NOFS);
8163                 else
8164                         set_extent_new(&root->dirty_log_pages, buf->start,
8165                                         buf->start + buf->len - 1);
8166         } else {
8167                 buf->log_index = -1;
8168                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
8169                          buf->start + buf->len - 1, GFP_NOFS);
8170         }
8171         trans->dirty = true;
8172         /* this returns a buffer locked for blocking */
8173         return buf;
8174 }
8175
8176 static struct btrfs_block_rsv *
8177 use_block_rsv(struct btrfs_trans_handle *trans,
8178               struct btrfs_root *root, u32 blocksize)
8179 {
8180         struct btrfs_fs_info *fs_info = root->fs_info;
8181         struct btrfs_block_rsv *block_rsv;
8182         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
8183         int ret;
8184         bool global_updated = false;
8185
8186         block_rsv = get_block_rsv(trans, root);
8187
8188         if (unlikely(block_rsv->size == 0))
8189                 goto try_reserve;
8190 again:
8191         ret = block_rsv_use_bytes(block_rsv, blocksize);
8192         if (!ret)
8193                 return block_rsv;
8194
8195         if (block_rsv->failfast)
8196                 return ERR_PTR(ret);
8197
8198         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
8199                 global_updated = true;
8200                 update_global_block_rsv(fs_info);
8201                 goto again;
8202         }
8203
8204         if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
8205                 static DEFINE_RATELIMIT_STATE(_rs,
8206                                 DEFAULT_RATELIMIT_INTERVAL * 10,
8207                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
8208                 if (__ratelimit(&_rs))
8209                         WARN(1, KERN_DEBUG
8210                                 "BTRFS: block rsv returned %d\n", ret);
8211         }
8212 try_reserve:
8213         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
8214                                      BTRFS_RESERVE_NO_FLUSH);
8215         if (!ret)
8216                 return block_rsv;
8217         /*
8218          * If we couldn't reserve metadata bytes try and use some from
8219          * the global reserve if its space type is the same as the global
8220          * reservation.
8221          */
8222         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
8223             block_rsv->space_info == global_rsv->space_info) {
8224                 ret = block_rsv_use_bytes(global_rsv, blocksize);
8225                 if (!ret)
8226                         return global_rsv;
8227         }
8228         return ERR_PTR(ret);
8229 }
8230
8231 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
8232                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
8233 {
8234         block_rsv_add_bytes(block_rsv, blocksize, 0);
8235         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0, NULL);
8236 }
8237
8238 /*
8239  * finds a free extent and does all the dirty work required for allocation
8240  * returns the tree buffer or an ERR_PTR on error.
8241  */
8242 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
8243                                              struct btrfs_root *root,
8244                                              u64 parent, u64 root_objectid,
8245                                              const struct btrfs_disk_key *key,
8246                                              int level, u64 hint,
8247                                              u64 empty_size)
8248 {
8249         struct btrfs_fs_info *fs_info = root->fs_info;
8250         struct btrfs_key ins;
8251         struct btrfs_block_rsv *block_rsv;
8252         struct extent_buffer *buf;
8253         struct btrfs_delayed_extent_op *extent_op;
8254         u64 flags = 0;
8255         int ret;
8256         u32 blocksize = fs_info->nodesize;
8257         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
8258
8259 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
8260         if (btrfs_is_testing(fs_info)) {
8261                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
8262                                             level, root_objectid);
8263                 if (!IS_ERR(buf))
8264                         root->alloc_bytenr += blocksize;
8265                 return buf;
8266         }
8267 #endif
8268
8269         block_rsv = use_block_rsv(trans, root, blocksize);
8270         if (IS_ERR(block_rsv))
8271                 return ERR_CAST(block_rsv);
8272
8273         ret = btrfs_reserve_extent(root, blocksize, blocksize, blocksize,
8274                                    empty_size, hint, &ins, 0, 0);
8275         if (ret)
8276                 goto out_unuse;
8277
8278         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level,
8279                                     root_objectid);
8280         if (IS_ERR(buf)) {
8281                 ret = PTR_ERR(buf);
8282                 goto out_free_reserved;
8283         }
8284
8285         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
8286                 if (parent == 0)
8287                         parent = ins.objectid;
8288                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
8289         } else
8290                 BUG_ON(parent > 0);
8291
8292         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
8293                 extent_op = btrfs_alloc_delayed_extent_op();
8294                 if (!extent_op) {
8295                         ret = -ENOMEM;
8296                         goto out_free_buf;
8297                 }
8298                 if (key)
8299                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
8300                 else
8301                         memset(&extent_op->key, 0, sizeof(extent_op->key));
8302                 extent_op->flags_to_set = flags;
8303                 extent_op->update_key = skinny_metadata ? false : true;
8304                 extent_op->update_flags = true;
8305                 extent_op->is_data = false;
8306                 extent_op->level = level;
8307
8308                 btrfs_ref_tree_mod(root, ins.objectid, ins.offset, parent,
8309                                    root_objectid, level, 0,
8310                                    BTRFS_ADD_DELAYED_EXTENT);
8311                 ret = btrfs_add_delayed_tree_ref(trans, ins.objectid,
8312                                                  ins.offset, parent,
8313                                                  root_objectid, level,
8314                                                  BTRFS_ADD_DELAYED_EXTENT,
8315                                                  extent_op, NULL, NULL);
8316                 if (ret)
8317                         goto out_free_delayed;
8318         }
8319         return buf;
8320
8321 out_free_delayed:
8322         btrfs_free_delayed_extent_op(extent_op);
8323 out_free_buf:
8324         free_extent_buffer(buf);
8325 out_free_reserved:
8326         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 0);
8327 out_unuse:
8328         unuse_block_rsv(fs_info, block_rsv, blocksize);
8329         return ERR_PTR(ret);
8330 }
8331
8332 struct walk_control {
8333         u64 refs[BTRFS_MAX_LEVEL];
8334         u64 flags[BTRFS_MAX_LEVEL];
8335         struct btrfs_key update_progress;
8336         int stage;
8337         int level;
8338         int shared_level;
8339         int update_ref;
8340         int keep_locks;
8341         int reada_slot;
8342         int reada_count;
8343         int for_reloc;
8344 };
8345
8346 #define DROP_REFERENCE  1
8347 #define UPDATE_BACKREF  2
8348
8349 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
8350                                      struct btrfs_root *root,
8351                                      struct walk_control *wc,
8352                                      struct btrfs_path *path)
8353 {
8354         struct btrfs_fs_info *fs_info = root->fs_info;
8355         u64 bytenr;
8356         u64 generation;
8357         u64 refs;
8358         u64 flags;
8359         u32 nritems;
8360         struct btrfs_key key;
8361         struct extent_buffer *eb;
8362         int ret;
8363         int slot;
8364         int nread = 0;
8365
8366         if (path->slots[wc->level] < wc->reada_slot) {
8367                 wc->reada_count = wc->reada_count * 2 / 3;
8368                 wc->reada_count = max(wc->reada_count, 2);
8369         } else {
8370                 wc->reada_count = wc->reada_count * 3 / 2;
8371                 wc->reada_count = min_t(int, wc->reada_count,
8372                                         BTRFS_NODEPTRS_PER_BLOCK(fs_info));
8373         }
8374
8375         eb = path->nodes[wc->level];
8376         nritems = btrfs_header_nritems(eb);
8377
8378         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
8379                 if (nread >= wc->reada_count)
8380                         break;
8381
8382                 cond_resched();
8383                 bytenr = btrfs_node_blockptr(eb, slot);
8384                 generation = btrfs_node_ptr_generation(eb, slot);
8385
8386                 if (slot == path->slots[wc->level])
8387                         goto reada;
8388
8389                 if (wc->stage == UPDATE_BACKREF &&
8390                     generation <= root->root_key.offset)
8391                         continue;
8392
8393                 /* We don't lock the tree block, it's OK to be racy here */
8394                 ret = btrfs_lookup_extent_info(trans, fs_info, bytenr,
8395                                                wc->level - 1, 1, &refs,
8396                                                &flags);
8397                 /* We don't care about errors in readahead. */
8398                 if (ret < 0)
8399                         continue;
8400                 BUG_ON(refs == 0);
8401
8402                 if (wc->stage == DROP_REFERENCE) {
8403                         if (refs == 1)
8404                                 goto reada;
8405
8406                         if (wc->level == 1 &&
8407                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8408                                 continue;
8409                         if (!wc->update_ref ||
8410                             generation <= root->root_key.offset)
8411                                 continue;
8412                         btrfs_node_key_to_cpu(eb, &key, slot);
8413                         ret = btrfs_comp_cpu_keys(&key,
8414                                                   &wc->update_progress);
8415                         if (ret < 0)
8416                                 continue;
8417                 } else {
8418                         if (wc->level == 1 &&
8419                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8420                                 continue;
8421                 }
8422 reada:
8423                 readahead_tree_block(fs_info, bytenr);
8424                 nread++;
8425         }
8426         wc->reada_slot = slot;
8427 }
8428
8429 /*
8430  * helper to process tree block while walking down the tree.
8431  *
8432  * when wc->stage == UPDATE_BACKREF, this function updates
8433  * back refs for pointers in the block.
8434  *
8435  * NOTE: return value 1 means we should stop walking down.
8436  */
8437 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
8438                                    struct btrfs_root *root,
8439                                    struct btrfs_path *path,
8440                                    struct walk_control *wc, int lookup_info)
8441 {
8442         struct btrfs_fs_info *fs_info = root->fs_info;
8443         int level = wc->level;
8444         struct extent_buffer *eb = path->nodes[level];
8445         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8446         int ret;
8447
8448         if (wc->stage == UPDATE_BACKREF &&
8449             btrfs_header_owner(eb) != root->root_key.objectid)
8450                 return 1;
8451
8452         /*
8453          * when reference count of tree block is 1, it won't increase
8454          * again. once full backref flag is set, we never clear it.
8455          */
8456         if (lookup_info &&
8457             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
8458              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
8459                 BUG_ON(!path->locks[level]);
8460                 ret = btrfs_lookup_extent_info(trans, fs_info,
8461                                                eb->start, level, 1,
8462                                                &wc->refs[level],
8463                                                &wc->flags[level]);
8464                 BUG_ON(ret == -ENOMEM);
8465                 if (ret)
8466                         return ret;
8467                 BUG_ON(wc->refs[level] == 0);
8468         }
8469
8470         if (wc->stage == DROP_REFERENCE) {
8471                 if (wc->refs[level] > 1)
8472                         return 1;
8473
8474                 if (path->locks[level] && !wc->keep_locks) {
8475                         btrfs_tree_unlock_rw(eb, path->locks[level]);
8476                         path->locks[level] = 0;
8477                 }
8478                 return 0;
8479         }
8480
8481         /* wc->stage == UPDATE_BACKREF */
8482         if (!(wc->flags[level] & flag)) {
8483                 BUG_ON(!path->locks[level]);
8484                 ret = btrfs_inc_ref(trans, root, eb, 1);
8485                 BUG_ON(ret); /* -ENOMEM */
8486                 ret = btrfs_dec_ref(trans, root, eb, 0);
8487                 BUG_ON(ret); /* -ENOMEM */
8488                 ret = btrfs_set_disk_extent_flags(trans, fs_info, eb->start,
8489                                                   eb->len, flag,
8490                                                   btrfs_header_level(eb), 0);
8491                 BUG_ON(ret); /* -ENOMEM */
8492                 wc->flags[level] |= flag;
8493         }
8494
8495         /*
8496          * the block is shared by multiple trees, so it's not good to
8497          * keep the tree lock
8498          */
8499         if (path->locks[level] && level > 0) {
8500                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8501                 path->locks[level] = 0;
8502         }
8503         return 0;
8504 }
8505
8506 /*
8507  * helper to process tree block pointer.
8508  *
8509  * when wc->stage == DROP_REFERENCE, this function checks
8510  * reference count of the block pointed to. if the block
8511  * is shared and we need update back refs for the subtree
8512  * rooted at the block, this function changes wc->stage to
8513  * UPDATE_BACKREF. if the block is shared and there is no
8514  * need to update back, this function drops the reference
8515  * to the block.
8516  *
8517  * NOTE: return value 1 means we should stop walking down.
8518  */
8519 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
8520                                  struct btrfs_root *root,
8521                                  struct btrfs_path *path,
8522                                  struct walk_control *wc, int *lookup_info)
8523 {
8524         struct btrfs_fs_info *fs_info = root->fs_info;
8525         u64 bytenr;
8526         u64 generation;
8527         u64 parent;
8528         u32 blocksize;
8529         struct btrfs_key key;
8530         struct btrfs_key first_key;
8531         struct extent_buffer *next;
8532         int level = wc->level;
8533         int reada = 0;
8534         int ret = 0;
8535         bool need_account = false;
8536
8537         generation = btrfs_node_ptr_generation(path->nodes[level],
8538                                                path->slots[level]);
8539         /*
8540          * if the lower level block was created before the snapshot
8541          * was created, we know there is no need to update back refs
8542          * for the subtree
8543          */
8544         if (wc->stage == UPDATE_BACKREF &&
8545             generation <= root->root_key.offset) {
8546                 *lookup_info = 1;
8547                 return 1;
8548         }
8549
8550         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
8551         btrfs_node_key_to_cpu(path->nodes[level], &first_key,
8552                               path->slots[level]);
8553         blocksize = fs_info->nodesize;
8554
8555         next = find_extent_buffer(fs_info, bytenr);
8556         if (!next) {
8557                 next = btrfs_find_create_tree_block(fs_info, bytenr);
8558                 if (IS_ERR(next))
8559                         return PTR_ERR(next);
8560
8561                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
8562                                                level - 1);
8563                 reada = 1;
8564         }
8565         btrfs_tree_lock(next);
8566         btrfs_set_lock_blocking(next);
8567
8568         ret = btrfs_lookup_extent_info(trans, fs_info, bytenr, level - 1, 1,
8569                                        &wc->refs[level - 1],
8570                                        &wc->flags[level - 1]);
8571         if (ret < 0)
8572                 goto out_unlock;
8573
8574         if (unlikely(wc->refs[level - 1] == 0)) {
8575                 btrfs_err(fs_info, "Missing references.");
8576                 ret = -EIO;
8577                 goto out_unlock;
8578         }
8579         *lookup_info = 0;
8580
8581         if (wc->stage == DROP_REFERENCE) {
8582                 if (wc->refs[level - 1] > 1) {
8583                         need_account = true;
8584                         if (level == 1 &&
8585                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8586                                 goto skip;
8587
8588                         if (!wc->update_ref ||
8589                             generation <= root->root_key.offset)
8590                                 goto skip;
8591
8592                         btrfs_node_key_to_cpu(path->nodes[level], &key,
8593                                               path->slots[level]);
8594                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
8595                         if (ret < 0)
8596                                 goto skip;
8597
8598                         wc->stage = UPDATE_BACKREF;
8599                         wc->shared_level = level - 1;
8600                 }
8601         } else {
8602                 if (level == 1 &&
8603                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8604                         goto skip;
8605         }
8606
8607         if (!btrfs_buffer_uptodate(next, generation, 0)) {
8608                 btrfs_tree_unlock(next);
8609                 free_extent_buffer(next);
8610                 next = NULL;
8611                 *lookup_info = 1;
8612         }
8613
8614         if (!next) {
8615                 if (reada && level == 1)
8616                         reada_walk_down(trans, root, wc, path);
8617                 next = read_tree_block(fs_info, bytenr, generation, level - 1,
8618                                        &first_key);
8619                 if (IS_ERR(next)) {
8620                         return PTR_ERR(next);
8621                 } else if (!extent_buffer_uptodate(next)) {
8622                         free_extent_buffer(next);
8623                         return -EIO;
8624                 }
8625                 btrfs_tree_lock(next);
8626                 btrfs_set_lock_blocking(next);
8627         }
8628
8629         level--;
8630         ASSERT(level == btrfs_header_level(next));
8631         if (level != btrfs_header_level(next)) {
8632                 btrfs_err(root->fs_info, "mismatched level");
8633                 ret = -EIO;
8634                 goto out_unlock;
8635         }
8636         path->nodes[level] = next;
8637         path->slots[level] = 0;
8638         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8639         wc->level = level;
8640         if (wc->level == 1)
8641                 wc->reada_slot = 0;
8642         return 0;
8643 skip:
8644         wc->refs[level - 1] = 0;
8645         wc->flags[level - 1] = 0;
8646         if (wc->stage == DROP_REFERENCE) {
8647                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
8648                         parent = path->nodes[level]->start;
8649                 } else {
8650                         ASSERT(root->root_key.objectid ==
8651                                btrfs_header_owner(path->nodes[level]));
8652                         if (root->root_key.objectid !=
8653                             btrfs_header_owner(path->nodes[level])) {
8654                                 btrfs_err(root->fs_info,
8655                                                 "mismatched block owner");
8656                                 ret = -EIO;
8657                                 goto out_unlock;
8658                         }
8659                         parent = 0;
8660                 }
8661
8662                 if (need_account) {
8663                         ret = btrfs_qgroup_trace_subtree(trans, next,
8664                                                          generation, level - 1);
8665                         if (ret) {
8666                                 btrfs_err_rl(fs_info,
8667                                              "Error %d accounting shared subtree. Quota is out of sync, rescan required.",
8668                                              ret);
8669                         }
8670                 }
8671                 ret = btrfs_free_extent(trans, root, bytenr, blocksize,
8672                                         parent, root->root_key.objectid,
8673                                         level - 1, 0);
8674                 if (ret)
8675                         goto out_unlock;
8676         }
8677
8678         *lookup_info = 1;
8679         ret = 1;
8680
8681 out_unlock:
8682         btrfs_tree_unlock(next);
8683         free_extent_buffer(next);
8684
8685         return ret;
8686 }
8687
8688 /*
8689  * helper to process tree block while walking up the tree.
8690  *
8691  * when wc->stage == DROP_REFERENCE, this function drops
8692  * reference count on the block.
8693  *
8694  * when wc->stage == UPDATE_BACKREF, this function changes
8695  * wc->stage back to DROP_REFERENCE if we changed wc->stage
8696  * to UPDATE_BACKREF previously while processing the block.
8697  *
8698  * NOTE: return value 1 means we should stop walking up.
8699  */
8700 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
8701                                  struct btrfs_root *root,
8702                                  struct btrfs_path *path,
8703                                  struct walk_control *wc)
8704 {
8705         struct btrfs_fs_info *fs_info = root->fs_info;
8706         int ret;
8707         int level = wc->level;
8708         struct extent_buffer *eb = path->nodes[level];
8709         u64 parent = 0;
8710
8711         if (wc->stage == UPDATE_BACKREF) {
8712                 BUG_ON(wc->shared_level < level);
8713                 if (level < wc->shared_level)
8714                         goto out;
8715
8716                 ret = find_next_key(path, level + 1, &wc->update_progress);
8717                 if (ret > 0)
8718                         wc->update_ref = 0;
8719
8720                 wc->stage = DROP_REFERENCE;
8721                 wc->shared_level = -1;
8722                 path->slots[level] = 0;
8723
8724                 /*
8725                  * check reference count again if the block isn't locked.
8726                  * we should start walking down the tree again if reference
8727                  * count is one.
8728                  */
8729                 if (!path->locks[level]) {
8730                         BUG_ON(level == 0);
8731                         btrfs_tree_lock(eb);
8732                         btrfs_set_lock_blocking(eb);
8733                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8734
8735                         ret = btrfs_lookup_extent_info(trans, fs_info,
8736                                                        eb->start, level, 1,
8737                                                        &wc->refs[level],
8738                                                        &wc->flags[level]);
8739                         if (ret < 0) {
8740                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8741                                 path->locks[level] = 0;
8742                                 return ret;
8743                         }
8744                         BUG_ON(wc->refs[level] == 0);
8745                         if (wc->refs[level] == 1) {
8746                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8747                                 path->locks[level] = 0;
8748                                 return 1;
8749                         }
8750                 }
8751         }
8752
8753         /* wc->stage == DROP_REFERENCE */
8754         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
8755
8756         if (wc->refs[level] == 1) {
8757                 if (level == 0) {
8758                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8759                                 ret = btrfs_dec_ref(trans, root, eb, 1);
8760                         else
8761                                 ret = btrfs_dec_ref(trans, root, eb, 0);
8762                         BUG_ON(ret); /* -ENOMEM */
8763                         ret = btrfs_qgroup_trace_leaf_items(trans, eb);
8764                         if (ret) {
8765                                 btrfs_err_rl(fs_info,
8766                                              "error %d accounting leaf items. Quota is out of sync, rescan required.",
8767                                              ret);
8768                         }
8769                 }
8770                 /* make block locked assertion in clean_tree_block happy */
8771                 if (!path->locks[level] &&
8772                     btrfs_header_generation(eb) == trans->transid) {
8773                         btrfs_tree_lock(eb);
8774                         btrfs_set_lock_blocking(eb);
8775                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8776                 }
8777                 clean_tree_block(fs_info, eb);
8778         }
8779
8780         if (eb == root->node) {
8781                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8782                         parent = eb->start;
8783                 else
8784                         BUG_ON(root->root_key.objectid !=
8785                                btrfs_header_owner(eb));
8786         } else {
8787                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8788                         parent = path->nodes[level + 1]->start;
8789                 else
8790                         BUG_ON(root->root_key.objectid !=
8791                                btrfs_header_owner(path->nodes[level + 1]));
8792         }
8793
8794         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
8795 out:
8796         wc->refs[level] = 0;
8797         wc->flags[level] = 0;
8798         return 0;
8799 }
8800
8801 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
8802                                    struct btrfs_root *root,
8803                                    struct btrfs_path *path,
8804                                    struct walk_control *wc)
8805 {
8806         int level = wc->level;
8807         int lookup_info = 1;
8808         int ret;
8809
8810         while (level >= 0) {
8811                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
8812                 if (ret > 0)
8813                         break;
8814
8815                 if (level == 0)
8816                         break;
8817
8818                 if (path->slots[level] >=
8819                     btrfs_header_nritems(path->nodes[level]))
8820                         break;
8821
8822                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
8823                 if (ret > 0) {
8824                         path->slots[level]++;
8825                         continue;
8826                 } else if (ret < 0)
8827                         return ret;
8828                 level = wc->level;
8829         }
8830         return 0;
8831 }
8832
8833 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
8834                                  struct btrfs_root *root,
8835                                  struct btrfs_path *path,
8836                                  struct walk_control *wc, int max_level)
8837 {
8838         int level = wc->level;
8839         int ret;
8840
8841         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
8842         while (level < max_level && path->nodes[level]) {
8843                 wc->level = level;
8844                 if (path->slots[level] + 1 <
8845                     btrfs_header_nritems(path->nodes[level])) {
8846                         path->slots[level]++;
8847                         return 0;
8848                 } else {
8849                         ret = walk_up_proc(trans, root, path, wc);
8850                         if (ret > 0)
8851                                 return 0;
8852
8853                         if (path->locks[level]) {
8854                                 btrfs_tree_unlock_rw(path->nodes[level],
8855                                                      path->locks[level]);
8856                                 path->locks[level] = 0;
8857                         }
8858                         free_extent_buffer(path->nodes[level]);
8859                         path->nodes[level] = NULL;
8860                         level++;
8861                 }
8862         }
8863         return 1;
8864 }
8865
8866 /*
8867  * drop a subvolume tree.
8868  *
8869  * this function traverses the tree freeing any blocks that only
8870  * referenced by the tree.
8871  *
8872  * when a shared tree block is found. this function decreases its
8873  * reference count by one. if update_ref is true, this function
8874  * also make sure backrefs for the shared block and all lower level
8875  * blocks are properly updated.
8876  *
8877  * If called with for_reloc == 0, may exit early with -EAGAIN
8878  */
8879 int btrfs_drop_snapshot(struct btrfs_root *root,
8880                          struct btrfs_block_rsv *block_rsv, int update_ref,
8881                          int for_reloc)
8882 {
8883         struct btrfs_fs_info *fs_info = root->fs_info;
8884         struct btrfs_path *path;
8885         struct btrfs_trans_handle *trans;
8886         struct btrfs_root *tree_root = fs_info->tree_root;
8887         struct btrfs_root_item *root_item = &root->root_item;
8888         struct walk_control *wc;
8889         struct btrfs_key key;
8890         int err = 0;
8891         int ret;
8892         int level;
8893         bool root_dropped = false;
8894
8895         btrfs_debug(fs_info, "Drop subvolume %llu", root->objectid);
8896
8897         path = btrfs_alloc_path();
8898         if (!path) {
8899                 err = -ENOMEM;
8900                 goto out;
8901         }
8902
8903         wc = kzalloc(sizeof(*wc), GFP_NOFS);
8904         if (!wc) {
8905                 btrfs_free_path(path);
8906                 err = -ENOMEM;
8907                 goto out;
8908         }
8909
8910         trans = btrfs_start_transaction(tree_root, 0);
8911         if (IS_ERR(trans)) {
8912                 err = PTR_ERR(trans);
8913                 goto out_free;
8914         }
8915
8916         if (block_rsv)
8917                 trans->block_rsv = block_rsv;
8918
8919         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
8920                 level = btrfs_header_level(root->node);
8921                 path->nodes[level] = btrfs_lock_root_node(root);
8922                 btrfs_set_lock_blocking(path->nodes[level]);
8923                 path->slots[level] = 0;
8924                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8925                 memset(&wc->update_progress, 0,
8926                        sizeof(wc->update_progress));
8927         } else {
8928                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
8929                 memcpy(&wc->update_progress, &key,
8930                        sizeof(wc->update_progress));
8931
8932                 level = root_item->drop_level;
8933                 BUG_ON(level == 0);
8934                 path->lowest_level = level;
8935                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
8936                 path->lowest_level = 0;
8937                 if (ret < 0) {
8938                         err = ret;
8939                         goto out_end_trans;
8940                 }
8941                 WARN_ON(ret > 0);
8942
8943                 /*
8944                  * unlock our path, this is safe because only this
8945                  * function is allowed to delete this snapshot
8946                  */
8947                 btrfs_unlock_up_safe(path, 0);
8948
8949                 level = btrfs_header_level(root->node);
8950                 while (1) {
8951                         btrfs_tree_lock(path->nodes[level]);
8952                         btrfs_set_lock_blocking(path->nodes[level]);
8953                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8954
8955                         ret = btrfs_lookup_extent_info(trans, fs_info,
8956                                                 path->nodes[level]->start,
8957                                                 level, 1, &wc->refs[level],
8958                                                 &wc->flags[level]);
8959                         if (ret < 0) {
8960                                 err = ret;
8961                                 goto out_end_trans;
8962                         }
8963                         BUG_ON(wc->refs[level] == 0);
8964
8965                         if (level == root_item->drop_level)
8966                                 break;
8967
8968                         btrfs_tree_unlock(path->nodes[level]);
8969                         path->locks[level] = 0;
8970                         WARN_ON(wc->refs[level] != 1);
8971                         level--;
8972                 }
8973         }
8974
8975         wc->level = level;
8976         wc->shared_level = -1;
8977         wc->stage = DROP_REFERENCE;
8978         wc->update_ref = update_ref;
8979         wc->keep_locks = 0;
8980         wc->for_reloc = for_reloc;
8981         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
8982
8983         while (1) {
8984
8985                 ret = walk_down_tree(trans, root, path, wc);
8986                 if (ret < 0) {
8987                         err = ret;
8988                         break;
8989                 }
8990
8991                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
8992                 if (ret < 0) {
8993                         err = ret;
8994                         break;
8995                 }
8996
8997                 if (ret > 0) {
8998                         BUG_ON(wc->stage != DROP_REFERENCE);
8999                         break;
9000                 }
9001
9002                 if (wc->stage == DROP_REFERENCE) {
9003                         level = wc->level;
9004                         btrfs_node_key(path->nodes[level],
9005                                        &root_item->drop_progress,
9006                                        path->slots[level]);
9007                         root_item->drop_level = level;
9008                 }
9009
9010                 BUG_ON(wc->level == 0);
9011                 if (btrfs_should_end_transaction(trans) ||
9012                     (!for_reloc && btrfs_need_cleaner_sleep(fs_info))) {
9013                         ret = btrfs_update_root(trans, tree_root,
9014                                                 &root->root_key,
9015                                                 root_item);
9016                         if (ret) {
9017                                 btrfs_abort_transaction(trans, ret);
9018                                 err = ret;
9019                                 goto out_end_trans;
9020                         }
9021
9022                         btrfs_end_transaction_throttle(trans);
9023                         if (!for_reloc && btrfs_need_cleaner_sleep(fs_info)) {
9024                                 btrfs_debug(fs_info,
9025                                             "drop snapshot early exit");
9026                                 err = -EAGAIN;
9027                                 goto out_free;
9028                         }
9029
9030                         trans = btrfs_start_transaction(tree_root, 0);
9031                         if (IS_ERR(trans)) {
9032                                 err = PTR_ERR(trans);
9033                                 goto out_free;
9034                         }
9035                         if (block_rsv)
9036                                 trans->block_rsv = block_rsv;
9037                 }
9038         }
9039         btrfs_release_path(path);
9040         if (err)
9041                 goto out_end_trans;
9042
9043         ret = btrfs_del_root(trans, fs_info, &root->root_key);
9044         if (ret) {
9045                 btrfs_abort_transaction(trans, ret);
9046                 err = ret;
9047                 goto out_end_trans;
9048         }
9049
9050         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
9051                 ret = btrfs_find_root(tree_root, &root->root_key, path,
9052                                       NULL, NULL);
9053                 if (ret < 0) {
9054                         btrfs_abort_transaction(trans, ret);
9055                         err = ret;
9056                         goto out_end_trans;
9057                 } else if (ret > 0) {
9058                         /* if we fail to delete the orphan item this time
9059                          * around, it'll get picked up the next time.
9060                          *
9061                          * The most common failure here is just -ENOENT.
9062                          */
9063                         btrfs_del_orphan_item(trans, tree_root,
9064                                               root->root_key.objectid);
9065                 }
9066         }
9067
9068         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
9069                 btrfs_add_dropped_root(trans, root);
9070         } else {
9071                 free_extent_buffer(root->node);
9072                 free_extent_buffer(root->commit_root);
9073                 btrfs_put_fs_root(root);
9074         }
9075         root_dropped = true;
9076 out_end_trans:
9077         btrfs_end_transaction_throttle(trans);
9078 out_free:
9079         kfree(wc);
9080         btrfs_free_path(path);
9081 out:
9082         /*
9083          * So if we need to stop dropping the snapshot for whatever reason we
9084          * need to make sure to add it back to the dead root list so that we
9085          * keep trying to do the work later.  This also cleans up roots if we
9086          * don't have it in the radix (like when we recover after a power fail
9087          * or unmount) so we don't leak memory.
9088          */
9089         if (!for_reloc && !root_dropped)
9090                 btrfs_add_dead_root(root);
9091         if (err && err != -EAGAIN)
9092                 btrfs_handle_fs_error(fs_info, err, NULL);
9093         return err;
9094 }
9095
9096 /*
9097  * drop subtree rooted at tree block 'node'.
9098  *
9099  * NOTE: this function will unlock and release tree block 'node'
9100  * only used by relocation code
9101  */
9102 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
9103                         struct btrfs_root *root,
9104                         struct extent_buffer *node,
9105                         struct extent_buffer *parent)
9106 {
9107         struct btrfs_fs_info *fs_info = root->fs_info;
9108         struct btrfs_path *path;
9109         struct walk_control *wc;
9110         int level;
9111         int parent_level;
9112         int ret = 0;
9113         int wret;
9114
9115         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
9116
9117         path = btrfs_alloc_path();
9118         if (!path)
9119                 return -ENOMEM;
9120
9121         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9122         if (!wc) {
9123                 btrfs_free_path(path);
9124                 return -ENOMEM;
9125         }
9126
9127         btrfs_assert_tree_locked(parent);
9128         parent_level = btrfs_header_level(parent);
9129         extent_buffer_get(parent);
9130         path->nodes[parent_level] = parent;
9131         path->slots[parent_level] = btrfs_header_nritems(parent);
9132
9133         btrfs_assert_tree_locked(node);
9134         level = btrfs_header_level(node);
9135         path->nodes[level] = node;
9136         path->slots[level] = 0;
9137         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9138
9139         wc->refs[parent_level] = 1;
9140         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
9141         wc->level = level;
9142         wc->shared_level = -1;
9143         wc->stage = DROP_REFERENCE;
9144         wc->update_ref = 0;
9145         wc->keep_locks = 1;
9146         wc->for_reloc = 1;
9147         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
9148
9149         while (1) {
9150                 wret = walk_down_tree(trans, root, path, wc);
9151                 if (wret < 0) {
9152                         ret = wret;
9153                         break;
9154                 }
9155
9156                 wret = walk_up_tree(trans, root, path, wc, parent_level);
9157                 if (wret < 0)
9158                         ret = wret;
9159                 if (wret != 0)
9160                         break;
9161         }
9162
9163         kfree(wc);
9164         btrfs_free_path(path);
9165         return ret;
9166 }
9167
9168 static u64 update_block_group_flags(struct btrfs_fs_info *fs_info, u64 flags)
9169 {
9170         u64 num_devices;
9171         u64 stripped;
9172
9173         /*
9174          * if restripe for this chunk_type is on pick target profile and
9175          * return, otherwise do the usual balance
9176          */
9177         stripped = get_restripe_target(fs_info, flags);
9178         if (stripped)
9179                 return extended_to_chunk(stripped);
9180
9181         num_devices = fs_info->fs_devices->rw_devices;
9182
9183         stripped = BTRFS_BLOCK_GROUP_RAID0 |
9184                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
9185                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
9186
9187         if (num_devices == 1) {
9188                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9189                 stripped = flags & ~stripped;
9190
9191                 /* turn raid0 into single device chunks */
9192                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
9193                         return stripped;
9194
9195                 /* turn mirroring into duplication */
9196                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
9197                              BTRFS_BLOCK_GROUP_RAID10))
9198                         return stripped | BTRFS_BLOCK_GROUP_DUP;
9199         } else {
9200                 /* they already had raid on here, just return */
9201                 if (flags & stripped)
9202                         return flags;
9203
9204                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9205                 stripped = flags & ~stripped;
9206
9207                 /* switch duplicated blocks with raid1 */
9208                 if (flags & BTRFS_BLOCK_GROUP_DUP)
9209                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
9210
9211                 /* this is drive concat, leave it alone */
9212         }
9213
9214         return flags;
9215 }
9216
9217 static int inc_block_group_ro(struct btrfs_block_group_cache *cache, int force)
9218 {
9219         struct btrfs_space_info *sinfo = cache->space_info;
9220         u64 num_bytes;
9221         u64 min_allocable_bytes;
9222         int ret = -ENOSPC;
9223
9224         /*
9225          * We need some metadata space and system metadata space for
9226          * allocating chunks in some corner cases until we force to set
9227          * it to be readonly.
9228          */
9229         if ((sinfo->flags &
9230              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
9231             !force)
9232                 min_allocable_bytes = SZ_1M;
9233         else
9234                 min_allocable_bytes = 0;
9235
9236         spin_lock(&sinfo->lock);
9237         spin_lock(&cache->lock);
9238
9239         if (cache->ro) {
9240                 cache->ro++;
9241                 ret = 0;
9242                 goto out;
9243         }
9244
9245         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
9246                     cache->bytes_super - btrfs_block_group_used(&cache->item);
9247
9248         if (btrfs_space_info_used(sinfo, true) + num_bytes +
9249             min_allocable_bytes <= sinfo->total_bytes) {
9250                 sinfo->bytes_readonly += num_bytes;
9251                 cache->ro++;
9252                 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
9253                 ret = 0;
9254         }
9255 out:
9256         spin_unlock(&cache->lock);
9257         spin_unlock(&sinfo->lock);
9258         return ret;
9259 }
9260
9261 int btrfs_inc_block_group_ro(struct btrfs_block_group_cache *cache)
9262
9263 {
9264         struct btrfs_fs_info *fs_info = cache->fs_info;
9265         struct btrfs_trans_handle *trans;
9266         u64 alloc_flags;
9267         int ret;
9268
9269 again:
9270         trans = btrfs_join_transaction(fs_info->extent_root);
9271         if (IS_ERR(trans))
9272                 return PTR_ERR(trans);
9273
9274         /*
9275          * we're not allowed to set block groups readonly after the dirty
9276          * block groups cache has started writing.  If it already started,
9277          * back off and let this transaction commit
9278          */
9279         mutex_lock(&fs_info->ro_block_group_mutex);
9280         if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
9281                 u64 transid = trans->transid;
9282
9283                 mutex_unlock(&fs_info->ro_block_group_mutex);
9284                 btrfs_end_transaction(trans);
9285
9286                 ret = btrfs_wait_for_commit(fs_info, transid);
9287                 if (ret)
9288                         return ret;
9289                 goto again;
9290         }
9291
9292         /*
9293          * if we are changing raid levels, try to allocate a corresponding
9294          * block group with the new raid level.
9295          */
9296         alloc_flags = update_block_group_flags(fs_info, cache->flags);
9297         if (alloc_flags != cache->flags) {
9298                 ret = do_chunk_alloc(trans, alloc_flags,
9299                                      CHUNK_ALLOC_FORCE);
9300                 /*
9301                  * ENOSPC is allowed here, we may have enough space
9302                  * already allocated at the new raid level to
9303                  * carry on
9304                  */
9305                 if (ret == -ENOSPC)
9306                         ret = 0;
9307                 if (ret < 0)
9308                         goto out;
9309         }
9310
9311         ret = inc_block_group_ro(cache, 0);
9312         if (!ret)
9313                 goto out;
9314         alloc_flags = get_alloc_profile(fs_info, cache->space_info->flags);
9315         ret = do_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
9316         if (ret < 0)
9317                 goto out;
9318         ret = inc_block_group_ro(cache, 0);
9319 out:
9320         if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
9321                 alloc_flags = update_block_group_flags(fs_info, cache->flags);
9322                 mutex_lock(&fs_info->chunk_mutex);
9323                 check_system_chunk(trans, alloc_flags);
9324                 mutex_unlock(&fs_info->chunk_mutex);
9325         }
9326         mutex_unlock(&fs_info->ro_block_group_mutex);
9327
9328         btrfs_end_transaction(trans);
9329         return ret;
9330 }
9331
9332 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans, u64 type)
9333 {
9334         u64 alloc_flags = get_alloc_profile(trans->fs_info, type);
9335
9336         return do_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
9337 }
9338
9339 /*
9340  * helper to account the unused space of all the readonly block group in the
9341  * space_info. takes mirrors into account.
9342  */
9343 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
9344 {
9345         struct btrfs_block_group_cache *block_group;
9346         u64 free_bytes = 0;
9347         int factor;
9348
9349         /* It's df, we don't care if it's racy */
9350         if (list_empty(&sinfo->ro_bgs))
9351                 return 0;
9352
9353         spin_lock(&sinfo->lock);
9354         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
9355                 spin_lock(&block_group->lock);
9356
9357                 if (!block_group->ro) {
9358                         spin_unlock(&block_group->lock);
9359                         continue;
9360                 }
9361
9362                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
9363                                           BTRFS_BLOCK_GROUP_RAID10 |
9364                                           BTRFS_BLOCK_GROUP_DUP))
9365                         factor = 2;
9366                 else
9367                         factor = 1;
9368
9369                 free_bytes += (block_group->key.offset -
9370                                btrfs_block_group_used(&block_group->item)) *
9371                                factor;
9372
9373                 spin_unlock(&block_group->lock);
9374         }
9375         spin_unlock(&sinfo->lock);
9376
9377         return free_bytes;
9378 }
9379
9380 void btrfs_dec_block_group_ro(struct btrfs_block_group_cache *cache)
9381 {
9382         struct btrfs_space_info *sinfo = cache->space_info;
9383         u64 num_bytes;
9384
9385         BUG_ON(!cache->ro);
9386
9387         spin_lock(&sinfo->lock);
9388         spin_lock(&cache->lock);
9389         if (!--cache->ro) {
9390                 num_bytes = cache->key.offset - cache->reserved -
9391                             cache->pinned - cache->bytes_super -
9392                             btrfs_block_group_used(&cache->item);
9393                 sinfo->bytes_readonly -= num_bytes;
9394                 list_del_init(&cache->ro_list);
9395         }
9396         spin_unlock(&cache->lock);
9397         spin_unlock(&sinfo->lock);
9398 }
9399
9400 /*
9401  * checks to see if its even possible to relocate this block group.
9402  *
9403  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
9404  * ok to go ahead and try.
9405  */
9406 int btrfs_can_relocate(struct btrfs_fs_info *fs_info, u64 bytenr)
9407 {
9408         struct btrfs_root *root = fs_info->extent_root;
9409         struct btrfs_block_group_cache *block_group;
9410         struct btrfs_space_info *space_info;
9411         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
9412         struct btrfs_device *device;
9413         struct btrfs_trans_handle *trans;
9414         u64 min_free;
9415         u64 dev_min = 1;
9416         u64 dev_nr = 0;
9417         u64 target;
9418         int debug;
9419         int index;
9420         int full = 0;
9421         int ret = 0;
9422
9423         debug = btrfs_test_opt(fs_info, ENOSPC_DEBUG);
9424
9425         block_group = btrfs_lookup_block_group(fs_info, bytenr);
9426
9427         /* odd, couldn't find the block group, leave it alone */
9428         if (!block_group) {
9429                 if (debug)
9430                         btrfs_warn(fs_info,
9431                                    "can't find block group for bytenr %llu",
9432                                    bytenr);
9433                 return -1;
9434         }
9435
9436         min_free = btrfs_block_group_used(&block_group->item);
9437
9438         /* no bytes used, we're good */
9439         if (!min_free)
9440                 goto out;
9441
9442         space_info = block_group->space_info;
9443         spin_lock(&space_info->lock);
9444
9445         full = space_info->full;
9446
9447         /*
9448          * if this is the last block group we have in this space, we can't
9449          * relocate it unless we're able to allocate a new chunk below.
9450          *
9451          * Otherwise, we need to make sure we have room in the space to handle
9452          * all of the extents from this block group.  If we can, we're good
9453          */
9454         if ((space_info->total_bytes != block_group->key.offset) &&
9455             (btrfs_space_info_used(space_info, false) + min_free <
9456              space_info->total_bytes)) {
9457                 spin_unlock(&space_info->lock);
9458                 goto out;
9459         }
9460         spin_unlock(&space_info->lock);
9461
9462         /*
9463          * ok we don't have enough space, but maybe we have free space on our
9464          * devices to allocate new chunks for relocation, so loop through our
9465          * alloc devices and guess if we have enough space.  if this block
9466          * group is going to be restriped, run checks against the target
9467          * profile instead of the current one.
9468          */
9469         ret = -1;
9470
9471         /*
9472          * index:
9473          *      0: raid10
9474          *      1: raid1
9475          *      2: dup
9476          *      3: raid0
9477          *      4: single
9478          */
9479         target = get_restripe_target(fs_info, block_group->flags);
9480         if (target) {
9481                 index = btrfs_bg_flags_to_raid_index(extended_to_chunk(target));
9482         } else {
9483                 /*
9484                  * this is just a balance, so if we were marked as full
9485                  * we know there is no space for a new chunk
9486                  */
9487                 if (full) {
9488                         if (debug)
9489                                 btrfs_warn(fs_info,
9490                                            "no space to alloc new chunk for block group %llu",
9491                                            block_group->key.objectid);
9492                         goto out;
9493                 }
9494
9495                 index = btrfs_bg_flags_to_raid_index(block_group->flags);
9496         }
9497
9498         if (index == BTRFS_RAID_RAID10) {
9499                 dev_min = 4;
9500                 /* Divide by 2 */
9501                 min_free >>= 1;
9502         } else if (index == BTRFS_RAID_RAID1) {
9503                 dev_min = 2;
9504         } else if (index == BTRFS_RAID_DUP) {
9505                 /* Multiply by 2 */
9506                 min_free <<= 1;
9507         } else if (index == BTRFS_RAID_RAID0) {
9508                 dev_min = fs_devices->rw_devices;
9509                 min_free = div64_u64(min_free, dev_min);
9510         }
9511
9512         /* We need to do this so that we can look at pending chunks */
9513         trans = btrfs_join_transaction(root);
9514         if (IS_ERR(trans)) {
9515                 ret = PTR_ERR(trans);
9516                 goto out;
9517         }
9518
9519         mutex_lock(&fs_info->chunk_mutex);
9520         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
9521                 u64 dev_offset;
9522
9523                 /*
9524                  * check to make sure we can actually find a chunk with enough
9525                  * space to fit our block group in.
9526                  */
9527                 if (device->total_bytes > device->bytes_used + min_free &&
9528                     !test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state)) {
9529                         ret = find_free_dev_extent(trans, device, min_free,
9530                                                    &dev_offset, NULL);
9531                         if (!ret)
9532                                 dev_nr++;
9533
9534                         if (dev_nr >= dev_min)
9535                                 break;
9536
9537                         ret = -1;
9538                 }
9539         }
9540         if (debug && ret == -1)
9541                 btrfs_warn(fs_info,
9542                            "no space to allocate a new chunk for block group %llu",
9543                            block_group->key.objectid);
9544         mutex_unlock(&fs_info->chunk_mutex);
9545         btrfs_end_transaction(trans);
9546 out:
9547         btrfs_put_block_group(block_group);
9548         return ret;
9549 }
9550
9551 static int find_first_block_group(struct btrfs_fs_info *fs_info,
9552                                   struct btrfs_path *path,
9553                                   struct btrfs_key *key)
9554 {
9555         struct btrfs_root *root = fs_info->extent_root;
9556         int ret = 0;
9557         struct btrfs_key found_key;
9558         struct extent_buffer *leaf;
9559         int slot;
9560
9561         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
9562         if (ret < 0)
9563                 goto out;
9564
9565         while (1) {
9566                 slot = path->slots[0];
9567                 leaf = path->nodes[0];
9568                 if (slot >= btrfs_header_nritems(leaf)) {
9569                         ret = btrfs_next_leaf(root, path);
9570                         if (ret == 0)
9571                                 continue;
9572                         if (ret < 0)
9573                                 goto out;
9574                         break;
9575                 }
9576                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
9577
9578                 if (found_key.objectid >= key->objectid &&
9579                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
9580                         struct extent_map_tree *em_tree;
9581                         struct extent_map *em;
9582
9583                         em_tree = &root->fs_info->mapping_tree.map_tree;
9584                         read_lock(&em_tree->lock);
9585                         em = lookup_extent_mapping(em_tree, found_key.objectid,
9586                                                    found_key.offset);
9587                         read_unlock(&em_tree->lock);
9588                         if (!em) {
9589                                 btrfs_err(fs_info,
9590                         "logical %llu len %llu found bg but no related chunk",
9591                                           found_key.objectid, found_key.offset);
9592                                 ret = -ENOENT;
9593                         } else {
9594                                 ret = 0;
9595                         }
9596                         free_extent_map(em);
9597                         goto out;
9598                 }
9599                 path->slots[0]++;
9600         }
9601 out:
9602         return ret;
9603 }
9604
9605 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
9606 {
9607         struct btrfs_block_group_cache *block_group;
9608         u64 last = 0;
9609
9610         while (1) {
9611                 struct inode *inode;
9612
9613                 block_group = btrfs_lookup_first_block_group(info, last);
9614                 while (block_group) {
9615                         spin_lock(&block_group->lock);
9616                         if (block_group->iref)
9617                                 break;
9618                         spin_unlock(&block_group->lock);
9619                         block_group = next_block_group(info, block_group);
9620                 }
9621                 if (!block_group) {
9622                         if (last == 0)
9623                                 break;
9624                         last = 0;
9625                         continue;
9626                 }
9627
9628                 inode = block_group->inode;
9629                 block_group->iref = 0;
9630                 block_group->inode = NULL;
9631                 spin_unlock(&block_group->lock);
9632                 ASSERT(block_group->io_ctl.inode == NULL);
9633                 iput(inode);
9634                 last = block_group->key.objectid + block_group->key.offset;
9635                 btrfs_put_block_group(block_group);
9636         }
9637 }
9638
9639 /*
9640  * Must be called only after stopping all workers, since we could have block
9641  * group caching kthreads running, and therefore they could race with us if we
9642  * freed the block groups before stopping them.
9643  */
9644 int btrfs_free_block_groups(struct btrfs_fs_info *info)
9645 {
9646         struct btrfs_block_group_cache *block_group;
9647         struct btrfs_space_info *space_info;
9648         struct btrfs_caching_control *caching_ctl;
9649         struct rb_node *n;
9650
9651         down_write(&info->commit_root_sem);
9652         while (!list_empty(&info->caching_block_groups)) {
9653                 caching_ctl = list_entry(info->caching_block_groups.next,
9654                                          struct btrfs_caching_control, list);
9655                 list_del(&caching_ctl->list);
9656                 put_caching_control(caching_ctl);
9657         }
9658         up_write(&info->commit_root_sem);
9659
9660         spin_lock(&info->unused_bgs_lock);
9661         while (!list_empty(&info->unused_bgs)) {
9662                 block_group = list_first_entry(&info->unused_bgs,
9663                                                struct btrfs_block_group_cache,
9664                                                bg_list);
9665                 list_del_init(&block_group->bg_list);
9666                 btrfs_put_block_group(block_group);
9667         }
9668         spin_unlock(&info->unused_bgs_lock);
9669
9670         spin_lock(&info->block_group_cache_lock);
9671         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
9672                 block_group = rb_entry(n, struct btrfs_block_group_cache,
9673                                        cache_node);
9674                 rb_erase(&block_group->cache_node,
9675                          &info->block_group_cache_tree);
9676                 RB_CLEAR_NODE(&block_group->cache_node);
9677                 spin_unlock(&info->block_group_cache_lock);
9678
9679                 down_write(&block_group->space_info->groups_sem);
9680                 list_del(&block_group->list);
9681                 up_write(&block_group->space_info->groups_sem);
9682
9683                 /*
9684                  * We haven't cached this block group, which means we could
9685                  * possibly have excluded extents on this block group.
9686                  */
9687                 if (block_group->cached == BTRFS_CACHE_NO ||
9688                     block_group->cached == BTRFS_CACHE_ERROR)
9689                         free_excluded_extents(block_group);
9690
9691                 btrfs_remove_free_space_cache(block_group);
9692                 ASSERT(block_group->cached != BTRFS_CACHE_STARTED);
9693                 ASSERT(list_empty(&block_group->dirty_list));
9694                 ASSERT(list_empty(&block_group->io_list));
9695                 ASSERT(list_empty(&block_group->bg_list));
9696                 ASSERT(atomic_read(&block_group->count) == 1);
9697                 btrfs_put_block_group(block_group);
9698
9699                 spin_lock(&info->block_group_cache_lock);
9700         }
9701         spin_unlock(&info->block_group_cache_lock);
9702
9703         /* now that all the block groups are freed, go through and
9704          * free all the space_info structs.  This is only called during
9705          * the final stages of unmount, and so we know nobody is
9706          * using them.  We call synchronize_rcu() once before we start,
9707          * just to be on the safe side.
9708          */
9709         synchronize_rcu();
9710
9711         release_global_block_rsv(info);
9712
9713         while (!list_empty(&info->space_info)) {
9714                 int i;
9715
9716                 space_info = list_entry(info->space_info.next,
9717                                         struct btrfs_space_info,
9718                                         list);
9719
9720                 /*
9721                  * Do not hide this behind enospc_debug, this is actually
9722                  * important and indicates a real bug if this happens.
9723                  */
9724                 if (WARN_ON(space_info->bytes_pinned > 0 ||
9725                             space_info->bytes_reserved > 0 ||
9726                             space_info->bytes_may_use > 0))
9727                         dump_space_info(info, space_info, 0, 0);
9728                 list_del(&space_info->list);
9729                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
9730                         struct kobject *kobj;
9731                         kobj = space_info->block_group_kobjs[i];
9732                         space_info->block_group_kobjs[i] = NULL;
9733                         if (kobj) {
9734                                 kobject_del(kobj);
9735                                 kobject_put(kobj);
9736                         }
9737                 }
9738                 kobject_del(&space_info->kobj);
9739                 kobject_put(&space_info->kobj);
9740         }
9741         return 0;
9742 }
9743
9744 /* link_block_group will queue up kobjects to add when we're reclaim-safe */
9745 void btrfs_add_raid_kobjects(struct btrfs_fs_info *fs_info)
9746 {
9747         struct btrfs_space_info *space_info;
9748         struct raid_kobject *rkobj;
9749         LIST_HEAD(list);
9750         int index;
9751         int ret = 0;
9752
9753         spin_lock(&fs_info->pending_raid_kobjs_lock);
9754         list_splice_init(&fs_info->pending_raid_kobjs, &list);
9755         spin_unlock(&fs_info->pending_raid_kobjs_lock);
9756
9757         list_for_each_entry(rkobj, &list, list) {
9758                 space_info = __find_space_info(fs_info, rkobj->flags);
9759                 index = btrfs_bg_flags_to_raid_index(rkobj->flags);
9760
9761                 ret = kobject_add(&rkobj->kobj, &space_info->kobj,
9762                                   "%s", get_raid_name(index));
9763                 if (ret) {
9764                         kobject_put(&rkobj->kobj);
9765                         break;
9766                 }
9767         }
9768         if (ret)
9769                 btrfs_warn(fs_info,
9770                            "failed to add kobject for block cache, ignoring");
9771 }
9772
9773 static void link_block_group(struct btrfs_block_group_cache *cache)
9774 {
9775         struct btrfs_space_info *space_info = cache->space_info;
9776         struct btrfs_fs_info *fs_info = cache->fs_info;
9777         int index = btrfs_bg_flags_to_raid_index(cache->flags);
9778         bool first = false;
9779
9780         down_write(&space_info->groups_sem);
9781         if (list_empty(&space_info->block_groups[index]))
9782                 first = true;
9783         list_add_tail(&cache->list, &space_info->block_groups[index]);
9784         up_write(&space_info->groups_sem);
9785
9786         if (first) {
9787                 struct raid_kobject *rkobj = kzalloc(sizeof(*rkobj), GFP_NOFS);
9788                 if (!rkobj) {
9789                         btrfs_warn(cache->fs_info,
9790                                 "couldn't alloc memory for raid level kobject");
9791                         return;
9792                 }
9793                 rkobj->flags = cache->flags;
9794                 kobject_init(&rkobj->kobj, &btrfs_raid_ktype);
9795
9796                 spin_lock(&fs_info->pending_raid_kobjs_lock);
9797                 list_add_tail(&rkobj->list, &fs_info->pending_raid_kobjs);
9798                 spin_unlock(&fs_info->pending_raid_kobjs_lock);
9799                 space_info->block_group_kobjs[index] = &rkobj->kobj;
9800         }
9801 }
9802
9803 static struct btrfs_block_group_cache *
9804 btrfs_create_block_group_cache(struct btrfs_fs_info *fs_info,
9805                                u64 start, u64 size)
9806 {
9807         struct btrfs_block_group_cache *cache;
9808
9809         cache = kzalloc(sizeof(*cache), GFP_NOFS);
9810         if (!cache)
9811                 return NULL;
9812
9813         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
9814                                         GFP_NOFS);
9815         if (!cache->free_space_ctl) {
9816                 kfree(cache);
9817                 return NULL;
9818         }
9819
9820         cache->key.objectid = start;
9821         cache->key.offset = size;
9822         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
9823
9824         cache->fs_info = fs_info;
9825         cache->full_stripe_len = btrfs_full_stripe_len(fs_info, start);
9826         set_free_space_tree_thresholds(cache);
9827
9828         atomic_set(&cache->count, 1);
9829         spin_lock_init(&cache->lock);
9830         init_rwsem(&cache->data_rwsem);
9831         INIT_LIST_HEAD(&cache->list);
9832         INIT_LIST_HEAD(&cache->cluster_list);
9833         INIT_LIST_HEAD(&cache->bg_list);
9834         INIT_LIST_HEAD(&cache->ro_list);
9835         INIT_LIST_HEAD(&cache->dirty_list);
9836         INIT_LIST_HEAD(&cache->io_list);
9837         btrfs_init_free_space_ctl(cache);
9838         atomic_set(&cache->trimming, 0);
9839         mutex_init(&cache->free_space_lock);
9840         btrfs_init_full_stripe_locks_tree(&cache->full_stripe_locks_root);
9841
9842         return cache;
9843 }
9844
9845 int btrfs_read_block_groups(struct btrfs_fs_info *info)
9846 {
9847         struct btrfs_path *path;
9848         int ret;
9849         struct btrfs_block_group_cache *cache;
9850         struct btrfs_space_info *space_info;
9851         struct btrfs_key key;
9852         struct btrfs_key found_key;
9853         struct extent_buffer *leaf;
9854         int need_clear = 0;
9855         u64 cache_gen;
9856         u64 feature;
9857         int mixed;
9858
9859         feature = btrfs_super_incompat_flags(info->super_copy);
9860         mixed = !!(feature & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS);
9861
9862         key.objectid = 0;
9863         key.offset = 0;
9864         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
9865         path = btrfs_alloc_path();
9866         if (!path)
9867                 return -ENOMEM;
9868         path->reada = READA_FORWARD;
9869
9870         cache_gen = btrfs_super_cache_generation(info->super_copy);
9871         if (btrfs_test_opt(info, SPACE_CACHE) &&
9872             btrfs_super_generation(info->super_copy) != cache_gen)
9873                 need_clear = 1;
9874         if (btrfs_test_opt(info, CLEAR_CACHE))
9875                 need_clear = 1;
9876
9877         while (1) {
9878                 ret = find_first_block_group(info, path, &key);
9879                 if (ret > 0)
9880                         break;
9881                 if (ret != 0)
9882                         goto error;
9883
9884                 leaf = path->nodes[0];
9885                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
9886
9887                 cache = btrfs_create_block_group_cache(info, found_key.objectid,
9888                                                        found_key.offset);
9889                 if (!cache) {
9890                         ret = -ENOMEM;
9891                         goto error;
9892                 }
9893
9894                 if (need_clear) {
9895                         /*
9896                          * When we mount with old space cache, we need to
9897                          * set BTRFS_DC_CLEAR and set dirty flag.
9898                          *
9899                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
9900                          *    truncate the old free space cache inode and
9901                          *    setup a new one.
9902                          * b) Setting 'dirty flag' makes sure that we flush
9903                          *    the new space cache info onto disk.
9904                          */
9905                         if (btrfs_test_opt(info, SPACE_CACHE))
9906                                 cache->disk_cache_state = BTRFS_DC_CLEAR;
9907                 }
9908
9909                 read_extent_buffer(leaf, &cache->item,
9910                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
9911                                    sizeof(cache->item));
9912                 cache->flags = btrfs_block_group_flags(&cache->item);
9913                 if (!mixed &&
9914                     ((cache->flags & BTRFS_BLOCK_GROUP_METADATA) &&
9915                     (cache->flags & BTRFS_BLOCK_GROUP_DATA))) {
9916                         btrfs_err(info,
9917 "bg %llu is a mixed block group but filesystem hasn't enabled mixed block groups",
9918                                   cache->key.objectid);
9919                         ret = -EINVAL;
9920                         goto error;
9921                 }
9922
9923                 key.objectid = found_key.objectid + found_key.offset;
9924                 btrfs_release_path(path);
9925
9926                 /*
9927                  * We need to exclude the super stripes now so that the space
9928                  * info has super bytes accounted for, otherwise we'll think
9929                  * we have more space than we actually do.
9930                  */
9931                 ret = exclude_super_stripes(cache);
9932                 if (ret) {
9933                         /*
9934                          * We may have excluded something, so call this just in
9935                          * case.
9936                          */
9937                         free_excluded_extents(cache);
9938                         btrfs_put_block_group(cache);
9939                         goto error;
9940                 }
9941
9942                 /*
9943                  * check for two cases, either we are full, and therefore
9944                  * don't need to bother with the caching work since we won't
9945                  * find any space, or we are empty, and we can just add all
9946                  * the space in and be done with it.  This saves us _alot_ of
9947                  * time, particularly in the full case.
9948                  */
9949                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
9950                         cache->last_byte_to_unpin = (u64)-1;
9951                         cache->cached = BTRFS_CACHE_FINISHED;
9952                         free_excluded_extents(cache);
9953                 } else if (btrfs_block_group_used(&cache->item) == 0) {
9954                         cache->last_byte_to_unpin = (u64)-1;
9955                         cache->cached = BTRFS_CACHE_FINISHED;
9956                         add_new_free_space(cache, found_key.objectid,
9957                                            found_key.objectid +
9958                                            found_key.offset);
9959                         free_excluded_extents(cache);
9960                 }
9961
9962                 ret = btrfs_add_block_group_cache(info, cache);
9963                 if (ret) {
9964                         btrfs_remove_free_space_cache(cache);
9965                         btrfs_put_block_group(cache);
9966                         goto error;
9967                 }
9968
9969                 trace_btrfs_add_block_group(info, cache, 0);
9970                 update_space_info(info, cache->flags, found_key.offset,
9971                                   btrfs_block_group_used(&cache->item),
9972                                   cache->bytes_super, &space_info);
9973
9974                 cache->space_info = space_info;
9975
9976                 link_block_group(cache);
9977
9978                 set_avail_alloc_bits(info, cache->flags);
9979                 if (btrfs_chunk_readonly(info, cache->key.objectid)) {
9980                         inc_block_group_ro(cache, 1);
9981                 } else if (btrfs_block_group_used(&cache->item) == 0) {
9982                         ASSERT(list_empty(&cache->bg_list));
9983                         btrfs_mark_bg_unused(cache);
9984                 }
9985         }
9986
9987         list_for_each_entry_rcu(space_info, &info->space_info, list) {
9988                 if (!(get_alloc_profile(info, space_info->flags) &
9989                       (BTRFS_BLOCK_GROUP_RAID10 |
9990                        BTRFS_BLOCK_GROUP_RAID1 |
9991                        BTRFS_BLOCK_GROUP_RAID5 |
9992                        BTRFS_BLOCK_GROUP_RAID6 |
9993                        BTRFS_BLOCK_GROUP_DUP)))
9994                         continue;
9995                 /*
9996                  * avoid allocating from un-mirrored block group if there are
9997                  * mirrored block groups.
9998                  */
9999                 list_for_each_entry(cache,
10000                                 &space_info->block_groups[BTRFS_RAID_RAID0],
10001                                 list)
10002                         inc_block_group_ro(cache, 1);
10003                 list_for_each_entry(cache,
10004                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
10005                                 list)
10006                         inc_block_group_ro(cache, 1);
10007         }
10008
10009         btrfs_add_raid_kobjects(info);
10010         init_global_block_rsv(info);
10011         ret = 0;
10012 error:
10013         btrfs_free_path(path);
10014         return ret;
10015 }
10016
10017 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans)
10018 {
10019         struct btrfs_fs_info *fs_info = trans->fs_info;
10020         struct btrfs_block_group_cache *block_group, *tmp;
10021         struct btrfs_root *extent_root = fs_info->extent_root;
10022         struct btrfs_block_group_item item;
10023         struct btrfs_key key;
10024         int ret = 0;
10025         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
10026
10027         trans->can_flush_pending_bgs = false;
10028         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) {
10029                 if (ret)
10030                         goto next;
10031
10032                 spin_lock(&block_group->lock);
10033                 memcpy(&item, &block_group->item, sizeof(item));
10034                 memcpy(&key, &block_group->key, sizeof(key));
10035                 spin_unlock(&block_group->lock);
10036
10037                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
10038                                         sizeof(item));
10039                 if (ret)
10040                         btrfs_abort_transaction(trans, ret);
10041                 ret = btrfs_finish_chunk_alloc(trans, fs_info, key.objectid,
10042                                                key.offset);
10043                 if (ret)
10044                         btrfs_abort_transaction(trans, ret);
10045                 add_block_group_free_space(trans, block_group);
10046                 /* already aborted the transaction if it failed. */
10047 next:
10048                 list_del_init(&block_group->bg_list);
10049         }
10050         trans->can_flush_pending_bgs = can_flush_pending_bgs;
10051 }
10052
10053 int btrfs_make_block_group(struct btrfs_trans_handle *trans, u64 bytes_used,
10054                            u64 type, u64 chunk_offset, u64 size)
10055 {
10056         struct btrfs_fs_info *fs_info = trans->fs_info;
10057         struct btrfs_block_group_cache *cache;
10058         int ret;
10059
10060         btrfs_set_log_full_commit(fs_info, trans);
10061
10062         cache = btrfs_create_block_group_cache(fs_info, chunk_offset, size);
10063         if (!cache)
10064                 return -ENOMEM;
10065
10066         btrfs_set_block_group_used(&cache->item, bytes_used);
10067         btrfs_set_block_group_chunk_objectid(&cache->item,
10068                                              BTRFS_FIRST_CHUNK_TREE_OBJECTID);
10069         btrfs_set_block_group_flags(&cache->item, type);
10070
10071         cache->flags = type;
10072         cache->last_byte_to_unpin = (u64)-1;
10073         cache->cached = BTRFS_CACHE_FINISHED;
10074         cache->needs_free_space = 1;
10075         ret = exclude_super_stripes(cache);
10076         if (ret) {
10077                 /*
10078                  * We may have excluded something, so call this just in
10079                  * case.
10080                  */
10081                 free_excluded_extents(cache);
10082                 btrfs_put_block_group(cache);
10083                 return ret;
10084         }
10085
10086         add_new_free_space(cache, chunk_offset, chunk_offset + size);
10087
10088         free_excluded_extents(cache);
10089
10090 #ifdef CONFIG_BTRFS_DEBUG
10091         if (btrfs_should_fragment_free_space(cache)) {
10092                 u64 new_bytes_used = size - bytes_used;
10093
10094                 bytes_used += new_bytes_used >> 1;
10095                 fragment_free_space(cache);
10096         }
10097 #endif
10098         /*
10099          * Ensure the corresponding space_info object is created and
10100          * assigned to our block group. We want our bg to be added to the rbtree
10101          * with its ->space_info set.
10102          */
10103         cache->space_info = __find_space_info(fs_info, cache->flags);
10104         ASSERT(cache->space_info);
10105
10106         ret = btrfs_add_block_group_cache(fs_info, cache);
10107         if (ret) {
10108                 btrfs_remove_free_space_cache(cache);
10109                 btrfs_put_block_group(cache);
10110                 return ret;
10111         }
10112
10113         /*
10114          * Now that our block group has its ->space_info set and is inserted in
10115          * the rbtree, update the space info's counters.
10116          */
10117         trace_btrfs_add_block_group(fs_info, cache, 1);
10118         update_space_info(fs_info, cache->flags, size, bytes_used,
10119                                 cache->bytes_super, &cache->space_info);
10120         update_global_block_rsv(fs_info);
10121
10122         link_block_group(cache);
10123
10124         list_add_tail(&cache->bg_list, &trans->new_bgs);
10125
10126         set_avail_alloc_bits(fs_info, type);
10127         return 0;
10128 }
10129
10130 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
10131 {
10132         u64 extra_flags = chunk_to_extended(flags) &
10133                                 BTRFS_EXTENDED_PROFILE_MASK;
10134
10135         write_seqlock(&fs_info->profiles_lock);
10136         if (flags & BTRFS_BLOCK_GROUP_DATA)
10137                 fs_info->avail_data_alloc_bits &= ~extra_flags;
10138         if (flags & BTRFS_BLOCK_GROUP_METADATA)
10139                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
10140         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
10141                 fs_info->avail_system_alloc_bits &= ~extra_flags;
10142         write_sequnlock(&fs_info->profiles_lock);
10143 }
10144
10145 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
10146                              u64 group_start, struct extent_map *em)
10147 {
10148         struct btrfs_fs_info *fs_info = trans->fs_info;
10149         struct btrfs_root *root = fs_info->extent_root;
10150         struct btrfs_path *path;
10151         struct btrfs_block_group_cache *block_group;
10152         struct btrfs_free_cluster *cluster;
10153         struct btrfs_root *tree_root = fs_info->tree_root;
10154         struct btrfs_key key;
10155         struct inode *inode;
10156         struct kobject *kobj = NULL;
10157         int ret;
10158         int index;
10159         int factor;
10160         struct btrfs_caching_control *caching_ctl = NULL;
10161         bool remove_em;
10162
10163         block_group = btrfs_lookup_block_group(fs_info, group_start);
10164         BUG_ON(!block_group);
10165         BUG_ON(!block_group->ro);
10166
10167         trace_btrfs_remove_block_group(block_group);
10168         /*
10169          * Free the reserved super bytes from this block group before
10170          * remove it.
10171          */
10172         free_excluded_extents(block_group);
10173         btrfs_free_ref_tree_range(fs_info, block_group->key.objectid,
10174                                   block_group->key.offset);
10175
10176         memcpy(&key, &block_group->key, sizeof(key));
10177         index = btrfs_bg_flags_to_raid_index(block_group->flags);
10178         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
10179                                   BTRFS_BLOCK_GROUP_RAID1 |
10180                                   BTRFS_BLOCK_GROUP_RAID10))
10181                 factor = 2;
10182         else
10183                 factor = 1;
10184
10185         /* make sure this block group isn't part of an allocation cluster */
10186         cluster = &fs_info->data_alloc_cluster;
10187         spin_lock(&cluster->refill_lock);
10188         btrfs_return_cluster_to_free_space(block_group, cluster);
10189         spin_unlock(&cluster->refill_lock);
10190
10191         /*
10192          * make sure this block group isn't part of a metadata
10193          * allocation cluster
10194          */
10195         cluster = &fs_info->meta_alloc_cluster;
10196         spin_lock(&cluster->refill_lock);
10197         btrfs_return_cluster_to_free_space(block_group, cluster);
10198         spin_unlock(&cluster->refill_lock);
10199
10200         path = btrfs_alloc_path();
10201         if (!path) {
10202                 ret = -ENOMEM;
10203                 goto out;
10204         }
10205
10206         /*
10207          * get the inode first so any iput calls done for the io_list
10208          * aren't the final iput (no unlinks allowed now)
10209          */
10210         inode = lookup_free_space_inode(fs_info, block_group, path);
10211
10212         mutex_lock(&trans->transaction->cache_write_mutex);
10213         /*
10214          * make sure our free spache cache IO is done before remove the
10215          * free space inode
10216          */
10217         spin_lock(&trans->transaction->dirty_bgs_lock);
10218         if (!list_empty(&block_group->io_list)) {
10219                 list_del_init(&block_group->io_list);
10220
10221                 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
10222
10223                 spin_unlock(&trans->transaction->dirty_bgs_lock);
10224                 btrfs_wait_cache_io(trans, block_group, path);
10225                 btrfs_put_block_group(block_group);
10226                 spin_lock(&trans->transaction->dirty_bgs_lock);
10227         }
10228
10229         if (!list_empty(&block_group->dirty_list)) {
10230                 list_del_init(&block_group->dirty_list);
10231                 btrfs_put_block_group(block_group);
10232         }
10233         spin_unlock(&trans->transaction->dirty_bgs_lock);
10234         mutex_unlock(&trans->transaction->cache_write_mutex);
10235
10236         if (!IS_ERR(inode)) {
10237                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
10238                 if (ret) {
10239                         btrfs_add_delayed_iput(inode);
10240                         goto out;
10241                 }
10242                 clear_nlink(inode);
10243                 /* One for the block groups ref */
10244                 spin_lock(&block_group->lock);
10245                 if (block_group->iref) {
10246                         block_group->iref = 0;
10247                         block_group->inode = NULL;
10248                         spin_unlock(&block_group->lock);
10249                         iput(inode);
10250                 } else {
10251                         spin_unlock(&block_group->lock);
10252                 }
10253                 /* One for our lookup ref */
10254                 btrfs_add_delayed_iput(inode);
10255         }
10256
10257         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
10258         key.offset = block_group->key.objectid;
10259         key.type = 0;
10260
10261         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
10262         if (ret < 0)
10263                 goto out;
10264         if (ret > 0)
10265                 btrfs_release_path(path);
10266         if (ret == 0) {
10267                 ret = btrfs_del_item(trans, tree_root, path);
10268                 if (ret)
10269                         goto out;
10270                 btrfs_release_path(path);
10271         }
10272
10273         spin_lock(&fs_info->block_group_cache_lock);
10274         rb_erase(&block_group->cache_node,
10275                  &fs_info->block_group_cache_tree);
10276         RB_CLEAR_NODE(&block_group->cache_node);
10277
10278         if (fs_info->first_logical_byte == block_group->key.objectid)
10279                 fs_info->first_logical_byte = (u64)-1;
10280         spin_unlock(&fs_info->block_group_cache_lock);
10281
10282         down_write(&block_group->space_info->groups_sem);
10283         /*
10284          * we must use list_del_init so people can check to see if they
10285          * are still on the list after taking the semaphore
10286          */
10287         list_del_init(&block_group->list);
10288         if (list_empty(&block_group->space_info->block_groups[index])) {
10289                 kobj = block_group->space_info->block_group_kobjs[index];
10290                 block_group->space_info->block_group_kobjs[index] = NULL;
10291                 clear_avail_alloc_bits(fs_info, block_group->flags);
10292         }
10293         up_write(&block_group->space_info->groups_sem);
10294         if (kobj) {
10295                 kobject_del(kobj);
10296                 kobject_put(kobj);
10297         }
10298
10299         if (block_group->has_caching_ctl)
10300                 caching_ctl = get_caching_control(block_group);
10301         if (block_group->cached == BTRFS_CACHE_STARTED)
10302                 wait_block_group_cache_done(block_group);
10303         if (block_group->has_caching_ctl) {
10304                 down_write(&fs_info->commit_root_sem);
10305                 if (!caching_ctl) {
10306                         struct btrfs_caching_control *ctl;
10307
10308                         list_for_each_entry(ctl,
10309                                     &fs_info->caching_block_groups, list)
10310                                 if (ctl->block_group == block_group) {
10311                                         caching_ctl = ctl;
10312                                         refcount_inc(&caching_ctl->count);
10313                                         break;
10314                                 }
10315                 }
10316                 if (caching_ctl)
10317                         list_del_init(&caching_ctl->list);
10318                 up_write(&fs_info->commit_root_sem);
10319                 if (caching_ctl) {
10320                         /* Once for the caching bgs list and once for us. */
10321                         put_caching_control(caching_ctl);
10322                         put_caching_control(caching_ctl);
10323                 }
10324         }
10325
10326         spin_lock(&trans->transaction->dirty_bgs_lock);
10327         if (!list_empty(&block_group->dirty_list)) {
10328                 WARN_ON(1);
10329         }
10330         if (!list_empty(&block_group->io_list)) {
10331                 WARN_ON(1);
10332         }
10333         spin_unlock(&trans->transaction->dirty_bgs_lock);
10334         btrfs_remove_free_space_cache(block_group);
10335
10336         spin_lock(&block_group->space_info->lock);
10337         list_del_init(&block_group->ro_list);
10338
10339         if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
10340                 WARN_ON(block_group->space_info->total_bytes
10341                         < block_group->key.offset);
10342                 WARN_ON(block_group->space_info->bytes_readonly
10343                         < block_group->key.offset);
10344                 WARN_ON(block_group->space_info->disk_total
10345                         < block_group->key.offset * factor);
10346         }
10347         block_group->space_info->total_bytes -= block_group->key.offset;
10348         block_group->space_info->bytes_readonly -= block_group->key.offset;
10349         block_group->space_info->disk_total -= block_group->key.offset * factor;
10350
10351         spin_unlock(&block_group->space_info->lock);
10352
10353         memcpy(&key, &block_group->key, sizeof(key));
10354
10355         mutex_lock(&fs_info->chunk_mutex);
10356         if (!list_empty(&em->list)) {
10357                 /* We're in the transaction->pending_chunks list. */
10358                 free_extent_map(em);
10359         }
10360         spin_lock(&block_group->lock);
10361         block_group->removed = 1;
10362         /*
10363          * At this point trimming can't start on this block group, because we
10364          * removed the block group from the tree fs_info->block_group_cache_tree
10365          * so no one can't find it anymore and even if someone already got this
10366          * block group before we removed it from the rbtree, they have already
10367          * incremented block_group->trimming - if they didn't, they won't find
10368          * any free space entries because we already removed them all when we
10369          * called btrfs_remove_free_space_cache().
10370          *
10371          * And we must not remove the extent map from the fs_info->mapping_tree
10372          * to prevent the same logical address range and physical device space
10373          * ranges from being reused for a new block group. This is because our
10374          * fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
10375          * completely transactionless, so while it is trimming a range the
10376          * currently running transaction might finish and a new one start,
10377          * allowing for new block groups to be created that can reuse the same
10378          * physical device locations unless we take this special care.
10379          *
10380          * There may also be an implicit trim operation if the file system
10381          * is mounted with -odiscard. The same protections must remain
10382          * in place until the extents have been discarded completely when
10383          * the transaction commit has completed.
10384          */
10385         remove_em = (atomic_read(&block_group->trimming) == 0);
10386         /*
10387          * Make sure a trimmer task always sees the em in the pinned_chunks list
10388          * if it sees block_group->removed == 1 (needs to lock block_group->lock
10389          * before checking block_group->removed).
10390          */
10391         if (!remove_em) {
10392                 /*
10393                  * Our em might be in trans->transaction->pending_chunks which
10394                  * is protected by fs_info->chunk_mutex ([lock|unlock]_chunks),
10395                  * and so is the fs_info->pinned_chunks list.
10396                  *
10397                  * So at this point we must be holding the chunk_mutex to avoid
10398                  * any races with chunk allocation (more specifically at
10399                  * volumes.c:contains_pending_extent()), to ensure it always
10400                  * sees the em, either in the pending_chunks list or in the
10401                  * pinned_chunks list.
10402                  */
10403                 list_move_tail(&em->list, &fs_info->pinned_chunks);
10404         }
10405         spin_unlock(&block_group->lock);
10406
10407         if (remove_em) {
10408                 struct extent_map_tree *em_tree;
10409
10410                 em_tree = &fs_info->mapping_tree.map_tree;
10411                 write_lock(&em_tree->lock);
10412                 /*
10413                  * The em might be in the pending_chunks list, so make sure the
10414                  * chunk mutex is locked, since remove_extent_mapping() will
10415                  * delete us from that list.
10416                  */
10417                 remove_extent_mapping(em_tree, em);
10418                 write_unlock(&em_tree->lock);
10419                 /* once for the tree */
10420                 free_extent_map(em);
10421         }
10422
10423         mutex_unlock(&fs_info->chunk_mutex);
10424
10425         ret = remove_block_group_free_space(trans, block_group);
10426         if (ret)
10427                 goto out;
10428
10429         btrfs_put_block_group(block_group);
10430         btrfs_put_block_group(block_group);
10431
10432         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
10433         if (ret > 0)
10434                 ret = -EIO;
10435         if (ret < 0)
10436                 goto out;
10437
10438         ret = btrfs_del_item(trans, root, path);
10439 out:
10440         btrfs_free_path(path);
10441         return ret;
10442 }
10443
10444 struct btrfs_trans_handle *
10445 btrfs_start_trans_remove_block_group(struct btrfs_fs_info *fs_info,
10446                                      const u64 chunk_offset)
10447 {
10448         struct extent_map_tree *em_tree = &fs_info->mapping_tree.map_tree;
10449         struct extent_map *em;
10450         struct map_lookup *map;
10451         unsigned int num_items;
10452
10453         read_lock(&em_tree->lock);
10454         em = lookup_extent_mapping(em_tree, chunk_offset, 1);
10455         read_unlock(&em_tree->lock);
10456         ASSERT(em && em->start == chunk_offset);
10457
10458         /*
10459          * We need to reserve 3 + N units from the metadata space info in order
10460          * to remove a block group (done at btrfs_remove_chunk() and at
10461          * btrfs_remove_block_group()), which are used for:
10462          *
10463          * 1 unit for adding the free space inode's orphan (located in the tree
10464          * of tree roots).
10465          * 1 unit for deleting the block group item (located in the extent
10466          * tree).
10467          * 1 unit for deleting the free space item (located in tree of tree
10468          * roots).
10469          * N units for deleting N device extent items corresponding to each
10470          * stripe (located in the device tree).
10471          *
10472          * In order to remove a block group we also need to reserve units in the
10473          * system space info in order to update the chunk tree (update one or
10474          * more device items and remove one chunk item), but this is done at
10475          * btrfs_remove_chunk() through a call to check_system_chunk().
10476          */
10477         map = em->map_lookup;
10478         num_items = 3 + map->num_stripes;
10479         free_extent_map(em);
10480
10481         return btrfs_start_transaction_fallback_global_rsv(fs_info->extent_root,
10482                                                            num_items, 1);
10483 }
10484
10485 /*
10486  * Process the unused_bgs list and remove any that don't have any allocated
10487  * space inside of them.
10488  */
10489 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
10490 {
10491         struct btrfs_block_group_cache *block_group;
10492         struct btrfs_space_info *space_info;
10493         struct btrfs_trans_handle *trans;
10494         int ret = 0;
10495
10496         if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
10497                 return;
10498
10499         spin_lock(&fs_info->unused_bgs_lock);
10500         while (!list_empty(&fs_info->unused_bgs)) {
10501                 u64 start, end;
10502                 int trimming;
10503
10504                 block_group = list_first_entry(&fs_info->unused_bgs,
10505                                                struct btrfs_block_group_cache,
10506                                                bg_list);
10507                 list_del_init(&block_group->bg_list);
10508
10509                 space_info = block_group->space_info;
10510
10511                 if (ret || btrfs_mixed_space_info(space_info)) {
10512                         btrfs_put_block_group(block_group);
10513                         continue;
10514                 }
10515                 spin_unlock(&fs_info->unused_bgs_lock);
10516
10517                 mutex_lock(&fs_info->delete_unused_bgs_mutex);
10518
10519                 /* Don't want to race with allocators so take the groups_sem */
10520                 down_write(&space_info->groups_sem);
10521                 spin_lock(&block_group->lock);
10522                 if (block_group->reserved || block_group->pinned ||
10523                     btrfs_block_group_used(&block_group->item) ||
10524                     block_group->ro ||
10525                     list_is_singular(&block_group->list)) {
10526                         /*
10527                          * We want to bail if we made new allocations or have
10528                          * outstanding allocations in this block group.  We do
10529                          * the ro check in case balance is currently acting on
10530                          * this block group.
10531                          */
10532                         trace_btrfs_skip_unused_block_group(block_group);
10533                         spin_unlock(&block_group->lock);
10534                         up_write(&space_info->groups_sem);
10535                         goto next;
10536                 }
10537                 spin_unlock(&block_group->lock);
10538
10539                 /* We don't want to force the issue, only flip if it's ok. */
10540                 ret = inc_block_group_ro(block_group, 0);
10541                 up_write(&space_info->groups_sem);
10542                 if (ret < 0) {
10543                         ret = 0;
10544                         goto next;
10545                 }
10546
10547                 /*
10548                  * Want to do this before we do anything else so we can recover
10549                  * properly if we fail to join the transaction.
10550                  */
10551                 trans = btrfs_start_trans_remove_block_group(fs_info,
10552                                                      block_group->key.objectid);
10553                 if (IS_ERR(trans)) {
10554                         btrfs_dec_block_group_ro(block_group);
10555                         ret = PTR_ERR(trans);
10556                         goto next;
10557                 }
10558
10559                 /*
10560                  * We could have pending pinned extents for this block group,
10561                  * just delete them, we don't care about them anymore.
10562                  */
10563                 start = block_group->key.objectid;
10564                 end = start + block_group->key.offset - 1;
10565                 /*
10566                  * Hold the unused_bg_unpin_mutex lock to avoid racing with
10567                  * btrfs_finish_extent_commit(). If we are at transaction N,
10568                  * another task might be running finish_extent_commit() for the
10569                  * previous transaction N - 1, and have seen a range belonging
10570                  * to the block group in freed_extents[] before we were able to
10571                  * clear the whole block group range from freed_extents[]. This
10572                  * means that task can lookup for the block group after we
10573                  * unpinned it from freed_extents[] and removed it, leading to
10574                  * a BUG_ON() at btrfs_unpin_extent_range().
10575                  */
10576                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
10577                 ret = clear_extent_bits(&fs_info->freed_extents[0], start, end,
10578                                   EXTENT_DIRTY);
10579                 if (ret) {
10580                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10581                         btrfs_dec_block_group_ro(block_group);
10582                         goto end_trans;
10583                 }
10584                 ret = clear_extent_bits(&fs_info->freed_extents[1], start, end,
10585                                   EXTENT_DIRTY);
10586                 if (ret) {
10587                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10588                         btrfs_dec_block_group_ro(block_group);
10589                         goto end_trans;
10590                 }
10591                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10592
10593                 /* Reset pinned so btrfs_put_block_group doesn't complain */
10594                 spin_lock(&space_info->lock);
10595                 spin_lock(&block_group->lock);
10596
10597                 space_info->bytes_pinned -= block_group->pinned;
10598                 space_info->bytes_readonly += block_group->pinned;
10599                 percpu_counter_add_batch(&space_info->total_bytes_pinned,
10600                                    -block_group->pinned,
10601                                    BTRFS_TOTAL_BYTES_PINNED_BATCH);
10602                 block_group->pinned = 0;
10603
10604                 spin_unlock(&block_group->lock);
10605                 spin_unlock(&space_info->lock);
10606
10607                 /* DISCARD can flip during remount */
10608                 trimming = btrfs_test_opt(fs_info, DISCARD);
10609
10610                 /* Implicit trim during transaction commit. */
10611                 if (trimming)
10612                         btrfs_get_block_group_trimming(block_group);
10613
10614                 /*
10615                  * Btrfs_remove_chunk will abort the transaction if things go
10616                  * horribly wrong.
10617                  */
10618                 ret = btrfs_remove_chunk(trans, fs_info,
10619                                          block_group->key.objectid);
10620
10621                 if (ret) {
10622                         if (trimming)
10623                                 btrfs_put_block_group_trimming(block_group);
10624                         goto end_trans;
10625                 }
10626
10627                 /*
10628                  * If we're not mounted with -odiscard, we can just forget
10629                  * about this block group. Otherwise we'll need to wait
10630                  * until transaction commit to do the actual discard.
10631                  */
10632                 if (trimming) {
10633                         spin_lock(&fs_info->unused_bgs_lock);
10634                         /*
10635                          * A concurrent scrub might have added us to the list
10636                          * fs_info->unused_bgs, so use a list_move operation
10637                          * to add the block group to the deleted_bgs list.
10638                          */
10639                         list_move(&block_group->bg_list,
10640                                   &trans->transaction->deleted_bgs);
10641                         spin_unlock(&fs_info->unused_bgs_lock);
10642                         btrfs_get_block_group(block_group);
10643                 }
10644 end_trans:
10645                 btrfs_end_transaction(trans);
10646 next:
10647                 mutex_unlock(&fs_info->delete_unused_bgs_mutex);
10648                 btrfs_put_block_group(block_group);
10649                 spin_lock(&fs_info->unused_bgs_lock);
10650         }
10651         spin_unlock(&fs_info->unused_bgs_lock);
10652 }
10653
10654 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
10655 {
10656         struct btrfs_super_block *disk_super;
10657         u64 features;
10658         u64 flags;
10659         int mixed = 0;
10660         int ret;
10661
10662         disk_super = fs_info->super_copy;
10663         if (!btrfs_super_root(disk_super))
10664                 return -EINVAL;
10665
10666         features = btrfs_super_incompat_flags(disk_super);
10667         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
10668                 mixed = 1;
10669
10670         flags = BTRFS_BLOCK_GROUP_SYSTEM;
10671         ret = create_space_info(fs_info, flags);
10672         if (ret)
10673                 goto out;
10674
10675         if (mixed) {
10676                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
10677                 ret = create_space_info(fs_info, flags);
10678         } else {
10679                 flags = BTRFS_BLOCK_GROUP_METADATA;
10680                 ret = create_space_info(fs_info, flags);
10681                 if (ret)
10682                         goto out;
10683
10684                 flags = BTRFS_BLOCK_GROUP_DATA;
10685                 ret = create_space_info(fs_info, flags);
10686         }
10687 out:
10688         return ret;
10689 }
10690
10691 int btrfs_error_unpin_extent_range(struct btrfs_fs_info *fs_info,
10692                                    u64 start, u64 end)
10693 {
10694         return unpin_extent_range(fs_info, start, end, false);
10695 }
10696
10697 /*
10698  * It used to be that old block groups would be left around forever.
10699  * Iterating over them would be enough to trim unused space.  Since we
10700  * now automatically remove them, we also need to iterate over unallocated
10701  * space.
10702  *
10703  * We don't want a transaction for this since the discard may take a
10704  * substantial amount of time.  We don't require that a transaction be
10705  * running, but we do need to take a running transaction into account
10706  * to ensure that we're not discarding chunks that were released in
10707  * the current transaction.
10708  *
10709  * Holding the chunks lock will prevent other threads from allocating
10710  * or releasing chunks, but it won't prevent a running transaction
10711  * from committing and releasing the memory that the pending chunks
10712  * list head uses.  For that, we need to take a reference to the
10713  * transaction.
10714  */
10715 static int btrfs_trim_free_extents(struct btrfs_device *device,
10716                                    u64 minlen, u64 *trimmed)
10717 {
10718         u64 start = 0, len = 0;
10719         int ret;
10720
10721         *trimmed = 0;
10722
10723         /* Not writeable = nothing to do. */
10724         if (!test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state))
10725                 return 0;
10726
10727         /* No free space = nothing to do. */
10728         if (device->total_bytes <= device->bytes_used)
10729                 return 0;
10730
10731         ret = 0;
10732
10733         while (1) {
10734                 struct btrfs_fs_info *fs_info = device->fs_info;
10735                 struct btrfs_transaction *trans;
10736                 u64 bytes;
10737
10738                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
10739                 if (ret)
10740                         return ret;
10741
10742                 down_read(&fs_info->commit_root_sem);
10743
10744                 spin_lock(&fs_info->trans_lock);
10745                 trans = fs_info->running_transaction;
10746                 if (trans)
10747                         refcount_inc(&trans->use_count);
10748                 spin_unlock(&fs_info->trans_lock);
10749
10750                 ret = find_free_dev_extent_start(trans, device, minlen, start,
10751                                                  &start, &len);
10752                 if (trans)
10753                         btrfs_put_transaction(trans);
10754
10755                 if (ret) {
10756                         up_read(&fs_info->commit_root_sem);
10757                         mutex_unlock(&fs_info->chunk_mutex);
10758                         if (ret == -ENOSPC)
10759                                 ret = 0;
10760                         break;
10761                 }
10762
10763                 ret = btrfs_issue_discard(device->bdev, start, len, &bytes);
10764                 up_read(&fs_info->commit_root_sem);
10765                 mutex_unlock(&fs_info->chunk_mutex);
10766
10767                 if (ret)
10768                         break;
10769
10770                 start += len;
10771                 *trimmed += bytes;
10772
10773                 if (fatal_signal_pending(current)) {
10774                         ret = -ERESTARTSYS;
10775                         break;
10776                 }
10777
10778                 cond_resched();
10779         }
10780
10781         return ret;
10782 }
10783
10784 int btrfs_trim_fs(struct btrfs_fs_info *fs_info, struct fstrim_range *range)
10785 {
10786         struct btrfs_block_group_cache *cache = NULL;
10787         struct btrfs_device *device;
10788         struct list_head *devices;
10789         u64 group_trimmed;
10790         u64 start;
10791         u64 end;
10792         u64 trimmed = 0;
10793         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
10794         int ret = 0;
10795
10796         /*
10797          * try to trim all FS space, our block group may start from non-zero.
10798          */
10799         if (range->len == total_bytes)
10800                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
10801         else
10802                 cache = btrfs_lookup_block_group(fs_info, range->start);
10803
10804         while (cache) {
10805                 if (cache->key.objectid >= (range->start + range->len)) {
10806                         btrfs_put_block_group(cache);
10807                         break;
10808                 }
10809
10810                 start = max(range->start, cache->key.objectid);
10811                 end = min(range->start + range->len,
10812                                 cache->key.objectid + cache->key.offset);
10813
10814                 if (end - start >= range->minlen) {
10815                         if (!block_group_cache_done(cache)) {
10816                                 ret = cache_block_group(cache, 0);
10817                                 if (ret) {
10818                                         btrfs_put_block_group(cache);
10819                                         break;
10820                                 }
10821                                 ret = wait_block_group_cache_done(cache);
10822                                 if (ret) {
10823                                         btrfs_put_block_group(cache);
10824                                         break;
10825                                 }
10826                         }
10827                         ret = btrfs_trim_block_group(cache,
10828                                                      &group_trimmed,
10829                                                      start,
10830                                                      end,
10831                                                      range->minlen);
10832
10833                         trimmed += group_trimmed;
10834                         if (ret) {
10835                                 btrfs_put_block_group(cache);
10836                                 break;
10837                         }
10838                 }
10839
10840                 cache = next_block_group(fs_info, cache);
10841         }
10842
10843         mutex_lock(&fs_info->fs_devices->device_list_mutex);
10844         devices = &fs_info->fs_devices->alloc_list;
10845         list_for_each_entry(device, devices, dev_alloc_list) {
10846                 ret = btrfs_trim_free_extents(device, range->minlen,
10847                                               &group_trimmed);
10848                 if (ret)
10849                         break;
10850
10851                 trimmed += group_trimmed;
10852         }
10853         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
10854
10855         range->len = trimmed;
10856         return ret;
10857 }
10858
10859 /*
10860  * btrfs_{start,end}_write_no_snapshotting() are similar to
10861  * mnt_{want,drop}_write(), they are used to prevent some tasks from writing
10862  * data into the page cache through nocow before the subvolume is snapshoted,
10863  * but flush the data into disk after the snapshot creation, or to prevent
10864  * operations while snapshotting is ongoing and that cause the snapshot to be
10865  * inconsistent (writes followed by expanding truncates for example).
10866  */
10867 void btrfs_end_write_no_snapshotting(struct btrfs_root *root)
10868 {
10869         percpu_counter_dec(&root->subv_writers->counter);
10870         cond_wake_up(&root->subv_writers->wait);
10871 }
10872
10873 int btrfs_start_write_no_snapshotting(struct btrfs_root *root)
10874 {
10875         if (atomic_read(&root->will_be_snapshotted))
10876                 return 0;
10877
10878         percpu_counter_inc(&root->subv_writers->counter);
10879         /*
10880          * Make sure counter is updated before we check for snapshot creation.
10881          */
10882         smp_mb();
10883         if (atomic_read(&root->will_be_snapshotted)) {
10884                 btrfs_end_write_no_snapshotting(root);
10885                 return 0;
10886         }
10887         return 1;
10888 }
10889
10890 void btrfs_wait_for_snapshot_creation(struct btrfs_root *root)
10891 {
10892         while (true) {
10893                 int ret;
10894
10895                 ret = btrfs_start_write_no_snapshotting(root);
10896                 if (ret)
10897                         break;
10898                 wait_var_event(&root->will_be_snapshotted,
10899                                !atomic_read(&root->will_be_snapshotted));
10900         }
10901 }
10902
10903 void btrfs_mark_bg_unused(struct btrfs_block_group_cache *bg)
10904 {
10905         struct btrfs_fs_info *fs_info = bg->fs_info;
10906
10907         spin_lock(&fs_info->unused_bgs_lock);
10908         if (list_empty(&bg->bg_list)) {
10909                 btrfs_get_block_group(bg);
10910                 trace_btrfs_add_unused_block_group(bg);
10911                 list_add_tail(&bg->bg_list, &fs_info->unused_bgs);
10912         }
10913         spin_unlock(&fs_info->unused_bgs_lock);
10914 }