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
3143         path = btrfs_alloc_path();
3144         if (!path)
3145                 return -ENOMEM;
3146
3147         do {
3148                 ret = check_committed_ref(root, path, objectid,
3149                                           offset, bytenr);
3150                 if (ret && ret != -ENOENT)
3151                         goto out;
3152
3153                 ret = check_delayed_ref(root, path, objectid, offset, bytenr);
3154         } while (ret == -EAGAIN);
3155
3156 out:
3157         btrfs_free_path(path);
3158         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
3159                 WARN_ON(ret > 0);
3160         return ret;
3161 }
3162
3163 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
3164                            struct btrfs_root *root,
3165                            struct extent_buffer *buf,
3166                            int full_backref, int inc)
3167 {
3168         struct btrfs_fs_info *fs_info = root->fs_info;
3169         u64 bytenr;
3170         u64 num_bytes;
3171         u64 parent;
3172         u64 ref_root;
3173         u32 nritems;
3174         struct btrfs_key key;
3175         struct btrfs_file_extent_item *fi;
3176         int i;
3177         int level;
3178         int ret = 0;
3179         int (*process_func)(struct btrfs_trans_handle *,
3180                             struct btrfs_root *,
3181                             u64, u64, u64, u64, u64, u64);
3182
3183
3184         if (btrfs_is_testing(fs_info))
3185                 return 0;
3186
3187         ref_root = btrfs_header_owner(buf);
3188         nritems = btrfs_header_nritems(buf);
3189         level = btrfs_header_level(buf);
3190
3191         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state) && level == 0)
3192                 return 0;
3193
3194         if (inc)
3195                 process_func = btrfs_inc_extent_ref;
3196         else
3197                 process_func = btrfs_free_extent;
3198
3199         if (full_backref)
3200                 parent = buf->start;
3201         else
3202                 parent = 0;
3203
3204         for (i = 0; i < nritems; i++) {
3205                 if (level == 0) {
3206                         btrfs_item_key_to_cpu(buf, &key, i);
3207                         if (key.type != BTRFS_EXTENT_DATA_KEY)
3208                                 continue;
3209                         fi = btrfs_item_ptr(buf, i,
3210                                             struct btrfs_file_extent_item);
3211                         if (btrfs_file_extent_type(buf, fi) ==
3212                             BTRFS_FILE_EXTENT_INLINE)
3213                                 continue;
3214                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3215                         if (bytenr == 0)
3216                                 continue;
3217
3218                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3219                         key.offset -= btrfs_file_extent_offset(buf, fi);
3220                         ret = process_func(trans, root, bytenr, num_bytes,
3221                                            parent, ref_root, key.objectid,
3222                                            key.offset);
3223                         if (ret)
3224                                 goto fail;
3225                 } else {
3226                         bytenr = btrfs_node_blockptr(buf, i);
3227                         num_bytes = fs_info->nodesize;
3228                         ret = process_func(trans, root, bytenr, num_bytes,
3229                                            parent, ref_root, level - 1, 0);
3230                         if (ret)
3231                                 goto fail;
3232                 }
3233         }
3234         return 0;
3235 fail:
3236         return ret;
3237 }
3238
3239 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3240                   struct extent_buffer *buf, int full_backref)
3241 {
3242         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
3243 }
3244
3245 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3246                   struct extent_buffer *buf, int full_backref)
3247 {
3248         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
3249 }
3250
3251 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3252                                  struct btrfs_fs_info *fs_info,
3253                                  struct btrfs_path *path,
3254                                  struct btrfs_block_group_cache *cache)
3255 {
3256         int ret;
3257         struct btrfs_root *extent_root = fs_info->extent_root;
3258         unsigned long bi;
3259         struct extent_buffer *leaf;
3260
3261         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3262         if (ret) {
3263                 if (ret > 0)
3264                         ret = -ENOENT;
3265                 goto fail;
3266         }
3267
3268         leaf = path->nodes[0];
3269         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3270         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3271         btrfs_mark_buffer_dirty(leaf);
3272 fail:
3273         btrfs_release_path(path);
3274         return ret;
3275
3276 }
3277
3278 static struct btrfs_block_group_cache *
3279 next_block_group(struct btrfs_fs_info *fs_info,
3280                  struct btrfs_block_group_cache *cache)
3281 {
3282         struct rb_node *node;
3283
3284         spin_lock(&fs_info->block_group_cache_lock);
3285
3286         /* If our block group was removed, we need a full search. */
3287         if (RB_EMPTY_NODE(&cache->cache_node)) {
3288                 const u64 next_bytenr = cache->key.objectid + cache->key.offset;
3289
3290                 spin_unlock(&fs_info->block_group_cache_lock);
3291                 btrfs_put_block_group(cache);
3292                 cache = btrfs_lookup_first_block_group(fs_info, next_bytenr); return cache;
3293         }
3294         node = rb_next(&cache->cache_node);
3295         btrfs_put_block_group(cache);
3296         if (node) {
3297                 cache = rb_entry(node, struct btrfs_block_group_cache,
3298                                  cache_node);
3299                 btrfs_get_block_group(cache);
3300         } else
3301                 cache = NULL;
3302         spin_unlock(&fs_info->block_group_cache_lock);
3303         return cache;
3304 }
3305
3306 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3307                             struct btrfs_trans_handle *trans,
3308                             struct btrfs_path *path)
3309 {
3310         struct btrfs_fs_info *fs_info = block_group->fs_info;
3311         struct btrfs_root *root = fs_info->tree_root;
3312         struct inode *inode = NULL;
3313         struct extent_changeset *data_reserved = NULL;
3314         u64 alloc_hint = 0;
3315         int dcs = BTRFS_DC_ERROR;
3316         u64 num_pages = 0;
3317         int retries = 0;
3318         int ret = 0;
3319
3320         /*
3321          * If this block group is smaller than 100 megs don't bother caching the
3322          * block group.
3323          */
3324         if (block_group->key.offset < (100 * SZ_1M)) {
3325                 spin_lock(&block_group->lock);
3326                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3327                 spin_unlock(&block_group->lock);
3328                 return 0;
3329         }
3330
3331         if (trans->aborted)
3332                 return 0;
3333 again:
3334         inode = lookup_free_space_inode(fs_info, block_group, path);
3335         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3336                 ret = PTR_ERR(inode);
3337                 btrfs_release_path(path);
3338                 goto out;
3339         }
3340
3341         if (IS_ERR(inode)) {
3342                 BUG_ON(retries);
3343                 retries++;
3344
3345                 if (block_group->ro)
3346                         goto out_free;
3347
3348                 ret = create_free_space_inode(fs_info, trans, block_group,
3349                                               path);
3350                 if (ret)
3351                         goto out_free;
3352                 goto again;
3353         }
3354
3355         /*
3356          * We want to set the generation to 0, that way if anything goes wrong
3357          * from here on out we know not to trust this cache when we load up next
3358          * time.
3359          */
3360         BTRFS_I(inode)->generation = 0;
3361         ret = btrfs_update_inode(trans, root, inode);
3362         if (ret) {
3363                 /*
3364                  * So theoretically we could recover from this, simply set the
3365                  * super cache generation to 0 so we know to invalidate the
3366                  * cache, but then we'd have to keep track of the block groups
3367                  * that fail this way so we know we _have_ to reset this cache
3368                  * before the next commit or risk reading stale cache.  So to
3369                  * limit our exposure to horrible edge cases lets just abort the
3370                  * transaction, this only happens in really bad situations
3371                  * anyway.
3372                  */
3373                 btrfs_abort_transaction(trans, ret);
3374                 goto out_put;
3375         }
3376         WARN_ON(ret);
3377
3378         /* We've already setup this transaction, go ahead and exit */
3379         if (block_group->cache_generation == trans->transid &&
3380             i_size_read(inode)) {
3381                 dcs = BTRFS_DC_SETUP;
3382                 goto out_put;
3383         }
3384
3385         if (i_size_read(inode) > 0) {
3386                 ret = btrfs_check_trunc_cache_free_space(fs_info,
3387                                         &fs_info->global_block_rsv);
3388                 if (ret)
3389                         goto out_put;
3390
3391                 ret = btrfs_truncate_free_space_cache(trans, NULL, inode);
3392                 if (ret)
3393                         goto out_put;
3394         }
3395
3396         spin_lock(&block_group->lock);
3397         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3398             !btrfs_test_opt(fs_info, SPACE_CACHE)) {
3399                 /*
3400                  * don't bother trying to write stuff out _if_
3401                  * a) we're not cached,
3402                  * b) we're with nospace_cache mount option,
3403                  * c) we're with v2 space_cache (FREE_SPACE_TREE).
3404                  */
3405                 dcs = BTRFS_DC_WRITTEN;
3406                 spin_unlock(&block_group->lock);
3407                 goto out_put;
3408         }
3409         spin_unlock(&block_group->lock);
3410
3411         /*
3412          * We hit an ENOSPC when setting up the cache in this transaction, just
3413          * skip doing the setup, we've already cleared the cache so we're safe.
3414          */
3415         if (test_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags)) {
3416                 ret = -ENOSPC;
3417                 goto out_put;
3418         }
3419
3420         /*
3421          * Try to preallocate enough space based on how big the block group is.
3422          * Keep in mind this has to include any pinned space which could end up
3423          * taking up quite a bit since it's not folded into the other space
3424          * cache.
3425          */
3426         num_pages = div_u64(block_group->key.offset, SZ_256M);
3427         if (!num_pages)
3428                 num_pages = 1;
3429
3430         num_pages *= 16;
3431         num_pages *= PAGE_SIZE;
3432
3433         ret = btrfs_check_data_free_space(inode, &data_reserved, 0, num_pages);
3434         if (ret)
3435                 goto out_put;
3436
3437         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3438                                               num_pages, num_pages,
3439                                               &alloc_hint);
3440         /*
3441          * Our cache requires contiguous chunks so that we don't modify a bunch
3442          * of metadata or split extents when writing the cache out, which means
3443          * we can enospc if we are heavily fragmented in addition to just normal
3444          * out of space conditions.  So if we hit this just skip setting up any
3445          * other block groups for this transaction, maybe we'll unpin enough
3446          * space the next time around.
3447          */
3448         if (!ret)
3449                 dcs = BTRFS_DC_SETUP;
3450         else if (ret == -ENOSPC)
3451                 set_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags);
3452
3453 out_put:
3454         iput(inode);
3455 out_free:
3456         btrfs_release_path(path);
3457 out:
3458         spin_lock(&block_group->lock);
3459         if (!ret && dcs == BTRFS_DC_SETUP)
3460                 block_group->cache_generation = trans->transid;
3461         block_group->disk_cache_state = dcs;
3462         spin_unlock(&block_group->lock);
3463
3464         extent_changeset_free(data_reserved);
3465         return ret;
3466 }
3467
3468 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans,
3469                             struct btrfs_fs_info *fs_info)
3470 {
3471         struct btrfs_block_group_cache *cache, *tmp;
3472         struct btrfs_transaction *cur_trans = trans->transaction;
3473         struct btrfs_path *path;
3474
3475         if (list_empty(&cur_trans->dirty_bgs) ||
3476             !btrfs_test_opt(fs_info, SPACE_CACHE))
3477                 return 0;
3478
3479         path = btrfs_alloc_path();
3480         if (!path)
3481                 return -ENOMEM;
3482
3483         /* Could add new block groups, use _safe just in case */
3484         list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3485                                  dirty_list) {
3486                 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3487                         cache_save_setup(cache, trans, path);
3488         }
3489
3490         btrfs_free_path(path);
3491         return 0;
3492 }
3493
3494 /*
3495  * transaction commit does final block group cache writeback during a
3496  * critical section where nothing is allowed to change the FS.  This is
3497  * required in order for the cache to actually match the block group,
3498  * but can introduce a lot of latency into the commit.
3499  *
3500  * So, btrfs_start_dirty_block_groups is here to kick off block group
3501  * cache IO.  There's a chance we'll have to redo some of it if the
3502  * block group changes again during the commit, but it greatly reduces
3503  * the commit latency by getting rid of the easy block groups while
3504  * we're still allowing others to join the commit.
3505  */
3506 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans)
3507 {
3508         struct btrfs_fs_info *fs_info = trans->fs_info;
3509         struct btrfs_block_group_cache *cache;
3510         struct btrfs_transaction *cur_trans = trans->transaction;
3511         int ret = 0;
3512         int should_put;
3513         struct btrfs_path *path = NULL;
3514         LIST_HEAD(dirty);
3515         struct list_head *io = &cur_trans->io_bgs;
3516         int num_started = 0;
3517         int loops = 0;
3518
3519         spin_lock(&cur_trans->dirty_bgs_lock);
3520         if (list_empty(&cur_trans->dirty_bgs)) {
3521                 spin_unlock(&cur_trans->dirty_bgs_lock);
3522                 return 0;
3523         }
3524         list_splice_init(&cur_trans->dirty_bgs, &dirty);
3525         spin_unlock(&cur_trans->dirty_bgs_lock);
3526
3527 again:
3528         /*
3529          * make sure all the block groups on our dirty list actually
3530          * exist
3531          */
3532         btrfs_create_pending_block_groups(trans);
3533
3534         if (!path) {
3535                 path = btrfs_alloc_path();
3536                 if (!path)
3537                         return -ENOMEM;
3538         }
3539
3540         /*
3541          * cache_write_mutex is here only to save us from balance or automatic
3542          * removal of empty block groups deleting this block group while we are
3543          * writing out the cache
3544          */
3545         mutex_lock(&trans->transaction->cache_write_mutex);
3546         while (!list_empty(&dirty)) {
3547                 cache = list_first_entry(&dirty,
3548                                          struct btrfs_block_group_cache,
3549                                          dirty_list);
3550                 /*
3551                  * this can happen if something re-dirties a block
3552                  * group that is already under IO.  Just wait for it to
3553                  * finish and then do it all again
3554                  */
3555                 if (!list_empty(&cache->io_list)) {
3556                         list_del_init(&cache->io_list);
3557                         btrfs_wait_cache_io(trans, cache, path);
3558                         btrfs_put_block_group(cache);
3559                 }
3560
3561
3562                 /*
3563                  * btrfs_wait_cache_io uses the cache->dirty_list to decide
3564                  * if it should update the cache_state.  Don't delete
3565                  * until after we wait.
3566                  *
3567                  * Since we're not running in the commit critical section
3568                  * we need the dirty_bgs_lock to protect from update_block_group
3569                  */
3570                 spin_lock(&cur_trans->dirty_bgs_lock);
3571                 list_del_init(&cache->dirty_list);
3572                 spin_unlock(&cur_trans->dirty_bgs_lock);
3573
3574                 should_put = 1;
3575
3576                 cache_save_setup(cache, trans, path);
3577
3578                 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
3579                         cache->io_ctl.inode = NULL;
3580                         ret = btrfs_write_out_cache(fs_info, trans,
3581                                                     cache, path);
3582                         if (ret == 0 && cache->io_ctl.inode) {
3583                                 num_started++;
3584                                 should_put = 0;
3585
3586                                 /*
3587                                  * The cache_write_mutex is protecting the
3588                                  * io_list, also refer to the definition of
3589                                  * btrfs_transaction::io_bgs for more details
3590                                  */
3591                                 list_add_tail(&cache->io_list, io);
3592                         } else {
3593                                 /*
3594                                  * if we failed to write the cache, the
3595                                  * generation will be bad and life goes on
3596                                  */
3597                                 ret = 0;
3598                         }
3599                 }
3600                 if (!ret) {
3601                         ret = write_one_cache_group(trans, fs_info,
3602                                                     path, cache);
3603                         /*
3604                          * Our block group might still be attached to the list
3605                          * of new block groups in the transaction handle of some
3606                          * other task (struct btrfs_trans_handle->new_bgs). This
3607                          * means its block group item isn't yet in the extent
3608                          * tree. If this happens ignore the error, as we will
3609                          * try again later in the critical section of the
3610                          * transaction commit.
3611                          */
3612                         if (ret == -ENOENT) {
3613                                 ret = 0;
3614                                 spin_lock(&cur_trans->dirty_bgs_lock);
3615                                 if (list_empty(&cache->dirty_list)) {
3616                                         list_add_tail(&cache->dirty_list,
3617                                                       &cur_trans->dirty_bgs);
3618                                         btrfs_get_block_group(cache);
3619                                 }
3620                                 spin_unlock(&cur_trans->dirty_bgs_lock);
3621                         } else if (ret) {
3622                                 btrfs_abort_transaction(trans, ret);
3623                         }
3624                 }
3625
3626                 /* if its not on the io list, we need to put the block group */
3627                 if (should_put)
3628                         btrfs_put_block_group(cache);
3629
3630                 if (ret)
3631                         break;
3632
3633                 /*
3634                  * Avoid blocking other tasks for too long. It might even save
3635                  * us from writing caches for block groups that are going to be
3636                  * removed.
3637                  */
3638                 mutex_unlock(&trans->transaction->cache_write_mutex);
3639                 mutex_lock(&trans->transaction->cache_write_mutex);
3640         }
3641         mutex_unlock(&trans->transaction->cache_write_mutex);
3642
3643         /*
3644          * go through delayed refs for all the stuff we've just kicked off
3645          * and then loop back (just once)
3646          */
3647         ret = btrfs_run_delayed_refs(trans, 0);
3648         if (!ret && loops == 0) {
3649                 loops++;
3650                 spin_lock(&cur_trans->dirty_bgs_lock);
3651                 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3652                 /*
3653                  * dirty_bgs_lock protects us from concurrent block group
3654                  * deletes too (not just cache_write_mutex).
3655                  */
3656                 if (!list_empty(&dirty)) {
3657                         spin_unlock(&cur_trans->dirty_bgs_lock);
3658                         goto again;
3659                 }
3660                 spin_unlock(&cur_trans->dirty_bgs_lock);
3661         } else if (ret < 0) {
3662                 btrfs_cleanup_dirty_bgs(cur_trans, fs_info);
3663         }
3664
3665         btrfs_free_path(path);
3666         return ret;
3667 }
3668
3669 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3670                                    struct btrfs_fs_info *fs_info)
3671 {
3672         struct btrfs_block_group_cache *cache;
3673         struct btrfs_transaction *cur_trans = trans->transaction;
3674         int ret = 0;
3675         int should_put;
3676         struct btrfs_path *path;
3677         struct list_head *io = &cur_trans->io_bgs;
3678         int num_started = 0;
3679
3680         path = btrfs_alloc_path();
3681         if (!path)
3682                 return -ENOMEM;
3683
3684         /*
3685          * Even though we are in the critical section of the transaction commit,
3686          * we can still have concurrent tasks adding elements to this
3687          * transaction's list of dirty block groups. These tasks correspond to
3688          * endio free space workers started when writeback finishes for a
3689          * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
3690          * allocate new block groups as a result of COWing nodes of the root
3691          * tree when updating the free space inode. The writeback for the space
3692          * caches is triggered by an earlier call to
3693          * btrfs_start_dirty_block_groups() and iterations of the following
3694          * loop.
3695          * Also we want to do the cache_save_setup first and then run the
3696          * delayed refs to make sure we have the best chance at doing this all
3697          * in one shot.
3698          */
3699         spin_lock(&cur_trans->dirty_bgs_lock);
3700         while (!list_empty(&cur_trans->dirty_bgs)) {
3701                 cache = list_first_entry(&cur_trans->dirty_bgs,
3702                                          struct btrfs_block_group_cache,
3703                                          dirty_list);
3704
3705                 /*
3706                  * this can happen if cache_save_setup re-dirties a block
3707                  * group that is already under IO.  Just wait for it to
3708                  * finish and then do it all again
3709                  */
3710                 if (!list_empty(&cache->io_list)) {
3711                         spin_unlock(&cur_trans->dirty_bgs_lock);
3712                         list_del_init(&cache->io_list);
3713                         btrfs_wait_cache_io(trans, cache, path);
3714                         btrfs_put_block_group(cache);
3715                         spin_lock(&cur_trans->dirty_bgs_lock);
3716                 }
3717
3718                 /*
3719                  * don't remove from the dirty list until after we've waited
3720                  * on any pending IO
3721                  */
3722                 list_del_init(&cache->dirty_list);
3723                 spin_unlock(&cur_trans->dirty_bgs_lock);
3724                 should_put = 1;
3725
3726                 cache_save_setup(cache, trans, path);
3727
3728                 if (!ret)
3729                         ret = btrfs_run_delayed_refs(trans,
3730                                                      (unsigned long) -1);
3731
3732                 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3733                         cache->io_ctl.inode = NULL;
3734                         ret = btrfs_write_out_cache(fs_info, trans,
3735                                                     cache, path);
3736                         if (ret == 0 && cache->io_ctl.inode) {
3737                                 num_started++;
3738                                 should_put = 0;
3739                                 list_add_tail(&cache->io_list, io);
3740                         } else {
3741                                 /*
3742                                  * if we failed to write the cache, the
3743                                  * generation will be bad and life goes on
3744                                  */
3745                                 ret = 0;
3746                         }
3747                 }
3748                 if (!ret) {
3749                         ret = write_one_cache_group(trans, fs_info,
3750                                                     path, cache);
3751                         /*
3752                          * One of the free space endio workers might have
3753                          * created a new block group while updating a free space
3754                          * cache's inode (at inode.c:btrfs_finish_ordered_io())
3755                          * and hasn't released its transaction handle yet, in
3756                          * which case the new block group is still attached to
3757                          * its transaction handle and its creation has not
3758                          * finished yet (no block group item in the extent tree
3759                          * yet, etc). If this is the case, wait for all free
3760                          * space endio workers to finish and retry. This is a
3761                          * a very rare case so no need for a more efficient and
3762                          * complex approach.
3763                          */
3764                         if (ret == -ENOENT) {
3765                                 wait_event(cur_trans->writer_wait,
3766                                    atomic_read(&cur_trans->num_writers) == 1);
3767                                 ret = write_one_cache_group(trans, fs_info,
3768                                                             path, cache);
3769                         }
3770                         if (ret)
3771                                 btrfs_abort_transaction(trans, ret);
3772                 }
3773
3774                 /* if its not on the io list, we need to put the block group */
3775                 if (should_put)
3776                         btrfs_put_block_group(cache);
3777                 spin_lock(&cur_trans->dirty_bgs_lock);
3778         }
3779         spin_unlock(&cur_trans->dirty_bgs_lock);
3780
3781         /*
3782          * Refer to the definition of io_bgs member for details why it's safe
3783          * to use it without any locking
3784          */
3785         while (!list_empty(io)) {
3786                 cache = list_first_entry(io, struct btrfs_block_group_cache,
3787                                          io_list);
3788                 list_del_init(&cache->io_list);
3789                 btrfs_wait_cache_io(trans, cache, path);
3790                 btrfs_put_block_group(cache);
3791         }
3792
3793         btrfs_free_path(path);
3794         return ret;
3795 }
3796
3797 int btrfs_extent_readonly(struct btrfs_fs_info *fs_info, u64 bytenr)
3798 {
3799         struct btrfs_block_group_cache *block_group;
3800         int readonly = 0;
3801
3802         block_group = btrfs_lookup_block_group(fs_info, bytenr);
3803         if (!block_group || block_group->ro)
3804                 readonly = 1;
3805         if (block_group)
3806                 btrfs_put_block_group(block_group);
3807         return readonly;
3808 }
3809
3810 bool btrfs_inc_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3811 {
3812         struct btrfs_block_group_cache *bg;
3813         bool ret = true;
3814
3815         bg = btrfs_lookup_block_group(fs_info, bytenr);
3816         if (!bg)
3817                 return false;
3818
3819         spin_lock(&bg->lock);
3820         if (bg->ro)
3821                 ret = false;
3822         else
3823                 atomic_inc(&bg->nocow_writers);
3824         spin_unlock(&bg->lock);
3825
3826         /* no put on block group, done by btrfs_dec_nocow_writers */
3827         if (!ret)
3828                 btrfs_put_block_group(bg);
3829
3830         return ret;
3831
3832 }
3833
3834 void btrfs_dec_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3835 {
3836         struct btrfs_block_group_cache *bg;
3837
3838         bg = btrfs_lookup_block_group(fs_info, bytenr);
3839         ASSERT(bg);
3840         if (atomic_dec_and_test(&bg->nocow_writers))
3841                 wake_up_var(&bg->nocow_writers);
3842         /*
3843          * Once for our lookup and once for the lookup done by a previous call
3844          * to btrfs_inc_nocow_writers()
3845          */
3846         btrfs_put_block_group(bg);
3847         btrfs_put_block_group(bg);
3848 }
3849
3850 void btrfs_wait_nocow_writers(struct btrfs_block_group_cache *bg)
3851 {
3852         wait_var_event(&bg->nocow_writers, !atomic_read(&bg->nocow_writers));
3853 }
3854
3855 static const char *alloc_name(u64 flags)
3856 {
3857         switch (flags) {
3858         case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA:
3859                 return "mixed";
3860         case BTRFS_BLOCK_GROUP_METADATA:
3861                 return "metadata";
3862         case BTRFS_BLOCK_GROUP_DATA:
3863                 return "data";
3864         case BTRFS_BLOCK_GROUP_SYSTEM:
3865                 return "system";
3866         default:
3867                 WARN_ON(1);
3868                 return "invalid-combination";
3869         };
3870 }
3871
3872 static int create_space_info(struct btrfs_fs_info *info, u64 flags)
3873 {
3874
3875         struct btrfs_space_info *space_info;
3876         int i;
3877         int ret;
3878
3879         space_info = kzalloc(sizeof(*space_info), GFP_NOFS);
3880         if (!space_info)
3881                 return -ENOMEM;
3882
3883         ret = percpu_counter_init(&space_info->total_bytes_pinned, 0,
3884                                  GFP_KERNEL);
3885         if (ret) {
3886                 kfree(space_info);
3887                 return ret;
3888         }
3889
3890         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3891                 INIT_LIST_HEAD(&space_info->block_groups[i]);
3892         init_rwsem(&space_info->groups_sem);
3893         spin_lock_init(&space_info->lock);
3894         space_info->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3895         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3896         init_waitqueue_head(&space_info->wait);
3897         INIT_LIST_HEAD(&space_info->ro_bgs);
3898         INIT_LIST_HEAD(&space_info->tickets);
3899         INIT_LIST_HEAD(&space_info->priority_tickets);
3900
3901         ret = kobject_init_and_add(&space_info->kobj, &space_info_ktype,
3902                                     info->space_info_kobj, "%s",
3903                                     alloc_name(space_info->flags));
3904         if (ret) {
3905                 percpu_counter_destroy(&space_info->total_bytes_pinned);
3906                 kfree(space_info);
3907                 return ret;
3908         }
3909
3910         list_add_rcu(&space_info->list, &info->space_info);
3911         if (flags & BTRFS_BLOCK_GROUP_DATA)
3912                 info->data_sinfo = space_info;
3913
3914         return ret;
3915 }
3916
3917 static void update_space_info(struct btrfs_fs_info *info, u64 flags,
3918                              u64 total_bytes, u64 bytes_used,
3919                              u64 bytes_readonly,
3920                              struct btrfs_space_info **space_info)
3921 {
3922         struct btrfs_space_info *found;
3923         int factor;
3924
3925         factor = btrfs_bg_type_to_factor(flags);
3926
3927         found = __find_space_info(info, flags);
3928         ASSERT(found);
3929         spin_lock(&found->lock);
3930         found->total_bytes += total_bytes;
3931         found->disk_total += total_bytes * factor;
3932         found->bytes_used += bytes_used;
3933         found->disk_used += bytes_used * factor;
3934         found->bytes_readonly += bytes_readonly;
3935         if (total_bytes > 0)
3936                 found->full = 0;
3937         space_info_add_new_bytes(info, found, total_bytes -
3938                                  bytes_used - bytes_readonly);
3939         spin_unlock(&found->lock);
3940         *space_info = found;
3941 }
3942
3943 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3944 {
3945         u64 extra_flags = chunk_to_extended(flags) &
3946                                 BTRFS_EXTENDED_PROFILE_MASK;
3947
3948         write_seqlock(&fs_info->profiles_lock);
3949         if (flags & BTRFS_BLOCK_GROUP_DATA)
3950                 fs_info->avail_data_alloc_bits |= extra_flags;
3951         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3952                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3953         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3954                 fs_info->avail_system_alloc_bits |= extra_flags;
3955         write_sequnlock(&fs_info->profiles_lock);
3956 }
3957
3958 /*
3959  * returns target flags in extended format or 0 if restripe for this
3960  * chunk_type is not in progress
3961  *
3962  * should be called with balance_lock held
3963  */
3964 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
3965 {
3966         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3967         u64 target = 0;
3968
3969         if (!bctl)
3970                 return 0;
3971
3972         if (flags & BTRFS_BLOCK_GROUP_DATA &&
3973             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3974                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
3975         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
3976                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3977                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
3978         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
3979                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3980                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
3981         }
3982
3983         return target;
3984 }
3985
3986 /*
3987  * @flags: available profiles in extended format (see ctree.h)
3988  *
3989  * Returns reduced profile in chunk format.  If profile changing is in
3990  * progress (either running or paused) picks the target profile (if it's
3991  * already available), otherwise falls back to plain reducing.
3992  */
3993 static u64 btrfs_reduce_alloc_profile(struct btrfs_fs_info *fs_info, u64 flags)
3994 {
3995         u64 num_devices = fs_info->fs_devices->rw_devices;
3996         u64 target;
3997         u64 raid_type;
3998         u64 allowed = 0;
3999
4000         /*
4001          * see if restripe for this chunk_type is in progress, if so
4002          * try to reduce to the target profile
4003          */
4004         spin_lock(&fs_info->balance_lock);
4005         target = get_restripe_target(fs_info, flags);
4006         if (target) {
4007                 /* pick target profile only if it's already available */
4008                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
4009                         spin_unlock(&fs_info->balance_lock);
4010                         return extended_to_chunk(target);
4011                 }
4012         }
4013         spin_unlock(&fs_info->balance_lock);
4014
4015         /* First, mask out the RAID levels which aren't possible */
4016         for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
4017                 if (num_devices >= btrfs_raid_array[raid_type].devs_min)
4018                         allowed |= btrfs_raid_array[raid_type].bg_flag;
4019         }
4020         allowed &= flags;
4021
4022         if (allowed & BTRFS_BLOCK_GROUP_RAID6)
4023                 allowed = BTRFS_BLOCK_GROUP_RAID6;
4024         else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
4025                 allowed = BTRFS_BLOCK_GROUP_RAID5;
4026         else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
4027                 allowed = BTRFS_BLOCK_GROUP_RAID10;
4028         else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
4029                 allowed = BTRFS_BLOCK_GROUP_RAID1;
4030         else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
4031                 allowed = BTRFS_BLOCK_GROUP_RAID0;
4032
4033         flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
4034
4035         return extended_to_chunk(flags | allowed);
4036 }
4037
4038 static u64 get_alloc_profile(struct btrfs_fs_info *fs_info, u64 orig_flags)
4039 {
4040         unsigned seq;
4041         u64 flags;
4042
4043         do {
4044                 flags = orig_flags;
4045                 seq = read_seqbegin(&fs_info->profiles_lock);
4046
4047                 if (flags & BTRFS_BLOCK_GROUP_DATA)
4048                         flags |= fs_info->avail_data_alloc_bits;
4049                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4050                         flags |= fs_info->avail_system_alloc_bits;
4051                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
4052                         flags |= fs_info->avail_metadata_alloc_bits;
4053         } while (read_seqretry(&fs_info->profiles_lock, seq));
4054
4055         return btrfs_reduce_alloc_profile(fs_info, flags);
4056 }
4057
4058 static u64 get_alloc_profile_by_root(struct btrfs_root *root, int data)
4059 {
4060         struct btrfs_fs_info *fs_info = root->fs_info;
4061         u64 flags;
4062         u64 ret;
4063
4064         if (data)
4065                 flags = BTRFS_BLOCK_GROUP_DATA;
4066         else if (root == fs_info->chunk_root)
4067                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
4068         else
4069                 flags = BTRFS_BLOCK_GROUP_METADATA;
4070
4071         ret = get_alloc_profile(fs_info, flags);
4072         return ret;
4073 }
4074
4075 u64 btrfs_data_alloc_profile(struct btrfs_fs_info *fs_info)
4076 {
4077         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_DATA);
4078 }
4079
4080 u64 btrfs_metadata_alloc_profile(struct btrfs_fs_info *fs_info)
4081 {
4082         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4083 }
4084
4085 u64 btrfs_system_alloc_profile(struct btrfs_fs_info *fs_info)
4086 {
4087         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4088 }
4089
4090 static u64 btrfs_space_info_used(struct btrfs_space_info *s_info,
4091                                  bool may_use_included)
4092 {
4093         ASSERT(s_info);
4094         return s_info->bytes_used + s_info->bytes_reserved +
4095                 s_info->bytes_pinned + s_info->bytes_readonly +
4096                 (may_use_included ? s_info->bytes_may_use : 0);
4097 }
4098
4099 int btrfs_alloc_data_chunk_ondemand(struct btrfs_inode *inode, u64 bytes)
4100 {
4101         struct btrfs_root *root = inode->root;
4102         struct btrfs_fs_info *fs_info = root->fs_info;
4103         struct btrfs_space_info *data_sinfo = fs_info->data_sinfo;
4104         u64 used;
4105         int ret = 0;
4106         int need_commit = 2;
4107         int have_pinned_space;
4108
4109         /* make sure bytes are sectorsize aligned */
4110         bytes = ALIGN(bytes, fs_info->sectorsize);
4111
4112         if (btrfs_is_free_space_inode(inode)) {
4113                 need_commit = 0;
4114                 ASSERT(current->journal_info);
4115         }
4116
4117 again:
4118         /* make sure we have enough space to handle the data first */
4119         spin_lock(&data_sinfo->lock);
4120         used = btrfs_space_info_used(data_sinfo, true);
4121
4122         if (used + bytes > data_sinfo->total_bytes) {
4123                 struct btrfs_trans_handle *trans;
4124
4125                 /*
4126                  * if we don't have enough free bytes in this space then we need
4127                  * to alloc a new chunk.
4128                  */
4129                 if (!data_sinfo->full) {
4130                         u64 alloc_target;
4131
4132                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
4133                         spin_unlock(&data_sinfo->lock);
4134
4135                         alloc_target = btrfs_data_alloc_profile(fs_info);
4136                         /*
4137                          * It is ugly that we don't call nolock join
4138                          * transaction for the free space inode case here.
4139                          * But it is safe because we only do the data space
4140                          * reservation for the free space cache in the
4141                          * transaction context, the common join transaction
4142                          * just increase the counter of the current transaction
4143                          * handler, doesn't try to acquire the trans_lock of
4144                          * the fs.
4145                          */
4146                         trans = btrfs_join_transaction(root);
4147                         if (IS_ERR(trans))
4148                                 return PTR_ERR(trans);
4149
4150                         ret = do_chunk_alloc(trans, alloc_target,
4151                                              CHUNK_ALLOC_NO_FORCE);
4152                         btrfs_end_transaction(trans);
4153                         if (ret < 0) {
4154                                 if (ret != -ENOSPC)
4155                                         return ret;
4156                                 else {
4157                                         have_pinned_space = 1;
4158                                         goto commit_trans;
4159                                 }
4160                         }
4161
4162                         goto again;
4163                 }
4164
4165                 /*
4166                  * If we don't have enough pinned space to deal with this
4167                  * allocation, and no removed chunk in current transaction,
4168                  * don't bother committing the transaction.
4169                  */
4170                 have_pinned_space = __percpu_counter_compare(
4171                         &data_sinfo->total_bytes_pinned,
4172                         used + bytes - data_sinfo->total_bytes,
4173                         BTRFS_TOTAL_BYTES_PINNED_BATCH);
4174                 spin_unlock(&data_sinfo->lock);
4175
4176                 /* commit the current transaction and try again */
4177 commit_trans:
4178                 if (need_commit) {
4179                         need_commit--;
4180
4181                         if (need_commit > 0) {
4182                                 btrfs_start_delalloc_roots(fs_info, -1);
4183                                 btrfs_wait_ordered_roots(fs_info, U64_MAX, 0,
4184                                                          (u64)-1);
4185                         }
4186
4187                         trans = btrfs_join_transaction(root);
4188                         if (IS_ERR(trans))
4189                                 return PTR_ERR(trans);
4190                         if (have_pinned_space >= 0 ||
4191                             test_bit(BTRFS_TRANS_HAVE_FREE_BGS,
4192                                      &trans->transaction->flags) ||
4193                             need_commit > 0) {
4194                                 ret = btrfs_commit_transaction(trans);
4195                                 if (ret)
4196                                         return ret;
4197                                 /*
4198                                  * The cleaner kthread might still be doing iput
4199                                  * operations. Wait for it to finish so that
4200                                  * more space is released.
4201                                  */
4202                                 mutex_lock(&fs_info->cleaner_delayed_iput_mutex);
4203                                 mutex_unlock(&fs_info->cleaner_delayed_iput_mutex);
4204                                 goto again;
4205                         } else {
4206                                 btrfs_end_transaction(trans);
4207                         }
4208                 }
4209
4210                 trace_btrfs_space_reservation(fs_info,
4211                                               "space_info:enospc",
4212                                               data_sinfo->flags, bytes, 1);
4213                 return -ENOSPC;
4214         }
4215         data_sinfo->bytes_may_use += bytes;
4216         trace_btrfs_space_reservation(fs_info, "space_info",
4217                                       data_sinfo->flags, bytes, 1);
4218         spin_unlock(&data_sinfo->lock);
4219
4220         return 0;
4221 }
4222
4223 int btrfs_check_data_free_space(struct inode *inode,
4224                         struct extent_changeset **reserved, u64 start, u64 len)
4225 {
4226         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4227         int ret;
4228
4229         /* align the range */
4230         len = round_up(start + len, fs_info->sectorsize) -
4231               round_down(start, fs_info->sectorsize);
4232         start = round_down(start, fs_info->sectorsize);
4233
4234         ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode), len);
4235         if (ret < 0)
4236                 return ret;
4237
4238         /* Use new btrfs_qgroup_reserve_data to reserve precious data space. */
4239         ret = btrfs_qgroup_reserve_data(inode, reserved, start, len);
4240         if (ret < 0)
4241                 btrfs_free_reserved_data_space_noquota(inode, start, len);
4242         else
4243                 ret = 0;
4244         return ret;
4245 }
4246
4247 /*
4248  * Called if we need to clear a data reservation for this inode
4249  * Normally in a error case.
4250  *
4251  * This one will *NOT* use accurate qgroup reserved space API, just for case
4252  * which we can't sleep and is sure it won't affect qgroup reserved space.
4253  * Like clear_bit_hook().
4254  */
4255 void btrfs_free_reserved_data_space_noquota(struct inode *inode, u64 start,
4256                                             u64 len)
4257 {
4258         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4259         struct btrfs_space_info *data_sinfo;
4260
4261         /* Make sure the range is aligned to sectorsize */
4262         len = round_up(start + len, fs_info->sectorsize) -
4263               round_down(start, fs_info->sectorsize);
4264         start = round_down(start, fs_info->sectorsize);
4265
4266         data_sinfo = fs_info->data_sinfo;
4267         spin_lock(&data_sinfo->lock);
4268         if (WARN_ON(data_sinfo->bytes_may_use < len))
4269                 data_sinfo->bytes_may_use = 0;
4270         else
4271                 data_sinfo->bytes_may_use -= len;
4272         trace_btrfs_space_reservation(fs_info, "space_info",
4273                                       data_sinfo->flags, len, 0);
4274         spin_unlock(&data_sinfo->lock);
4275 }
4276
4277 /*
4278  * Called if we need to clear a data reservation for this inode
4279  * Normally in a error case.
4280  *
4281  * This one will handle the per-inode data rsv map for accurate reserved
4282  * space framework.
4283  */
4284 void btrfs_free_reserved_data_space(struct inode *inode,
4285                         struct extent_changeset *reserved, u64 start, u64 len)
4286 {
4287         struct btrfs_root *root = BTRFS_I(inode)->root;
4288
4289         /* Make sure the range is aligned to sectorsize */
4290         len = round_up(start + len, root->fs_info->sectorsize) -
4291               round_down(start, root->fs_info->sectorsize);
4292         start = round_down(start, root->fs_info->sectorsize);
4293
4294         btrfs_free_reserved_data_space_noquota(inode, start, len);
4295         btrfs_qgroup_free_data(inode, reserved, start, len);
4296 }
4297
4298 static void force_metadata_allocation(struct btrfs_fs_info *info)
4299 {
4300         struct list_head *head = &info->space_info;
4301         struct btrfs_space_info *found;
4302
4303         rcu_read_lock();
4304         list_for_each_entry_rcu(found, head, list) {
4305                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
4306                         found->force_alloc = CHUNK_ALLOC_FORCE;
4307         }
4308         rcu_read_unlock();
4309 }
4310
4311 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
4312 {
4313         return (global->size << 1);
4314 }
4315
4316 static int should_alloc_chunk(struct btrfs_fs_info *fs_info,
4317                               struct btrfs_space_info *sinfo, int force)
4318 {
4319         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4320         u64 bytes_used = btrfs_space_info_used(sinfo, false);
4321         u64 thresh;
4322
4323         if (force == CHUNK_ALLOC_FORCE)
4324                 return 1;
4325
4326         /*
4327          * We need to take into account the global rsv because for all intents
4328          * and purposes it's used space.  Don't worry about locking the
4329          * global_rsv, it doesn't change except when the transaction commits.
4330          */
4331         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
4332                 bytes_used += calc_global_rsv_need_space(global_rsv);
4333
4334         /*
4335          * in limited mode, we want to have some free space up to
4336          * about 1% of the FS size.
4337          */
4338         if (force == CHUNK_ALLOC_LIMITED) {
4339                 thresh = btrfs_super_total_bytes(fs_info->super_copy);
4340                 thresh = max_t(u64, SZ_64M, div_factor_fine(thresh, 1));
4341
4342                 if (sinfo->total_bytes - bytes_used < thresh)
4343                         return 1;
4344         }
4345
4346         if (bytes_used + SZ_2M < div_factor(sinfo->total_bytes, 8))
4347                 return 0;
4348         return 1;
4349 }
4350
4351 static u64 get_profile_num_devs(struct btrfs_fs_info *fs_info, u64 type)
4352 {
4353         u64 num_dev;
4354
4355         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
4356                     BTRFS_BLOCK_GROUP_RAID0 |
4357                     BTRFS_BLOCK_GROUP_RAID5 |
4358                     BTRFS_BLOCK_GROUP_RAID6))
4359                 num_dev = fs_info->fs_devices->rw_devices;
4360         else if (type & BTRFS_BLOCK_GROUP_RAID1)
4361                 num_dev = 2;
4362         else
4363                 num_dev = 1;    /* DUP or single */
4364
4365         return num_dev;
4366 }
4367
4368 /*
4369  * If @is_allocation is true, reserve space in the system space info necessary
4370  * for allocating a chunk, otherwise if it's false, reserve space necessary for
4371  * removing a chunk.
4372  */
4373 void check_system_chunk(struct btrfs_trans_handle *trans, u64 type)
4374 {
4375         struct btrfs_fs_info *fs_info = trans->fs_info;
4376         struct btrfs_space_info *info;
4377         u64 left;
4378         u64 thresh;
4379         int ret = 0;
4380         u64 num_devs;
4381
4382         /*
4383          * Needed because we can end up allocating a system chunk and for an
4384          * atomic and race free space reservation in the chunk block reserve.
4385          */
4386         lockdep_assert_held(&fs_info->chunk_mutex);
4387
4388         info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4389         spin_lock(&info->lock);
4390         left = info->total_bytes - btrfs_space_info_used(info, true);
4391         spin_unlock(&info->lock);
4392
4393         num_devs = get_profile_num_devs(fs_info, type);
4394
4395         /* num_devs device items to update and 1 chunk item to add or remove */
4396         thresh = btrfs_calc_trunc_metadata_size(fs_info, num_devs) +
4397                 btrfs_calc_trans_metadata_size(fs_info, 1);
4398
4399         if (left < thresh && btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
4400                 btrfs_info(fs_info, "left=%llu, need=%llu, flags=%llu",
4401                            left, thresh, type);
4402                 dump_space_info(fs_info, info, 0, 0);
4403         }
4404
4405         if (left < thresh) {
4406                 u64 flags = btrfs_system_alloc_profile(fs_info);
4407
4408                 /*
4409                  * Ignore failure to create system chunk. We might end up not
4410                  * needing it, as we might not need to COW all nodes/leafs from
4411                  * the paths we visit in the chunk tree (they were already COWed
4412                  * or created in the current transaction for example).
4413                  */
4414                 ret = btrfs_alloc_chunk(trans, flags);
4415         }
4416
4417         if (!ret) {
4418                 ret = btrfs_block_rsv_add(fs_info->chunk_root,
4419                                           &fs_info->chunk_block_rsv,
4420                                           thresh, BTRFS_RESERVE_NO_FLUSH);
4421                 if (!ret)
4422                         trans->chunk_bytes_reserved += thresh;
4423         }
4424 }
4425
4426 /*
4427  * If force is CHUNK_ALLOC_FORCE:
4428  *    - return 1 if it successfully allocates a chunk,
4429  *    - return errors including -ENOSPC otherwise.
4430  * If force is NOT CHUNK_ALLOC_FORCE:
4431  *    - return 0 if it doesn't need to allocate a new chunk,
4432  *    - return 1 if it successfully allocates a chunk,
4433  *    - return errors including -ENOSPC otherwise.
4434  */
4435 static int do_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags,
4436                           int force)
4437 {
4438         struct btrfs_fs_info *fs_info = trans->fs_info;
4439         struct btrfs_space_info *space_info;
4440         bool wait_for_alloc = false;
4441         bool should_alloc = false;
4442         int ret = 0;
4443
4444         /* Don't re-enter if we're already allocating a chunk */
4445         if (trans->allocating_chunk)
4446                 return -ENOSPC;
4447
4448         space_info = __find_space_info(fs_info, flags);
4449         ASSERT(space_info);
4450
4451         do {
4452                 spin_lock(&space_info->lock);
4453                 if (force < space_info->force_alloc)
4454                         force = space_info->force_alloc;
4455                 should_alloc = should_alloc_chunk(fs_info, space_info, force);
4456                 if (space_info->full) {
4457                         /* No more free physical space */
4458                         if (should_alloc)
4459                                 ret = -ENOSPC;
4460                         else
4461                                 ret = 0;
4462                         spin_unlock(&space_info->lock);
4463                         return ret;
4464                 } else if (!should_alloc) {
4465                         spin_unlock(&space_info->lock);
4466                         return 0;
4467                 } else if (space_info->chunk_alloc) {
4468                         /*
4469                          * Someone is already allocating, so we need to block
4470                          * until this someone is finished and then loop to
4471                          * recheck if we should continue with our allocation
4472                          * attempt.
4473                          */
4474                         wait_for_alloc = true;
4475                         spin_unlock(&space_info->lock);
4476                         mutex_lock(&fs_info->chunk_mutex);
4477                         mutex_unlock(&fs_info->chunk_mutex);
4478                 } else {
4479                         /* Proceed with allocation */
4480                         space_info->chunk_alloc = 1;
4481                         wait_for_alloc = false;
4482                         spin_unlock(&space_info->lock);
4483                 }
4484
4485                 cond_resched();
4486         } while (wait_for_alloc);
4487
4488         mutex_lock(&fs_info->chunk_mutex);
4489         trans->allocating_chunk = true;
4490
4491         /*
4492          * If we have mixed data/metadata chunks we want to make sure we keep
4493          * allocating mixed chunks instead of individual chunks.
4494          */
4495         if (btrfs_mixed_space_info(space_info))
4496                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4497
4498         /*
4499          * if we're doing a data chunk, go ahead and make sure that
4500          * we keep a reasonable number of metadata chunks allocated in the
4501          * FS as well.
4502          */
4503         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4504                 fs_info->data_chunk_allocations++;
4505                 if (!(fs_info->data_chunk_allocations %
4506                       fs_info->metadata_ratio))
4507                         force_metadata_allocation(fs_info);
4508         }
4509
4510         /*
4511          * Check if we have enough space in SYSTEM chunk because we may need
4512          * to update devices.
4513          */
4514         check_system_chunk(trans, flags);
4515
4516         ret = btrfs_alloc_chunk(trans, flags);
4517         trans->allocating_chunk = false;
4518
4519         spin_lock(&space_info->lock);
4520         if (ret < 0) {
4521                 if (ret == -ENOSPC)
4522                         space_info->full = 1;
4523                 else
4524                         goto out;
4525         } else {
4526                 ret = 1;
4527         }
4528
4529         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4530 out:
4531         space_info->chunk_alloc = 0;
4532         spin_unlock(&space_info->lock);
4533         mutex_unlock(&fs_info->chunk_mutex);
4534         /*
4535          * When we allocate a new chunk we reserve space in the chunk block
4536          * reserve to make sure we can COW nodes/leafs in the chunk tree or
4537          * add new nodes/leafs to it if we end up needing to do it when
4538          * inserting the chunk item and updating device items as part of the
4539          * second phase of chunk allocation, performed by
4540          * btrfs_finish_chunk_alloc(). So make sure we don't accumulate a
4541          * large number of new block groups to create in our transaction
4542          * handle's new_bgs list to avoid exhausting the chunk block reserve
4543          * in extreme cases - like having a single transaction create many new
4544          * block groups when starting to write out the free space caches of all
4545          * the block groups that were made dirty during the lifetime of the
4546          * transaction.
4547          */
4548         if (trans->can_flush_pending_bgs &&
4549             trans->chunk_bytes_reserved >= (u64)SZ_2M) {
4550                 btrfs_create_pending_block_groups(trans);
4551                 btrfs_trans_release_chunk_metadata(trans);
4552         }
4553         return ret;
4554 }
4555
4556 static int can_overcommit(struct btrfs_fs_info *fs_info,
4557                           struct btrfs_space_info *space_info, u64 bytes,
4558                           enum btrfs_reserve_flush_enum flush,
4559                           bool system_chunk)
4560 {
4561         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4562         u64 profile;
4563         u64 space_size;
4564         u64 avail;
4565         u64 used;
4566         int factor;
4567
4568         /* Don't overcommit when in mixed mode. */
4569         if (space_info->flags & BTRFS_BLOCK_GROUP_DATA)
4570                 return 0;
4571
4572         if (system_chunk)
4573                 profile = btrfs_system_alloc_profile(fs_info);
4574         else
4575                 profile = btrfs_metadata_alloc_profile(fs_info);
4576
4577         used = btrfs_space_info_used(space_info, false);
4578
4579         /*
4580          * We only want to allow over committing if we have lots of actual space
4581          * free, but if we don't have enough space to handle the global reserve
4582          * space then we could end up having a real enospc problem when trying
4583          * to allocate a chunk or some other such important allocation.
4584          */
4585         spin_lock(&global_rsv->lock);
4586         space_size = calc_global_rsv_need_space(global_rsv);
4587         spin_unlock(&global_rsv->lock);
4588         if (used + space_size >= space_info->total_bytes)
4589                 return 0;
4590
4591         used += space_info->bytes_may_use;
4592
4593         avail = atomic64_read(&fs_info->free_chunk_space);
4594
4595         /*
4596          * If we have dup, raid1 or raid10 then only half of the free
4597          * space is actually useable.  For raid56, the space info used
4598          * doesn't include the parity drive, so we don't have to
4599          * change the math
4600          */
4601         factor = btrfs_bg_type_to_factor(profile);
4602         avail = div_u64(avail, factor);
4603
4604         /*
4605          * If we aren't flushing all things, let us overcommit up to
4606          * 1/2th of the space. If we can flush, don't let us overcommit
4607          * too much, let it overcommit up to 1/8 of the space.
4608          */
4609         if (flush == BTRFS_RESERVE_FLUSH_ALL)
4610                 avail >>= 3;
4611         else
4612                 avail >>= 1;
4613
4614         if (used + bytes < space_info->total_bytes + avail)
4615                 return 1;
4616         return 0;
4617 }
4618
4619 static void btrfs_writeback_inodes_sb_nr(struct btrfs_fs_info *fs_info,
4620                                          unsigned long nr_pages, int nr_items)
4621 {
4622         struct super_block *sb = fs_info->sb;
4623
4624         if (down_read_trylock(&sb->s_umount)) {
4625                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
4626                 up_read(&sb->s_umount);
4627         } else {
4628                 /*
4629                  * We needn't worry the filesystem going from r/w to r/o though
4630                  * we don't acquire ->s_umount mutex, because the filesystem
4631                  * should guarantee the delalloc inodes list be empty after
4632                  * the filesystem is readonly(all dirty pages are written to
4633                  * the disk).
4634                  */
4635                 btrfs_start_delalloc_roots(fs_info, nr_items);
4636                 if (!current->journal_info)
4637                         btrfs_wait_ordered_roots(fs_info, nr_items, 0, (u64)-1);
4638         }
4639 }
4640
4641 static inline u64 calc_reclaim_items_nr(struct btrfs_fs_info *fs_info,
4642                                         u64 to_reclaim)
4643 {
4644         u64 bytes;
4645         u64 nr;
4646
4647         bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
4648         nr = div64_u64(to_reclaim, bytes);
4649         if (!nr)
4650                 nr = 1;
4651         return nr;
4652 }
4653
4654 #define EXTENT_SIZE_PER_ITEM    SZ_256K
4655
4656 /*
4657  * shrink metadata reservation for delalloc
4658  */
4659 static void shrink_delalloc(struct btrfs_fs_info *fs_info, u64 to_reclaim,
4660                             u64 orig, bool wait_ordered)
4661 {
4662         struct btrfs_space_info *space_info;
4663         struct btrfs_trans_handle *trans;
4664         u64 delalloc_bytes;
4665         u64 max_reclaim;
4666         u64 items;
4667         long time_left;
4668         unsigned long nr_pages;
4669         int loops;
4670
4671         /* Calc the number of the pages we need flush for space reservation */
4672         items = calc_reclaim_items_nr(fs_info, to_reclaim);
4673         to_reclaim = items * EXTENT_SIZE_PER_ITEM;
4674
4675         trans = (struct btrfs_trans_handle *)current->journal_info;
4676         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4677
4678         delalloc_bytes = percpu_counter_sum_positive(
4679                                                 &fs_info->delalloc_bytes);
4680         if (delalloc_bytes == 0) {
4681                 if (trans)
4682                         return;
4683                 if (wait_ordered)
4684                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
4685                 return;
4686         }
4687
4688         loops = 0;
4689         while (delalloc_bytes && loops < 3) {
4690                 max_reclaim = min(delalloc_bytes, to_reclaim);
4691                 nr_pages = max_reclaim >> PAGE_SHIFT;
4692                 btrfs_writeback_inodes_sb_nr(fs_info, nr_pages, items);
4693                 /*
4694                  * We need to wait for the async pages to actually start before
4695                  * we do anything.
4696                  */
4697                 max_reclaim = atomic_read(&fs_info->async_delalloc_pages);
4698                 if (!max_reclaim)
4699                         goto skip_async;
4700
4701                 if (max_reclaim <= nr_pages)
4702                         max_reclaim = 0;
4703                 else
4704                         max_reclaim -= nr_pages;
4705
4706                 wait_event(fs_info->async_submit_wait,
4707                            atomic_read(&fs_info->async_delalloc_pages) <=
4708                            (int)max_reclaim);
4709 skip_async:
4710                 spin_lock(&space_info->lock);
4711                 if (list_empty(&space_info->tickets) &&
4712                     list_empty(&space_info->priority_tickets)) {
4713                         spin_unlock(&space_info->lock);
4714                         break;
4715                 }
4716                 spin_unlock(&space_info->lock);
4717
4718                 loops++;
4719                 if (wait_ordered && !trans) {
4720                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
4721                 } else {
4722                         time_left = schedule_timeout_killable(1);
4723                         if (time_left)
4724                                 break;
4725                 }
4726                 delalloc_bytes = percpu_counter_sum_positive(
4727                                                 &fs_info->delalloc_bytes);
4728         }
4729 }
4730
4731 struct reserve_ticket {
4732         u64 bytes;
4733         int error;
4734         struct list_head list;
4735         wait_queue_head_t wait;
4736 };
4737
4738 /**
4739  * maybe_commit_transaction - possibly commit the transaction if its ok to
4740  * @root - the root we're allocating for
4741  * @bytes - the number of bytes we want to reserve
4742  * @force - force the commit
4743  *
4744  * This will check to make sure that committing the transaction will actually
4745  * get us somewhere and then commit the transaction if it does.  Otherwise it
4746  * will return -ENOSPC.
4747  */
4748 static int may_commit_transaction(struct btrfs_fs_info *fs_info,
4749                                   struct btrfs_space_info *space_info)
4750 {
4751         struct reserve_ticket *ticket = NULL;
4752         struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_block_rsv;
4753         struct btrfs_trans_handle *trans;
4754         u64 bytes;
4755
4756         trans = (struct btrfs_trans_handle *)current->journal_info;
4757         if (trans)
4758                 return -EAGAIN;
4759
4760         spin_lock(&space_info->lock);
4761         if (!list_empty(&space_info->priority_tickets))
4762                 ticket = list_first_entry(&space_info->priority_tickets,
4763                                           struct reserve_ticket, list);
4764         else if (!list_empty(&space_info->tickets))
4765                 ticket = list_first_entry(&space_info->tickets,
4766                                           struct reserve_ticket, list);
4767         bytes = (ticket) ? ticket->bytes : 0;
4768         spin_unlock(&space_info->lock);
4769
4770         if (!bytes)
4771                 return 0;
4772
4773         /* See if there is enough pinned space to make this reservation */
4774         if (__percpu_counter_compare(&space_info->total_bytes_pinned,
4775                                    bytes,
4776                                    BTRFS_TOTAL_BYTES_PINNED_BATCH) >= 0)
4777                 goto commit;
4778
4779         /*
4780          * See if there is some space in the delayed insertion reservation for
4781          * this reservation.
4782          */
4783         if (space_info != delayed_rsv->space_info)
4784                 return -ENOSPC;
4785
4786         spin_lock(&delayed_rsv->lock);
4787         if (delayed_rsv->size > bytes)
4788                 bytes = 0;
4789         else
4790                 bytes -= delayed_rsv->size;
4791         spin_unlock(&delayed_rsv->lock);
4792
4793         if (__percpu_counter_compare(&space_info->total_bytes_pinned,
4794                                    bytes,
4795                                    BTRFS_TOTAL_BYTES_PINNED_BATCH) < 0) {
4796                 return -ENOSPC;
4797         }
4798
4799 commit:
4800         trans = btrfs_join_transaction(fs_info->extent_root);
4801         if (IS_ERR(trans))
4802                 return -ENOSPC;
4803
4804         return btrfs_commit_transaction(trans);
4805 }
4806
4807 /*
4808  * Try to flush some data based on policy set by @state. This is only advisory
4809  * and may fail for various reasons. The caller is supposed to examine the
4810  * state of @space_info to detect the outcome.
4811  */
4812 static void flush_space(struct btrfs_fs_info *fs_info,
4813                        struct btrfs_space_info *space_info, u64 num_bytes,
4814                        int state)
4815 {
4816         struct btrfs_root *root = fs_info->extent_root;
4817         struct btrfs_trans_handle *trans;
4818         int nr;
4819         int ret = 0;
4820
4821         switch (state) {
4822         case FLUSH_DELAYED_ITEMS_NR:
4823         case FLUSH_DELAYED_ITEMS:
4824                 if (state == FLUSH_DELAYED_ITEMS_NR)
4825                         nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2;
4826                 else
4827                         nr = -1;
4828
4829                 trans = btrfs_join_transaction(root);
4830                 if (IS_ERR(trans)) {
4831                         ret = PTR_ERR(trans);
4832                         break;
4833                 }
4834                 ret = btrfs_run_delayed_items_nr(trans, nr);
4835                 btrfs_end_transaction(trans);
4836                 break;
4837         case FLUSH_DELALLOC:
4838         case FLUSH_DELALLOC_WAIT:
4839                 shrink_delalloc(fs_info, num_bytes * 2, num_bytes,
4840                                 state == FLUSH_DELALLOC_WAIT);
4841                 break;
4842         case ALLOC_CHUNK:
4843                 trans = btrfs_join_transaction(root);
4844                 if (IS_ERR(trans)) {
4845                         ret = PTR_ERR(trans);
4846                         break;
4847                 }
4848                 ret = do_chunk_alloc(trans,
4849                                      btrfs_metadata_alloc_profile(fs_info),
4850                                      CHUNK_ALLOC_NO_FORCE);
4851                 btrfs_end_transaction(trans);
4852                 if (ret > 0 || ret == -ENOSPC)
4853                         ret = 0;
4854                 break;
4855         case COMMIT_TRANS:
4856                 ret = may_commit_transaction(fs_info, space_info);
4857                 break;
4858         default:
4859                 ret = -ENOSPC;
4860                 break;
4861         }
4862
4863         trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state,
4864                                 ret);
4865         return;
4866 }
4867
4868 static inline u64
4869 btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info,
4870                                  struct btrfs_space_info *space_info,
4871                                  bool system_chunk)
4872 {
4873         struct reserve_ticket *ticket;
4874         u64 used;
4875         u64 expected;
4876         u64 to_reclaim = 0;
4877
4878         list_for_each_entry(ticket, &space_info->tickets, list)
4879                 to_reclaim += ticket->bytes;
4880         list_for_each_entry(ticket, &space_info->priority_tickets, list)
4881                 to_reclaim += ticket->bytes;
4882         if (to_reclaim)
4883                 return to_reclaim;
4884
4885         to_reclaim = min_t(u64, num_online_cpus() * SZ_1M, SZ_16M);
4886         if (can_overcommit(fs_info, space_info, to_reclaim,
4887                            BTRFS_RESERVE_FLUSH_ALL, system_chunk))
4888                 return 0;
4889
4890         used = btrfs_space_info_used(space_info, true);
4891
4892         if (can_overcommit(fs_info, space_info, SZ_1M,
4893                            BTRFS_RESERVE_FLUSH_ALL, system_chunk))
4894                 expected = div_factor_fine(space_info->total_bytes, 95);
4895         else
4896                 expected = div_factor_fine(space_info->total_bytes, 90);
4897
4898         if (used > expected)
4899                 to_reclaim = used - expected;
4900         else
4901                 to_reclaim = 0;
4902         to_reclaim = min(to_reclaim, space_info->bytes_may_use +
4903                                      space_info->bytes_reserved);
4904         return to_reclaim;
4905 }
4906
4907 static inline int need_do_async_reclaim(struct btrfs_fs_info *fs_info,
4908                                         struct btrfs_space_info *space_info,
4909                                         u64 used, bool system_chunk)
4910 {
4911         u64 thresh = div_factor_fine(space_info->total_bytes, 98);
4912
4913         /* If we're just plain full then async reclaim just slows us down. */
4914         if ((space_info->bytes_used + space_info->bytes_reserved) >= thresh)
4915                 return 0;
4916
4917         if (!btrfs_calc_reclaim_metadata_size(fs_info, space_info,
4918                                               system_chunk))
4919                 return 0;
4920
4921         return (used >= thresh && !btrfs_fs_closing(fs_info) &&
4922                 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
4923 }
4924
4925 static void wake_all_tickets(struct list_head *head)
4926 {
4927         struct reserve_ticket *ticket;
4928
4929         while (!list_empty(head)) {
4930                 ticket = list_first_entry(head, struct reserve_ticket, list);
4931                 list_del_init(&ticket->list);
4932                 ticket->error = -ENOSPC;
4933                 wake_up(&ticket->wait);
4934         }
4935 }
4936
4937 /*
4938  * This is for normal flushers, we can wait all goddamned day if we want to.  We
4939  * will loop and continuously try to flush as long as we are making progress.
4940  * We count progress as clearing off tickets each time we have to loop.
4941  */
4942 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
4943 {
4944         struct btrfs_fs_info *fs_info;
4945         struct btrfs_space_info *space_info;
4946         u64 to_reclaim;
4947         int flush_state;
4948         int commit_cycles = 0;
4949         u64 last_tickets_id;
4950
4951         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
4952         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4953
4954         spin_lock(&space_info->lock);
4955         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info,
4956                                                       false);
4957         if (!to_reclaim) {
4958                 space_info->flush = 0;
4959                 spin_unlock(&space_info->lock);
4960                 return;
4961         }
4962         last_tickets_id = space_info->tickets_id;
4963         spin_unlock(&space_info->lock);
4964
4965         flush_state = FLUSH_DELAYED_ITEMS_NR;
4966         do {
4967                 flush_space(fs_info, space_info, to_reclaim, flush_state);
4968                 spin_lock(&space_info->lock);
4969                 if (list_empty(&space_info->tickets)) {
4970                         space_info->flush = 0;
4971                         spin_unlock(&space_info->lock);
4972                         return;
4973                 }
4974                 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info,
4975                                                               space_info,
4976                                                               false);
4977                 if (last_tickets_id == space_info->tickets_id) {
4978                         flush_state++;
4979                 } else {
4980                         last_tickets_id = space_info->tickets_id;
4981                         flush_state = FLUSH_DELAYED_ITEMS_NR;
4982                         if (commit_cycles)
4983                                 commit_cycles--;
4984                 }
4985
4986                 if (flush_state > COMMIT_TRANS) {
4987                         commit_cycles++;
4988                         if (commit_cycles > 2) {
4989                                 wake_all_tickets(&space_info->tickets);
4990                                 space_info->flush = 0;
4991                         } else {
4992                                 flush_state = FLUSH_DELAYED_ITEMS_NR;
4993                         }
4994                 }
4995                 spin_unlock(&space_info->lock);
4996         } while (flush_state <= COMMIT_TRANS);
4997 }
4998
4999 void btrfs_init_async_reclaim_work(struct work_struct *work)
5000 {
5001         INIT_WORK(work, btrfs_async_reclaim_metadata_space);
5002 }
5003
5004 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info,
5005                                             struct btrfs_space_info *space_info,
5006                                             struct reserve_ticket *ticket)
5007 {
5008         u64 to_reclaim;
5009         int flush_state = FLUSH_DELAYED_ITEMS_NR;
5010
5011         spin_lock(&space_info->lock);
5012         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5013                                                       false);
5014         if (!to_reclaim) {
5015                 spin_unlock(&space_info->lock);
5016                 return;
5017         }
5018         spin_unlock(&space_info->lock);
5019
5020         do {
5021                 flush_space(fs_info, space_info, to_reclaim, flush_state);
5022                 flush_state++;
5023                 spin_lock(&space_info->lock);
5024                 if (ticket->bytes == 0) {
5025                         spin_unlock(&space_info->lock);
5026                         return;
5027                 }
5028                 spin_unlock(&space_info->lock);
5029
5030                 /*
5031                  * Priority flushers can't wait on delalloc without
5032                  * deadlocking.
5033                  */
5034                 if (flush_state == FLUSH_DELALLOC ||
5035                     flush_state == FLUSH_DELALLOC_WAIT)
5036                         flush_state = ALLOC_CHUNK;
5037         } while (flush_state < COMMIT_TRANS);
5038 }
5039
5040 static int wait_reserve_ticket(struct btrfs_fs_info *fs_info,
5041                                struct btrfs_space_info *space_info,
5042                                struct reserve_ticket *ticket, u64 orig_bytes)
5043
5044 {
5045         DEFINE_WAIT(wait);
5046         int ret = 0;
5047
5048         spin_lock(&space_info->lock);
5049         while (ticket->bytes > 0 && ticket->error == 0) {
5050                 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE);
5051                 if (ret) {
5052                         ret = -EINTR;
5053                         break;
5054                 }
5055                 spin_unlock(&space_info->lock);
5056
5057                 schedule();
5058
5059                 finish_wait(&ticket->wait, &wait);
5060                 spin_lock(&space_info->lock);
5061         }
5062         if (!ret)
5063                 ret = ticket->error;
5064         if (!list_empty(&ticket->list))
5065                 list_del_init(&ticket->list);
5066         if (ticket->bytes && ticket->bytes < orig_bytes) {
5067                 u64 num_bytes = orig_bytes - ticket->bytes;
5068                 space_info->bytes_may_use -= num_bytes;
5069                 trace_btrfs_space_reservation(fs_info, "space_info",
5070                                               space_info->flags, num_bytes, 0);
5071         }
5072         spin_unlock(&space_info->lock);
5073
5074         return ret;
5075 }
5076
5077 /**
5078  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5079  * @root - the root we're allocating for
5080  * @space_info - the space info we want to allocate from
5081  * @orig_bytes - the number of bytes we want
5082  * @flush - whether or not we can flush to make our reservation
5083  *
5084  * This will reserve orig_bytes number of bytes from the space info associated
5085  * with the block_rsv.  If there is not enough space it will make an attempt to
5086  * flush out space to make room.  It will do this by flushing delalloc if
5087  * possible or committing the transaction.  If flush is 0 then no attempts to
5088  * regain reservations will be made and this will fail if there is not enough
5089  * space already.
5090  */
5091 static int __reserve_metadata_bytes(struct btrfs_fs_info *fs_info,
5092                                     struct btrfs_space_info *space_info,
5093                                     u64 orig_bytes,
5094                                     enum btrfs_reserve_flush_enum flush,
5095                                     bool system_chunk)
5096 {
5097         struct reserve_ticket ticket;
5098         u64 used;
5099         int ret = 0;
5100
5101         ASSERT(orig_bytes);
5102         ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL);
5103
5104         spin_lock(&space_info->lock);
5105         ret = -ENOSPC;
5106         used = btrfs_space_info_used(space_info, true);
5107
5108         /*
5109          * If we have enough space then hooray, make our reservation and carry
5110          * on.  If not see if we can overcommit, and if we can, hooray carry on.
5111          * If not things get more complicated.
5112          */
5113         if (used + orig_bytes <= space_info->total_bytes) {
5114                 space_info->bytes_may_use += orig_bytes;
5115                 trace_btrfs_space_reservation(fs_info, "space_info",
5116                                               space_info->flags, orig_bytes, 1);
5117                 ret = 0;
5118         } else if (can_overcommit(fs_info, space_info, orig_bytes, flush,
5119                                   system_chunk)) {
5120                 space_info->bytes_may_use += orig_bytes;
5121                 trace_btrfs_space_reservation(fs_info, "space_info",
5122                                               space_info->flags, orig_bytes, 1);
5123                 ret = 0;
5124         }
5125
5126         /*
5127          * If we couldn't make a reservation then setup our reservation ticket
5128          * and kick the async worker if it's not already running.
5129          *
5130          * If we are a priority flusher then we just need to add our ticket to
5131          * the list and we will do our own flushing further down.
5132          */
5133         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
5134                 ticket.bytes = orig_bytes;
5135                 ticket.error = 0;
5136                 init_waitqueue_head(&ticket.wait);
5137                 if (flush == BTRFS_RESERVE_FLUSH_ALL) {
5138                         list_add_tail(&ticket.list, &space_info->tickets);
5139                         if (!space_info->flush) {
5140                                 space_info->flush = 1;
5141                                 trace_btrfs_trigger_flush(fs_info,
5142                                                           space_info->flags,
5143                                                           orig_bytes, flush,
5144                                                           "enospc");
5145                                 queue_work(system_unbound_wq,
5146                                            &fs_info->async_reclaim_work);
5147                         }
5148                 } else {
5149                         list_add_tail(&ticket.list,
5150                                       &space_info->priority_tickets);
5151                 }
5152         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
5153                 used += orig_bytes;
5154                 /*
5155                  * We will do the space reservation dance during log replay,
5156                  * which means we won't have fs_info->fs_root set, so don't do
5157                  * the async reclaim as we will panic.
5158                  */
5159                 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) &&
5160                     need_do_async_reclaim(fs_info, space_info,
5161                                           used, system_chunk) &&
5162                     !work_busy(&fs_info->async_reclaim_work)) {
5163                         trace_btrfs_trigger_flush(fs_info, space_info->flags,
5164                                                   orig_bytes, flush, "preempt");
5165                         queue_work(system_unbound_wq,
5166                                    &fs_info->async_reclaim_work);
5167                 }
5168         }
5169         spin_unlock(&space_info->lock);
5170         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
5171                 return ret;
5172
5173         if (flush == BTRFS_RESERVE_FLUSH_ALL)
5174                 return wait_reserve_ticket(fs_info, space_info, &ticket,
5175                                            orig_bytes);
5176
5177         ret = 0;
5178         priority_reclaim_metadata_space(fs_info, space_info, &ticket);
5179         spin_lock(&space_info->lock);
5180         if (ticket.bytes) {
5181                 if (ticket.bytes < orig_bytes) {
5182                         u64 num_bytes = orig_bytes - ticket.bytes;
5183                         space_info->bytes_may_use -= num_bytes;
5184                         trace_btrfs_space_reservation(fs_info, "space_info",
5185                                                       space_info->flags,
5186                                                       num_bytes, 0);
5187
5188                 }
5189                 list_del_init(&ticket.list);
5190                 ret = -ENOSPC;
5191         }
5192         spin_unlock(&space_info->lock);
5193         ASSERT(list_empty(&ticket.list));
5194         return ret;
5195 }
5196
5197 /**
5198  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5199  * @root - the root we're allocating for
5200  * @block_rsv - the block_rsv we're allocating for
5201  * @orig_bytes - the number of bytes we want
5202  * @flush - whether or not we can flush to make our reservation
5203  *
5204  * This will reserve orgi_bytes number of bytes from the space info associated
5205  * with the block_rsv.  If there is not enough space it will make an attempt to
5206  * flush out space to make room.  It will do this by flushing delalloc if
5207  * possible or committing the transaction.  If flush is 0 then no attempts to
5208  * regain reservations will be made and this will fail if there is not enough
5209  * space already.
5210  */
5211 static int reserve_metadata_bytes(struct btrfs_root *root,
5212                                   struct btrfs_block_rsv *block_rsv,
5213                                   u64 orig_bytes,
5214                                   enum btrfs_reserve_flush_enum flush)
5215 {
5216         struct btrfs_fs_info *fs_info = root->fs_info;
5217         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5218         int ret;
5219         bool system_chunk = (root == fs_info->chunk_root);
5220
5221         ret = __reserve_metadata_bytes(fs_info, block_rsv->space_info,
5222                                        orig_bytes, flush, system_chunk);
5223         if (ret == -ENOSPC &&
5224             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
5225                 if (block_rsv != global_rsv &&
5226                     !block_rsv_use_bytes(global_rsv, orig_bytes))
5227                         ret = 0;
5228         }
5229         if (ret == -ENOSPC) {
5230                 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
5231                                               block_rsv->space_info->flags,
5232                                               orig_bytes, 1);
5233
5234                 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
5235                         dump_space_info(fs_info, block_rsv->space_info,
5236                                         orig_bytes, 0);
5237         }
5238         return ret;
5239 }
5240
5241 static struct btrfs_block_rsv *get_block_rsv(
5242                                         const struct btrfs_trans_handle *trans,
5243                                         const struct btrfs_root *root)
5244 {
5245         struct btrfs_fs_info *fs_info = root->fs_info;
5246         struct btrfs_block_rsv *block_rsv = NULL;
5247
5248         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
5249             (root == fs_info->csum_root && trans->adding_csums) ||
5250             (root == fs_info->uuid_root))
5251                 block_rsv = trans->block_rsv;
5252
5253         if (!block_rsv)
5254                 block_rsv = root->block_rsv;
5255
5256         if (!block_rsv)
5257                 block_rsv = &fs_info->empty_block_rsv;
5258
5259         return block_rsv;
5260 }
5261
5262 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
5263                                u64 num_bytes)
5264 {
5265         int ret = -ENOSPC;
5266         spin_lock(&block_rsv->lock);
5267         if (block_rsv->reserved >= num_bytes) {
5268                 block_rsv->reserved -= num_bytes;
5269                 if (block_rsv->reserved < block_rsv->size)
5270                         block_rsv->full = 0;
5271                 ret = 0;
5272         }
5273         spin_unlock(&block_rsv->lock);
5274         return ret;
5275 }
5276
5277 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
5278                                 u64 num_bytes, bool update_size)
5279 {
5280         spin_lock(&block_rsv->lock);
5281         block_rsv->reserved += num_bytes;
5282         if (update_size)
5283                 block_rsv->size += num_bytes;
5284         else if (block_rsv->reserved >= block_rsv->size)
5285                 block_rsv->full = 1;
5286         spin_unlock(&block_rsv->lock);
5287 }
5288
5289 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
5290                              struct btrfs_block_rsv *dest, u64 num_bytes,
5291                              int min_factor)
5292 {
5293         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5294         u64 min_bytes;
5295
5296         if (global_rsv->space_info != dest->space_info)
5297                 return -ENOSPC;
5298
5299         spin_lock(&global_rsv->lock);
5300         min_bytes = div_factor(global_rsv->size, min_factor);
5301         if (global_rsv->reserved < min_bytes + num_bytes) {
5302                 spin_unlock(&global_rsv->lock);
5303                 return -ENOSPC;
5304         }
5305         global_rsv->reserved -= num_bytes;
5306         if (global_rsv->reserved < global_rsv->size)
5307                 global_rsv->full = 0;
5308         spin_unlock(&global_rsv->lock);
5309
5310         block_rsv_add_bytes(dest, num_bytes, true);
5311         return 0;
5312 }
5313
5314 /*
5315  * This is for space we already have accounted in space_info->bytes_may_use, so
5316  * basically when we're returning space from block_rsv's.
5317  */
5318 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
5319                                      struct btrfs_space_info *space_info,
5320                                      u64 num_bytes)
5321 {
5322         struct reserve_ticket *ticket;
5323         struct list_head *head;
5324         u64 used;
5325         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH;
5326         bool check_overcommit = false;
5327
5328         spin_lock(&space_info->lock);
5329         head = &space_info->priority_tickets;
5330
5331         /*
5332          * If we are over our limit then we need to check and see if we can
5333          * overcommit, and if we can't then we just need to free up our space
5334          * and not satisfy any requests.
5335          */
5336         used = btrfs_space_info_used(space_info, true);
5337         if (used - num_bytes >= space_info->total_bytes)
5338                 check_overcommit = true;
5339 again:
5340         while (!list_empty(head) && num_bytes) {
5341                 ticket = list_first_entry(head, struct reserve_ticket,
5342                                           list);
5343                 /*
5344                  * We use 0 bytes because this space is already reserved, so
5345                  * adding the ticket space would be a double count.
5346                  */
5347                 if (check_overcommit &&
5348                     !can_overcommit(fs_info, space_info, 0, flush, false))
5349                         break;
5350                 if (num_bytes >= ticket->bytes) {
5351                         list_del_init(&ticket->list);
5352                         num_bytes -= ticket->bytes;
5353                         ticket->bytes = 0;
5354                         space_info->tickets_id++;
5355                         wake_up(&ticket->wait);
5356                 } else {
5357                         ticket->bytes -= num_bytes;
5358                         num_bytes = 0;
5359                 }
5360         }
5361
5362         if (num_bytes && head == &space_info->priority_tickets) {
5363                 head = &space_info->tickets;
5364                 flush = BTRFS_RESERVE_FLUSH_ALL;
5365                 goto again;
5366         }
5367         space_info->bytes_may_use -= num_bytes;
5368         trace_btrfs_space_reservation(fs_info, "space_info",
5369                                       space_info->flags, num_bytes, 0);
5370         spin_unlock(&space_info->lock);
5371 }
5372
5373 /*
5374  * This is for newly allocated space that isn't accounted in
5375  * space_info->bytes_may_use yet.  So if we allocate a chunk or unpin an extent
5376  * we use this helper.
5377  */
5378 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
5379                                      struct btrfs_space_info *space_info,
5380                                      u64 num_bytes)
5381 {
5382         struct reserve_ticket *ticket;
5383         struct list_head *head = &space_info->priority_tickets;
5384
5385 again:
5386         while (!list_empty(head) && num_bytes) {
5387                 ticket = list_first_entry(head, struct reserve_ticket,
5388                                           list);
5389                 if (num_bytes >= ticket->bytes) {
5390                         trace_btrfs_space_reservation(fs_info, "space_info",
5391                                                       space_info->flags,
5392                                                       ticket->bytes, 1);
5393                         list_del_init(&ticket->list);
5394                         num_bytes -= ticket->bytes;
5395                         space_info->bytes_may_use += ticket->bytes;
5396                         ticket->bytes = 0;
5397                         space_info->tickets_id++;
5398                         wake_up(&ticket->wait);
5399                 } else {
5400                         trace_btrfs_space_reservation(fs_info, "space_info",
5401                                                       space_info->flags,
5402                                                       num_bytes, 1);
5403                         space_info->bytes_may_use += num_bytes;
5404                         ticket->bytes -= num_bytes;
5405                         num_bytes = 0;
5406                 }
5407         }
5408
5409         if (num_bytes && head == &space_info->priority_tickets) {
5410                 head = &space_info->tickets;
5411                 goto again;
5412         }
5413 }
5414
5415 static u64 block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
5416                                     struct btrfs_block_rsv *block_rsv,
5417                                     struct btrfs_block_rsv *dest, u64 num_bytes,
5418                                     u64 *qgroup_to_release_ret)
5419 {
5420         struct btrfs_space_info *space_info = block_rsv->space_info;
5421         u64 qgroup_to_release = 0;
5422         u64 ret;
5423
5424         spin_lock(&block_rsv->lock);
5425         if (num_bytes == (u64)-1) {
5426                 num_bytes = block_rsv->size;
5427                 qgroup_to_release = block_rsv->qgroup_rsv_size;
5428         }
5429         block_rsv->size -= num_bytes;
5430         if (block_rsv->reserved >= block_rsv->size) {
5431                 num_bytes = block_rsv->reserved - block_rsv->size;
5432                 block_rsv->reserved = block_rsv->size;
5433                 block_rsv->full = 1;
5434         } else {
5435                 num_bytes = 0;
5436         }
5437         if (block_rsv->qgroup_rsv_reserved >= block_rsv->qgroup_rsv_size) {
5438                 qgroup_to_release = block_rsv->qgroup_rsv_reserved -
5439                                     block_rsv->qgroup_rsv_size;
5440                 block_rsv->qgroup_rsv_reserved = block_rsv->qgroup_rsv_size;
5441         } else {
5442                 qgroup_to_release = 0;
5443         }
5444         spin_unlock(&block_rsv->lock);
5445
5446         ret = num_bytes;
5447         if (num_bytes > 0) {
5448                 if (dest) {
5449                         spin_lock(&dest->lock);
5450                         if (!dest->full) {
5451                                 u64 bytes_to_add;
5452
5453                                 bytes_to_add = dest->size - dest->reserved;
5454                                 bytes_to_add = min(num_bytes, bytes_to_add);
5455                                 dest->reserved += bytes_to_add;
5456                                 if (dest->reserved >= dest->size)
5457                                         dest->full = 1;
5458                                 num_bytes -= bytes_to_add;
5459                         }
5460                         spin_unlock(&dest->lock);
5461                 }
5462                 if (num_bytes)
5463                         space_info_add_old_bytes(fs_info, space_info,
5464                                                  num_bytes);
5465         }
5466         if (qgroup_to_release_ret)
5467                 *qgroup_to_release_ret = qgroup_to_release;
5468         return ret;
5469 }
5470
5471 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src,
5472                             struct btrfs_block_rsv *dst, u64 num_bytes,
5473                             bool update_size)
5474 {
5475         int ret;
5476
5477         ret = block_rsv_use_bytes(src, num_bytes);
5478         if (ret)
5479                 return ret;
5480
5481         block_rsv_add_bytes(dst, num_bytes, update_size);
5482         return 0;
5483 }
5484
5485 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
5486 {
5487         memset(rsv, 0, sizeof(*rsv));
5488         spin_lock_init(&rsv->lock);
5489         rsv->type = type;
5490 }
5491
5492 void btrfs_init_metadata_block_rsv(struct btrfs_fs_info *fs_info,
5493                                    struct btrfs_block_rsv *rsv,
5494                                    unsigned short type)
5495 {
5496         btrfs_init_block_rsv(rsv, type);
5497         rsv->space_info = __find_space_info(fs_info,
5498                                             BTRFS_BLOCK_GROUP_METADATA);
5499 }
5500
5501 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_fs_info *fs_info,
5502                                               unsigned short type)
5503 {
5504         struct btrfs_block_rsv *block_rsv;
5505
5506         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
5507         if (!block_rsv)
5508                 return NULL;
5509
5510         btrfs_init_metadata_block_rsv(fs_info, block_rsv, type);
5511         return block_rsv;
5512 }
5513
5514 void btrfs_free_block_rsv(struct btrfs_fs_info *fs_info,
5515                           struct btrfs_block_rsv *rsv)
5516 {
5517         if (!rsv)
5518                 return;
5519         btrfs_block_rsv_release(fs_info, rsv, (u64)-1);
5520         kfree(rsv);
5521 }
5522
5523 int btrfs_block_rsv_add(struct btrfs_root *root,
5524                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
5525                         enum btrfs_reserve_flush_enum flush)
5526 {
5527         int ret;
5528
5529         if (num_bytes == 0)
5530                 return 0;
5531
5532         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5533         if (!ret)
5534                 block_rsv_add_bytes(block_rsv, num_bytes, true);
5535
5536         return ret;
5537 }
5538
5539 int btrfs_block_rsv_check(struct btrfs_block_rsv *block_rsv, int min_factor)
5540 {
5541         u64 num_bytes = 0;
5542         int ret = -ENOSPC;
5543
5544         if (!block_rsv)
5545                 return 0;
5546
5547         spin_lock(&block_rsv->lock);
5548         num_bytes = div_factor(block_rsv->size, min_factor);
5549         if (block_rsv->reserved >= num_bytes)
5550                 ret = 0;
5551         spin_unlock(&block_rsv->lock);
5552
5553         return ret;
5554 }
5555
5556 int btrfs_block_rsv_refill(struct btrfs_root *root,
5557                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
5558                            enum btrfs_reserve_flush_enum flush)
5559 {
5560         u64 num_bytes = 0;
5561         int ret = -ENOSPC;
5562
5563         if (!block_rsv)
5564                 return 0;
5565
5566         spin_lock(&block_rsv->lock);
5567         num_bytes = min_reserved;
5568         if (block_rsv->reserved >= num_bytes)
5569                 ret = 0;
5570         else
5571                 num_bytes -= block_rsv->reserved;
5572         spin_unlock(&block_rsv->lock);
5573
5574         if (!ret)
5575                 return 0;
5576
5577         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5578         if (!ret) {
5579                 block_rsv_add_bytes(block_rsv, num_bytes, false);
5580                 return 0;
5581         }
5582
5583         return ret;
5584 }
5585
5586 /**
5587  * btrfs_inode_rsv_refill - refill the inode block rsv.
5588  * @inode - the inode we are refilling.
5589  * @flush - the flusing restriction.
5590  *
5591  * Essentially the same as btrfs_block_rsv_refill, except it uses the
5592  * block_rsv->size as the minimum size.  We'll either refill the missing amount
5593  * or return if we already have enough space.  This will also handle the resreve
5594  * tracepoint for the reserved amount.
5595  */
5596 static int btrfs_inode_rsv_refill(struct btrfs_inode *inode,
5597                                   enum btrfs_reserve_flush_enum flush)
5598 {
5599         struct btrfs_root *root = inode->root;
5600         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
5601         u64 num_bytes = 0;
5602         u64 qgroup_num_bytes = 0;
5603         int ret = -ENOSPC;
5604
5605         spin_lock(&block_rsv->lock);
5606         if (block_rsv->reserved < block_rsv->size)
5607                 num_bytes = block_rsv->size - block_rsv->reserved;
5608         if (block_rsv->qgroup_rsv_reserved < block_rsv->qgroup_rsv_size)
5609                 qgroup_num_bytes = block_rsv->qgroup_rsv_size -
5610                                    block_rsv->qgroup_rsv_reserved;
5611         spin_unlock(&block_rsv->lock);
5612
5613         if (num_bytes == 0)
5614                 return 0;
5615
5616         ret = btrfs_qgroup_reserve_meta_prealloc(root, qgroup_num_bytes, true);
5617         if (ret)
5618                 return ret;
5619         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5620         if (!ret) {
5621                 block_rsv_add_bytes(block_rsv, num_bytes, false);
5622                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5623                                               btrfs_ino(inode), num_bytes, 1);
5624
5625                 /* Don't forget to increase qgroup_rsv_reserved */
5626                 spin_lock(&block_rsv->lock);
5627                 block_rsv->qgroup_rsv_reserved += qgroup_num_bytes;
5628                 spin_unlock(&block_rsv->lock);
5629         } else
5630                 btrfs_qgroup_free_meta_prealloc(root, qgroup_num_bytes);
5631         return ret;
5632 }
5633
5634 /**
5635  * btrfs_inode_rsv_release - release any excessive reservation.
5636  * @inode - the inode we need to release from.
5637  * @qgroup_free - free or convert qgroup meta.
5638  *   Unlike normal operation, qgroup meta reservation needs to know if we are
5639  *   freeing qgroup reservation or just converting it into per-trans.  Normally
5640  *   @qgroup_free is true for error handling, and false for normal release.
5641  *
5642  * This is the same as btrfs_block_rsv_release, except that it handles the
5643  * tracepoint for the reservation.
5644  */
5645 static void btrfs_inode_rsv_release(struct btrfs_inode *inode, bool qgroup_free)
5646 {
5647         struct btrfs_fs_info *fs_info = inode->root->fs_info;
5648         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5649         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
5650         u64 released = 0;
5651         u64 qgroup_to_release = 0;
5652
5653         /*
5654          * Since we statically set the block_rsv->size we just want to say we
5655          * are releasing 0 bytes, and then we'll just get the reservation over
5656          * the size free'd.
5657          */
5658         released = block_rsv_release_bytes(fs_info, block_rsv, global_rsv, 0,
5659                                            &qgroup_to_release);
5660         if (released > 0)
5661                 trace_btrfs_space_reservation(fs_info, "delalloc",
5662                                               btrfs_ino(inode), released, 0);
5663         if (qgroup_free)
5664                 btrfs_qgroup_free_meta_prealloc(inode->root, qgroup_to_release);
5665         else
5666                 btrfs_qgroup_convert_reserved_meta(inode->root,
5667                                                    qgroup_to_release);
5668 }
5669
5670 void btrfs_block_rsv_release(struct btrfs_fs_info *fs_info,
5671                              struct btrfs_block_rsv *block_rsv,
5672                              u64 num_bytes)
5673 {
5674         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5675
5676         if (global_rsv == block_rsv ||
5677             block_rsv->space_info != global_rsv->space_info)
5678                 global_rsv = NULL;
5679         block_rsv_release_bytes(fs_info, block_rsv, global_rsv, num_bytes, NULL);
5680 }
5681
5682 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
5683 {
5684         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
5685         struct btrfs_space_info *sinfo = block_rsv->space_info;
5686         u64 num_bytes;
5687
5688         /*
5689          * The global block rsv is based on the size of the extent tree, the
5690          * checksum tree and the root tree.  If the fs is empty we want to set
5691          * it to a minimal amount for safety.
5692          */
5693         num_bytes = btrfs_root_used(&fs_info->extent_root->root_item) +
5694                 btrfs_root_used(&fs_info->csum_root->root_item) +
5695                 btrfs_root_used(&fs_info->tree_root->root_item);
5696         num_bytes = max_t(u64, num_bytes, SZ_16M);
5697
5698         spin_lock(&sinfo->lock);
5699         spin_lock(&block_rsv->lock);
5700
5701         block_rsv->size = min_t(u64, num_bytes, SZ_512M);
5702
5703         if (block_rsv->reserved < block_rsv->size) {
5704                 num_bytes = btrfs_space_info_used(sinfo, true);
5705                 if (sinfo->total_bytes > num_bytes) {
5706                         num_bytes = sinfo->total_bytes - num_bytes;
5707                         num_bytes = min(num_bytes,
5708                                         block_rsv->size - block_rsv->reserved);
5709                         block_rsv->reserved += num_bytes;
5710                         sinfo->bytes_may_use += num_bytes;
5711                         trace_btrfs_space_reservation(fs_info, "space_info",
5712                                                       sinfo->flags, num_bytes,
5713                                                       1);
5714                 }
5715         } else if (block_rsv->reserved > block_rsv->size) {
5716                 num_bytes = block_rsv->reserved - block_rsv->size;
5717                 sinfo->bytes_may_use -= num_bytes;
5718                 trace_btrfs_space_reservation(fs_info, "space_info",
5719                                       sinfo->flags, num_bytes, 0);
5720                 block_rsv->reserved = block_rsv->size;
5721         }
5722
5723         if (block_rsv->reserved == block_rsv->size)
5724                 block_rsv->full = 1;
5725         else
5726                 block_rsv->full = 0;
5727
5728         spin_unlock(&block_rsv->lock);
5729         spin_unlock(&sinfo->lock);
5730 }
5731
5732 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
5733 {
5734         struct btrfs_space_info *space_info;
5735
5736         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
5737         fs_info->chunk_block_rsv.space_info = space_info;
5738
5739         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5740         fs_info->global_block_rsv.space_info = space_info;
5741         fs_info->trans_block_rsv.space_info = space_info;
5742         fs_info->empty_block_rsv.space_info = space_info;
5743         fs_info->delayed_block_rsv.space_info = space_info;
5744
5745         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
5746         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
5747         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
5748         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
5749         if (fs_info->quota_root)
5750                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
5751         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
5752
5753         update_global_block_rsv(fs_info);
5754 }
5755
5756 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
5757 {
5758         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
5759                                 (u64)-1, NULL);
5760         WARN_ON(fs_info->trans_block_rsv.size > 0);
5761         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
5762         WARN_ON(fs_info->chunk_block_rsv.size > 0);
5763         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
5764         WARN_ON(fs_info->delayed_block_rsv.size > 0);
5765         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
5766 }
5767
5768
5769 /*
5770  * To be called after all the new block groups attached to the transaction
5771  * handle have been created (btrfs_create_pending_block_groups()).
5772  */
5773 void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans)
5774 {
5775         struct btrfs_fs_info *fs_info = trans->fs_info;
5776
5777         if (!trans->chunk_bytes_reserved)
5778                 return;
5779
5780         WARN_ON_ONCE(!list_empty(&trans->new_bgs));
5781
5782         block_rsv_release_bytes(fs_info, &fs_info->chunk_block_rsv, NULL,
5783                                 trans->chunk_bytes_reserved, NULL);
5784         trans->chunk_bytes_reserved = 0;
5785 }
5786
5787 /*
5788  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
5789  * root: the root of the parent directory
5790  * rsv: block reservation
5791  * items: the number of items that we need do reservation
5792  * use_global_rsv: allow fallback to the global block reservation
5793  *
5794  * This function is used to reserve the space for snapshot/subvolume
5795  * creation and deletion. Those operations are different with the
5796  * common file/directory operations, they change two fs/file trees
5797  * and root tree, the number of items that the qgroup reserves is
5798  * different with the free space reservation. So we can not use
5799  * the space reservation mechanism in start_transaction().
5800  */
5801 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
5802                                      struct btrfs_block_rsv *rsv, int items,
5803                                      bool use_global_rsv)
5804 {
5805         u64 qgroup_num_bytes = 0;
5806         u64 num_bytes;
5807         int ret;
5808         struct btrfs_fs_info *fs_info = root->fs_info;
5809         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5810
5811         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)) {
5812                 /* One for parent inode, two for dir entries */
5813                 qgroup_num_bytes = 3 * fs_info->nodesize;
5814                 ret = btrfs_qgroup_reserve_meta_prealloc(root,
5815                                 qgroup_num_bytes, true);
5816                 if (ret)
5817                         return ret;
5818         }
5819
5820         num_bytes = btrfs_calc_trans_metadata_size(fs_info, items);
5821         rsv->space_info = __find_space_info(fs_info,
5822                                             BTRFS_BLOCK_GROUP_METADATA);
5823         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
5824                                   BTRFS_RESERVE_FLUSH_ALL);
5825
5826         if (ret == -ENOSPC && use_global_rsv)
5827                 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes, true);
5828
5829         if (ret && qgroup_num_bytes)
5830                 btrfs_qgroup_free_meta_prealloc(root, qgroup_num_bytes);
5831
5832         return ret;
5833 }
5834
5835 void btrfs_subvolume_release_metadata(struct btrfs_fs_info *fs_info,
5836                                       struct btrfs_block_rsv *rsv)
5837 {
5838         btrfs_block_rsv_release(fs_info, rsv, (u64)-1);
5839 }
5840
5841 static void btrfs_calculate_inode_block_rsv_size(struct btrfs_fs_info *fs_info,
5842                                                  struct btrfs_inode *inode)
5843 {
5844         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
5845         u64 reserve_size = 0;
5846         u64 qgroup_rsv_size = 0;
5847         u64 csum_leaves;
5848         unsigned outstanding_extents;
5849
5850         lockdep_assert_held(&inode->lock);
5851         outstanding_extents = inode->outstanding_extents;
5852         if (outstanding_extents)
5853                 reserve_size = btrfs_calc_trans_metadata_size(fs_info,
5854                                                 outstanding_extents + 1);
5855         csum_leaves = btrfs_csum_bytes_to_leaves(fs_info,
5856                                                  inode->csum_bytes);
5857         reserve_size += btrfs_calc_trans_metadata_size(fs_info,
5858                                                        csum_leaves);
5859         /*
5860          * For qgroup rsv, the calculation is very simple:
5861          * account one nodesize for each outstanding extent
5862          *
5863          * This is overestimating in most cases.
5864          */
5865         qgroup_rsv_size = outstanding_extents * fs_info->nodesize;
5866
5867         spin_lock(&block_rsv->lock);
5868         block_rsv->size = reserve_size;
5869         block_rsv->qgroup_rsv_size = qgroup_rsv_size;
5870         spin_unlock(&block_rsv->lock);
5871 }
5872
5873 int btrfs_delalloc_reserve_metadata(struct btrfs_inode *inode, u64 num_bytes)
5874 {
5875         struct btrfs_fs_info *fs_info = inode->root->fs_info;
5876         unsigned nr_extents;
5877         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
5878         int ret = 0;
5879         bool delalloc_lock = true;
5880
5881         /* If we are a free space inode we need to not flush since we will be in
5882          * the middle of a transaction commit.  We also don't need the delalloc
5883          * mutex since we won't race with anybody.  We need this mostly to make
5884          * lockdep shut its filthy mouth.
5885          *
5886          * If we have a transaction open (can happen if we call truncate_block
5887          * from truncate), then we need FLUSH_LIMIT so we don't deadlock.
5888          */
5889         if (btrfs_is_free_space_inode(inode)) {
5890                 flush = BTRFS_RESERVE_NO_FLUSH;
5891                 delalloc_lock = false;
5892         } else {
5893                 if (current->journal_info)
5894                         flush = BTRFS_RESERVE_FLUSH_LIMIT;
5895
5896                 if (btrfs_transaction_in_commit(fs_info))
5897                         schedule_timeout(1);
5898         }
5899
5900         if (delalloc_lock)
5901                 mutex_lock(&inode->delalloc_mutex);
5902
5903         num_bytes = ALIGN(num_bytes, fs_info->sectorsize);
5904
5905         /* Add our new extents and calculate the new rsv size. */
5906         spin_lock(&inode->lock);
5907         nr_extents = count_max_extents(num_bytes);
5908         btrfs_mod_outstanding_extents(inode, nr_extents);
5909         inode->csum_bytes += num_bytes;
5910         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
5911         spin_unlock(&inode->lock);
5912
5913         ret = btrfs_inode_rsv_refill(inode, flush);
5914         if (unlikely(ret))
5915                 goto out_fail;
5916
5917         if (delalloc_lock)
5918                 mutex_unlock(&inode->delalloc_mutex);
5919         return 0;
5920
5921 out_fail:
5922         spin_lock(&inode->lock);
5923         nr_extents = count_max_extents(num_bytes);
5924         btrfs_mod_outstanding_extents(inode, -nr_extents);
5925         inode->csum_bytes -= num_bytes;
5926         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
5927         spin_unlock(&inode->lock);
5928
5929         btrfs_inode_rsv_release(inode, true);
5930         if (delalloc_lock)
5931                 mutex_unlock(&inode->delalloc_mutex);
5932         return ret;
5933 }
5934
5935 /**
5936  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
5937  * @inode: the inode to release the reservation for.
5938  * @num_bytes: the number of bytes we are releasing.
5939  * @qgroup_free: free qgroup reservation or convert it to per-trans reservation
5940  *
5941  * This will release the metadata reservation for an inode.  This can be called
5942  * once we complete IO for a given set of bytes to release their metadata
5943  * reservations, or on error for the same reason.
5944  */
5945 void btrfs_delalloc_release_metadata(struct btrfs_inode *inode, u64 num_bytes,
5946                                      bool qgroup_free)
5947 {
5948         struct btrfs_fs_info *fs_info = inode->root->fs_info;
5949
5950         num_bytes = ALIGN(num_bytes, fs_info->sectorsize);
5951         spin_lock(&inode->lock);
5952         inode->csum_bytes -= num_bytes;
5953         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
5954         spin_unlock(&inode->lock);
5955
5956         if (btrfs_is_testing(fs_info))
5957                 return;
5958
5959         btrfs_inode_rsv_release(inode, qgroup_free);
5960 }
5961
5962 /**
5963  * btrfs_delalloc_release_extents - release our outstanding_extents
5964  * @inode: the inode to balance the reservation for.
5965  * @num_bytes: the number of bytes we originally reserved with
5966  * @qgroup_free: do we need to free qgroup meta reservation or convert them.
5967  *
5968  * When we reserve space we increase outstanding_extents for the extents we may
5969  * add.  Once we've set the range as delalloc or created our ordered extents we
5970  * have outstanding_extents to track the real usage, so we use this to free our
5971  * temporarily tracked outstanding_extents.  This _must_ be used in conjunction
5972  * with btrfs_delalloc_reserve_metadata.
5973  */
5974 void btrfs_delalloc_release_extents(struct btrfs_inode *inode, u64 num_bytes,
5975                                     bool qgroup_free)
5976 {
5977         struct btrfs_fs_info *fs_info = inode->root->fs_info;
5978         unsigned num_extents;
5979
5980         spin_lock(&inode->lock);
5981         num_extents = count_max_extents(num_bytes);
5982         btrfs_mod_outstanding_extents(inode, -num_extents);
5983         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
5984         spin_unlock(&inode->lock);
5985
5986         if (btrfs_is_testing(fs_info))
5987                 return;
5988
5989         btrfs_inode_rsv_release(inode, qgroup_free);
5990 }
5991
5992 /**
5993  * btrfs_delalloc_reserve_space - reserve data and metadata space for
5994  * delalloc
5995  * @inode: inode we're writing to
5996  * @start: start range we are writing to
5997  * @len: how long the range we are writing to
5998  * @reserved: mandatory parameter, record actually reserved qgroup ranges of
5999  *            current reservation.
6000  *
6001  * This will do the following things
6002  *
6003  * o reserve space in data space info for num bytes
6004  *   and reserve precious corresponding qgroup space
6005  *   (Done in check_data_free_space)
6006  *
6007  * o reserve space for metadata space, based on the number of outstanding
6008  *   extents and how much csums will be needed
6009  *   also reserve metadata space in a per root over-reserve method.
6010  * o add to the inodes->delalloc_bytes
6011  * o add it to the fs_info's delalloc inodes list.
6012  *   (Above 3 all done in delalloc_reserve_metadata)
6013  *
6014  * Return 0 for success
6015  * Return <0 for error(-ENOSPC or -EQUOT)
6016  */
6017 int btrfs_delalloc_reserve_space(struct inode *inode,
6018                         struct extent_changeset **reserved, u64 start, u64 len)
6019 {
6020         int ret;
6021
6022         ret = btrfs_check_data_free_space(inode, reserved, start, len);
6023         if (ret < 0)
6024                 return ret;
6025         ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode), len);
6026         if (ret < 0)
6027                 btrfs_free_reserved_data_space(inode, *reserved, start, len);
6028         return ret;
6029 }
6030
6031 /**
6032  * btrfs_delalloc_release_space - release data and metadata space for delalloc
6033  * @inode: inode we're releasing space for
6034  * @start: start position of the space already reserved
6035  * @len: the len of the space already reserved
6036  * @release_bytes: the len of the space we consumed or didn't use
6037  *
6038  * This function will release the metadata space that was not used and will
6039  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
6040  * list if there are no delalloc bytes left.
6041  * Also it will handle the qgroup reserved space.
6042  */
6043 void btrfs_delalloc_release_space(struct inode *inode,
6044                                   struct extent_changeset *reserved,
6045                                   u64 start, u64 len, bool qgroup_free)
6046 {
6047         btrfs_delalloc_release_metadata(BTRFS_I(inode), len, qgroup_free);
6048         btrfs_free_reserved_data_space(inode, reserved, start, len);
6049 }
6050
6051 static int update_block_group(struct btrfs_trans_handle *trans,
6052                               struct btrfs_fs_info *info, u64 bytenr,
6053                               u64 num_bytes, int alloc)
6054 {
6055         struct btrfs_block_group_cache *cache = NULL;
6056         u64 total = num_bytes;
6057         u64 old_val;
6058         u64 byte_in_group;
6059         int factor;
6060
6061         /* block accounting for super block */
6062         spin_lock(&info->delalloc_root_lock);
6063         old_val = btrfs_super_bytes_used(info->super_copy);
6064         if (alloc)
6065                 old_val += num_bytes;
6066         else
6067                 old_val -= num_bytes;
6068         btrfs_set_super_bytes_used(info->super_copy, old_val);
6069         spin_unlock(&info->delalloc_root_lock);
6070
6071         while (total) {
6072                 cache = btrfs_lookup_block_group(info, bytenr);
6073                 if (!cache)
6074                         return -ENOENT;
6075                 factor = btrfs_bg_type_to_factor(cache->flags);
6076
6077                 /*
6078                  * If this block group has free space cache written out, we
6079                  * need to make sure to load it if we are removing space.  This
6080                  * is because we need the unpinning stage to actually add the
6081                  * space back to the block group, otherwise we will leak space.
6082                  */
6083                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
6084                         cache_block_group(cache, 1);
6085
6086                 byte_in_group = bytenr - cache->key.objectid;
6087                 WARN_ON(byte_in_group > cache->key.offset);
6088
6089                 spin_lock(&cache->space_info->lock);
6090                 spin_lock(&cache->lock);
6091
6092                 if (btrfs_test_opt(info, SPACE_CACHE) &&
6093                     cache->disk_cache_state < BTRFS_DC_CLEAR)
6094                         cache->disk_cache_state = BTRFS_DC_CLEAR;
6095
6096                 old_val = btrfs_block_group_used(&cache->item);
6097                 num_bytes = min(total, cache->key.offset - byte_in_group);
6098                 if (alloc) {
6099                         old_val += num_bytes;
6100                         btrfs_set_block_group_used(&cache->item, old_val);
6101                         cache->reserved -= num_bytes;
6102                         cache->space_info->bytes_reserved -= num_bytes;
6103                         cache->space_info->bytes_used += num_bytes;
6104                         cache->space_info->disk_used += num_bytes * factor;
6105                         spin_unlock(&cache->lock);
6106                         spin_unlock(&cache->space_info->lock);
6107                 } else {
6108                         old_val -= num_bytes;
6109                         btrfs_set_block_group_used(&cache->item, old_val);
6110                         cache->pinned += num_bytes;
6111                         cache->space_info->bytes_pinned += num_bytes;
6112                         cache->space_info->bytes_used -= num_bytes;
6113                         cache->space_info->disk_used -= num_bytes * factor;
6114                         spin_unlock(&cache->lock);
6115                         spin_unlock(&cache->space_info->lock);
6116
6117                         trace_btrfs_space_reservation(info, "pinned",
6118                                                       cache->space_info->flags,
6119                                                       num_bytes, 1);
6120                         percpu_counter_add_batch(&cache->space_info->total_bytes_pinned,
6121                                            num_bytes,
6122                                            BTRFS_TOTAL_BYTES_PINNED_BATCH);
6123                         set_extent_dirty(info->pinned_extents,
6124                                          bytenr, bytenr + num_bytes - 1,
6125                                          GFP_NOFS | __GFP_NOFAIL);
6126                 }
6127
6128                 spin_lock(&trans->transaction->dirty_bgs_lock);
6129                 if (list_empty(&cache->dirty_list)) {
6130                         list_add_tail(&cache->dirty_list,
6131                                       &trans->transaction->dirty_bgs);
6132                         trans->transaction->num_dirty_bgs++;
6133                         btrfs_get_block_group(cache);
6134                 }
6135                 spin_unlock(&trans->transaction->dirty_bgs_lock);
6136
6137                 /*
6138                  * No longer have used bytes in this block group, queue it for
6139                  * deletion. We do this after adding the block group to the
6140                  * dirty list to avoid races between cleaner kthread and space
6141                  * cache writeout.
6142                  */
6143                 if (!alloc && old_val == 0)
6144                         btrfs_mark_bg_unused(cache);
6145
6146                 btrfs_put_block_group(cache);
6147                 total -= num_bytes;
6148                 bytenr += num_bytes;
6149         }
6150         return 0;
6151 }
6152
6153 static u64 first_logical_byte(struct btrfs_fs_info *fs_info, u64 search_start)
6154 {
6155         struct btrfs_block_group_cache *cache;
6156         u64 bytenr;
6157
6158         spin_lock(&fs_info->block_group_cache_lock);
6159         bytenr = fs_info->first_logical_byte;
6160         spin_unlock(&fs_info->block_group_cache_lock);
6161
6162         if (bytenr < (u64)-1)
6163                 return bytenr;
6164
6165         cache = btrfs_lookup_first_block_group(fs_info, search_start);
6166         if (!cache)
6167                 return 0;
6168
6169         bytenr = cache->key.objectid;
6170         btrfs_put_block_group(cache);
6171
6172         return bytenr;
6173 }
6174
6175 static int pin_down_extent(struct btrfs_fs_info *fs_info,
6176                            struct btrfs_block_group_cache *cache,
6177                            u64 bytenr, u64 num_bytes, int reserved)
6178 {
6179         spin_lock(&cache->space_info->lock);
6180         spin_lock(&cache->lock);
6181         cache->pinned += num_bytes;
6182         cache->space_info->bytes_pinned += num_bytes;
6183         if (reserved) {
6184                 cache->reserved -= num_bytes;
6185                 cache->space_info->bytes_reserved -= num_bytes;
6186         }
6187         spin_unlock(&cache->lock);
6188         spin_unlock(&cache->space_info->lock);
6189
6190         trace_btrfs_space_reservation(fs_info, "pinned",
6191                                       cache->space_info->flags, num_bytes, 1);
6192         percpu_counter_add_batch(&cache->space_info->total_bytes_pinned,
6193                     num_bytes, BTRFS_TOTAL_BYTES_PINNED_BATCH);
6194         set_extent_dirty(fs_info->pinned_extents, bytenr,
6195                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
6196         return 0;
6197 }
6198
6199 /*
6200  * this function must be called within transaction
6201  */
6202 int btrfs_pin_extent(struct btrfs_fs_info *fs_info,
6203                      u64 bytenr, u64 num_bytes, int reserved)
6204 {
6205         struct btrfs_block_group_cache *cache;
6206
6207         cache = btrfs_lookup_block_group(fs_info, bytenr);
6208         BUG_ON(!cache); /* Logic error */
6209
6210         pin_down_extent(fs_info, cache, bytenr, num_bytes, reserved);
6211
6212         btrfs_put_block_group(cache);
6213         return 0;
6214 }
6215
6216 /*
6217  * this function must be called within transaction
6218  */
6219 int btrfs_pin_extent_for_log_replay(struct btrfs_fs_info *fs_info,
6220                                     u64 bytenr, u64 num_bytes)
6221 {
6222         struct btrfs_block_group_cache *cache;
6223         int ret;
6224
6225         cache = btrfs_lookup_block_group(fs_info, bytenr);
6226         if (!cache)
6227                 return -EINVAL;
6228
6229         /*
6230          * pull in the free space cache (if any) so that our pin
6231          * removes the free space from the cache.  We have load_only set
6232          * to one because the slow code to read in the free extents does check
6233          * the pinned extents.
6234          */
6235         cache_block_group(cache, 1);
6236
6237         pin_down_extent(fs_info, cache, bytenr, num_bytes, 0);
6238
6239         /* remove us from the free space cache (if we're there at all) */
6240         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
6241         btrfs_put_block_group(cache);
6242         return ret;
6243 }
6244
6245 static int __exclude_logged_extent(struct btrfs_fs_info *fs_info,
6246                                    u64 start, u64 num_bytes)
6247 {
6248         int ret;
6249         struct btrfs_block_group_cache *block_group;
6250         struct btrfs_caching_control *caching_ctl;
6251
6252         block_group = btrfs_lookup_block_group(fs_info, start);
6253         if (!block_group)
6254                 return -EINVAL;
6255
6256         cache_block_group(block_group, 0);
6257         caching_ctl = get_caching_control(block_group);
6258
6259         if (!caching_ctl) {
6260                 /* Logic error */
6261                 BUG_ON(!block_group_cache_done(block_group));
6262                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6263         } else {
6264                 mutex_lock(&caching_ctl->mutex);
6265
6266                 if (start >= caching_ctl->progress) {
6267                         ret = add_excluded_extent(fs_info, start, num_bytes);
6268                 } else if (start + num_bytes <= caching_ctl->progress) {
6269                         ret = btrfs_remove_free_space(block_group,
6270                                                       start, num_bytes);
6271                 } else {
6272                         num_bytes = caching_ctl->progress - start;
6273                         ret = btrfs_remove_free_space(block_group,
6274                                                       start, num_bytes);
6275                         if (ret)
6276                                 goto out_lock;
6277
6278                         num_bytes = (start + num_bytes) -
6279                                 caching_ctl->progress;
6280                         start = caching_ctl->progress;
6281                         ret = add_excluded_extent(fs_info, start, num_bytes);
6282                 }
6283 out_lock:
6284                 mutex_unlock(&caching_ctl->mutex);
6285                 put_caching_control(caching_ctl);
6286         }
6287         btrfs_put_block_group(block_group);
6288         return ret;
6289 }
6290
6291 int btrfs_exclude_logged_extents(struct btrfs_fs_info *fs_info,
6292                                  struct extent_buffer *eb)
6293 {
6294         struct btrfs_file_extent_item *item;
6295         struct btrfs_key key;
6296         int found_type;
6297         int i;
6298         int ret = 0;
6299
6300         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS))
6301                 return 0;
6302
6303         for (i = 0; i < btrfs_header_nritems(eb); i++) {
6304                 btrfs_item_key_to_cpu(eb, &key, i);
6305                 if (key.type != BTRFS_EXTENT_DATA_KEY)
6306                         continue;
6307                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
6308                 found_type = btrfs_file_extent_type(eb, item);
6309                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
6310                         continue;
6311                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
6312                         continue;
6313                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
6314                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
6315                 ret = __exclude_logged_extent(fs_info, key.objectid, key.offset);
6316                 if (ret)
6317                         break;
6318         }
6319
6320         return ret;
6321 }
6322
6323 static void
6324 btrfs_inc_block_group_reservations(struct btrfs_block_group_cache *bg)
6325 {
6326         atomic_inc(&bg->reservations);
6327 }
6328
6329 void btrfs_dec_block_group_reservations(struct btrfs_fs_info *fs_info,
6330                                         const u64 start)
6331 {
6332         struct btrfs_block_group_cache *bg;
6333
6334         bg = btrfs_lookup_block_group(fs_info, start);
6335         ASSERT(bg);
6336         if (atomic_dec_and_test(&bg->reservations))
6337                 wake_up_var(&bg->reservations);
6338         btrfs_put_block_group(bg);
6339 }
6340
6341 void btrfs_wait_block_group_reservations(struct btrfs_block_group_cache *bg)
6342 {
6343         struct btrfs_space_info *space_info = bg->space_info;
6344
6345         ASSERT(bg->ro);
6346
6347         if (!(bg->flags & BTRFS_BLOCK_GROUP_DATA))
6348                 return;
6349
6350         /*
6351          * Our block group is read only but before we set it to read only,
6352          * some task might have had allocated an extent from it already, but it
6353          * has not yet created a respective ordered extent (and added it to a
6354          * root's list of ordered extents).
6355          * Therefore wait for any task currently allocating extents, since the
6356          * block group's reservations counter is incremented while a read lock
6357          * on the groups' semaphore is held and decremented after releasing
6358          * the read access on that semaphore and creating the ordered extent.
6359          */
6360         down_write(&space_info->groups_sem);
6361         up_write(&space_info->groups_sem);
6362
6363         wait_var_event(&bg->reservations, !atomic_read(&bg->reservations));
6364 }
6365
6366 /**
6367  * btrfs_add_reserved_bytes - update the block_group and space info counters
6368  * @cache:      The cache we are manipulating
6369  * @ram_bytes:  The number of bytes of file content, and will be same to
6370  *              @num_bytes except for the compress path.
6371  * @num_bytes:  The number of bytes in question
6372  * @delalloc:   The blocks are allocated for the delalloc write
6373  *
6374  * This is called by the allocator when it reserves space. If this is a
6375  * reservation and the block group has become read only we cannot make the
6376  * reservation and return -EAGAIN, otherwise this function always succeeds.
6377  */
6378 static int btrfs_add_reserved_bytes(struct btrfs_block_group_cache *cache,
6379                                     u64 ram_bytes, u64 num_bytes, int delalloc)
6380 {
6381         struct btrfs_space_info *space_info = cache->space_info;
6382         int ret = 0;
6383
6384         spin_lock(&space_info->lock);
6385         spin_lock(&cache->lock);
6386         if (cache->ro) {
6387                 ret = -EAGAIN;
6388         } else {
6389                 cache->reserved += num_bytes;
6390                 space_info->bytes_reserved += num_bytes;
6391                 space_info->bytes_may_use -= ram_bytes;
6392                 if (delalloc)
6393                         cache->delalloc_bytes += num_bytes;
6394         }
6395         spin_unlock(&cache->lock);
6396         spin_unlock(&space_info->lock);
6397         return ret;
6398 }
6399
6400 /**
6401  * btrfs_free_reserved_bytes - update the block_group and space info counters
6402  * @cache:      The cache we are manipulating
6403  * @num_bytes:  The number of bytes in question
6404  * @delalloc:   The blocks are allocated for the delalloc write
6405  *
6406  * This is called by somebody who is freeing space that was never actually used
6407  * on disk.  For example if you reserve some space for a new leaf in transaction
6408  * A and before transaction A commits you free that leaf, you call this with
6409  * reserve set to 0 in order to clear the reservation.
6410  */
6411
6412 static void btrfs_free_reserved_bytes(struct btrfs_block_group_cache *cache,
6413                                       u64 num_bytes, int delalloc)
6414 {
6415         struct btrfs_space_info *space_info = cache->space_info;
6416
6417         spin_lock(&space_info->lock);
6418         spin_lock(&cache->lock);
6419         if (cache->ro)
6420                 space_info->bytes_readonly += num_bytes;
6421         cache->reserved -= num_bytes;
6422         space_info->bytes_reserved -= num_bytes;
6423
6424         if (delalloc)
6425                 cache->delalloc_bytes -= num_bytes;
6426         spin_unlock(&cache->lock);
6427         spin_unlock(&space_info->lock);
6428 }
6429 void btrfs_prepare_extent_commit(struct btrfs_fs_info *fs_info)
6430 {
6431         struct btrfs_caching_control *next;
6432         struct btrfs_caching_control *caching_ctl;
6433         struct btrfs_block_group_cache *cache;
6434
6435         down_write(&fs_info->commit_root_sem);
6436
6437         list_for_each_entry_safe(caching_ctl, next,
6438                                  &fs_info->caching_block_groups, list) {
6439                 cache = caching_ctl->block_group;
6440                 if (block_group_cache_done(cache)) {
6441                         cache->last_byte_to_unpin = (u64)-1;
6442                         list_del_init(&caching_ctl->list);
6443                         put_caching_control(caching_ctl);
6444                 } else {
6445                         cache->last_byte_to_unpin = caching_ctl->progress;
6446                 }
6447         }
6448
6449         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6450                 fs_info->pinned_extents = &fs_info->freed_extents[1];
6451         else
6452                 fs_info->pinned_extents = &fs_info->freed_extents[0];
6453
6454         up_write(&fs_info->commit_root_sem);
6455
6456         update_global_block_rsv(fs_info);
6457 }
6458
6459 /*
6460  * Returns the free cluster for the given space info and sets empty_cluster to
6461  * what it should be based on the mount options.
6462  */
6463 static struct btrfs_free_cluster *
6464 fetch_cluster_info(struct btrfs_fs_info *fs_info,
6465                    struct btrfs_space_info *space_info, u64 *empty_cluster)
6466 {
6467         struct btrfs_free_cluster *ret = NULL;
6468
6469         *empty_cluster = 0;
6470         if (btrfs_mixed_space_info(space_info))
6471                 return ret;
6472
6473         if (space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
6474                 ret = &fs_info->meta_alloc_cluster;
6475                 if (btrfs_test_opt(fs_info, SSD))
6476                         *empty_cluster = SZ_2M;
6477                 else
6478                         *empty_cluster = SZ_64K;
6479         } else if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) &&
6480                    btrfs_test_opt(fs_info, SSD_SPREAD)) {
6481                 *empty_cluster = SZ_2M;
6482                 ret = &fs_info->data_alloc_cluster;
6483         }
6484
6485         return ret;
6486 }
6487
6488 static int unpin_extent_range(struct btrfs_fs_info *fs_info,
6489                               u64 start, u64 end,
6490                               const bool return_free_space)
6491 {
6492         struct btrfs_block_group_cache *cache = NULL;
6493         struct btrfs_space_info *space_info;
6494         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
6495         struct btrfs_free_cluster *cluster = NULL;
6496         u64 len;
6497         u64 total_unpinned = 0;
6498         u64 empty_cluster = 0;
6499         bool readonly;
6500
6501         while (start <= end) {
6502                 readonly = false;
6503                 if (!cache ||
6504                     start >= cache->key.objectid + cache->key.offset) {
6505                         if (cache)
6506                                 btrfs_put_block_group(cache);
6507                         total_unpinned = 0;
6508                         cache = btrfs_lookup_block_group(fs_info, start);
6509                         BUG_ON(!cache); /* Logic error */
6510
6511                         cluster = fetch_cluster_info(fs_info,
6512                                                      cache->space_info,
6513                                                      &empty_cluster);
6514                         empty_cluster <<= 1;
6515                 }
6516
6517                 len = cache->key.objectid + cache->key.offset - start;
6518                 len = min(len, end + 1 - start);
6519
6520                 if (start < cache->last_byte_to_unpin) {
6521                         len = min(len, cache->last_byte_to_unpin - start);
6522                         if (return_free_space)
6523                                 btrfs_add_free_space(cache, start, len);
6524                 }
6525
6526                 start += len;
6527                 total_unpinned += len;
6528                 space_info = cache->space_info;
6529
6530                 /*
6531                  * If this space cluster has been marked as fragmented and we've
6532                  * unpinned enough in this block group to potentially allow a
6533                  * cluster to be created inside of it go ahead and clear the
6534                  * fragmented check.
6535                  */
6536                 if (cluster && cluster->fragmented &&
6537                     total_unpinned > empty_cluster) {
6538                         spin_lock(&cluster->lock);
6539                         cluster->fragmented = 0;
6540                         spin_unlock(&cluster->lock);
6541                 }
6542
6543                 spin_lock(&space_info->lock);
6544                 spin_lock(&cache->lock);
6545                 cache->pinned -= len;
6546                 space_info->bytes_pinned -= len;
6547
6548                 trace_btrfs_space_reservation(fs_info, "pinned",
6549                                               space_info->flags, len, 0);
6550                 space_info->max_extent_size = 0;
6551                 percpu_counter_add_batch(&space_info->total_bytes_pinned,
6552                             -len, BTRFS_TOTAL_BYTES_PINNED_BATCH);
6553                 if (cache->ro) {
6554                         space_info->bytes_readonly += len;
6555                         readonly = true;
6556                 }
6557                 spin_unlock(&cache->lock);
6558                 if (!readonly && return_free_space &&
6559                     global_rsv->space_info == space_info) {
6560                         u64 to_add = len;
6561
6562                         spin_lock(&global_rsv->lock);
6563                         if (!global_rsv->full) {
6564                                 to_add = min(len, global_rsv->size -
6565                                              global_rsv->reserved);
6566                                 global_rsv->reserved += to_add;
6567                                 space_info->bytes_may_use += to_add;
6568                                 if (global_rsv->reserved >= global_rsv->size)
6569                                         global_rsv->full = 1;
6570                                 trace_btrfs_space_reservation(fs_info,
6571                                                               "space_info",
6572                                                               space_info->flags,
6573                                                               to_add, 1);
6574                                 len -= to_add;
6575                         }
6576                         spin_unlock(&global_rsv->lock);
6577                         /* Add to any tickets we may have */
6578                         if (len)
6579                                 space_info_add_new_bytes(fs_info, space_info,
6580                                                          len);
6581                 }
6582                 spin_unlock(&space_info->lock);
6583         }
6584
6585         if (cache)
6586                 btrfs_put_block_group(cache);
6587         return 0;
6588 }
6589
6590 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans)
6591 {
6592         struct btrfs_fs_info *fs_info = trans->fs_info;
6593         struct btrfs_block_group_cache *block_group, *tmp;
6594         struct list_head *deleted_bgs;
6595         struct extent_io_tree *unpin;
6596         u64 start;
6597         u64 end;
6598         int ret;
6599
6600         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6601                 unpin = &fs_info->freed_extents[1];
6602         else
6603                 unpin = &fs_info->freed_extents[0];
6604
6605         while (!trans->aborted) {
6606                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
6607                 ret = find_first_extent_bit(unpin, 0, &start, &end,
6608                                             EXTENT_DIRTY, NULL);
6609                 if (ret) {
6610                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6611                         break;
6612                 }
6613
6614                 if (btrfs_test_opt(fs_info, DISCARD))
6615                         ret = btrfs_discard_extent(fs_info, start,
6616                                                    end + 1 - start, NULL);
6617
6618                 clear_extent_dirty(unpin, start, end);
6619                 unpin_extent_range(fs_info, start, end, true);
6620                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6621                 cond_resched();
6622         }
6623
6624         /*
6625          * Transaction is finished.  We don't need the lock anymore.  We
6626          * do need to clean up the block groups in case of a transaction
6627          * abort.
6628          */
6629         deleted_bgs = &trans->transaction->deleted_bgs;
6630         list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
6631                 u64 trimmed = 0;
6632
6633                 ret = -EROFS;
6634                 if (!trans->aborted)
6635                         ret = btrfs_discard_extent(fs_info,
6636                                                    block_group->key.objectid,
6637                                                    block_group->key.offset,
6638                                                    &trimmed);
6639
6640                 list_del_init(&block_group->bg_list);
6641                 btrfs_put_block_group_trimming(block_group);
6642                 btrfs_put_block_group(block_group);
6643
6644                 if (ret) {
6645                         const char *errstr = btrfs_decode_error(ret);
6646                         btrfs_warn(fs_info,
6647                            "discard failed while removing blockgroup: errno=%d %s",
6648                                    ret, errstr);
6649                 }
6650         }
6651
6652         return 0;
6653 }
6654
6655 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
6656                                struct btrfs_delayed_ref_node *node, u64 parent,
6657                                u64 root_objectid, u64 owner_objectid,
6658                                u64 owner_offset, int refs_to_drop,
6659                                struct btrfs_delayed_extent_op *extent_op)
6660 {
6661         struct btrfs_fs_info *info = trans->fs_info;
6662         struct btrfs_key key;
6663         struct btrfs_path *path;
6664         struct btrfs_root *extent_root = info->extent_root;
6665         struct extent_buffer *leaf;
6666         struct btrfs_extent_item *ei;
6667         struct btrfs_extent_inline_ref *iref;
6668         int ret;
6669         int is_data;
6670         int extent_slot = 0;
6671         int found_extent = 0;
6672         int num_to_del = 1;
6673         u32 item_size;
6674         u64 refs;
6675         u64 bytenr = node->bytenr;
6676         u64 num_bytes = node->num_bytes;
6677         int last_ref = 0;
6678         bool skinny_metadata = btrfs_fs_incompat(info, SKINNY_METADATA);
6679
6680         path = btrfs_alloc_path();
6681         if (!path)
6682                 return -ENOMEM;
6683
6684         path->reada = READA_FORWARD;
6685         path->leave_spinning = 1;
6686
6687         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
6688         BUG_ON(!is_data && refs_to_drop != 1);
6689
6690         if (is_data)
6691                 skinny_metadata = false;
6692
6693         ret = lookup_extent_backref(trans, path, &iref, bytenr, num_bytes,
6694                                     parent, root_objectid, owner_objectid,
6695                                     owner_offset);
6696         if (ret == 0) {
6697                 extent_slot = path->slots[0];
6698                 while (extent_slot >= 0) {
6699                         btrfs_item_key_to_cpu(path->nodes[0], &key,
6700                                               extent_slot);
6701                         if (key.objectid != bytenr)
6702                                 break;
6703                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
6704                             key.offset == num_bytes) {
6705                                 found_extent = 1;
6706                                 break;
6707                         }
6708                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
6709                             key.offset == owner_objectid) {
6710                                 found_extent = 1;
6711                                 break;
6712                         }
6713                         if (path->slots[0] - extent_slot > 5)
6714                                 break;
6715                         extent_slot--;
6716                 }
6717
6718                 if (!found_extent) {
6719                         BUG_ON(iref);
6720                         ret = remove_extent_backref(trans, path, NULL,
6721                                                     refs_to_drop,
6722                                                     is_data, &last_ref);
6723                         if (ret) {
6724                                 btrfs_abort_transaction(trans, ret);
6725                                 goto out;
6726                         }
6727                         btrfs_release_path(path);
6728                         path->leave_spinning = 1;
6729
6730                         key.objectid = bytenr;
6731                         key.type = BTRFS_EXTENT_ITEM_KEY;
6732                         key.offset = num_bytes;
6733
6734                         if (!is_data && skinny_metadata) {
6735                                 key.type = BTRFS_METADATA_ITEM_KEY;
6736                                 key.offset = owner_objectid;
6737                         }
6738
6739                         ret = btrfs_search_slot(trans, extent_root,
6740                                                 &key, path, -1, 1);
6741                         if (ret > 0 && skinny_metadata && path->slots[0]) {
6742                                 /*
6743                                  * Couldn't find our skinny metadata item,
6744                                  * see if we have ye olde extent item.
6745                                  */
6746                                 path->slots[0]--;
6747                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
6748                                                       path->slots[0]);
6749                                 if (key.objectid == bytenr &&
6750                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
6751                                     key.offset == num_bytes)
6752                                         ret = 0;
6753                         }
6754
6755                         if (ret > 0 && skinny_metadata) {
6756                                 skinny_metadata = false;
6757                                 key.objectid = bytenr;
6758                                 key.type = BTRFS_EXTENT_ITEM_KEY;
6759                                 key.offset = num_bytes;
6760                                 btrfs_release_path(path);
6761                                 ret = btrfs_search_slot(trans, extent_root,
6762                                                         &key, path, -1, 1);
6763                         }
6764
6765                         if (ret) {
6766                                 btrfs_err(info,
6767                                           "umm, got %d back from search, was looking for %llu",
6768                                           ret, bytenr);
6769                                 if (ret > 0)
6770                                         btrfs_print_leaf(path->nodes[0]);
6771                         }
6772                         if (ret < 0) {
6773                                 btrfs_abort_transaction(trans, ret);
6774                                 goto out;
6775                         }
6776                         extent_slot = path->slots[0];
6777                 }
6778         } else if (WARN_ON(ret == -ENOENT)) {
6779                 btrfs_print_leaf(path->nodes[0]);
6780                 btrfs_err(info,
6781                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
6782                         bytenr, parent, root_objectid, owner_objectid,
6783                         owner_offset);
6784                 btrfs_abort_transaction(trans, ret);
6785                 goto out;
6786         } else {
6787                 btrfs_abort_transaction(trans, ret);
6788                 goto out;
6789         }
6790
6791         leaf = path->nodes[0];
6792         item_size = btrfs_item_size_nr(leaf, extent_slot);
6793         if (unlikely(item_size < sizeof(*ei))) {
6794                 ret = -EINVAL;
6795                 btrfs_print_v0_err(info);
6796                 btrfs_abort_transaction(trans, ret);
6797                 goto out;
6798         }
6799         ei = btrfs_item_ptr(leaf, extent_slot,
6800                             struct btrfs_extent_item);
6801         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
6802             key.type == BTRFS_EXTENT_ITEM_KEY) {
6803                 struct btrfs_tree_block_info *bi;
6804                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
6805                 bi = (struct btrfs_tree_block_info *)(ei + 1);
6806                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
6807         }
6808
6809         refs = btrfs_extent_refs(leaf, ei);
6810         if (refs < refs_to_drop) {
6811                 btrfs_err(info,
6812                           "trying to drop %d refs but we only have %Lu for bytenr %Lu",
6813                           refs_to_drop, refs, bytenr);
6814                 ret = -EINVAL;
6815                 btrfs_abort_transaction(trans, ret);
6816                 goto out;
6817         }
6818         refs -= refs_to_drop;
6819
6820         if (refs > 0) {
6821                 if (extent_op)
6822                         __run_delayed_extent_op(extent_op, leaf, ei);
6823                 /*
6824                  * In the case of inline back ref, reference count will
6825                  * be updated by remove_extent_backref
6826                  */
6827                 if (iref) {
6828                         BUG_ON(!found_extent);
6829                 } else {
6830                         btrfs_set_extent_refs(leaf, ei, refs);
6831                         btrfs_mark_buffer_dirty(leaf);
6832                 }
6833                 if (found_extent) {
6834                         ret = remove_extent_backref(trans, path, iref,
6835                                                     refs_to_drop, is_data,
6836                                                     &last_ref);
6837                         if (ret) {
6838                                 btrfs_abort_transaction(trans, ret);
6839                                 goto out;
6840                         }
6841                 }
6842         } else {
6843                 if (found_extent) {
6844                         BUG_ON(is_data && refs_to_drop !=
6845                                extent_data_ref_count(path, iref));
6846                         if (iref) {
6847                                 BUG_ON(path->slots[0] != extent_slot);
6848                         } else {
6849                                 BUG_ON(path->slots[0] != extent_slot + 1);
6850                                 path->slots[0] = extent_slot;
6851                                 num_to_del = 2;
6852                         }
6853                 }
6854
6855                 last_ref = 1;
6856                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
6857                                       num_to_del);
6858                 if (ret) {
6859                         btrfs_abort_transaction(trans, ret);
6860                         goto out;
6861                 }
6862                 btrfs_release_path(path);
6863
6864                 if (is_data) {
6865                         ret = btrfs_del_csums(trans, info, bytenr, num_bytes);
6866                         if (ret) {
6867                                 btrfs_abort_transaction(trans, ret);
6868                                 goto out;
6869                         }
6870                 }
6871
6872                 ret = add_to_free_space_tree(trans, bytenr, num_bytes);
6873                 if (ret) {
6874                         btrfs_abort_transaction(trans, ret);
6875                         goto out;
6876                 }
6877
6878                 ret = update_block_group(trans, info, bytenr, num_bytes, 0);
6879                 if (ret) {
6880                         btrfs_abort_transaction(trans, ret);
6881                         goto out;
6882                 }
6883         }
6884         btrfs_release_path(path);
6885
6886 out:
6887         btrfs_free_path(path);
6888         return ret;
6889 }
6890
6891 /*
6892  * when we free an block, it is possible (and likely) that we free the last
6893  * delayed ref for that extent as well.  This searches the delayed ref tree for
6894  * a given extent, and if there are no other delayed refs to be processed, it
6895  * removes it from the tree.
6896  */
6897 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
6898                                       u64 bytenr)
6899 {
6900         struct btrfs_delayed_ref_head *head;
6901         struct btrfs_delayed_ref_root *delayed_refs;
6902         int ret = 0;
6903
6904         delayed_refs = &trans->transaction->delayed_refs;
6905         spin_lock(&delayed_refs->lock);
6906         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
6907         if (!head)
6908                 goto out_delayed_unlock;
6909
6910         spin_lock(&head->lock);
6911         if (!RB_EMPTY_ROOT(&head->ref_tree))
6912                 goto out;
6913
6914         if (head->extent_op) {
6915                 if (!head->must_insert_reserved)
6916                         goto out;
6917                 btrfs_free_delayed_extent_op(head->extent_op);
6918                 head->extent_op = NULL;
6919         }
6920
6921         /*
6922          * waiting for the lock here would deadlock.  If someone else has it
6923          * locked they are already in the process of dropping it anyway
6924          */
6925         if (!mutex_trylock(&head->mutex))
6926                 goto out;
6927
6928         /*
6929          * at this point we have a head with no other entries.  Go
6930          * ahead and process it.
6931          */
6932         rb_erase(&head->href_node, &delayed_refs->href_root);
6933         RB_CLEAR_NODE(&head->href_node);
6934         atomic_dec(&delayed_refs->num_entries);
6935
6936         /*
6937          * we don't take a ref on the node because we're removing it from the
6938          * tree, so we just steal the ref the tree was holding.
6939          */
6940         delayed_refs->num_heads--;
6941         if (head->processing == 0)
6942                 delayed_refs->num_heads_ready--;
6943         head->processing = 0;
6944         spin_unlock(&head->lock);
6945         spin_unlock(&delayed_refs->lock);
6946
6947         BUG_ON(head->extent_op);
6948         if (head->must_insert_reserved)
6949                 ret = 1;
6950
6951         mutex_unlock(&head->mutex);
6952         btrfs_put_delayed_ref_head(head);
6953         return ret;
6954 out:
6955         spin_unlock(&head->lock);
6956
6957 out_delayed_unlock:
6958         spin_unlock(&delayed_refs->lock);
6959         return 0;
6960 }
6961
6962 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
6963                            struct btrfs_root *root,
6964                            struct extent_buffer *buf,
6965                            u64 parent, int last_ref)
6966 {
6967         struct btrfs_fs_info *fs_info = root->fs_info;
6968         int pin = 1;
6969         int ret;
6970
6971         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
6972                 int old_ref_mod, new_ref_mod;
6973
6974                 btrfs_ref_tree_mod(root, buf->start, buf->len, parent,
6975                                    root->root_key.objectid,
6976                                    btrfs_header_level(buf), 0,
6977                                    BTRFS_DROP_DELAYED_REF);
6978                 ret = btrfs_add_delayed_tree_ref(trans, buf->start,
6979                                                  buf->len, parent,
6980                                                  root->root_key.objectid,
6981                                                  btrfs_header_level(buf),
6982                                                  BTRFS_DROP_DELAYED_REF, NULL,
6983                                                  &old_ref_mod, &new_ref_mod);
6984                 BUG_ON(ret); /* -ENOMEM */
6985                 pin = old_ref_mod >= 0 && new_ref_mod < 0;
6986         }
6987
6988         if (last_ref && btrfs_header_generation(buf) == trans->transid) {
6989                 struct btrfs_block_group_cache *cache;
6990
6991                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
6992                         ret = check_ref_cleanup(trans, buf->start);
6993                         if (!ret)
6994                                 goto out;
6995                 }
6996
6997                 pin = 0;
6998                 cache = btrfs_lookup_block_group(fs_info, buf->start);
6999
7000                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
7001                         pin_down_extent(fs_info, cache, buf->start,
7002                                         buf->len, 1);
7003                         btrfs_put_block_group(cache);
7004                         goto out;
7005                 }
7006
7007                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
7008
7009                 btrfs_add_free_space(cache, buf->start, buf->len);
7010                 btrfs_free_reserved_bytes(cache, buf->len, 0);
7011                 btrfs_put_block_group(cache);
7012                 trace_btrfs_reserved_extent_free(fs_info, buf->start, buf->len);
7013         }
7014 out:
7015         if (pin)
7016                 add_pinned_bytes(fs_info, buf->len, true,
7017                                  root->root_key.objectid);
7018
7019         if (last_ref) {
7020                 /*
7021                  * Deleting the buffer, clear the corrupt flag since it doesn't
7022                  * matter anymore.
7023                  */
7024                 clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
7025         }
7026 }
7027
7028 /* Can return -ENOMEM */
7029 int btrfs_free_extent(struct btrfs_trans_handle *trans,
7030                       struct btrfs_root *root,
7031                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
7032                       u64 owner, u64 offset)
7033 {
7034         struct btrfs_fs_info *fs_info = root->fs_info;
7035         int old_ref_mod, new_ref_mod;
7036         int ret;
7037
7038         if (btrfs_is_testing(fs_info))
7039                 return 0;
7040
7041         if (root_objectid != BTRFS_TREE_LOG_OBJECTID)
7042                 btrfs_ref_tree_mod(root, bytenr, num_bytes, parent,
7043                                    root_objectid, owner, offset,
7044                                    BTRFS_DROP_DELAYED_REF);
7045
7046         /*
7047          * tree log blocks never actually go into the extent allocation
7048          * tree, just update pinning info and exit early.
7049          */
7050         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
7051                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
7052                 /* unlocks the pinned mutex */
7053                 btrfs_pin_extent(fs_info, bytenr, num_bytes, 1);
7054                 old_ref_mod = new_ref_mod = 0;
7055                 ret = 0;
7056         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
7057                 ret = btrfs_add_delayed_tree_ref(trans, bytenr,
7058                                                  num_bytes, parent,
7059                                                  root_objectid, (int)owner,
7060                                                  BTRFS_DROP_DELAYED_REF, NULL,
7061                                                  &old_ref_mod, &new_ref_mod);
7062         } else {
7063                 ret = btrfs_add_delayed_data_ref(trans, bytenr,
7064                                                  num_bytes, parent,
7065                                                  root_objectid, owner, offset,
7066                                                  0, BTRFS_DROP_DELAYED_REF,
7067                                                  &old_ref_mod, &new_ref_mod);
7068         }
7069
7070         if (ret == 0 && old_ref_mod >= 0 && new_ref_mod < 0) {
7071                 bool metadata = owner < BTRFS_FIRST_FREE_OBJECTID;
7072
7073                 add_pinned_bytes(fs_info, num_bytes, metadata, root_objectid);
7074         }
7075
7076         return ret;
7077 }
7078
7079 /*
7080  * when we wait for progress in the block group caching, its because
7081  * our allocation attempt failed at least once.  So, we must sleep
7082  * and let some progress happen before we try again.
7083  *
7084  * This function will sleep at least once waiting for new free space to
7085  * show up, and then it will check the block group free space numbers
7086  * for our min num_bytes.  Another option is to have it go ahead
7087  * and look in the rbtree for a free extent of a given size, but this
7088  * is a good start.
7089  *
7090  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
7091  * any of the information in this block group.
7092  */
7093 static noinline void
7094 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
7095                                 u64 num_bytes)
7096 {
7097         struct btrfs_caching_control *caching_ctl;
7098
7099         caching_ctl = get_caching_control(cache);
7100         if (!caching_ctl)
7101                 return;
7102
7103         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
7104                    (cache->free_space_ctl->free_space >= num_bytes));
7105
7106         put_caching_control(caching_ctl);
7107 }
7108
7109 static noinline int
7110 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
7111 {
7112         struct btrfs_caching_control *caching_ctl;
7113         int ret = 0;
7114
7115         caching_ctl = get_caching_control(cache);
7116         if (!caching_ctl)
7117                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
7118
7119         wait_event(caching_ctl->wait, block_group_cache_done(cache));
7120         if (cache->cached == BTRFS_CACHE_ERROR)
7121                 ret = -EIO;
7122         put_caching_control(caching_ctl);
7123         return ret;
7124 }
7125
7126 enum btrfs_loop_type {
7127         LOOP_CACHING_NOWAIT = 0,
7128         LOOP_CACHING_WAIT = 1,
7129         LOOP_ALLOC_CHUNK = 2,
7130         LOOP_NO_EMPTY_SIZE = 3,
7131 };
7132
7133 static inline void
7134 btrfs_lock_block_group(struct btrfs_block_group_cache *cache,
7135                        int delalloc)
7136 {
7137         if (delalloc)
7138                 down_read(&cache->data_rwsem);
7139 }
7140
7141 static inline void
7142 btrfs_grab_block_group(struct btrfs_block_group_cache *cache,
7143                        int delalloc)
7144 {
7145         btrfs_get_block_group(cache);
7146         if (delalloc)
7147                 down_read(&cache->data_rwsem);
7148 }
7149
7150 static struct btrfs_block_group_cache *
7151 btrfs_lock_cluster(struct btrfs_block_group_cache *block_group,
7152                    struct btrfs_free_cluster *cluster,
7153                    int delalloc)
7154 {
7155         struct btrfs_block_group_cache *used_bg = NULL;
7156
7157         spin_lock(&cluster->refill_lock);
7158         while (1) {
7159                 used_bg = cluster->block_group;
7160                 if (!used_bg)
7161                         return NULL;
7162
7163                 if (used_bg == block_group)
7164                         return used_bg;
7165
7166                 btrfs_get_block_group(used_bg);
7167
7168                 if (!delalloc)
7169                         return used_bg;
7170
7171                 if (down_read_trylock(&used_bg->data_rwsem))
7172                         return used_bg;
7173
7174                 spin_unlock(&cluster->refill_lock);
7175
7176                 /* We should only have one-level nested. */
7177                 down_read_nested(&used_bg->data_rwsem, SINGLE_DEPTH_NESTING);
7178
7179                 spin_lock(&cluster->refill_lock);
7180                 if (used_bg == cluster->block_group)
7181                         return used_bg;
7182
7183                 up_read(&used_bg->data_rwsem);
7184                 btrfs_put_block_group(used_bg);
7185         }
7186 }
7187
7188 static inline void
7189 btrfs_release_block_group(struct btrfs_block_group_cache *cache,
7190                          int delalloc)
7191 {
7192         if (delalloc)
7193                 up_read(&cache->data_rwsem);
7194         btrfs_put_block_group(cache);
7195 }
7196
7197 /*
7198  * walks the btree of allocated extents and find a hole of a given size.
7199  * The key ins is changed to record the hole:
7200  * ins->objectid == start position
7201  * ins->flags = BTRFS_EXTENT_ITEM_KEY
7202  * ins->offset == the size of the hole.
7203  * Any available blocks before search_start are skipped.
7204  *
7205  * If there is no suitable free space, we will record the max size of
7206  * the free space extent currently.
7207  */
7208 static noinline int find_free_extent(struct btrfs_fs_info *fs_info,
7209                                 u64 ram_bytes, u64 num_bytes, u64 empty_size,
7210                                 u64 hint_byte, struct btrfs_key *ins,
7211                                 u64 flags, int delalloc)
7212 {
7213         int ret = 0;
7214         struct btrfs_root *root = fs_info->extent_root;
7215         struct btrfs_free_cluster *last_ptr = NULL;
7216         struct btrfs_block_group_cache *block_group = NULL;
7217         u64 search_start = 0;
7218         u64 max_extent_size = 0;
7219         u64 empty_cluster = 0;
7220         struct btrfs_space_info *space_info;
7221         int loop = 0;
7222         int index = btrfs_bg_flags_to_raid_index(flags);
7223         bool failed_cluster_refill = false;
7224         bool failed_alloc = false;
7225         bool use_cluster = true;
7226         bool have_caching_bg = false;
7227         bool orig_have_caching_bg = false;
7228         bool full_search = false;
7229
7230         WARN_ON(num_bytes < fs_info->sectorsize);
7231         ins->type = BTRFS_EXTENT_ITEM_KEY;
7232         ins->objectid = 0;
7233         ins->offset = 0;
7234
7235         trace_find_free_extent(fs_info, num_bytes, empty_size, flags);
7236
7237         space_info = __find_space_info(fs_info, flags);
7238         if (!space_info) {
7239                 btrfs_err(fs_info, "No space info for %llu", flags);
7240                 return -ENOSPC;
7241         }
7242
7243         /*
7244          * If our free space is heavily fragmented we may not be able to make
7245          * big contiguous allocations, so instead of doing the expensive search
7246          * for free space, simply return ENOSPC with our max_extent_size so we
7247          * can go ahead and search for a more manageable chunk.
7248          *
7249          * If our max_extent_size is large enough for our allocation simply
7250          * disable clustering since we will likely not be able to find enough
7251          * space to create a cluster and induce latency trying.
7252          */
7253         if (unlikely(space_info->max_extent_size)) {
7254                 spin_lock(&space_info->lock);
7255                 if (space_info->max_extent_size &&
7256                     num_bytes > space_info->max_extent_size) {
7257                         ins->offset = space_info->max_extent_size;
7258                         spin_unlock(&space_info->lock);
7259                         return -ENOSPC;
7260                 } else if (space_info->max_extent_size) {
7261                         use_cluster = false;
7262                 }
7263                 spin_unlock(&space_info->lock);
7264         }
7265
7266         last_ptr = fetch_cluster_info(fs_info, space_info, &empty_cluster);
7267         if (last_ptr) {
7268                 spin_lock(&last_ptr->lock);
7269                 if (last_ptr->block_group)
7270                         hint_byte = last_ptr->window_start;
7271                 if (last_ptr->fragmented) {
7272                         /*
7273                          * We still set window_start so we can keep track of the
7274                          * last place we found an allocation to try and save
7275                          * some time.
7276                          */
7277                         hint_byte = last_ptr->window_start;
7278                         use_cluster = false;
7279                 }
7280                 spin_unlock(&last_ptr->lock);
7281         }
7282
7283         search_start = max(search_start, first_logical_byte(fs_info, 0));
7284         search_start = max(search_start, hint_byte);
7285         if (search_start == hint_byte) {
7286                 block_group = btrfs_lookup_block_group(fs_info, search_start);
7287                 /*
7288                  * we don't want to use the block group if it doesn't match our
7289                  * allocation bits, or if its not cached.
7290                  *
7291                  * However if we are re-searching with an ideal block group
7292                  * picked out then we don't care that the block group is cached.
7293                  */
7294                 if (block_group && block_group_bits(block_group, flags) &&
7295                     block_group->cached != BTRFS_CACHE_NO) {
7296                         down_read(&space_info->groups_sem);
7297                         if (list_empty(&block_group->list) ||
7298                             block_group->ro) {
7299                                 /*
7300                                  * someone is removing this block group,
7301                                  * we can't jump into the have_block_group
7302                                  * target because our list pointers are not
7303                                  * valid
7304                                  */
7305                                 btrfs_put_block_group(block_group);
7306                                 up_read(&space_info->groups_sem);
7307                         } else {
7308                                 index = btrfs_bg_flags_to_raid_index(
7309                                                 block_group->flags);
7310                                 btrfs_lock_block_group(block_group, delalloc);
7311                                 goto have_block_group;
7312                         }
7313                 } else if (block_group) {
7314                         btrfs_put_block_group(block_group);
7315                 }
7316         }
7317 search:
7318         have_caching_bg = false;
7319         if (index == 0 || index == btrfs_bg_flags_to_raid_index(flags))
7320                 full_search = true;
7321         down_read(&space_info->groups_sem);
7322         list_for_each_entry(block_group, &space_info->block_groups[index],
7323                             list) {
7324                 u64 offset;
7325                 int cached;
7326
7327                 /* If the block group is read-only, we can skip it entirely. */
7328                 if (unlikely(block_group->ro))
7329                         continue;
7330
7331                 btrfs_grab_block_group(block_group, delalloc);
7332                 search_start = block_group->key.objectid;
7333
7334                 /*
7335                  * this can happen if we end up cycling through all the
7336                  * raid types, but we want to make sure we only allocate
7337                  * for the proper type.
7338                  */
7339                 if (!block_group_bits(block_group, flags)) {
7340                         u64 extra = BTRFS_BLOCK_GROUP_DUP |
7341                                 BTRFS_BLOCK_GROUP_RAID1 |
7342                                 BTRFS_BLOCK_GROUP_RAID5 |
7343                                 BTRFS_BLOCK_GROUP_RAID6 |
7344                                 BTRFS_BLOCK_GROUP_RAID10;
7345
7346                         /*
7347                          * if they asked for extra copies and this block group
7348                          * doesn't provide them, bail.  This does allow us to
7349                          * fill raid0 from raid1.
7350                          */
7351                         if ((flags & extra) && !(block_group->flags & extra))
7352                                 goto loop;
7353                 }
7354
7355 have_block_group:
7356                 cached = block_group_cache_done(block_group);
7357                 if (unlikely(!cached)) {
7358                         have_caching_bg = true;
7359                         ret = cache_block_group(block_group, 0);
7360                         BUG_ON(ret < 0);
7361                         ret = 0;
7362                 }
7363
7364                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
7365                         goto loop;
7366
7367                 /*
7368                  * Ok we want to try and use the cluster allocator, so
7369                  * lets look there
7370                  */
7371                 if (last_ptr && use_cluster) {
7372                         struct btrfs_block_group_cache *used_block_group;
7373                         unsigned long aligned_cluster;
7374                         /*
7375                          * the refill lock keeps out other
7376                          * people trying to start a new cluster
7377                          */
7378                         used_block_group = btrfs_lock_cluster(block_group,
7379                                                               last_ptr,
7380                                                               delalloc);
7381                         if (!used_block_group)
7382                                 goto refill_cluster;
7383
7384                         if (used_block_group != block_group &&
7385                             (used_block_group->ro ||
7386                              !block_group_bits(used_block_group, flags)))
7387                                 goto release_cluster;
7388
7389                         offset = btrfs_alloc_from_cluster(used_block_group,
7390                                                 last_ptr,
7391                                                 num_bytes,
7392                                                 used_block_group->key.objectid,
7393                                                 &max_extent_size);
7394                         if (offset) {
7395                                 /* we have a block, we're done */
7396                                 spin_unlock(&last_ptr->refill_lock);
7397                                 trace_btrfs_reserve_extent_cluster(
7398                                                 used_block_group,
7399                                                 search_start, num_bytes);
7400                                 if (used_block_group != block_group) {
7401                                         btrfs_release_block_group(block_group,
7402                                                                   delalloc);
7403                                         block_group = used_block_group;
7404                                 }
7405                                 goto checks;
7406                         }
7407
7408                         WARN_ON(last_ptr->block_group != used_block_group);
7409 release_cluster:
7410                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
7411                          * set up a new clusters, so lets just skip it
7412                          * and let the allocator find whatever block
7413                          * it can find.  If we reach this point, we
7414                          * will have tried the cluster allocator
7415                          * plenty of times and not have found
7416                          * anything, so we are likely way too
7417                          * fragmented for the clustering stuff to find
7418                          * anything.
7419                          *
7420                          * However, if the cluster is taken from the
7421                          * current block group, release the cluster
7422                          * first, so that we stand a better chance of
7423                          * succeeding in the unclustered
7424                          * allocation.  */
7425                         if (loop >= LOOP_NO_EMPTY_SIZE &&
7426                             used_block_group != block_group) {
7427                                 spin_unlock(&last_ptr->refill_lock);
7428                                 btrfs_release_block_group(used_block_group,
7429                                                           delalloc);
7430                                 goto unclustered_alloc;
7431                         }
7432
7433                         /*
7434                          * this cluster didn't work out, free it and
7435                          * start over
7436                          */
7437                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7438
7439                         if (used_block_group != block_group)
7440                                 btrfs_release_block_group(used_block_group,
7441                                                           delalloc);
7442 refill_cluster:
7443                         if (loop >= LOOP_NO_EMPTY_SIZE) {
7444                                 spin_unlock(&last_ptr->refill_lock);
7445                                 goto unclustered_alloc;
7446                         }
7447
7448                         aligned_cluster = max_t(unsigned long,
7449                                                 empty_cluster + empty_size,
7450                                               block_group->full_stripe_len);
7451
7452                         /* allocate a cluster in this block group */
7453                         ret = btrfs_find_space_cluster(fs_info, block_group,
7454                                                        last_ptr, search_start,
7455                                                        num_bytes,
7456                                                        aligned_cluster);
7457                         if (ret == 0) {
7458                                 /*
7459                                  * now pull our allocation out of this
7460                                  * cluster
7461                                  */
7462                                 offset = btrfs_alloc_from_cluster(block_group,
7463                                                         last_ptr,
7464                                                         num_bytes,
7465                                                         search_start,
7466                                                         &max_extent_size);
7467                                 if (offset) {
7468                                         /* we found one, proceed */
7469                                         spin_unlock(&last_ptr->refill_lock);
7470                                         trace_btrfs_reserve_extent_cluster(
7471                                                 block_group, search_start,
7472                                                 num_bytes);
7473                                         goto checks;
7474                                 }
7475                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
7476                                    && !failed_cluster_refill) {
7477                                 spin_unlock(&last_ptr->refill_lock);
7478
7479                                 failed_cluster_refill = true;
7480                                 wait_block_group_cache_progress(block_group,
7481                                        num_bytes + empty_cluster + empty_size);
7482                                 goto have_block_group;
7483                         }
7484
7485                         /*
7486                          * at this point we either didn't find a cluster
7487                          * or we weren't able to allocate a block from our
7488                          * cluster.  Free the cluster we've been trying
7489                          * to use, and go to the next block group
7490                          */
7491                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7492                         spin_unlock(&last_ptr->refill_lock);
7493                         goto loop;
7494                 }
7495
7496 unclustered_alloc:
7497                 /*
7498                  * We are doing an unclustered alloc, set the fragmented flag so
7499                  * we don't bother trying to setup a cluster again until we get
7500                  * more space.
7501                  */
7502                 if (unlikely(last_ptr)) {
7503                         spin_lock(&last_ptr->lock);
7504                         last_ptr->fragmented = 1;
7505                         spin_unlock(&last_ptr->lock);
7506                 }
7507                 if (cached) {
7508                         struct btrfs_free_space_ctl *ctl =
7509                                 block_group->free_space_ctl;
7510
7511                         spin_lock(&ctl->tree_lock);
7512                         if (ctl->free_space <
7513                             num_bytes + empty_cluster + empty_size) {
7514                                 if (ctl->free_space > max_extent_size)
7515                                         max_extent_size = ctl->free_space;
7516                                 spin_unlock(&ctl->tree_lock);
7517                                 goto loop;
7518                         }
7519                         spin_unlock(&ctl->tree_lock);
7520                 }
7521
7522                 offset = btrfs_find_space_for_alloc(block_group, search_start,
7523                                                     num_bytes, empty_size,
7524                                                     &max_extent_size);
7525                 /*
7526                  * If we didn't find a chunk, and we haven't failed on this
7527                  * block group before, and this block group is in the middle of
7528                  * caching and we are ok with waiting, then go ahead and wait
7529                  * for progress to be made, and set failed_alloc to true.
7530                  *
7531                  * If failed_alloc is true then we've already waited on this
7532                  * block group once and should move on to the next block group.
7533                  */
7534                 if (!offset && !failed_alloc && !cached &&
7535                     loop > LOOP_CACHING_NOWAIT) {
7536                         wait_block_group_cache_progress(block_group,
7537                                                 num_bytes + empty_size);
7538                         failed_alloc = true;
7539                         goto have_block_group;
7540                 } else if (!offset) {
7541                         goto loop;
7542                 }
7543 checks:
7544                 search_start = round_up(offset, fs_info->stripesize);
7545
7546                 /* move on to the next group */
7547                 if (search_start + num_bytes >
7548                     block_group->key.objectid + block_group->key.offset) {
7549                         btrfs_add_free_space(block_group, offset, num_bytes);
7550                         goto loop;
7551                 }
7552
7553                 if (offset < search_start)
7554                         btrfs_add_free_space(block_group, offset,
7555                                              search_start - offset);
7556
7557                 ret = btrfs_add_reserved_bytes(block_group, ram_bytes,
7558                                 num_bytes, delalloc);
7559                 if (ret == -EAGAIN) {
7560                         btrfs_add_free_space(block_group, offset, num_bytes);
7561                         goto loop;
7562                 }
7563                 btrfs_inc_block_group_reservations(block_group);
7564
7565                 /* we are all good, lets return */
7566                 ins->objectid = search_start;
7567                 ins->offset = num_bytes;
7568
7569                 trace_btrfs_reserve_extent(block_group, search_start, num_bytes);
7570                 btrfs_release_block_group(block_group, delalloc);
7571                 break;
7572 loop:
7573                 failed_cluster_refill = false;
7574                 failed_alloc = false;
7575                 BUG_ON(btrfs_bg_flags_to_raid_index(block_group->flags) !=
7576                        index);
7577                 btrfs_release_block_group(block_group, delalloc);
7578                 cond_resched();
7579         }
7580         up_read(&space_info->groups_sem);
7581
7582         if ((loop == LOOP_CACHING_NOWAIT) && have_caching_bg
7583                 && !orig_have_caching_bg)
7584                 orig_have_caching_bg = true;
7585
7586         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
7587                 goto search;
7588
7589         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
7590                 goto search;
7591
7592         /*
7593          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
7594          *                      caching kthreads as we move along
7595          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
7596          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
7597          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
7598          *                      again
7599          */
7600         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
7601                 index = 0;
7602                 if (loop == LOOP_CACHING_NOWAIT) {
7603                         /*
7604                          * We want to skip the LOOP_CACHING_WAIT step if we
7605                          * don't have any uncached bgs and we've already done a
7606                          * full search through.
7607                          */
7608                         if (orig_have_caching_bg || !full_search)
7609                                 loop = LOOP_CACHING_WAIT;
7610                         else
7611                                 loop = LOOP_ALLOC_CHUNK;
7612                 } else {
7613                         loop++;
7614                 }
7615
7616                 if (loop == LOOP_ALLOC_CHUNK) {
7617                         struct btrfs_trans_handle *trans;
7618                         int exist = 0;
7619
7620                         trans = current->journal_info;
7621                         if (trans)
7622                                 exist = 1;
7623                         else
7624                                 trans = btrfs_join_transaction(root);
7625
7626                         if (IS_ERR(trans)) {
7627                                 ret = PTR_ERR(trans);
7628                                 goto out;
7629                         }
7630
7631                         ret = do_chunk_alloc(trans, flags, CHUNK_ALLOC_FORCE);
7632
7633                         /*
7634                          * If we can't allocate a new chunk we've already looped
7635                          * through at least once, move on to the NO_EMPTY_SIZE
7636                          * case.
7637                          */
7638                         if (ret == -ENOSPC)
7639                                 loop = LOOP_NO_EMPTY_SIZE;
7640
7641                         /*
7642                          * Do not bail out on ENOSPC since we
7643                          * can do more things.
7644                          */
7645                         if (ret < 0 && ret != -ENOSPC)
7646                                 btrfs_abort_transaction(trans, ret);
7647                         else
7648                                 ret = 0;
7649                         if (!exist)
7650                                 btrfs_end_transaction(trans);
7651                         if (ret)
7652                                 goto out;
7653                 }
7654
7655                 if (loop == LOOP_NO_EMPTY_SIZE) {
7656                         /*
7657                          * Don't loop again if we already have no empty_size and
7658                          * no empty_cluster.
7659                          */
7660                         if (empty_size == 0 &&
7661                             empty_cluster == 0) {
7662                                 ret = -ENOSPC;
7663                                 goto out;
7664                         }
7665                         empty_size = 0;
7666                         empty_cluster = 0;
7667                 }
7668
7669                 goto search;
7670         } else if (!ins->objectid) {
7671                 ret = -ENOSPC;
7672         } else if (ins->objectid) {
7673                 if (!use_cluster && last_ptr) {
7674                         spin_lock(&last_ptr->lock);
7675                         last_ptr->window_start = ins->objectid;
7676                         spin_unlock(&last_ptr->lock);
7677                 }
7678                 ret = 0;
7679         }
7680 out:
7681         if (ret == -ENOSPC) {
7682                 spin_lock(&space_info->lock);
7683                 space_info->max_extent_size = max_extent_size;
7684                 spin_unlock(&space_info->lock);
7685                 ins->offset = max_extent_size;
7686         }
7687         return ret;
7688 }
7689
7690 static void dump_space_info(struct btrfs_fs_info *fs_info,
7691                             struct btrfs_space_info *info, u64 bytes,
7692                             int dump_block_groups)
7693 {
7694         struct btrfs_block_group_cache *cache;
7695         int index = 0;
7696
7697         spin_lock(&info->lock);
7698         btrfs_info(fs_info, "space_info %llu has %llu free, is %sfull",
7699                    info->flags,
7700                    info->total_bytes - btrfs_space_info_used(info, true),
7701                    info->full ? "" : "not ");
7702         btrfs_info(fs_info,
7703                 "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu",
7704                 info->total_bytes, info->bytes_used, info->bytes_pinned,
7705                 info->bytes_reserved, info->bytes_may_use,
7706                 info->bytes_readonly);
7707         spin_unlock(&info->lock);
7708
7709         if (!dump_block_groups)
7710                 return;
7711
7712         down_read(&info->groups_sem);
7713 again:
7714         list_for_each_entry(cache, &info->block_groups[index], list) {
7715                 spin_lock(&cache->lock);
7716                 btrfs_info(fs_info,
7717                         "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s",
7718                         cache->key.objectid, cache->key.offset,
7719                         btrfs_block_group_used(&cache->item), cache->pinned,
7720                         cache->reserved, cache->ro ? "[readonly]" : "");
7721                 btrfs_dump_free_space(cache, bytes);
7722                 spin_unlock(&cache->lock);
7723         }
7724         if (++index < BTRFS_NR_RAID_TYPES)
7725                 goto again;
7726         up_read(&info->groups_sem);
7727 }
7728
7729 /*
7730  * btrfs_reserve_extent - entry point to the extent allocator. Tries to find a
7731  *                        hole that is at least as big as @num_bytes.
7732  *
7733  * @root           -    The root that will contain this extent
7734  *
7735  * @ram_bytes      -    The amount of space in ram that @num_bytes take. This
7736  *                      is used for accounting purposes. This value differs
7737  *                      from @num_bytes only in the case of compressed extents.
7738  *
7739  * @num_bytes      -    Number of bytes to allocate on-disk.
7740  *
7741  * @min_alloc_size -    Indicates the minimum amount of space that the
7742  *                      allocator should try to satisfy. In some cases
7743  *                      @num_bytes may be larger than what is required and if
7744  *                      the filesystem is fragmented then allocation fails.
7745  *                      However, the presence of @min_alloc_size gives a
7746  *                      chance to try and satisfy the smaller allocation.
7747  *
7748  * @empty_size     -    A hint that you plan on doing more COW. This is the
7749  *                      size in bytes the allocator should try to find free
7750  *                      next to the block it returns.  This is just a hint and
7751  *                      may be ignored by the allocator.
7752  *
7753  * @hint_byte      -    Hint to the allocator to start searching above the byte
7754  *                      address passed. It might be ignored.
7755  *
7756  * @ins            -    This key is modified to record the found hole. It will
7757  *                      have the following values:
7758  *                      ins->objectid == start position
7759  *                      ins->flags = BTRFS_EXTENT_ITEM_KEY
7760  *                      ins->offset == the size of the hole.
7761  *
7762  * @is_data        -    Boolean flag indicating whether an extent is
7763  *                      allocated for data (true) or metadata (false)
7764  *
7765  * @delalloc       -    Boolean flag indicating whether this allocation is for
7766  *                      delalloc or not. If 'true' data_rwsem of block groups
7767  *                      is going to be acquired.
7768  *
7769  *
7770  * Returns 0 when an allocation succeeded or < 0 when an error occurred. In
7771  * case -ENOSPC is returned then @ins->offset will contain the size of the
7772  * largest available hole the allocator managed to find.
7773  */
7774 int btrfs_reserve_extent(struct btrfs_root *root, u64 ram_bytes,
7775                          u64 num_bytes, u64 min_alloc_size,
7776                          u64 empty_size, u64 hint_byte,
7777                          struct btrfs_key *ins, int is_data, int delalloc)
7778 {
7779         struct btrfs_fs_info *fs_info = root->fs_info;
7780         bool final_tried = num_bytes == min_alloc_size;
7781         u64 flags;
7782         int ret;
7783
7784         flags = get_alloc_profile_by_root(root, is_data);
7785 again:
7786         WARN_ON(num_bytes < fs_info->sectorsize);
7787         ret = find_free_extent(fs_info, ram_bytes, num_bytes, empty_size,
7788                                hint_byte, ins, flags, delalloc);
7789         if (!ret && !is_data) {
7790                 btrfs_dec_block_group_reservations(fs_info, ins->objectid);
7791         } else if (ret == -ENOSPC) {
7792                 if (!final_tried && ins->offset) {
7793                         num_bytes = min(num_bytes >> 1, ins->offset);
7794                         num_bytes = round_down(num_bytes,
7795                                                fs_info->sectorsize);
7796                         num_bytes = max(num_bytes, min_alloc_size);
7797                         ram_bytes = num_bytes;
7798                         if (num_bytes == min_alloc_size)
7799                                 final_tried = true;
7800                         goto again;
7801                 } else if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
7802                         struct btrfs_space_info *sinfo;
7803
7804                         sinfo = __find_space_info(fs_info, flags);
7805                         btrfs_err(fs_info,
7806                                   "allocation failed flags %llu, wanted %llu",
7807                                   flags, num_bytes);
7808                         if (sinfo)
7809                                 dump_space_info(fs_info, sinfo, num_bytes, 1);
7810                 }
7811         }
7812
7813         return ret;
7814 }
7815
7816 static int __btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
7817                                         u64 start, u64 len,
7818                                         int pin, int delalloc)
7819 {
7820         struct btrfs_block_group_cache *cache;
7821         int ret = 0;
7822
7823         cache = btrfs_lookup_block_group(fs_info, start);
7824         if (!cache) {
7825                 btrfs_err(fs_info, "Unable to find block group for %llu",
7826                           start);
7827                 return -ENOSPC;
7828         }
7829
7830         if (pin)
7831                 pin_down_extent(fs_info, cache, start, len, 1);
7832         else {
7833                 if (btrfs_test_opt(fs_info, DISCARD))
7834                         ret = btrfs_discard_extent(fs_info, start, len, NULL);
7835                 btrfs_add_free_space(cache, start, len);
7836                 btrfs_free_reserved_bytes(cache, len, delalloc);
7837                 trace_btrfs_reserved_extent_free(fs_info, start, len);
7838         }
7839
7840         btrfs_put_block_group(cache);
7841         return ret;
7842 }
7843
7844 int btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
7845                                u64 start, u64 len, int delalloc)
7846 {
7847         return __btrfs_free_reserved_extent(fs_info, start, len, 0, delalloc);
7848 }
7849
7850 int btrfs_free_and_pin_reserved_extent(struct btrfs_fs_info *fs_info,
7851                                        u64 start, u64 len)
7852 {
7853         return __btrfs_free_reserved_extent(fs_info, start, len, 1, 0);
7854 }
7855
7856 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
7857                                       u64 parent, u64 root_objectid,
7858                                       u64 flags, u64 owner, u64 offset,
7859                                       struct btrfs_key *ins, int ref_mod)
7860 {
7861         struct btrfs_fs_info *fs_info = trans->fs_info;
7862         int ret;
7863         struct btrfs_extent_item *extent_item;
7864         struct btrfs_extent_inline_ref *iref;
7865         struct btrfs_path *path;
7866         struct extent_buffer *leaf;
7867         int type;
7868         u32 size;
7869
7870         if (parent > 0)
7871                 type = BTRFS_SHARED_DATA_REF_KEY;
7872         else
7873                 type = BTRFS_EXTENT_DATA_REF_KEY;
7874
7875         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
7876
7877         path = btrfs_alloc_path();
7878         if (!path)
7879                 return -ENOMEM;
7880
7881         path->leave_spinning = 1;
7882         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
7883                                       ins, size);
7884         if (ret) {
7885                 btrfs_free_path(path);
7886                 return ret;
7887         }
7888
7889         leaf = path->nodes[0];
7890         extent_item = btrfs_item_ptr(leaf, path->slots[0],
7891                                      struct btrfs_extent_item);
7892         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
7893         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
7894         btrfs_set_extent_flags(leaf, extent_item,
7895                                flags | BTRFS_EXTENT_FLAG_DATA);
7896
7897         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
7898         btrfs_set_extent_inline_ref_type(leaf, iref, type);
7899         if (parent > 0) {
7900                 struct btrfs_shared_data_ref *ref;
7901                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
7902                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
7903                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
7904         } else {
7905                 struct btrfs_extent_data_ref *ref;
7906                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
7907                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
7908                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
7909                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
7910                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
7911         }
7912
7913         btrfs_mark_buffer_dirty(path->nodes[0]);
7914         btrfs_free_path(path);
7915
7916         ret = remove_from_free_space_tree(trans, ins->objectid, ins->offset);
7917         if (ret)
7918                 return ret;
7919
7920         ret = update_block_group(trans, fs_info, ins->objectid, ins->offset, 1);
7921         if (ret) { /* -ENOENT, logic error */
7922                 btrfs_err(fs_info, "update block group failed for %llu %llu",
7923                         ins->objectid, ins->offset);
7924                 BUG();
7925         }
7926         trace_btrfs_reserved_extent_alloc(fs_info, ins->objectid, ins->offset);
7927         return ret;
7928 }
7929
7930 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
7931                                      struct btrfs_delayed_ref_node *node,
7932                                      struct btrfs_delayed_extent_op *extent_op)
7933 {
7934         struct btrfs_fs_info *fs_info = trans->fs_info;
7935         int ret;
7936         struct btrfs_extent_item *extent_item;
7937         struct btrfs_key extent_key;
7938         struct btrfs_tree_block_info *block_info;
7939         struct btrfs_extent_inline_ref *iref;
7940         struct btrfs_path *path;
7941         struct extent_buffer *leaf;
7942         struct btrfs_delayed_tree_ref *ref;
7943         u32 size = sizeof(*extent_item) + sizeof(*iref);
7944         u64 num_bytes;
7945         u64 flags = extent_op->flags_to_set;
7946         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
7947
7948         ref = btrfs_delayed_node_to_tree_ref(node);
7949
7950         extent_key.objectid = node->bytenr;
7951         if (skinny_metadata) {
7952                 extent_key.offset = ref->level;
7953                 extent_key.type = BTRFS_METADATA_ITEM_KEY;
7954                 num_bytes = fs_info->nodesize;
7955         } else {
7956                 extent_key.offset = node->num_bytes;
7957                 extent_key.type = BTRFS_EXTENT_ITEM_KEY;
7958                 size += sizeof(*block_info);
7959                 num_bytes = node->num_bytes;
7960         }
7961
7962         path = btrfs_alloc_path();
7963         if (!path) {
7964                 btrfs_free_and_pin_reserved_extent(fs_info,
7965                                                    extent_key.objectid,
7966                                                    fs_info->nodesize);
7967                 return -ENOMEM;
7968         }
7969
7970         path->leave_spinning = 1;
7971         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
7972                                       &extent_key, size);
7973         if (ret) {
7974                 btrfs_free_path(path);
7975                 btrfs_free_and_pin_reserved_extent(fs_info,
7976                                                    extent_key.objectid,
7977                                                    fs_info->nodesize);
7978                 return ret;
7979         }
7980
7981         leaf = path->nodes[0];
7982         extent_item = btrfs_item_ptr(leaf, path->slots[0],
7983                                      struct btrfs_extent_item);
7984         btrfs_set_extent_refs(leaf, extent_item, 1);
7985         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
7986         btrfs_set_extent_flags(leaf, extent_item,
7987                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
7988
7989         if (skinny_metadata) {
7990                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
7991         } else {
7992                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
7993                 btrfs_set_tree_block_key(leaf, block_info, &extent_op->key);
7994                 btrfs_set_tree_block_level(leaf, block_info, ref->level);
7995                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
7996         }
7997
7998         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY) {
7999                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
8000                 btrfs_set_extent_inline_ref_type(leaf, iref,
8001                                                  BTRFS_SHARED_BLOCK_REF_KEY);
8002                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->parent);
8003         } else {
8004                 btrfs_set_extent_inline_ref_type(leaf, iref,
8005                                                  BTRFS_TREE_BLOCK_REF_KEY);
8006                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->root);
8007         }
8008
8009         btrfs_mark_buffer_dirty(leaf);
8010         btrfs_free_path(path);
8011
8012         ret = remove_from_free_space_tree(trans, extent_key.objectid,
8013                                           num_bytes);
8014         if (ret)
8015                 return ret;
8016
8017         ret = update_block_group(trans, fs_info, extent_key.objectid,
8018                                  fs_info->nodesize, 1);
8019         if (ret) { /* -ENOENT, logic error */
8020                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8021                         extent_key.objectid, extent_key.offset);
8022                 BUG();
8023         }
8024
8025         trace_btrfs_reserved_extent_alloc(fs_info, extent_key.objectid,
8026                                           fs_info->nodesize);
8027         return ret;
8028 }
8029
8030 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8031                                      struct btrfs_root *root, u64 owner,
8032                                      u64 offset, u64 ram_bytes,
8033                                      struct btrfs_key *ins)
8034 {
8035         int ret;
8036
8037         BUG_ON(root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
8038
8039         btrfs_ref_tree_mod(root, ins->objectid, ins->offset, 0,
8040                            root->root_key.objectid, owner, offset,
8041                            BTRFS_ADD_DELAYED_EXTENT);
8042
8043         ret = btrfs_add_delayed_data_ref(trans, ins->objectid,
8044                                          ins->offset, 0,
8045                                          root->root_key.objectid, owner,
8046                                          offset, ram_bytes,
8047                                          BTRFS_ADD_DELAYED_EXTENT, NULL, NULL);
8048         return ret;
8049 }
8050
8051 /*
8052  * this is used by the tree logging recovery code.  It records that
8053  * an extent has been allocated and makes sure to clear the free
8054  * space cache bits as well
8055  */
8056 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
8057                                    u64 root_objectid, u64 owner, u64 offset,
8058                                    struct btrfs_key *ins)
8059 {
8060         struct btrfs_fs_info *fs_info = trans->fs_info;
8061         int ret;
8062         struct btrfs_block_group_cache *block_group;
8063         struct btrfs_space_info *space_info;
8064
8065         /*
8066          * Mixed block groups will exclude before processing the log so we only
8067          * need to do the exclude dance if this fs isn't mixed.
8068          */
8069         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS)) {
8070                 ret = __exclude_logged_extent(fs_info, ins->objectid,
8071                                               ins->offset);
8072                 if (ret)
8073                         return ret;
8074         }
8075
8076         block_group = btrfs_lookup_block_group(fs_info, ins->objectid);
8077         if (!block_group)
8078                 return -EINVAL;
8079
8080         space_info = block_group->space_info;
8081         spin_lock(&space_info->lock);
8082         spin_lock(&block_group->lock);
8083         space_info->bytes_reserved += ins->offset;
8084         block_group->reserved += ins->offset;
8085         spin_unlock(&block_group->lock);
8086         spin_unlock(&space_info->lock);
8087
8088         ret = alloc_reserved_file_extent(trans, 0, root_objectid, 0, owner,
8089                                          offset, ins, 1);
8090         btrfs_put_block_group(block_group);
8091         return ret;
8092 }
8093
8094 static struct extent_buffer *
8095 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
8096                       u64 bytenr, int level, u64 owner)
8097 {
8098         struct btrfs_fs_info *fs_info = root->fs_info;
8099         struct extent_buffer *buf;
8100
8101         buf = btrfs_find_create_tree_block(fs_info, bytenr);
8102         if (IS_ERR(buf))
8103                 return buf;
8104
8105         /*
8106          * Extra safety check in case the extent tree is corrupted and extent
8107          * allocator chooses to use a tree block which is already used and
8108          * locked.
8109          */
8110         if (buf->lock_owner == current->pid) {
8111                 btrfs_err_rl(fs_info,
8112 "tree block %llu owner %llu already locked by pid=%d, extent tree corruption detected",
8113                         buf->start, btrfs_header_owner(buf), current->pid);
8114                 free_extent_buffer(buf);
8115                 return ERR_PTR(-EUCLEAN);
8116         }
8117
8118         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
8119         btrfs_tree_lock(buf);
8120         clean_tree_block(fs_info, buf);
8121         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
8122
8123         btrfs_set_lock_blocking(buf);
8124         set_extent_buffer_uptodate(buf);
8125
8126         memzero_extent_buffer(buf, 0, sizeof(struct btrfs_header));
8127         btrfs_set_header_level(buf, level);
8128         btrfs_set_header_bytenr(buf, buf->start);
8129         btrfs_set_header_generation(buf, trans->transid);
8130         btrfs_set_header_backref_rev(buf, BTRFS_MIXED_BACKREF_REV);
8131         btrfs_set_header_owner(buf, owner);
8132         write_extent_buffer_fsid(buf, fs_info->fsid);
8133         write_extent_buffer_chunk_tree_uuid(buf, fs_info->chunk_tree_uuid);
8134         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
8135                 buf->log_index = root->log_transid % 2;
8136                 /*
8137                  * we allow two log transactions at a time, use different
8138                  * EXENT bit to differentiate dirty pages.
8139                  */
8140                 if (buf->log_index == 0)
8141                         set_extent_dirty(&root->dirty_log_pages, buf->start,
8142                                         buf->start + buf->len - 1, GFP_NOFS);
8143                 else
8144                         set_extent_new(&root->dirty_log_pages, buf->start,
8145                                         buf->start + buf->len - 1);
8146         } else {
8147                 buf->log_index = -1;
8148                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
8149                          buf->start + buf->len - 1, GFP_NOFS);
8150         }
8151         trans->dirty = true;
8152         /* this returns a buffer locked for blocking */
8153         return buf;
8154 }
8155
8156 static struct btrfs_block_rsv *
8157 use_block_rsv(struct btrfs_trans_handle *trans,
8158               struct btrfs_root *root, u32 blocksize)
8159 {
8160         struct btrfs_fs_info *fs_info = root->fs_info;
8161         struct btrfs_block_rsv *block_rsv;
8162         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
8163         int ret;
8164         bool global_updated = false;
8165
8166         block_rsv = get_block_rsv(trans, root);
8167
8168         if (unlikely(block_rsv->size == 0))
8169                 goto try_reserve;
8170 again:
8171         ret = block_rsv_use_bytes(block_rsv, blocksize);
8172         if (!ret)
8173                 return block_rsv;
8174
8175         if (block_rsv->failfast)
8176                 return ERR_PTR(ret);
8177
8178         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
8179                 global_updated = true;
8180                 update_global_block_rsv(fs_info);
8181                 goto again;
8182         }
8183
8184         if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
8185                 static DEFINE_RATELIMIT_STATE(_rs,
8186                                 DEFAULT_RATELIMIT_INTERVAL * 10,
8187                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
8188                 if (__ratelimit(&_rs))
8189                         WARN(1, KERN_DEBUG
8190                                 "BTRFS: block rsv returned %d\n", ret);
8191         }
8192 try_reserve:
8193         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
8194                                      BTRFS_RESERVE_NO_FLUSH);
8195         if (!ret)
8196                 return block_rsv;
8197         /*
8198          * If we couldn't reserve metadata bytes try and use some from
8199          * the global reserve if its space type is the same as the global
8200          * reservation.
8201          */
8202         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
8203             block_rsv->space_info == global_rsv->space_info) {
8204                 ret = block_rsv_use_bytes(global_rsv, blocksize);
8205                 if (!ret)
8206                         return global_rsv;
8207         }
8208         return ERR_PTR(ret);
8209 }
8210
8211 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
8212                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
8213 {
8214         block_rsv_add_bytes(block_rsv, blocksize, false);
8215         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0, NULL);
8216 }
8217
8218 /*
8219  * finds a free extent and does all the dirty work required for allocation
8220  * returns the tree buffer or an ERR_PTR on error.
8221  */
8222 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
8223                                              struct btrfs_root *root,
8224                                              u64 parent, u64 root_objectid,
8225                                              const struct btrfs_disk_key *key,
8226                                              int level, u64 hint,
8227                                              u64 empty_size)
8228 {
8229         struct btrfs_fs_info *fs_info = root->fs_info;
8230         struct btrfs_key ins;
8231         struct btrfs_block_rsv *block_rsv;
8232         struct extent_buffer *buf;
8233         struct btrfs_delayed_extent_op *extent_op;
8234         u64 flags = 0;
8235         int ret;
8236         u32 blocksize = fs_info->nodesize;
8237         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
8238
8239 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
8240         if (btrfs_is_testing(fs_info)) {
8241                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
8242                                             level, root_objectid);
8243                 if (!IS_ERR(buf))
8244                         root->alloc_bytenr += blocksize;
8245                 return buf;
8246         }
8247 #endif
8248
8249         block_rsv = use_block_rsv(trans, root, blocksize);
8250         if (IS_ERR(block_rsv))
8251                 return ERR_CAST(block_rsv);
8252
8253         ret = btrfs_reserve_extent(root, blocksize, blocksize, blocksize,
8254                                    empty_size, hint, &ins, 0, 0);
8255         if (ret)
8256                 goto out_unuse;
8257
8258         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level,
8259                                     root_objectid);
8260         if (IS_ERR(buf)) {
8261                 ret = PTR_ERR(buf);
8262                 goto out_free_reserved;
8263         }
8264
8265         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
8266                 if (parent == 0)
8267                         parent = ins.objectid;
8268                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
8269         } else
8270                 BUG_ON(parent > 0);
8271
8272         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
8273                 extent_op = btrfs_alloc_delayed_extent_op();
8274                 if (!extent_op) {
8275                         ret = -ENOMEM;
8276                         goto out_free_buf;
8277                 }
8278                 if (key)
8279                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
8280                 else
8281                         memset(&extent_op->key, 0, sizeof(extent_op->key));
8282                 extent_op->flags_to_set = flags;
8283                 extent_op->update_key = skinny_metadata ? false : true;
8284                 extent_op->update_flags = true;
8285                 extent_op->is_data = false;
8286                 extent_op->level = level;
8287
8288                 btrfs_ref_tree_mod(root, ins.objectid, ins.offset, parent,
8289                                    root_objectid, level, 0,
8290                                    BTRFS_ADD_DELAYED_EXTENT);
8291                 ret = btrfs_add_delayed_tree_ref(trans, ins.objectid,
8292                                                  ins.offset, parent,
8293                                                  root_objectid, level,
8294                                                  BTRFS_ADD_DELAYED_EXTENT,
8295                                                  extent_op, NULL, NULL);
8296                 if (ret)
8297                         goto out_free_delayed;
8298         }
8299         return buf;
8300
8301 out_free_delayed:
8302         btrfs_free_delayed_extent_op(extent_op);
8303 out_free_buf:
8304         free_extent_buffer(buf);
8305 out_free_reserved:
8306         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 0);
8307 out_unuse:
8308         unuse_block_rsv(fs_info, block_rsv, blocksize);
8309         return ERR_PTR(ret);
8310 }
8311
8312 struct walk_control {
8313         u64 refs[BTRFS_MAX_LEVEL];
8314         u64 flags[BTRFS_MAX_LEVEL];
8315         struct btrfs_key update_progress;
8316         int stage;
8317         int level;
8318         int shared_level;
8319         int update_ref;
8320         int keep_locks;
8321         int reada_slot;
8322         int reada_count;
8323 };
8324
8325 #define DROP_REFERENCE  1
8326 #define UPDATE_BACKREF  2
8327
8328 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
8329                                      struct btrfs_root *root,
8330                                      struct walk_control *wc,
8331                                      struct btrfs_path *path)
8332 {
8333         struct btrfs_fs_info *fs_info = root->fs_info;
8334         u64 bytenr;
8335         u64 generation;
8336         u64 refs;
8337         u64 flags;
8338         u32 nritems;
8339         struct btrfs_key key;
8340         struct extent_buffer *eb;
8341         int ret;
8342         int slot;
8343         int nread = 0;
8344
8345         if (path->slots[wc->level] < wc->reada_slot) {
8346                 wc->reada_count = wc->reada_count * 2 / 3;
8347                 wc->reada_count = max(wc->reada_count, 2);
8348         } else {
8349                 wc->reada_count = wc->reada_count * 3 / 2;
8350                 wc->reada_count = min_t(int, wc->reada_count,
8351                                         BTRFS_NODEPTRS_PER_BLOCK(fs_info));
8352         }
8353
8354         eb = path->nodes[wc->level];
8355         nritems = btrfs_header_nritems(eb);
8356
8357         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
8358                 if (nread >= wc->reada_count)
8359                         break;
8360
8361                 cond_resched();
8362                 bytenr = btrfs_node_blockptr(eb, slot);
8363                 generation = btrfs_node_ptr_generation(eb, slot);
8364
8365                 if (slot == path->slots[wc->level])
8366                         goto reada;
8367
8368                 if (wc->stage == UPDATE_BACKREF &&
8369                     generation <= root->root_key.offset)
8370                         continue;
8371
8372                 /* We don't lock the tree block, it's OK to be racy here */
8373                 ret = btrfs_lookup_extent_info(trans, fs_info, bytenr,
8374                                                wc->level - 1, 1, &refs,
8375                                                &flags);
8376                 /* We don't care about errors in readahead. */
8377                 if (ret < 0)
8378                         continue;
8379                 BUG_ON(refs == 0);
8380
8381                 if (wc->stage == DROP_REFERENCE) {
8382                         if (refs == 1)
8383                                 goto reada;
8384
8385                         if (wc->level == 1 &&
8386                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8387                                 continue;
8388                         if (!wc->update_ref ||
8389                             generation <= root->root_key.offset)
8390                                 continue;
8391                         btrfs_node_key_to_cpu(eb, &key, slot);
8392                         ret = btrfs_comp_cpu_keys(&key,
8393                                                   &wc->update_progress);
8394                         if (ret < 0)
8395                                 continue;
8396                 } else {
8397                         if (wc->level == 1 &&
8398                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8399                                 continue;
8400                 }
8401 reada:
8402                 readahead_tree_block(fs_info, bytenr);
8403                 nread++;
8404         }
8405         wc->reada_slot = slot;
8406 }
8407
8408 /*
8409  * helper to process tree block while walking down the tree.
8410  *
8411  * when wc->stage == UPDATE_BACKREF, this function updates
8412  * back refs for pointers in the block.
8413  *
8414  * NOTE: return value 1 means we should stop walking down.
8415  */
8416 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
8417                                    struct btrfs_root *root,
8418                                    struct btrfs_path *path,
8419                                    struct walk_control *wc, int lookup_info)
8420 {
8421         struct btrfs_fs_info *fs_info = root->fs_info;
8422         int level = wc->level;
8423         struct extent_buffer *eb = path->nodes[level];
8424         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8425         int ret;
8426
8427         if (wc->stage == UPDATE_BACKREF &&
8428             btrfs_header_owner(eb) != root->root_key.objectid)
8429                 return 1;
8430
8431         /*
8432          * when reference count of tree block is 1, it won't increase
8433          * again. once full backref flag is set, we never clear it.
8434          */
8435         if (lookup_info &&
8436             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
8437              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
8438                 BUG_ON(!path->locks[level]);
8439                 ret = btrfs_lookup_extent_info(trans, fs_info,
8440                                                eb->start, level, 1,
8441                                                &wc->refs[level],
8442                                                &wc->flags[level]);
8443                 BUG_ON(ret == -ENOMEM);
8444                 if (ret)
8445                         return ret;
8446                 BUG_ON(wc->refs[level] == 0);
8447         }
8448
8449         if (wc->stage == DROP_REFERENCE) {
8450                 if (wc->refs[level] > 1)
8451                         return 1;
8452
8453                 if (path->locks[level] && !wc->keep_locks) {
8454                         btrfs_tree_unlock_rw(eb, path->locks[level]);
8455                         path->locks[level] = 0;
8456                 }
8457                 return 0;
8458         }
8459
8460         /* wc->stage == UPDATE_BACKREF */
8461         if (!(wc->flags[level] & flag)) {
8462                 BUG_ON(!path->locks[level]);
8463                 ret = btrfs_inc_ref(trans, root, eb, 1);
8464                 BUG_ON(ret); /* -ENOMEM */
8465                 ret = btrfs_dec_ref(trans, root, eb, 0);
8466                 BUG_ON(ret); /* -ENOMEM */
8467                 ret = btrfs_set_disk_extent_flags(trans, fs_info, eb->start,
8468                                                   eb->len, flag,
8469                                                   btrfs_header_level(eb), 0);
8470                 BUG_ON(ret); /* -ENOMEM */
8471                 wc->flags[level] |= flag;
8472         }
8473
8474         /*
8475          * the block is shared by multiple trees, so it's not good to
8476          * keep the tree lock
8477          */
8478         if (path->locks[level] && level > 0) {
8479                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8480                 path->locks[level] = 0;
8481         }
8482         return 0;
8483 }
8484
8485 /*
8486  * helper to process tree block pointer.
8487  *
8488  * when wc->stage == DROP_REFERENCE, this function checks
8489  * reference count of the block pointed to. if the block
8490  * is shared and we need update back refs for the subtree
8491  * rooted at the block, this function changes wc->stage to
8492  * UPDATE_BACKREF. if the block is shared and there is no
8493  * need to update back, this function drops the reference
8494  * to the block.
8495  *
8496  * NOTE: return value 1 means we should stop walking down.
8497  */
8498 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
8499                                  struct btrfs_root *root,
8500                                  struct btrfs_path *path,
8501                                  struct walk_control *wc, int *lookup_info)
8502 {
8503         struct btrfs_fs_info *fs_info = root->fs_info;
8504         u64 bytenr;
8505         u64 generation;
8506         u64 parent;
8507         u32 blocksize;
8508         struct btrfs_key key;
8509         struct btrfs_key first_key;
8510         struct extent_buffer *next;
8511         int level = wc->level;
8512         int reada = 0;
8513         int ret = 0;
8514         bool need_account = false;
8515
8516         generation = btrfs_node_ptr_generation(path->nodes[level],
8517                                                path->slots[level]);
8518         /*
8519          * if the lower level block was created before the snapshot
8520          * was created, we know there is no need to update back refs
8521          * for the subtree
8522          */
8523         if (wc->stage == UPDATE_BACKREF &&
8524             generation <= root->root_key.offset) {
8525                 *lookup_info = 1;
8526                 return 1;
8527         }
8528
8529         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
8530         btrfs_node_key_to_cpu(path->nodes[level], &first_key,
8531                               path->slots[level]);
8532         blocksize = fs_info->nodesize;
8533
8534         next = find_extent_buffer(fs_info, bytenr);
8535         if (!next) {
8536                 next = btrfs_find_create_tree_block(fs_info, bytenr);
8537                 if (IS_ERR(next))
8538                         return PTR_ERR(next);
8539
8540                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
8541                                                level - 1);
8542                 reada = 1;
8543         }
8544         btrfs_tree_lock(next);
8545         btrfs_set_lock_blocking(next);
8546
8547         ret = btrfs_lookup_extent_info(trans, fs_info, bytenr, level - 1, 1,
8548                                        &wc->refs[level - 1],
8549                                        &wc->flags[level - 1]);
8550         if (ret < 0)
8551                 goto out_unlock;
8552
8553         if (unlikely(wc->refs[level - 1] == 0)) {
8554                 btrfs_err(fs_info, "Missing references.");
8555                 ret = -EIO;
8556                 goto out_unlock;
8557         }
8558         *lookup_info = 0;
8559
8560         if (wc->stage == DROP_REFERENCE) {
8561                 if (wc->refs[level - 1] > 1) {
8562                         need_account = true;
8563                         if (level == 1 &&
8564                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8565                                 goto skip;
8566
8567                         if (!wc->update_ref ||
8568                             generation <= root->root_key.offset)
8569                                 goto skip;
8570
8571                         btrfs_node_key_to_cpu(path->nodes[level], &key,
8572                                               path->slots[level]);
8573                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
8574                         if (ret < 0)
8575                                 goto skip;
8576
8577                         wc->stage = UPDATE_BACKREF;
8578                         wc->shared_level = level - 1;
8579                 }
8580         } else {
8581                 if (level == 1 &&
8582                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8583                         goto skip;
8584         }
8585
8586         if (!btrfs_buffer_uptodate(next, generation, 0)) {
8587                 btrfs_tree_unlock(next);
8588                 free_extent_buffer(next);
8589                 next = NULL;
8590                 *lookup_info = 1;
8591         }
8592
8593         if (!next) {
8594                 if (reada && level == 1)
8595                         reada_walk_down(trans, root, wc, path);
8596                 next = read_tree_block(fs_info, bytenr, generation, level - 1,
8597                                        &first_key);
8598                 if (IS_ERR(next)) {
8599                         return PTR_ERR(next);
8600                 } else if (!extent_buffer_uptodate(next)) {
8601                         free_extent_buffer(next);
8602                         return -EIO;
8603                 }
8604                 btrfs_tree_lock(next);
8605                 btrfs_set_lock_blocking(next);
8606         }
8607
8608         level--;
8609         ASSERT(level == btrfs_header_level(next));
8610         if (level != btrfs_header_level(next)) {
8611                 btrfs_err(root->fs_info, "mismatched level");
8612                 ret = -EIO;
8613                 goto out_unlock;
8614         }
8615         path->nodes[level] = next;
8616         path->slots[level] = 0;
8617         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8618         wc->level = level;
8619         if (wc->level == 1)
8620                 wc->reada_slot = 0;
8621         return 0;
8622 skip:
8623         wc->refs[level - 1] = 0;
8624         wc->flags[level - 1] = 0;
8625         if (wc->stage == DROP_REFERENCE) {
8626                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
8627                         parent = path->nodes[level]->start;
8628                 } else {
8629                         ASSERT(root->root_key.objectid ==
8630                                btrfs_header_owner(path->nodes[level]));
8631                         if (root->root_key.objectid !=
8632                             btrfs_header_owner(path->nodes[level])) {
8633                                 btrfs_err(root->fs_info,
8634                                                 "mismatched block owner");
8635                                 ret = -EIO;
8636                                 goto out_unlock;
8637                         }
8638                         parent = 0;
8639                 }
8640
8641                 if (need_account) {
8642                         ret = btrfs_qgroup_trace_subtree(trans, next,
8643                                                          generation, level - 1);
8644                         if (ret) {
8645                                 btrfs_err_rl(fs_info,
8646                                              "Error %d accounting shared subtree. Quota is out of sync, rescan required.",
8647                                              ret);
8648                         }
8649                 }
8650                 ret = btrfs_free_extent(trans, root, bytenr, blocksize,
8651                                         parent, root->root_key.objectid,
8652                                         level - 1, 0);
8653                 if (ret)
8654                         goto out_unlock;
8655         }
8656
8657         *lookup_info = 1;
8658         ret = 1;
8659
8660 out_unlock:
8661         btrfs_tree_unlock(next);
8662         free_extent_buffer(next);
8663
8664         return ret;
8665 }
8666
8667 /*
8668  * helper to process tree block while walking up the tree.
8669  *
8670  * when wc->stage == DROP_REFERENCE, this function drops
8671  * reference count on the block.
8672  *
8673  * when wc->stage == UPDATE_BACKREF, this function changes
8674  * wc->stage back to DROP_REFERENCE if we changed wc->stage
8675  * to UPDATE_BACKREF previously while processing the block.
8676  *
8677  * NOTE: return value 1 means we should stop walking up.
8678  */
8679 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
8680                                  struct btrfs_root *root,
8681                                  struct btrfs_path *path,
8682                                  struct walk_control *wc)
8683 {
8684         struct btrfs_fs_info *fs_info = root->fs_info;
8685         int ret;
8686         int level = wc->level;
8687         struct extent_buffer *eb = path->nodes[level];
8688         u64 parent = 0;
8689
8690         if (wc->stage == UPDATE_BACKREF) {
8691                 BUG_ON(wc->shared_level < level);
8692                 if (level < wc->shared_level)
8693                         goto out;
8694
8695                 ret = find_next_key(path, level + 1, &wc->update_progress);
8696                 if (ret > 0)
8697                         wc->update_ref = 0;
8698
8699                 wc->stage = DROP_REFERENCE;
8700                 wc->shared_level = -1;
8701                 path->slots[level] = 0;
8702
8703                 /*
8704                  * check reference count again if the block isn't locked.
8705                  * we should start walking down the tree again if reference
8706                  * count is one.
8707                  */
8708                 if (!path->locks[level]) {
8709                         BUG_ON(level == 0);
8710                         btrfs_tree_lock(eb);
8711                         btrfs_set_lock_blocking(eb);
8712                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8713
8714                         ret = btrfs_lookup_extent_info(trans, fs_info,
8715                                                        eb->start, level, 1,
8716                                                        &wc->refs[level],
8717                                                        &wc->flags[level]);
8718                         if (ret < 0) {
8719                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8720                                 path->locks[level] = 0;
8721                                 return ret;
8722                         }
8723                         BUG_ON(wc->refs[level] == 0);
8724                         if (wc->refs[level] == 1) {
8725                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8726                                 path->locks[level] = 0;
8727                                 return 1;
8728                         }
8729                 }
8730         }
8731
8732         /* wc->stage == DROP_REFERENCE */
8733         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
8734
8735         if (wc->refs[level] == 1) {
8736                 if (level == 0) {
8737                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8738                                 ret = btrfs_dec_ref(trans, root, eb, 1);
8739                         else
8740                                 ret = btrfs_dec_ref(trans, root, eb, 0);
8741                         BUG_ON(ret); /* -ENOMEM */
8742                         ret = btrfs_qgroup_trace_leaf_items(trans, eb);
8743                         if (ret) {
8744                                 btrfs_err_rl(fs_info,
8745                                              "error %d accounting leaf items. Quota is out of sync, rescan required.",
8746                                              ret);
8747                         }
8748                 }
8749                 /* make block locked assertion in clean_tree_block happy */
8750                 if (!path->locks[level] &&
8751                     btrfs_header_generation(eb) == trans->transid) {
8752                         btrfs_tree_lock(eb);
8753                         btrfs_set_lock_blocking(eb);
8754                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8755                 }
8756                 clean_tree_block(fs_info, eb);
8757         }
8758
8759         if (eb == root->node) {
8760                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8761                         parent = eb->start;
8762                 else if (root->root_key.objectid != btrfs_header_owner(eb))
8763                         goto owner_mismatch;
8764         } else {
8765                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8766                         parent = path->nodes[level + 1]->start;
8767                 else if (root->root_key.objectid !=
8768                          btrfs_header_owner(path->nodes[level + 1]))
8769                         goto owner_mismatch;
8770         }
8771
8772         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
8773 out:
8774         wc->refs[level] = 0;
8775         wc->flags[level] = 0;
8776         return 0;
8777
8778 owner_mismatch:
8779         btrfs_err_rl(fs_info, "unexpected tree owner, have %llu expect %llu",
8780                      btrfs_header_owner(eb), root->root_key.objectid);
8781         return -EUCLEAN;
8782 }
8783
8784 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
8785                                    struct btrfs_root *root,
8786                                    struct btrfs_path *path,
8787                                    struct walk_control *wc)
8788 {
8789         int level = wc->level;
8790         int lookup_info = 1;
8791         int ret;
8792
8793         while (level >= 0) {
8794                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
8795                 if (ret > 0)
8796                         break;
8797
8798                 if (level == 0)
8799                         break;
8800
8801                 if (path->slots[level] >=
8802                     btrfs_header_nritems(path->nodes[level]))
8803                         break;
8804
8805                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
8806                 if (ret > 0) {
8807                         path->slots[level]++;
8808                         continue;
8809                 } else if (ret < 0)
8810                         return ret;
8811                 level = wc->level;
8812         }
8813         return 0;
8814 }
8815
8816 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
8817                                  struct btrfs_root *root,
8818                                  struct btrfs_path *path,
8819                                  struct walk_control *wc, int max_level)
8820 {
8821         int level = wc->level;
8822         int ret;
8823
8824         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
8825         while (level < max_level && path->nodes[level]) {
8826                 wc->level = level;
8827                 if (path->slots[level] + 1 <
8828                     btrfs_header_nritems(path->nodes[level])) {
8829                         path->slots[level]++;
8830                         return 0;
8831                 } else {
8832                         ret = walk_up_proc(trans, root, path, wc);
8833                         if (ret > 0)
8834                                 return 0;
8835                         if (ret < 0)
8836                                 return ret;
8837
8838                         if (path->locks[level]) {
8839                                 btrfs_tree_unlock_rw(path->nodes[level],
8840                                                      path->locks[level]);
8841                                 path->locks[level] = 0;
8842                         }
8843                         free_extent_buffer(path->nodes[level]);
8844                         path->nodes[level] = NULL;
8845                         level++;
8846                 }
8847         }
8848         return 1;
8849 }
8850
8851 /*
8852  * drop a subvolume tree.
8853  *
8854  * this function traverses the tree freeing any blocks that only
8855  * referenced by the tree.
8856  *
8857  * when a shared tree block is found. this function decreases its
8858  * reference count by one. if update_ref is true, this function
8859  * also make sure backrefs for the shared block and all lower level
8860  * blocks are properly updated.
8861  *
8862  * If called with for_reloc == 0, may exit early with -EAGAIN
8863  */
8864 int btrfs_drop_snapshot(struct btrfs_root *root,
8865                          struct btrfs_block_rsv *block_rsv, int update_ref,
8866                          int for_reloc)
8867 {
8868         struct btrfs_fs_info *fs_info = root->fs_info;
8869         struct btrfs_path *path;
8870         struct btrfs_trans_handle *trans;
8871         struct btrfs_root *tree_root = fs_info->tree_root;
8872         struct btrfs_root_item *root_item = &root->root_item;
8873         struct walk_control *wc;
8874         struct btrfs_key key;
8875         int err = 0;
8876         int ret;
8877         int level;
8878         bool root_dropped = false;
8879
8880         btrfs_debug(fs_info, "Drop subvolume %llu", root->root_key.objectid);
8881
8882         path = btrfs_alloc_path();
8883         if (!path) {
8884                 err = -ENOMEM;
8885                 goto out;
8886         }
8887
8888         wc = kzalloc(sizeof(*wc), GFP_NOFS);
8889         if (!wc) {
8890                 btrfs_free_path(path);
8891                 err = -ENOMEM;
8892                 goto out;
8893         }
8894
8895         trans = btrfs_start_transaction(tree_root, 0);
8896         if (IS_ERR(trans)) {
8897                 err = PTR_ERR(trans);
8898                 goto out_free;
8899         }
8900
8901         if (block_rsv)
8902                 trans->block_rsv = block_rsv;
8903
8904         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
8905                 level = btrfs_header_level(root->node);
8906                 path->nodes[level] = btrfs_lock_root_node(root);
8907                 btrfs_set_lock_blocking(path->nodes[level]);
8908                 path->slots[level] = 0;
8909                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8910                 memset(&wc->update_progress, 0,
8911                        sizeof(wc->update_progress));
8912         } else {
8913                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
8914                 memcpy(&wc->update_progress, &key,
8915                        sizeof(wc->update_progress));
8916
8917                 level = root_item->drop_level;
8918                 BUG_ON(level == 0);
8919                 path->lowest_level = level;
8920                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
8921                 path->lowest_level = 0;
8922                 if (ret < 0) {
8923                         err = ret;
8924                         goto out_end_trans;
8925                 }
8926                 WARN_ON(ret > 0);
8927
8928                 /*
8929                  * unlock our path, this is safe because only this
8930                  * function is allowed to delete this snapshot
8931                  */
8932                 btrfs_unlock_up_safe(path, 0);
8933
8934                 level = btrfs_header_level(root->node);
8935                 while (1) {
8936                         btrfs_tree_lock(path->nodes[level]);
8937                         btrfs_set_lock_blocking(path->nodes[level]);
8938                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8939
8940                         ret = btrfs_lookup_extent_info(trans, fs_info,
8941                                                 path->nodes[level]->start,
8942                                                 level, 1, &wc->refs[level],
8943                                                 &wc->flags[level]);
8944                         if (ret < 0) {
8945                                 err = ret;
8946                                 goto out_end_trans;
8947                         }
8948                         BUG_ON(wc->refs[level] == 0);
8949
8950                         if (level == root_item->drop_level)
8951                                 break;
8952
8953                         btrfs_tree_unlock(path->nodes[level]);
8954                         path->locks[level] = 0;
8955                         WARN_ON(wc->refs[level] != 1);
8956                         level--;
8957                 }
8958         }
8959
8960         wc->level = level;
8961         wc->shared_level = -1;
8962         wc->stage = DROP_REFERENCE;
8963         wc->update_ref = update_ref;
8964         wc->keep_locks = 0;
8965         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
8966
8967         while (1) {
8968
8969                 ret = walk_down_tree(trans, root, path, wc);
8970                 if (ret < 0) {
8971                         err = ret;
8972                         break;
8973                 }
8974
8975                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
8976                 if (ret < 0) {
8977                         err = ret;
8978                         break;
8979                 }
8980
8981                 if (ret > 0) {
8982                         BUG_ON(wc->stage != DROP_REFERENCE);
8983                         break;
8984                 }
8985
8986                 if (wc->stage == DROP_REFERENCE) {
8987                         level = wc->level;
8988                         btrfs_node_key(path->nodes[level],
8989                                        &root_item->drop_progress,
8990                                        path->slots[level]);
8991                         root_item->drop_level = level;
8992                 }
8993
8994                 BUG_ON(wc->level == 0);
8995                 if (btrfs_should_end_transaction(trans) ||
8996                     (!for_reloc && btrfs_need_cleaner_sleep(fs_info))) {
8997                         ret = btrfs_update_root(trans, tree_root,
8998                                                 &root->root_key,
8999                                                 root_item);
9000                         if (ret) {
9001                                 btrfs_abort_transaction(trans, ret);
9002                                 err = ret;
9003                                 goto out_end_trans;
9004                         }
9005
9006                         btrfs_end_transaction_throttle(trans);
9007                         if (!for_reloc && btrfs_need_cleaner_sleep(fs_info)) {
9008                                 btrfs_debug(fs_info,
9009                                             "drop snapshot early exit");
9010                                 err = -EAGAIN;
9011                                 goto out_free;
9012                         }
9013
9014                         trans = btrfs_start_transaction(tree_root, 0);
9015                         if (IS_ERR(trans)) {
9016                                 err = PTR_ERR(trans);
9017                                 goto out_free;
9018                         }
9019                         if (block_rsv)
9020                                 trans->block_rsv = block_rsv;
9021                 }
9022         }
9023         btrfs_release_path(path);
9024         if (err)
9025                 goto out_end_trans;
9026
9027         ret = btrfs_del_root(trans, &root->root_key);
9028         if (ret) {
9029                 btrfs_abort_transaction(trans, ret);
9030                 err = ret;
9031                 goto out_end_trans;
9032         }
9033
9034         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
9035                 ret = btrfs_find_root(tree_root, &root->root_key, path,
9036                                       NULL, NULL);
9037                 if (ret < 0) {
9038                         btrfs_abort_transaction(trans, ret);
9039                         err = ret;
9040                         goto out_end_trans;
9041                 } else if (ret > 0) {
9042                         /* if we fail to delete the orphan item this time
9043                          * around, it'll get picked up the next time.
9044                          *
9045                          * The most common failure here is just -ENOENT.
9046                          */
9047                         btrfs_del_orphan_item(trans, tree_root,
9048                                               root->root_key.objectid);
9049                 }
9050         }
9051
9052         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
9053                 btrfs_add_dropped_root(trans, root);
9054         } else {
9055                 free_extent_buffer(root->node);
9056                 free_extent_buffer(root->commit_root);
9057                 btrfs_put_fs_root(root);
9058         }
9059         root_dropped = true;
9060 out_end_trans:
9061         btrfs_end_transaction_throttle(trans);
9062 out_free:
9063         kfree(wc);
9064         btrfs_free_path(path);
9065 out:
9066         /*
9067          * So if we need to stop dropping the snapshot for whatever reason we
9068          * need to make sure to add it back to the dead root list so that we
9069          * keep trying to do the work later.  This also cleans up roots if we
9070          * don't have it in the radix (like when we recover after a power fail
9071          * or unmount) so we don't leak memory.
9072          */
9073         if (!for_reloc && !root_dropped)
9074                 btrfs_add_dead_root(root);
9075         if (err && err != -EAGAIN)
9076                 btrfs_handle_fs_error(fs_info, err, NULL);
9077         return err;
9078 }
9079
9080 /*
9081  * drop subtree rooted at tree block 'node'.
9082  *
9083  * NOTE: this function will unlock and release tree block 'node'
9084  * only used by relocation code
9085  */
9086 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
9087                         struct btrfs_root *root,
9088                         struct extent_buffer *node,
9089                         struct extent_buffer *parent)
9090 {
9091         struct btrfs_fs_info *fs_info = root->fs_info;
9092         struct btrfs_path *path;
9093         struct walk_control *wc;
9094         int level;
9095         int parent_level;
9096         int ret = 0;
9097         int wret;
9098
9099         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
9100
9101         path = btrfs_alloc_path();
9102         if (!path)
9103                 return -ENOMEM;
9104
9105         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9106         if (!wc) {
9107                 btrfs_free_path(path);
9108                 return -ENOMEM;
9109         }
9110
9111         btrfs_assert_tree_locked(parent);
9112         parent_level = btrfs_header_level(parent);
9113         extent_buffer_get(parent);
9114         path->nodes[parent_level] = parent;
9115         path->slots[parent_level] = btrfs_header_nritems(parent);
9116
9117         btrfs_assert_tree_locked(node);
9118         level = btrfs_header_level(node);
9119         path->nodes[level] = node;
9120         path->slots[level] = 0;
9121         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9122
9123         wc->refs[parent_level] = 1;
9124         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
9125         wc->level = level;
9126         wc->shared_level = -1;
9127         wc->stage = DROP_REFERENCE;
9128         wc->update_ref = 0;
9129         wc->keep_locks = 1;
9130         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
9131
9132         while (1) {
9133                 wret = walk_down_tree(trans, root, path, wc);
9134                 if (wret < 0) {
9135                         ret = wret;
9136                         break;
9137                 }
9138
9139                 wret = walk_up_tree(trans, root, path, wc, parent_level);
9140                 if (wret < 0)
9141                         ret = wret;
9142                 if (wret != 0)
9143                         break;
9144         }
9145
9146         kfree(wc);
9147         btrfs_free_path(path);
9148         return ret;
9149 }
9150
9151 static u64 update_block_group_flags(struct btrfs_fs_info *fs_info, u64 flags)
9152 {
9153         u64 num_devices;
9154         u64 stripped;
9155
9156         /*
9157          * if restripe for this chunk_type is on pick target profile and
9158          * return, otherwise do the usual balance
9159          */
9160         stripped = get_restripe_target(fs_info, flags);
9161         if (stripped)
9162                 return extended_to_chunk(stripped);
9163
9164         num_devices = fs_info->fs_devices->rw_devices;
9165
9166         stripped = BTRFS_BLOCK_GROUP_RAID0 |
9167                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
9168                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
9169
9170         if (num_devices == 1) {
9171                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9172                 stripped = flags & ~stripped;
9173
9174                 /* turn raid0 into single device chunks */
9175                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
9176                         return stripped;
9177
9178                 /* turn mirroring into duplication */
9179                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
9180                              BTRFS_BLOCK_GROUP_RAID10))
9181                         return stripped | BTRFS_BLOCK_GROUP_DUP;
9182         } else {
9183                 /* they already had raid on here, just return */
9184                 if (flags & stripped)
9185                         return flags;
9186
9187                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9188                 stripped = flags & ~stripped;
9189
9190                 /* switch duplicated blocks with raid1 */
9191                 if (flags & BTRFS_BLOCK_GROUP_DUP)
9192                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
9193
9194                 /* this is drive concat, leave it alone */
9195         }
9196
9197         return flags;
9198 }
9199
9200 static int inc_block_group_ro(struct btrfs_block_group_cache *cache, int force)
9201 {
9202         struct btrfs_space_info *sinfo = cache->space_info;
9203         u64 num_bytes;
9204         u64 min_allocable_bytes;
9205         int ret = -ENOSPC;
9206
9207         /*
9208          * We need some metadata space and system metadata space for
9209          * allocating chunks in some corner cases until we force to set
9210          * it to be readonly.
9211          */
9212         if ((sinfo->flags &
9213              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
9214             !force)
9215                 min_allocable_bytes = SZ_1M;
9216         else
9217                 min_allocable_bytes = 0;
9218
9219         spin_lock(&sinfo->lock);
9220         spin_lock(&cache->lock);
9221
9222         if (cache->ro) {
9223                 cache->ro++;
9224                 ret = 0;
9225                 goto out;
9226         }
9227
9228         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
9229                     cache->bytes_super - btrfs_block_group_used(&cache->item);
9230
9231         if (btrfs_space_info_used(sinfo, true) + num_bytes +
9232             min_allocable_bytes <= sinfo->total_bytes) {
9233                 sinfo->bytes_readonly += num_bytes;
9234                 cache->ro++;
9235                 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
9236                 ret = 0;
9237         }
9238 out:
9239         spin_unlock(&cache->lock);
9240         spin_unlock(&sinfo->lock);
9241         return ret;
9242 }
9243
9244 int btrfs_inc_block_group_ro(struct btrfs_block_group_cache *cache)
9245
9246 {
9247         struct btrfs_fs_info *fs_info = cache->fs_info;
9248         struct btrfs_trans_handle *trans;
9249         u64 alloc_flags;
9250         int ret;
9251
9252 again:
9253         trans = btrfs_join_transaction(fs_info->extent_root);
9254         if (IS_ERR(trans))
9255                 return PTR_ERR(trans);
9256
9257         /*
9258          * we're not allowed to set block groups readonly after the dirty
9259          * block groups cache has started writing.  If it already started,
9260          * back off and let this transaction commit
9261          */
9262         mutex_lock(&fs_info->ro_block_group_mutex);
9263         if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
9264                 u64 transid = trans->transid;
9265
9266                 mutex_unlock(&fs_info->ro_block_group_mutex);
9267                 btrfs_end_transaction(trans);
9268
9269                 ret = btrfs_wait_for_commit(fs_info, transid);
9270                 if (ret)
9271                         return ret;
9272                 goto again;
9273         }
9274
9275         /*
9276          * if we are changing raid levels, try to allocate a corresponding
9277          * block group with the new raid level.
9278          */
9279         alloc_flags = update_block_group_flags(fs_info, cache->flags);
9280         if (alloc_flags != cache->flags) {
9281                 ret = do_chunk_alloc(trans, alloc_flags,
9282                                      CHUNK_ALLOC_FORCE);
9283                 /*
9284                  * ENOSPC is allowed here, we may have enough space
9285                  * already allocated at the new raid level to
9286                  * carry on
9287                  */
9288                 if (ret == -ENOSPC)
9289                         ret = 0;
9290                 if (ret < 0)
9291                         goto out;
9292         }
9293
9294         ret = inc_block_group_ro(cache, 0);
9295         if (!ret)
9296                 goto out;
9297         alloc_flags = get_alloc_profile(fs_info, cache->space_info->flags);
9298         ret = do_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
9299         if (ret < 0)
9300                 goto out;
9301         ret = inc_block_group_ro(cache, 0);
9302 out:
9303         if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
9304                 alloc_flags = update_block_group_flags(fs_info, cache->flags);
9305                 mutex_lock(&fs_info->chunk_mutex);
9306                 check_system_chunk(trans, alloc_flags);
9307                 mutex_unlock(&fs_info->chunk_mutex);
9308         }
9309         mutex_unlock(&fs_info->ro_block_group_mutex);
9310
9311         btrfs_end_transaction(trans);
9312         return ret;
9313 }
9314
9315 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans, u64 type)
9316 {
9317         u64 alloc_flags = get_alloc_profile(trans->fs_info, type);
9318
9319         return do_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
9320 }
9321
9322 /*
9323  * helper to account the unused space of all the readonly block group in the
9324  * space_info. takes mirrors into account.
9325  */
9326 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
9327 {
9328         struct btrfs_block_group_cache *block_group;
9329         u64 free_bytes = 0;
9330         int factor;
9331
9332         /* It's df, we don't care if it's racy */
9333         if (list_empty(&sinfo->ro_bgs))
9334                 return 0;
9335
9336         spin_lock(&sinfo->lock);
9337         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
9338                 spin_lock(&block_group->lock);
9339
9340                 if (!block_group->ro) {
9341                         spin_unlock(&block_group->lock);
9342                         continue;
9343                 }
9344
9345                 factor = btrfs_bg_type_to_factor(block_group->flags);
9346                 free_bytes += (block_group->key.offset -
9347                                btrfs_block_group_used(&block_group->item)) *
9348                                factor;
9349
9350                 spin_unlock(&block_group->lock);
9351         }
9352         spin_unlock(&sinfo->lock);
9353
9354         return free_bytes;
9355 }
9356
9357 void btrfs_dec_block_group_ro(struct btrfs_block_group_cache *cache)
9358 {
9359         struct btrfs_space_info *sinfo = cache->space_info;
9360         u64 num_bytes;
9361
9362         BUG_ON(!cache->ro);
9363
9364         spin_lock(&sinfo->lock);
9365         spin_lock(&cache->lock);
9366         if (!--cache->ro) {
9367                 num_bytes = cache->key.offset - cache->reserved -
9368                             cache->pinned - cache->bytes_super -
9369                             btrfs_block_group_used(&cache->item);
9370                 sinfo->bytes_readonly -= num_bytes;
9371                 list_del_init(&cache->ro_list);
9372         }
9373         spin_unlock(&cache->lock);
9374         spin_unlock(&sinfo->lock);
9375 }
9376
9377 /*
9378  * checks to see if its even possible to relocate this block group.
9379  *
9380  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
9381  * ok to go ahead and try.
9382  */
9383 int btrfs_can_relocate(struct btrfs_fs_info *fs_info, u64 bytenr)
9384 {
9385         struct btrfs_root *root = fs_info->extent_root;
9386         struct btrfs_block_group_cache *block_group;
9387         struct btrfs_space_info *space_info;
9388         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
9389         struct btrfs_device *device;
9390         struct btrfs_trans_handle *trans;
9391         u64 min_free;
9392         u64 dev_min = 1;
9393         u64 dev_nr = 0;
9394         u64 target;
9395         int debug;
9396         int index;
9397         int full = 0;
9398         int ret = 0;
9399
9400         debug = btrfs_test_opt(fs_info, ENOSPC_DEBUG);
9401
9402         block_group = btrfs_lookup_block_group(fs_info, bytenr);
9403
9404         /* odd, couldn't find the block group, leave it alone */
9405         if (!block_group) {
9406                 if (debug)
9407                         btrfs_warn(fs_info,
9408                                    "can't find block group for bytenr %llu",
9409                                    bytenr);
9410                 return -1;
9411         }
9412
9413         min_free = btrfs_block_group_used(&block_group->item);
9414
9415         /* no bytes used, we're good */
9416         if (!min_free)
9417                 goto out;
9418
9419         space_info = block_group->space_info;
9420         spin_lock(&space_info->lock);
9421
9422         full = space_info->full;
9423
9424         /*
9425          * if this is the last block group we have in this space, we can't
9426          * relocate it unless we're able to allocate a new chunk below.
9427          *
9428          * Otherwise, we need to make sure we have room in the space to handle
9429          * all of the extents from this block group.  If we can, we're good
9430          */
9431         if ((space_info->total_bytes != block_group->key.offset) &&
9432             (btrfs_space_info_used(space_info, false) + min_free <
9433              space_info->total_bytes)) {
9434                 spin_unlock(&space_info->lock);
9435                 goto out;
9436         }
9437         spin_unlock(&space_info->lock);
9438
9439         /*
9440          * ok we don't have enough space, but maybe we have free space on our
9441          * devices to allocate new chunks for relocation, so loop through our
9442          * alloc devices and guess if we have enough space.  if this block
9443          * group is going to be restriped, run checks against the target
9444          * profile instead of the current one.
9445          */
9446         ret = -1;
9447
9448         /*
9449          * index:
9450          *      0: raid10
9451          *      1: raid1
9452          *      2: dup
9453          *      3: raid0
9454          *      4: single
9455          */
9456         target = get_restripe_target(fs_info, block_group->flags);
9457         if (target) {
9458                 index = btrfs_bg_flags_to_raid_index(extended_to_chunk(target));
9459         } else {
9460                 /*
9461                  * this is just a balance, so if we were marked as full
9462                  * we know there is no space for a new chunk
9463                  */
9464                 if (full) {
9465                         if (debug)
9466                                 btrfs_warn(fs_info,
9467                                            "no space to alloc new chunk for block group %llu",
9468                                            block_group->key.objectid);
9469                         goto out;
9470                 }
9471
9472                 index = btrfs_bg_flags_to_raid_index(block_group->flags);
9473         }
9474
9475         if (index == BTRFS_RAID_RAID10) {
9476                 dev_min = 4;
9477                 /* Divide by 2 */
9478                 min_free >>= 1;
9479         } else if (index == BTRFS_RAID_RAID1) {
9480                 dev_min = 2;
9481         } else if (index == BTRFS_RAID_DUP) {
9482                 /* Multiply by 2 */
9483                 min_free <<= 1;
9484         } else if (index == BTRFS_RAID_RAID0) {
9485                 dev_min = fs_devices->rw_devices;
9486                 min_free = div64_u64(min_free, dev_min);
9487         }
9488
9489         /* We need to do this so that we can look at pending chunks */
9490         trans = btrfs_join_transaction(root);
9491         if (IS_ERR(trans)) {
9492                 ret = PTR_ERR(trans);
9493                 goto out;
9494         }
9495
9496         mutex_lock(&fs_info->chunk_mutex);
9497         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
9498                 u64 dev_offset;
9499
9500                 /*
9501                  * check to make sure we can actually find a chunk with enough
9502                  * space to fit our block group in.
9503                  */
9504                 if (device->total_bytes > device->bytes_used + min_free &&
9505                     !test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state)) {
9506                         ret = find_free_dev_extent(trans, device, min_free,
9507                                                    &dev_offset, NULL);
9508                         if (!ret)
9509                                 dev_nr++;
9510
9511                         if (dev_nr >= dev_min)
9512                                 break;
9513
9514                         ret = -1;
9515                 }
9516         }
9517         if (debug && ret == -1)
9518                 btrfs_warn(fs_info,
9519                            "no space to allocate a new chunk for block group %llu",
9520                            block_group->key.objectid);
9521         mutex_unlock(&fs_info->chunk_mutex);
9522         btrfs_end_transaction(trans);
9523 out:
9524         btrfs_put_block_group(block_group);
9525         return ret;
9526 }
9527
9528 static int find_first_block_group(struct btrfs_fs_info *fs_info,
9529                                   struct btrfs_path *path,
9530                                   struct btrfs_key *key)
9531 {
9532         struct btrfs_root *root = fs_info->extent_root;
9533         int ret = 0;
9534         struct btrfs_key found_key;
9535         struct extent_buffer *leaf;
9536         struct btrfs_block_group_item bg;
9537         u64 flags;
9538         int slot;
9539
9540         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
9541         if (ret < 0)
9542                 goto out;
9543
9544         while (1) {
9545                 slot = path->slots[0];
9546                 leaf = path->nodes[0];
9547                 if (slot >= btrfs_header_nritems(leaf)) {
9548                         ret = btrfs_next_leaf(root, path);
9549                         if (ret == 0)
9550                                 continue;
9551                         if (ret < 0)
9552                                 goto out;
9553                         break;
9554                 }
9555                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
9556
9557                 if (found_key.objectid >= key->objectid &&
9558                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
9559                         struct extent_map_tree *em_tree;
9560                         struct extent_map *em;
9561
9562                         em_tree = &root->fs_info->mapping_tree.map_tree;
9563                         read_lock(&em_tree->lock);
9564                         em = lookup_extent_mapping(em_tree, found_key.objectid,
9565                                                    found_key.offset);
9566                         read_unlock(&em_tree->lock);
9567                         if (!em) {
9568                                 btrfs_err(fs_info,
9569                         "logical %llu len %llu found bg but no related chunk",
9570                                           found_key.objectid, found_key.offset);
9571                                 ret = -ENOENT;
9572                         } else if (em->start != found_key.objectid ||
9573                                    em->len != found_key.offset) {
9574                                 btrfs_err(fs_info,
9575                 "block group %llu len %llu mismatch with chunk %llu len %llu",
9576                                           found_key.objectid, found_key.offset,
9577                                           em->start, em->len);
9578                                 ret = -EUCLEAN;
9579                         } else {
9580                                 read_extent_buffer(leaf, &bg,
9581                                         btrfs_item_ptr_offset(leaf, slot),
9582                                         sizeof(bg));
9583                                 flags = btrfs_block_group_flags(&bg) &
9584                                         BTRFS_BLOCK_GROUP_TYPE_MASK;
9585
9586                                 if (flags != (em->map_lookup->type &
9587                                               BTRFS_BLOCK_GROUP_TYPE_MASK)) {
9588                                         btrfs_err(fs_info,
9589 "block group %llu len %llu type flags 0x%llx mismatch with chunk type flags 0x%llx",
9590                                                 found_key.objectid,
9591                                                 found_key.offset, flags,
9592                                                 (BTRFS_BLOCK_GROUP_TYPE_MASK &
9593                                                  em->map_lookup->type));
9594                                         ret = -EUCLEAN;
9595                                 } else {
9596                                         ret = 0;
9597                                 }
9598                         }
9599                         free_extent_map(em);
9600                         goto out;
9601                 }
9602                 path->slots[0]++;
9603         }
9604 out:
9605         return ret;
9606 }
9607
9608 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
9609 {
9610         struct btrfs_block_group_cache *block_group;
9611         u64 last = 0;
9612
9613         while (1) {
9614                 struct inode *inode;
9615
9616                 block_group = btrfs_lookup_first_block_group(info, last);
9617                 while (block_group) {
9618                         spin_lock(&block_group->lock);
9619                         if (block_group->iref)
9620                                 break;
9621                         spin_unlock(&block_group->lock);
9622                         block_group = next_block_group(info, block_group);
9623                 }
9624                 if (!block_group) {
9625                         if (last == 0)
9626                                 break;
9627                         last = 0;
9628                         continue;
9629                 }
9630
9631                 inode = block_group->inode;
9632                 block_group->iref = 0;
9633                 block_group->inode = NULL;
9634                 spin_unlock(&block_group->lock);
9635                 ASSERT(block_group->io_ctl.inode == NULL);
9636                 iput(inode);
9637                 last = block_group->key.objectid + block_group->key.offset;
9638                 btrfs_put_block_group(block_group);
9639         }
9640 }
9641
9642 /*
9643  * Must be called only after stopping all workers, since we could have block
9644  * group caching kthreads running, and therefore they could race with us if we
9645  * freed the block groups before stopping them.
9646  */
9647 int btrfs_free_block_groups(struct btrfs_fs_info *info)
9648 {
9649         struct btrfs_block_group_cache *block_group;
9650         struct btrfs_space_info *space_info;
9651         struct btrfs_caching_control *caching_ctl;
9652         struct rb_node *n;
9653
9654         down_write(&info->commit_root_sem);
9655         while (!list_empty(&info->caching_block_groups)) {
9656                 caching_ctl = list_entry(info->caching_block_groups.next,
9657                                          struct btrfs_caching_control, list);
9658                 list_del(&caching_ctl->list);
9659                 put_caching_control(caching_ctl);
9660         }
9661         up_write(&info->commit_root_sem);
9662
9663         spin_lock(&info->unused_bgs_lock);
9664         while (!list_empty(&info->unused_bgs)) {
9665                 block_group = list_first_entry(&info->unused_bgs,
9666                                                struct btrfs_block_group_cache,
9667                                                bg_list);
9668                 list_del_init(&block_group->bg_list);
9669                 btrfs_put_block_group(block_group);
9670         }
9671         spin_unlock(&info->unused_bgs_lock);
9672
9673         spin_lock(&info->block_group_cache_lock);
9674         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
9675                 block_group = rb_entry(n, struct btrfs_block_group_cache,
9676                                        cache_node);
9677                 rb_erase(&block_group->cache_node,
9678                          &info->block_group_cache_tree);
9679                 RB_CLEAR_NODE(&block_group->cache_node);
9680                 spin_unlock(&info->block_group_cache_lock);
9681
9682                 down_write(&block_group->space_info->groups_sem);
9683                 list_del(&block_group->list);
9684                 up_write(&block_group->space_info->groups_sem);
9685
9686                 /*
9687                  * We haven't cached this block group, which means we could
9688                  * possibly have excluded extents on this block group.
9689                  */
9690                 if (block_group->cached == BTRFS_CACHE_NO ||
9691                     block_group->cached == BTRFS_CACHE_ERROR)
9692                         free_excluded_extents(block_group);
9693
9694                 btrfs_remove_free_space_cache(block_group);
9695                 ASSERT(block_group->cached != BTRFS_CACHE_STARTED);
9696                 ASSERT(list_empty(&block_group->dirty_list));
9697                 ASSERT(list_empty(&block_group->io_list));
9698                 ASSERT(list_empty(&block_group->bg_list));
9699                 ASSERT(atomic_read(&block_group->count) == 1);
9700                 btrfs_put_block_group(block_group);
9701
9702                 spin_lock(&info->block_group_cache_lock);
9703         }
9704         spin_unlock(&info->block_group_cache_lock);
9705
9706         /* now that all the block groups are freed, go through and
9707          * free all the space_info structs.  This is only called during
9708          * the final stages of unmount, and so we know nobody is
9709          * using them.  We call synchronize_rcu() once before we start,
9710          * just to be on the safe side.
9711          */
9712         synchronize_rcu();
9713
9714         release_global_block_rsv(info);
9715
9716         while (!list_empty(&info->space_info)) {
9717                 int i;
9718
9719                 space_info = list_entry(info->space_info.next,
9720                                         struct btrfs_space_info,
9721                                         list);
9722
9723                 /*
9724                  * Do not hide this behind enospc_debug, this is actually
9725                  * important and indicates a real bug if this happens.
9726                  */
9727                 if (WARN_ON(space_info->bytes_pinned > 0 ||
9728                             space_info->bytes_reserved > 0 ||
9729                             space_info->bytes_may_use > 0))
9730                         dump_space_info(info, space_info, 0, 0);
9731                 list_del(&space_info->list);
9732                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
9733                         struct kobject *kobj;
9734                         kobj = space_info->block_group_kobjs[i];
9735                         space_info->block_group_kobjs[i] = NULL;
9736                         if (kobj) {
9737                                 kobject_del(kobj);
9738                                 kobject_put(kobj);
9739                         }
9740                 }
9741                 kobject_del(&space_info->kobj);
9742                 kobject_put(&space_info->kobj);
9743         }
9744         return 0;
9745 }
9746
9747 /* link_block_group will queue up kobjects to add when we're reclaim-safe */
9748 void btrfs_add_raid_kobjects(struct btrfs_fs_info *fs_info)
9749 {
9750         struct btrfs_space_info *space_info;
9751         struct raid_kobject *rkobj;
9752         LIST_HEAD(list);
9753         int index;
9754         int ret = 0;
9755
9756         spin_lock(&fs_info->pending_raid_kobjs_lock);
9757         list_splice_init(&fs_info->pending_raid_kobjs, &list);
9758         spin_unlock(&fs_info->pending_raid_kobjs_lock);
9759
9760         list_for_each_entry(rkobj, &list, list) {
9761                 space_info = __find_space_info(fs_info, rkobj->flags);
9762                 index = btrfs_bg_flags_to_raid_index(rkobj->flags);
9763
9764                 ret = kobject_add(&rkobj->kobj, &space_info->kobj,
9765                                   "%s", get_raid_name(index));
9766                 if (ret) {
9767                         kobject_put(&rkobj->kobj);
9768                         break;
9769                 }
9770         }
9771         if (ret)
9772                 btrfs_warn(fs_info,
9773                            "failed to add kobject for block cache, ignoring");
9774 }
9775
9776 static void link_block_group(struct btrfs_block_group_cache *cache)
9777 {
9778         struct btrfs_space_info *space_info = cache->space_info;
9779         struct btrfs_fs_info *fs_info = cache->fs_info;
9780         int index = btrfs_bg_flags_to_raid_index(cache->flags);
9781         bool first = false;
9782
9783         down_write(&space_info->groups_sem);
9784         if (list_empty(&space_info->block_groups[index]))
9785                 first = true;
9786         list_add_tail(&cache->list, &space_info->block_groups[index]);
9787         up_write(&space_info->groups_sem);
9788
9789         if (first) {
9790                 struct raid_kobject *rkobj = kzalloc(sizeof(*rkobj), GFP_NOFS);
9791                 if (!rkobj) {
9792                         btrfs_warn(cache->fs_info,
9793                                 "couldn't alloc memory for raid level kobject");
9794                         return;
9795                 }
9796                 rkobj->flags = cache->flags;
9797                 kobject_init(&rkobj->kobj, &btrfs_raid_ktype);
9798
9799                 spin_lock(&fs_info->pending_raid_kobjs_lock);
9800                 list_add_tail(&rkobj->list, &fs_info->pending_raid_kobjs);
9801                 spin_unlock(&fs_info->pending_raid_kobjs_lock);
9802                 space_info->block_group_kobjs[index] = &rkobj->kobj;
9803         }
9804 }
9805
9806 static struct btrfs_block_group_cache *
9807 btrfs_create_block_group_cache(struct btrfs_fs_info *fs_info,
9808                                u64 start, u64 size)
9809 {
9810         struct btrfs_block_group_cache *cache;
9811
9812         cache = kzalloc(sizeof(*cache), GFP_NOFS);
9813         if (!cache)
9814                 return NULL;
9815
9816         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
9817                                         GFP_NOFS);
9818         if (!cache->free_space_ctl) {
9819                 kfree(cache);
9820                 return NULL;
9821         }
9822
9823         cache->key.objectid = start;
9824         cache->key.offset = size;
9825         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
9826
9827         cache->fs_info = fs_info;
9828         cache->full_stripe_len = btrfs_full_stripe_len(fs_info, start);
9829         set_free_space_tree_thresholds(cache);
9830
9831         atomic_set(&cache->count, 1);
9832         spin_lock_init(&cache->lock);
9833         init_rwsem(&cache->data_rwsem);
9834         INIT_LIST_HEAD(&cache->list);
9835         INIT_LIST_HEAD(&cache->cluster_list);
9836         INIT_LIST_HEAD(&cache->bg_list);
9837         INIT_LIST_HEAD(&cache->ro_list);
9838         INIT_LIST_HEAD(&cache->dirty_list);
9839         INIT_LIST_HEAD(&cache->io_list);
9840         btrfs_init_free_space_ctl(cache);
9841         atomic_set(&cache->trimming, 0);
9842         mutex_init(&cache->free_space_lock);
9843         btrfs_init_full_stripe_locks_tree(&cache->full_stripe_locks_root);
9844
9845         return cache;
9846 }
9847
9848
9849 /*
9850  * Iterate all chunks and verify that each of them has the corresponding block
9851  * group
9852  */
9853 static int check_chunk_block_group_mappings(struct btrfs_fs_info *fs_info)
9854 {
9855         struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
9856         struct extent_map *em;
9857         struct btrfs_block_group_cache *bg;
9858         u64 start = 0;
9859         int ret = 0;
9860
9861         while (1) {
9862                 read_lock(&map_tree->map_tree.lock);
9863                 /*
9864                  * lookup_extent_mapping will return the first extent map
9865                  * intersecting the range, so setting @len to 1 is enough to
9866                  * get the first chunk.
9867                  */
9868                 em = lookup_extent_mapping(&map_tree->map_tree, start, 1);
9869                 read_unlock(&map_tree->map_tree.lock);
9870                 if (!em)
9871                         break;
9872
9873                 bg = btrfs_lookup_block_group(fs_info, em->start);
9874                 if (!bg) {
9875                         btrfs_err(fs_info,
9876         "chunk start=%llu len=%llu doesn't have corresponding block group",
9877                                      em->start, em->len);
9878                         ret = -EUCLEAN;
9879                         free_extent_map(em);
9880                         break;
9881                 }
9882                 if (bg->key.objectid != em->start ||
9883                     bg->key.offset != em->len ||
9884                     (bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK) !=
9885                     (em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK)) {
9886                         btrfs_err(fs_info,
9887 "chunk start=%llu len=%llu flags=0x%llx doesn't match block group start=%llu len=%llu flags=0x%llx",
9888                                 em->start, em->len,
9889                                 em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK,
9890                                 bg->key.objectid, bg->key.offset,
9891                                 bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK);
9892                         ret = -EUCLEAN;
9893                         free_extent_map(em);
9894                         btrfs_put_block_group(bg);
9895                         break;
9896                 }
9897                 start = em->start + em->len;
9898                 free_extent_map(em);
9899                 btrfs_put_block_group(bg);
9900         }
9901         return ret;
9902 }
9903
9904 int btrfs_read_block_groups(struct btrfs_fs_info *info)
9905 {
9906         struct btrfs_path *path;
9907         int ret;
9908         struct btrfs_block_group_cache *cache;
9909         struct btrfs_space_info *space_info;
9910         struct btrfs_key key;
9911         struct btrfs_key found_key;
9912         struct extent_buffer *leaf;
9913         int need_clear = 0;
9914         u64 cache_gen;
9915         u64 feature;
9916         int mixed;
9917
9918         feature = btrfs_super_incompat_flags(info->super_copy);
9919         mixed = !!(feature & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS);
9920
9921         key.objectid = 0;
9922         key.offset = 0;
9923         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
9924         path = btrfs_alloc_path();
9925         if (!path)
9926                 return -ENOMEM;
9927         path->reada = READA_FORWARD;
9928
9929         cache_gen = btrfs_super_cache_generation(info->super_copy);
9930         if (btrfs_test_opt(info, SPACE_CACHE) &&
9931             btrfs_super_generation(info->super_copy) != cache_gen)
9932                 need_clear = 1;
9933         if (btrfs_test_opt(info, CLEAR_CACHE))
9934                 need_clear = 1;
9935
9936         while (1) {
9937                 ret = find_first_block_group(info, path, &key);
9938                 if (ret > 0)
9939                         break;
9940                 if (ret != 0)
9941                         goto error;
9942
9943                 leaf = path->nodes[0];
9944                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
9945
9946                 cache = btrfs_create_block_group_cache(info, found_key.objectid,
9947                                                        found_key.offset);
9948                 if (!cache) {
9949                         ret = -ENOMEM;
9950                         goto error;
9951                 }
9952
9953                 if (need_clear) {
9954                         /*
9955                          * When we mount with old space cache, we need to
9956                          * set BTRFS_DC_CLEAR and set dirty flag.
9957                          *
9958                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
9959                          *    truncate the old free space cache inode and
9960                          *    setup a new one.
9961                          * b) Setting 'dirty flag' makes sure that we flush
9962                          *    the new space cache info onto disk.
9963                          */
9964                         if (btrfs_test_opt(info, SPACE_CACHE))
9965                                 cache->disk_cache_state = BTRFS_DC_CLEAR;
9966                 }
9967
9968                 read_extent_buffer(leaf, &cache->item,
9969                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
9970                                    sizeof(cache->item));
9971                 cache->flags = btrfs_block_group_flags(&cache->item);
9972                 if (!mixed &&
9973                     ((cache->flags & BTRFS_BLOCK_GROUP_METADATA) &&
9974                     (cache->flags & BTRFS_BLOCK_GROUP_DATA))) {
9975                         btrfs_err(info,
9976 "bg %llu is a mixed block group but filesystem hasn't enabled mixed block groups",
9977                                   cache->key.objectid);
9978                         ret = -EINVAL;
9979                         goto error;
9980                 }
9981
9982                 key.objectid = found_key.objectid + found_key.offset;
9983                 btrfs_release_path(path);
9984
9985                 /*
9986                  * We need to exclude the super stripes now so that the space
9987                  * info has super bytes accounted for, otherwise we'll think
9988                  * we have more space than we actually do.
9989                  */
9990                 ret = exclude_super_stripes(cache);
9991                 if (ret) {
9992                         /*
9993                          * We may have excluded something, so call this just in
9994                          * case.
9995                          */
9996                         free_excluded_extents(cache);
9997                         btrfs_put_block_group(cache);
9998                         goto error;
9999                 }
10000
10001                 /*
10002                  * check for two cases, either we are full, and therefore
10003                  * don't need to bother with the caching work since we won't
10004                  * find any space, or we are empty, and we can just add all
10005                  * the space in and be done with it.  This saves us _alot_ of
10006                  * time, particularly in the full case.
10007                  */
10008                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
10009                         cache->last_byte_to_unpin = (u64)-1;
10010                         cache->cached = BTRFS_CACHE_FINISHED;
10011                         free_excluded_extents(cache);
10012                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10013                         cache->last_byte_to_unpin = (u64)-1;
10014                         cache->cached = BTRFS_CACHE_FINISHED;
10015                         add_new_free_space(cache, found_key.objectid,
10016                                            found_key.objectid +
10017                                            found_key.offset);
10018                         free_excluded_extents(cache);
10019                 }
10020
10021                 ret = btrfs_add_block_group_cache(info, cache);
10022                 if (ret) {
10023                         btrfs_remove_free_space_cache(cache);
10024                         btrfs_put_block_group(cache);
10025                         goto error;
10026                 }
10027
10028                 trace_btrfs_add_block_group(info, cache, 0);
10029                 update_space_info(info, cache->flags, found_key.offset,
10030                                   btrfs_block_group_used(&cache->item),
10031                                   cache->bytes_super, &space_info);
10032
10033                 cache->space_info = space_info;
10034
10035                 link_block_group(cache);
10036
10037                 set_avail_alloc_bits(info, cache->flags);
10038                 if (btrfs_chunk_readonly(info, cache->key.objectid)) {
10039                         inc_block_group_ro(cache, 1);
10040                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10041                         ASSERT(list_empty(&cache->bg_list));
10042                         btrfs_mark_bg_unused(cache);
10043                 }
10044         }
10045
10046         list_for_each_entry_rcu(space_info, &info->space_info, list) {
10047                 if (!(get_alloc_profile(info, space_info->flags) &
10048                       (BTRFS_BLOCK_GROUP_RAID10 |
10049                        BTRFS_BLOCK_GROUP_RAID1 |
10050                        BTRFS_BLOCK_GROUP_RAID5 |
10051                        BTRFS_BLOCK_GROUP_RAID6 |
10052                        BTRFS_BLOCK_GROUP_DUP)))
10053                         continue;
10054                 /*
10055                  * avoid allocating from un-mirrored block group if there are
10056                  * mirrored block groups.
10057                  */
10058                 list_for_each_entry(cache,
10059                                 &space_info->block_groups[BTRFS_RAID_RAID0],
10060                                 list)
10061                         inc_block_group_ro(cache, 1);
10062                 list_for_each_entry(cache,
10063                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
10064                                 list)
10065                         inc_block_group_ro(cache, 1);
10066         }
10067
10068         btrfs_add_raid_kobjects(info);
10069         init_global_block_rsv(info);
10070         ret = check_chunk_block_group_mappings(info);
10071 error:
10072         btrfs_free_path(path);
10073         return ret;
10074 }
10075
10076 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans)
10077 {
10078         struct btrfs_fs_info *fs_info = trans->fs_info;
10079         struct btrfs_block_group_cache *block_group, *tmp;
10080         struct btrfs_root *extent_root = fs_info->extent_root;
10081         struct btrfs_block_group_item item;
10082         struct btrfs_key key;
10083         int ret = 0;
10084         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
10085
10086         trans->can_flush_pending_bgs = false;
10087         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) {
10088                 if (ret)
10089                         goto next;
10090
10091                 spin_lock(&block_group->lock);
10092                 memcpy(&item, &block_group->item, sizeof(item));
10093                 memcpy(&key, &block_group->key, sizeof(key));
10094                 spin_unlock(&block_group->lock);
10095
10096                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
10097                                         sizeof(item));
10098                 if (ret)
10099                         btrfs_abort_transaction(trans, ret);
10100                 ret = btrfs_finish_chunk_alloc(trans, key.objectid, key.offset);
10101                 if (ret)
10102                         btrfs_abort_transaction(trans, ret);
10103                 add_block_group_free_space(trans, block_group);
10104                 /* already aborted the transaction if it failed. */
10105 next:
10106                 list_del_init(&block_group->bg_list);
10107         }
10108         trans->can_flush_pending_bgs = can_flush_pending_bgs;
10109 }
10110
10111 int btrfs_make_block_group(struct btrfs_trans_handle *trans, u64 bytes_used,
10112                            u64 type, u64 chunk_offset, u64 size)
10113 {
10114         struct btrfs_fs_info *fs_info = trans->fs_info;
10115         struct btrfs_block_group_cache *cache;
10116         int ret;
10117
10118         btrfs_set_log_full_commit(fs_info, trans);
10119
10120         cache = btrfs_create_block_group_cache(fs_info, chunk_offset, size);
10121         if (!cache)
10122                 return -ENOMEM;
10123
10124         btrfs_set_block_group_used(&cache->item, bytes_used);
10125         btrfs_set_block_group_chunk_objectid(&cache->item,
10126                                              BTRFS_FIRST_CHUNK_TREE_OBJECTID);
10127         btrfs_set_block_group_flags(&cache->item, type);
10128
10129         cache->flags = type;
10130         cache->last_byte_to_unpin = (u64)-1;
10131         cache->cached = BTRFS_CACHE_FINISHED;
10132         cache->needs_free_space = 1;
10133         ret = exclude_super_stripes(cache);
10134         if (ret) {
10135                 /*
10136                  * We may have excluded something, so call this just in
10137                  * case.
10138                  */
10139                 free_excluded_extents(cache);
10140                 btrfs_put_block_group(cache);
10141                 return ret;
10142         }
10143
10144         add_new_free_space(cache, chunk_offset, chunk_offset + size);
10145
10146         free_excluded_extents(cache);
10147
10148 #ifdef CONFIG_BTRFS_DEBUG
10149         if (btrfs_should_fragment_free_space(cache)) {
10150                 u64 new_bytes_used = size - bytes_used;
10151
10152                 bytes_used += new_bytes_used >> 1;
10153                 fragment_free_space(cache);
10154         }
10155 #endif
10156         /*
10157          * Ensure the corresponding space_info object is created and
10158          * assigned to our block group. We want our bg to be added to the rbtree
10159          * with its ->space_info set.
10160          */
10161         cache->space_info = __find_space_info(fs_info, cache->flags);
10162         ASSERT(cache->space_info);
10163
10164         ret = btrfs_add_block_group_cache(fs_info, cache);
10165         if (ret) {
10166                 btrfs_remove_free_space_cache(cache);
10167                 btrfs_put_block_group(cache);
10168                 return ret;
10169         }
10170
10171         /*
10172          * Now that our block group has its ->space_info set and is inserted in
10173          * the rbtree, update the space info's counters.
10174          */
10175         trace_btrfs_add_block_group(fs_info, cache, 1);
10176         update_space_info(fs_info, cache->flags, size, bytes_used,
10177                                 cache->bytes_super, &cache->space_info);
10178         update_global_block_rsv(fs_info);
10179
10180         link_block_group(cache);
10181
10182         list_add_tail(&cache->bg_list, &trans->new_bgs);
10183
10184         set_avail_alloc_bits(fs_info, type);
10185         return 0;
10186 }
10187
10188 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
10189 {
10190         u64 extra_flags = chunk_to_extended(flags) &
10191                                 BTRFS_EXTENDED_PROFILE_MASK;
10192
10193         write_seqlock(&fs_info->profiles_lock);
10194         if (flags & BTRFS_BLOCK_GROUP_DATA)
10195                 fs_info->avail_data_alloc_bits &= ~extra_flags;
10196         if (flags & BTRFS_BLOCK_GROUP_METADATA)
10197                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
10198         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
10199                 fs_info->avail_system_alloc_bits &= ~extra_flags;
10200         write_sequnlock(&fs_info->profiles_lock);
10201 }
10202
10203 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
10204                              u64 group_start, struct extent_map *em)
10205 {
10206         struct btrfs_fs_info *fs_info = trans->fs_info;
10207         struct btrfs_root *root = fs_info->extent_root;
10208         struct btrfs_path *path;
10209         struct btrfs_block_group_cache *block_group;
10210         struct btrfs_free_cluster *cluster;
10211         struct btrfs_root *tree_root = fs_info->tree_root;
10212         struct btrfs_key key;
10213         struct inode *inode;
10214         struct kobject *kobj = NULL;
10215         int ret;
10216         int index;
10217         int factor;
10218         struct btrfs_caching_control *caching_ctl = NULL;
10219         bool remove_em;
10220
10221         block_group = btrfs_lookup_block_group(fs_info, group_start);
10222         BUG_ON(!block_group);
10223         BUG_ON(!block_group->ro);
10224
10225         trace_btrfs_remove_block_group(block_group);
10226         /*
10227          * Free the reserved super bytes from this block group before
10228          * remove it.
10229          */
10230         free_excluded_extents(block_group);
10231         btrfs_free_ref_tree_range(fs_info, block_group->key.objectid,
10232                                   block_group->key.offset);
10233
10234         memcpy(&key, &block_group->key, sizeof(key));
10235         index = btrfs_bg_flags_to_raid_index(block_group->flags);
10236         factor = btrfs_bg_type_to_factor(block_group->flags);
10237
10238         /* make sure this block group isn't part of an allocation cluster */
10239         cluster = &fs_info->data_alloc_cluster;
10240         spin_lock(&cluster->refill_lock);
10241         btrfs_return_cluster_to_free_space(block_group, cluster);
10242         spin_unlock(&cluster->refill_lock);
10243
10244         /*
10245          * make sure this block group isn't part of a metadata
10246          * allocation cluster
10247          */
10248         cluster = &fs_info->meta_alloc_cluster;
10249         spin_lock(&cluster->refill_lock);
10250         btrfs_return_cluster_to_free_space(block_group, cluster);
10251         spin_unlock(&cluster->refill_lock);
10252
10253         path = btrfs_alloc_path();
10254         if (!path) {
10255                 ret = -ENOMEM;
10256                 goto out;
10257         }
10258
10259         /*
10260          * get the inode first so any iput calls done for the io_list
10261          * aren't the final iput (no unlinks allowed now)
10262          */
10263         inode = lookup_free_space_inode(fs_info, block_group, path);
10264
10265         mutex_lock(&trans->transaction->cache_write_mutex);
10266         /*
10267          * make sure our free spache cache IO is done before remove the
10268          * free space inode
10269          */
10270         spin_lock(&trans->transaction->dirty_bgs_lock);
10271         if (!list_empty(&block_group->io_list)) {
10272                 list_del_init(&block_group->io_list);
10273
10274                 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
10275
10276                 spin_unlock(&trans->transaction->dirty_bgs_lock);
10277                 btrfs_wait_cache_io(trans, block_group, path);
10278                 btrfs_put_block_group(block_group);
10279                 spin_lock(&trans->transaction->dirty_bgs_lock);
10280         }
10281
10282         if (!list_empty(&block_group->dirty_list)) {
10283                 list_del_init(&block_group->dirty_list);
10284                 btrfs_put_block_group(block_group);
10285         }
10286         spin_unlock(&trans->transaction->dirty_bgs_lock);
10287         mutex_unlock(&trans->transaction->cache_write_mutex);
10288
10289         if (!IS_ERR(inode)) {
10290                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
10291                 if (ret) {
10292                         btrfs_add_delayed_iput(inode);
10293                         goto out;
10294                 }
10295                 clear_nlink(inode);
10296                 /* One for the block groups ref */
10297                 spin_lock(&block_group->lock);
10298                 if (block_group->iref) {
10299                         block_group->iref = 0;
10300                         block_group->inode = NULL;
10301                         spin_unlock(&block_group->lock);
10302                         iput(inode);
10303                 } else {
10304                         spin_unlock(&block_group->lock);
10305                 }
10306                 /* One for our lookup ref */
10307                 btrfs_add_delayed_iput(inode);
10308         }
10309
10310         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
10311         key.offset = block_group->key.objectid;
10312         key.type = 0;
10313
10314         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
10315         if (ret < 0)
10316                 goto out;
10317         if (ret > 0)
10318                 btrfs_release_path(path);
10319         if (ret == 0) {
10320                 ret = btrfs_del_item(trans, tree_root, path);
10321                 if (ret)
10322                         goto out;
10323                 btrfs_release_path(path);
10324         }
10325
10326         spin_lock(&fs_info->block_group_cache_lock);
10327         rb_erase(&block_group->cache_node,
10328                  &fs_info->block_group_cache_tree);
10329         RB_CLEAR_NODE(&block_group->cache_node);
10330
10331         if (fs_info->first_logical_byte == block_group->key.objectid)
10332                 fs_info->first_logical_byte = (u64)-1;
10333         spin_unlock(&fs_info->block_group_cache_lock);
10334
10335         down_write(&block_group->space_info->groups_sem);
10336         /*
10337          * we must use list_del_init so people can check to see if they
10338          * are still on the list after taking the semaphore
10339          */
10340         list_del_init(&block_group->list);
10341         if (list_empty(&block_group->space_info->block_groups[index])) {
10342                 kobj = block_group->space_info->block_group_kobjs[index];
10343                 block_group->space_info->block_group_kobjs[index] = NULL;
10344                 clear_avail_alloc_bits(fs_info, block_group->flags);
10345         }
10346         up_write(&block_group->space_info->groups_sem);
10347         if (kobj) {
10348                 kobject_del(kobj);
10349                 kobject_put(kobj);
10350         }
10351
10352         if (block_group->has_caching_ctl)
10353                 caching_ctl = get_caching_control(block_group);
10354         if (block_group->cached == BTRFS_CACHE_STARTED)
10355                 wait_block_group_cache_done(block_group);
10356         if (block_group->has_caching_ctl) {
10357                 down_write(&fs_info->commit_root_sem);
10358                 if (!caching_ctl) {
10359                         struct btrfs_caching_control *ctl;
10360
10361                         list_for_each_entry(ctl,
10362                                     &fs_info->caching_block_groups, list)
10363                                 if (ctl->block_group == block_group) {
10364                                         caching_ctl = ctl;
10365                                         refcount_inc(&caching_ctl->count);
10366                                         break;
10367                                 }
10368                 }
10369                 if (caching_ctl)
10370                         list_del_init(&caching_ctl->list);
10371                 up_write(&fs_info->commit_root_sem);
10372                 if (caching_ctl) {
10373                         /* Once for the caching bgs list and once for us. */
10374                         put_caching_control(caching_ctl);
10375                         put_caching_control(caching_ctl);
10376                 }
10377         }
10378
10379         spin_lock(&trans->transaction->dirty_bgs_lock);
10380         if (!list_empty(&block_group->dirty_list)) {
10381                 WARN_ON(1);
10382         }
10383         if (!list_empty(&block_group->io_list)) {
10384                 WARN_ON(1);
10385         }
10386         spin_unlock(&trans->transaction->dirty_bgs_lock);
10387         btrfs_remove_free_space_cache(block_group);
10388
10389         spin_lock(&block_group->space_info->lock);
10390         list_del_init(&block_group->ro_list);
10391
10392         if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
10393                 WARN_ON(block_group->space_info->total_bytes
10394                         < block_group->key.offset);
10395                 WARN_ON(block_group->space_info->bytes_readonly
10396                         < block_group->key.offset);
10397                 WARN_ON(block_group->space_info->disk_total
10398                         < block_group->key.offset * factor);
10399         }
10400         block_group->space_info->total_bytes -= block_group->key.offset;
10401         block_group->space_info->bytes_readonly -= block_group->key.offset;
10402         block_group->space_info->disk_total -= block_group->key.offset * factor;
10403
10404         spin_unlock(&block_group->space_info->lock);
10405
10406         memcpy(&key, &block_group->key, sizeof(key));
10407
10408         mutex_lock(&fs_info->chunk_mutex);
10409         if (!list_empty(&em->list)) {
10410                 /* We're in the transaction->pending_chunks list. */
10411                 free_extent_map(em);
10412         }
10413         spin_lock(&block_group->lock);
10414         block_group->removed = 1;
10415         /*
10416          * At this point trimming can't start on this block group, because we
10417          * removed the block group from the tree fs_info->block_group_cache_tree
10418          * so no one can't find it anymore and even if someone already got this
10419          * block group before we removed it from the rbtree, they have already
10420          * incremented block_group->trimming - if they didn't, they won't find
10421          * any free space entries because we already removed them all when we
10422          * called btrfs_remove_free_space_cache().
10423          *
10424          * And we must not remove the extent map from the fs_info->mapping_tree
10425          * to prevent the same logical address range and physical device space
10426          * ranges from being reused for a new block group. This is because our
10427          * fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
10428          * completely transactionless, so while it is trimming a range the
10429          * currently running transaction might finish and a new one start,
10430          * allowing for new block groups to be created that can reuse the same
10431          * physical device locations unless we take this special care.
10432          *
10433          * There may also be an implicit trim operation if the file system
10434          * is mounted with -odiscard. The same protections must remain
10435          * in place until the extents have been discarded completely when
10436          * the transaction commit has completed.
10437          */
10438         remove_em = (atomic_read(&block_group->trimming) == 0);
10439         /*
10440          * Make sure a trimmer task always sees the em in the pinned_chunks list
10441          * if it sees block_group->removed == 1 (needs to lock block_group->lock
10442          * before checking block_group->removed).
10443          */
10444         if (!remove_em) {
10445                 /*
10446                  * Our em might be in trans->transaction->pending_chunks which
10447                  * is protected by fs_info->chunk_mutex ([lock|unlock]_chunks),
10448                  * and so is the fs_info->pinned_chunks list.
10449                  *
10450                  * So at this point we must be holding the chunk_mutex to avoid
10451                  * any races with chunk allocation (more specifically at
10452                  * volumes.c:contains_pending_extent()), to ensure it always
10453                  * sees the em, either in the pending_chunks list or in the
10454                  * pinned_chunks list.
10455                  */
10456                 list_move_tail(&em->list, &fs_info->pinned_chunks);
10457         }
10458         spin_unlock(&block_group->lock);
10459
10460         if (remove_em) {
10461                 struct extent_map_tree *em_tree;
10462
10463                 em_tree = &fs_info->mapping_tree.map_tree;
10464                 write_lock(&em_tree->lock);
10465                 /*
10466                  * The em might be in the pending_chunks list, so make sure the
10467                  * chunk mutex is locked, since remove_extent_mapping() will
10468                  * delete us from that list.
10469                  */
10470                 remove_extent_mapping(em_tree, em);
10471                 write_unlock(&em_tree->lock);
10472                 /* once for the tree */
10473                 free_extent_map(em);
10474         }
10475
10476         mutex_unlock(&fs_info->chunk_mutex);
10477
10478         ret = remove_block_group_free_space(trans, block_group);
10479         if (ret)
10480                 goto out;
10481
10482         btrfs_put_block_group(block_group);
10483         btrfs_put_block_group(block_group);
10484
10485         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
10486         if (ret > 0)
10487                 ret = -EIO;
10488         if (ret < 0)
10489                 goto out;
10490
10491         ret = btrfs_del_item(trans, root, path);
10492 out:
10493         btrfs_free_path(path);
10494         return ret;
10495 }
10496
10497 struct btrfs_trans_handle *
10498 btrfs_start_trans_remove_block_group(struct btrfs_fs_info *fs_info,
10499                                      const u64 chunk_offset)
10500 {
10501         struct extent_map_tree *em_tree = &fs_info->mapping_tree.map_tree;
10502         struct extent_map *em;
10503         struct map_lookup *map;
10504         unsigned int num_items;
10505
10506         read_lock(&em_tree->lock);
10507         em = lookup_extent_mapping(em_tree, chunk_offset, 1);
10508         read_unlock(&em_tree->lock);
10509         ASSERT(em && em->start == chunk_offset);
10510
10511         /*
10512          * We need to reserve 3 + N units from the metadata space info in order
10513          * to remove a block group (done at btrfs_remove_chunk() and at
10514          * btrfs_remove_block_group()), which are used for:
10515          *
10516          * 1 unit for adding the free space inode's orphan (located in the tree
10517          * of tree roots).
10518          * 1 unit for deleting the block group item (located in the extent
10519          * tree).
10520          * 1 unit for deleting the free space item (located in tree of tree
10521          * roots).
10522          * N units for deleting N device extent items corresponding to each
10523          * stripe (located in the device tree).
10524          *
10525          * In order to remove a block group we also need to reserve units in the
10526          * system space info in order to update the chunk tree (update one or
10527          * more device items and remove one chunk item), but this is done at
10528          * btrfs_remove_chunk() through a call to check_system_chunk().
10529          */
10530         map = em->map_lookup;
10531         num_items = 3 + map->num_stripes;
10532         free_extent_map(em);
10533
10534         return btrfs_start_transaction_fallback_global_rsv(fs_info->extent_root,
10535                                                            num_items, 1);
10536 }
10537
10538 /*
10539  * Process the unused_bgs list and remove any that don't have any allocated
10540  * space inside of them.
10541  */
10542 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
10543 {
10544         struct btrfs_block_group_cache *block_group;
10545         struct btrfs_space_info *space_info;
10546         struct btrfs_trans_handle *trans;
10547         int ret = 0;
10548
10549         if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
10550                 return;
10551
10552         spin_lock(&fs_info->unused_bgs_lock);
10553         while (!list_empty(&fs_info->unused_bgs)) {
10554                 u64 start, end;
10555                 int trimming;
10556
10557                 block_group = list_first_entry(&fs_info->unused_bgs,
10558                                                struct btrfs_block_group_cache,
10559                                                bg_list);
10560                 list_del_init(&block_group->bg_list);
10561
10562                 space_info = block_group->space_info;
10563
10564                 if (ret || btrfs_mixed_space_info(space_info)) {
10565                         btrfs_put_block_group(block_group);
10566                         continue;
10567                 }
10568                 spin_unlock(&fs_info->unused_bgs_lock);
10569
10570                 mutex_lock(&fs_info->delete_unused_bgs_mutex);
10571
10572                 /* Don't want to race with allocators so take the groups_sem */
10573                 down_write(&space_info->groups_sem);
10574                 spin_lock(&block_group->lock);
10575                 if (block_group->reserved || block_group->pinned ||
10576                     btrfs_block_group_used(&block_group->item) ||
10577                     block_group->ro ||
10578                     list_is_singular(&block_group->list)) {
10579                         /*
10580                          * We want to bail if we made new allocations or have
10581                          * outstanding allocations in this block group.  We do
10582                          * the ro check in case balance is currently acting on
10583                          * this block group.
10584                          */
10585                         trace_btrfs_skip_unused_block_group(block_group);
10586                         spin_unlock(&block_group->lock);
10587                         up_write(&space_info->groups_sem);
10588                         goto next;
10589                 }
10590                 spin_unlock(&block_group->lock);
10591
10592                 /* We don't want to force the issue, only flip if it's ok. */
10593                 ret = inc_block_group_ro(block_group, 0);
10594                 up_write(&space_info->groups_sem);
10595                 if (ret < 0) {
10596                         ret = 0;
10597                         goto next;
10598                 }
10599
10600                 /*
10601                  * Want to do this before we do anything else so we can recover
10602                  * properly if we fail to join the transaction.
10603                  */
10604                 trans = btrfs_start_trans_remove_block_group(fs_info,
10605                                                      block_group->key.objectid);
10606                 if (IS_ERR(trans)) {
10607                         btrfs_dec_block_group_ro(block_group);
10608                         ret = PTR_ERR(trans);
10609                         goto next;
10610                 }
10611
10612                 /*
10613                  * We could have pending pinned extents for this block group,
10614                  * just delete them, we don't care about them anymore.
10615                  */
10616                 start = block_group->key.objectid;
10617                 end = start + block_group->key.offset - 1;
10618                 /*
10619                  * Hold the unused_bg_unpin_mutex lock to avoid racing with
10620                  * btrfs_finish_extent_commit(). If we are at transaction N,
10621                  * another task might be running finish_extent_commit() for the
10622                  * previous transaction N - 1, and have seen a range belonging
10623                  * to the block group in freed_extents[] before we were able to
10624                  * clear the whole block group range from freed_extents[]. This
10625                  * means that task can lookup for the block group after we
10626                  * unpinned it from freed_extents[] and removed it, leading to
10627                  * a BUG_ON() at btrfs_unpin_extent_range().
10628                  */
10629                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
10630                 ret = clear_extent_bits(&fs_info->freed_extents[0], start, end,
10631                                   EXTENT_DIRTY);
10632                 if (ret) {
10633                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10634                         btrfs_dec_block_group_ro(block_group);
10635                         goto end_trans;
10636                 }
10637                 ret = clear_extent_bits(&fs_info->freed_extents[1], start, end,
10638                                   EXTENT_DIRTY);
10639                 if (ret) {
10640                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10641                         btrfs_dec_block_group_ro(block_group);
10642                         goto end_trans;
10643                 }
10644                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10645
10646                 /* Reset pinned so btrfs_put_block_group doesn't complain */
10647                 spin_lock(&space_info->lock);
10648                 spin_lock(&block_group->lock);
10649
10650                 space_info->bytes_pinned -= block_group->pinned;
10651                 space_info->bytes_readonly += block_group->pinned;
10652                 percpu_counter_add_batch(&space_info->total_bytes_pinned,
10653                                    -block_group->pinned,
10654                                    BTRFS_TOTAL_BYTES_PINNED_BATCH);
10655                 block_group->pinned = 0;
10656
10657                 spin_unlock(&block_group->lock);
10658                 spin_unlock(&space_info->lock);
10659
10660                 /* DISCARD can flip during remount */
10661                 trimming = btrfs_test_opt(fs_info, DISCARD);
10662
10663                 /* Implicit trim during transaction commit. */
10664                 if (trimming)
10665                         btrfs_get_block_group_trimming(block_group);
10666
10667                 /*
10668                  * Btrfs_remove_chunk will abort the transaction if things go
10669                  * horribly wrong.
10670                  */
10671                 ret = btrfs_remove_chunk(trans, block_group->key.objectid);
10672
10673                 if (ret) {
10674                         if (trimming)
10675                                 btrfs_put_block_group_trimming(block_group);
10676                         goto end_trans;
10677                 }
10678
10679                 /*
10680                  * If we're not mounted with -odiscard, we can just forget
10681                  * about this block group. Otherwise we'll need to wait
10682                  * until transaction commit to do the actual discard.
10683                  */
10684                 if (trimming) {
10685                         spin_lock(&fs_info->unused_bgs_lock);
10686                         /*
10687                          * A concurrent scrub might have added us to the list
10688                          * fs_info->unused_bgs, so use a list_move operation
10689                          * to add the block group to the deleted_bgs list.
10690                          */
10691                         list_move(&block_group->bg_list,
10692                                   &trans->transaction->deleted_bgs);
10693                         spin_unlock(&fs_info->unused_bgs_lock);
10694                         btrfs_get_block_group(block_group);
10695                 }
10696 end_trans:
10697                 btrfs_end_transaction(trans);
10698 next:
10699                 mutex_unlock(&fs_info->delete_unused_bgs_mutex);
10700                 btrfs_put_block_group(block_group);
10701                 spin_lock(&fs_info->unused_bgs_lock);
10702         }
10703         spin_unlock(&fs_info->unused_bgs_lock);
10704 }
10705
10706 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
10707 {
10708         struct btrfs_super_block *disk_super;
10709         u64 features;
10710         u64 flags;
10711         int mixed = 0;
10712         int ret;
10713
10714         disk_super = fs_info->super_copy;
10715         if (!btrfs_super_root(disk_super))
10716                 return -EINVAL;
10717
10718         features = btrfs_super_incompat_flags(disk_super);
10719         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
10720                 mixed = 1;
10721
10722         flags = BTRFS_BLOCK_GROUP_SYSTEM;
10723         ret = create_space_info(fs_info, flags);
10724         if (ret)
10725                 goto out;
10726
10727         if (mixed) {
10728                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
10729                 ret = create_space_info(fs_info, flags);
10730         } else {
10731                 flags = BTRFS_BLOCK_GROUP_METADATA;
10732                 ret = create_space_info(fs_info, flags);
10733                 if (ret)
10734                         goto out;
10735
10736                 flags = BTRFS_BLOCK_GROUP_DATA;
10737                 ret = create_space_info(fs_info, flags);
10738         }
10739 out:
10740         return ret;
10741 }
10742
10743 int btrfs_error_unpin_extent_range(struct btrfs_fs_info *fs_info,
10744                                    u64 start, u64 end)
10745 {
10746         return unpin_extent_range(fs_info, start, end, false);
10747 }
10748
10749 /*
10750  * It used to be that old block groups would be left around forever.
10751  * Iterating over them would be enough to trim unused space.  Since we
10752  * now automatically remove them, we also need to iterate over unallocated
10753  * space.
10754  *
10755  * We don't want a transaction for this since the discard may take a
10756  * substantial amount of time.  We don't require that a transaction be
10757  * running, but we do need to take a running transaction into account
10758  * to ensure that we're not discarding chunks that were released in
10759  * the current transaction.
10760  *
10761  * Holding the chunks lock will prevent other threads from allocating
10762  * or releasing chunks, but it won't prevent a running transaction
10763  * from committing and releasing the memory that the pending chunks
10764  * list head uses.  For that, we need to take a reference to the
10765  * transaction.
10766  */
10767 static int btrfs_trim_free_extents(struct btrfs_device *device,
10768                                    u64 minlen, u64 *trimmed)
10769 {
10770         u64 start = 0, len = 0;
10771         int ret;
10772
10773         *trimmed = 0;
10774
10775         /* Not writeable = nothing to do. */
10776         if (!test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state))
10777                 return 0;
10778
10779         /* No free space = nothing to do. */
10780         if (device->total_bytes <= device->bytes_used)
10781                 return 0;
10782
10783         ret = 0;
10784
10785         while (1) {
10786                 struct btrfs_fs_info *fs_info = device->fs_info;
10787                 struct btrfs_transaction *trans;
10788                 u64 bytes;
10789
10790                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
10791                 if (ret)
10792                         return ret;
10793
10794                 down_read(&fs_info->commit_root_sem);
10795
10796                 spin_lock(&fs_info->trans_lock);
10797                 trans = fs_info->running_transaction;
10798                 if (trans)
10799                         refcount_inc(&trans->use_count);
10800                 spin_unlock(&fs_info->trans_lock);
10801
10802                 ret = find_free_dev_extent_start(trans, device, minlen, start,
10803                                                  &start, &len);
10804                 if (trans)
10805                         btrfs_put_transaction(trans);
10806
10807                 if (ret) {
10808                         up_read(&fs_info->commit_root_sem);
10809                         mutex_unlock(&fs_info->chunk_mutex);
10810                         if (ret == -ENOSPC)
10811                                 ret = 0;
10812                         break;
10813                 }
10814
10815                 ret = btrfs_issue_discard(device->bdev, start, len, &bytes);
10816                 up_read(&fs_info->commit_root_sem);
10817                 mutex_unlock(&fs_info->chunk_mutex);
10818
10819                 if (ret)
10820                         break;
10821
10822                 start += len;
10823                 *trimmed += bytes;
10824
10825                 if (fatal_signal_pending(current)) {
10826                         ret = -ERESTARTSYS;
10827                         break;
10828                 }
10829
10830                 cond_resched();
10831         }
10832
10833         return ret;
10834 }
10835
10836 /*
10837  * Trim the whole filesystem by:
10838  * 1) trimming the free space in each block group
10839  * 2) trimming the unallocated space on each device
10840  *
10841  * This will also continue trimming even if a block group or device encounters
10842  * an error.  The return value will be the last error, or 0 if nothing bad
10843  * happens.
10844  */
10845 int btrfs_trim_fs(struct btrfs_fs_info *fs_info, struct fstrim_range *range)
10846 {
10847         struct btrfs_block_group_cache *cache = NULL;
10848         struct btrfs_device *device;
10849         struct list_head *devices;
10850         u64 group_trimmed;
10851         u64 start;
10852         u64 end;
10853         u64 trimmed = 0;
10854         u64 bg_failed = 0;
10855         u64 dev_failed = 0;
10856         int bg_ret = 0;
10857         int dev_ret = 0;
10858         int ret = 0;
10859
10860         cache = btrfs_lookup_first_block_group(fs_info, range->start);
10861         for (; cache; cache = next_block_group(fs_info, cache)) {
10862                 if (cache->key.objectid >= (range->start + range->len)) {
10863                         btrfs_put_block_group(cache);
10864                         break;
10865                 }
10866
10867                 start = max(range->start, cache->key.objectid);
10868                 end = min(range->start + range->len,
10869                                 cache->key.objectid + cache->key.offset);
10870
10871                 if (end - start >= range->minlen) {
10872                         if (!block_group_cache_done(cache)) {
10873                                 ret = cache_block_group(cache, 0);
10874                                 if (ret) {
10875                                         bg_failed++;
10876                                         bg_ret = ret;
10877                                         continue;
10878                                 }
10879                                 ret = wait_block_group_cache_done(cache);
10880                                 if (ret) {
10881                                         bg_failed++;
10882                                         bg_ret = ret;
10883                                         continue;
10884                                 }
10885                         }
10886                         ret = btrfs_trim_block_group(cache,
10887                                                      &group_trimmed,
10888                                                      start,
10889                                                      end,
10890                                                      range->minlen);
10891
10892                         trimmed += group_trimmed;
10893                         if (ret) {
10894                                 bg_failed++;
10895                                 bg_ret = ret;
10896                                 continue;
10897                         }
10898                 }
10899         }
10900
10901         if (bg_failed)
10902                 btrfs_warn(fs_info,
10903                         "failed to trim %llu block group(s), last error %d",
10904                         bg_failed, bg_ret);
10905         mutex_lock(&fs_info->fs_devices->device_list_mutex);
10906         devices = &fs_info->fs_devices->devices;
10907         list_for_each_entry(device, devices, dev_list) {
10908                 ret = btrfs_trim_free_extents(device, range->minlen,
10909                                               &group_trimmed);
10910                 if (ret) {
10911                         dev_failed++;
10912                         dev_ret = ret;
10913                         break;
10914                 }
10915
10916                 trimmed += group_trimmed;
10917         }
10918         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
10919
10920         if (dev_failed)
10921                 btrfs_warn(fs_info,
10922                         "failed to trim %llu device(s), last error %d",
10923                         dev_failed, dev_ret);
10924         range->len = trimmed;
10925         if (bg_ret)
10926                 return bg_ret;
10927         return dev_ret;
10928 }
10929
10930 /*
10931  * btrfs_{start,end}_write_no_snapshotting() are similar to
10932  * mnt_{want,drop}_write(), they are used to prevent some tasks from writing
10933  * data into the page cache through nocow before the subvolume is snapshoted,
10934  * but flush the data into disk after the snapshot creation, or to prevent
10935  * operations while snapshotting is ongoing and that cause the snapshot to be
10936  * inconsistent (writes followed by expanding truncates for example).
10937  */
10938 void btrfs_end_write_no_snapshotting(struct btrfs_root *root)
10939 {
10940         percpu_counter_dec(&root->subv_writers->counter);
10941         cond_wake_up(&root->subv_writers->wait);
10942 }
10943
10944 int btrfs_start_write_no_snapshotting(struct btrfs_root *root)
10945 {
10946         if (atomic_read(&root->will_be_snapshotted))
10947                 return 0;
10948
10949         percpu_counter_inc(&root->subv_writers->counter);
10950         /*
10951          * Make sure counter is updated before we check for snapshot creation.
10952          */
10953         smp_mb();
10954         if (atomic_read(&root->will_be_snapshotted)) {
10955                 btrfs_end_write_no_snapshotting(root);
10956                 return 0;
10957         }
10958         return 1;
10959 }
10960
10961 void btrfs_wait_for_snapshot_creation(struct btrfs_root *root)
10962 {
10963         while (true) {
10964                 int ret;
10965
10966                 ret = btrfs_start_write_no_snapshotting(root);
10967                 if (ret)
10968                         break;
10969                 wait_var_event(&root->will_be_snapshotted,
10970                                !atomic_read(&root->will_be_snapshotted));
10971         }
10972 }
10973
10974 void btrfs_mark_bg_unused(struct btrfs_block_group_cache *bg)
10975 {
10976         struct btrfs_fs_info *fs_info = bg->fs_info;
10977
10978         spin_lock(&fs_info->unused_bgs_lock);
10979         if (list_empty(&bg->bg_list)) {
10980                 btrfs_get_block_group(bg);
10981                 trace_btrfs_add_unused_block_group(bg);
10982                 list_add_tail(&bg->bg_list, &fs_info->unused_bgs);
10983         }
10984         spin_unlock(&fs_info->unused_bgs_lock);
10985 }