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_fs_info *fs_info,
  56                                 struct btrfs_delayed_ref_node *node, u64 parent,
  57                                 u64 root_objectid, u64 owner_objectid,
  58                                 u64 owner_offset, int refs_to_drop,
  59                                 struct btrfs_delayed_extent_op *extra_op);
  60 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  61                                     struct extent_buffer *leaf,
  62                                     struct btrfs_extent_item *ei);
  63 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  64                                       struct btrfs_fs_info *fs_info,
  65                                       u64 parent, u64 root_objectid,
  66                                       u64 flags, u64 owner, u64 offset,
  67                                       struct btrfs_key *ins, int ref_mod);
  68 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  69                                      struct btrfs_fs_info *fs_info,
  70                                      u64 parent, u64 root_objectid,
  71                                      u64 flags, struct btrfs_disk_key *key,
  72                                      int level, struct btrfs_key *ins);
  73 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
  74                           struct btrfs_fs_info *fs_info, u64 flags,
  75                           int force);
  76 static int find_next_key(struct btrfs_path *path, int level,
  77                          struct btrfs_key *key);
  78 static void dump_space_info(struct btrfs_fs_info *fs_info,
  79                             struct btrfs_space_info *info, u64 bytes,
  80                             int dump_block_groups);
  81 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
  82                                u64 num_bytes);
  83 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
  84                                      struct btrfs_space_info *space_info,
  85                                      u64 num_bytes);
  86 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
  87                                      struct btrfs_space_info *space_info,
  88                                      u64 num_bytes);
  89
  90 static noinline int
  91 block_group_cache_done(struct btrfs_block_group_cache *cache)
  92 {
  93         smp_mb();
  94         return cache->cached == BTRFS_CACHE_FINISHED ||
  95                 cache->cached == BTRFS_CACHE_ERROR;
  96 }
  97
  98 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
  99 {
 100         return (cache->flags & bits) == bits;
 101 }
 102
 103 void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 104 {
 105         atomic_inc(&cache->count);
 106 }
 107
 108 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 109 {
 110         if (atomic_dec_and_test(&cache->count)) {
 111                 WARN_ON(cache->pinned > 0);
 112                 WARN_ON(cache->reserved > 0);
 113
 114                 /*
 115                  * If not empty, someone is still holding mutex of
 116                  * full_stripe_lock, which can only be released by caller.
 117                  * And it will definitely cause use-after-free when caller
 118                  * tries to release full stripe lock.
 119                  *
 120                  * No better way to resolve, but only to warn.
 121                  */
 122                 WARN_ON(!RB_EMPTY_ROOT(&cache->full_stripe_locks_root.root));
 123                 kfree(cache->free_space_ctl);
 124                 kfree(cache);
 125         }
 126 }
 127
 128 /*
 129  * this adds the block group to the fs_info rb tree for the block group
 130  * cache
 131  */
 132 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 133                                 struct btrfs_block_group_cache *block_group)
 134 {
 135         struct rb_node **p;
 136         struct rb_node *parent = NULL;
 137         struct btrfs_block_group_cache *cache;
 138
 139         spin_lock(&info->block_group_cache_lock);
 140         p = &info->block_group_cache_tree.rb_node;
 141
 142         while (*p) {
 143                 parent = *p;
 144                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 145                                  cache_node);
 146                 if (block_group->key.objectid < cache->key.objectid) {
 147                         p = &(*p)->rb_left;
 148                 } else if (block_group->key.objectid > cache->key.objectid) {
 149                         p = &(*p)->rb_right;
 150                 } else {
 151                         spin_unlock(&info->block_group_cache_lock);
 152                         return -EEXIST;
 153                 }
 154         }
 155
 156         rb_link_node(&block_group->cache_node, parent, p);
 157         rb_insert_color(&block_group->cache_node,
 158                         &info->block_group_cache_tree);
 159
 160         if (info->first_logical_byte > block_group->key.objectid)
 161                 info->first_logical_byte = block_group->key.objectid;
 162
 163         spin_unlock(&info->block_group_cache_lock);
 164
 165         return 0;
 166 }
 167
 168 /*
 169  * This will return the block group at or after bytenr if contains is 0, else
 170  * it will return the block group that contains the bytenr
 171  */
 172 static struct btrfs_block_group_cache *
 173 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 174                               int contains)
 175 {
 176         struct btrfs_block_group_cache *cache, *ret = NULL;
 177         struct rb_node *n;
 178         u64 end, start;
 179
 180         spin_lock(&info->block_group_cache_lock);
 181         n = info->block_group_cache_tree.rb_node;
 182
 183         while (n) {
 184                 cache = rb_entry(n, struct btrfs_block_group_cache,
 185                                  cache_node);
 186                 end = cache->key.objectid + cache->key.offset - 1;
 187                 start = cache->key.objectid;
 188
 189                 if (bytenr < start) {
 190                         if (!contains && (!ret || start < ret->key.objectid))
 191                                 ret = cache;
 192                         n = n->rb_left;
 193                 } else if (bytenr > start) {
 194                         if (contains && bytenr <= end) {
 195                                 ret = cache;
 196                                 break;
 197                         }
 198                         n = n->rb_right;
 199                 } else {
 200                         ret = cache;
 201                         break;
 202                 }
 203         }
 204         if (ret) {
 205                 btrfs_get_block_group(ret);
 206                 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
 207                         info->first_logical_byte = ret->key.objectid;
 208         }
 209         spin_unlock(&info->block_group_cache_lock);
 210
 211         return ret;
 212 }
 213
 214 static int add_excluded_extent(struct btrfs_fs_info *fs_info,
 215                                u64 start, u64 num_bytes)
 216 {
 217         u64 end = start + num_bytes - 1;
 218         set_extent_bits(&fs_info->freed_extents[0],
 219                         start, end, EXTENT_UPTODATE);
 220         set_extent_bits(&fs_info->freed_extents[1],
 221                         start, end, EXTENT_UPTODATE);
 222         return 0;
 223 }
 224
 225 static void free_excluded_extents(struct btrfs_fs_info *fs_info,
 226                                   struct btrfs_block_group_cache *cache)
 227 {
 228         u64 start, end;
 229
 230         start = cache->key.objectid;
 231         end = start + cache->key.offset - 1;
 232
 233         clear_extent_bits(&fs_info->freed_extents[0],
 234                           start, end, EXTENT_UPTODATE);
 235         clear_extent_bits(&fs_info->freed_extents[1],
 236                           start, end, EXTENT_UPTODATE);
 237 }
 238
 239 static int exclude_super_stripes(struct btrfs_fs_info *fs_info,
 240                                  struct btrfs_block_group_cache *cache)
 241 {
 242         u64 bytenr;
 243         u64 *logical;
 244         int stripe_len;
 245         int i, nr, ret;
 246
 247         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 248                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 249                 cache->bytes_super += stripe_len;
 250                 ret = add_excluded_extent(fs_info, cache->key.objectid,
 251                                           stripe_len);
 252                 if (ret)
 253                         return ret;
 254         }
 255
 256         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 257                 bytenr = btrfs_sb_offset(i);
 258                 ret = btrfs_rmap_block(fs_info, cache->key.objectid,
 259                                        bytenr, 0, &logical, &nr, &stripe_len);
 260                 if (ret)
 261                         return ret;
 262
 263                 while (nr--) {
 264                         u64 start, len;
 265
 266                         if (logical[nr] > cache->key.objectid +
 267                             cache->key.offset)
 268                                 continue;
 269
 270                         if (logical[nr] + stripe_len <= cache->key.objectid)
 271                                 continue;
 272
 273                         start = logical[nr];
 274                         if (start < cache->key.objectid) {
 275                                 start = cache->key.objectid;
 276                                 len = (logical[nr] + stripe_len) - start;
 277                         } else {
 278                                 len = min_t(u64, stripe_len,
 279                                             cache->key.objectid +
 280                                             cache->key.offset - start);
 281                         }
 282
 283                         cache->bytes_super += len;
 284                         ret = add_excluded_extent(fs_info, start, len);
 285                         if (ret) {
 286                                 kfree(logical);
 287                                 return ret;
 288                         }
 289                 }
 290
 291                 kfree(logical);
 292         }
 293         return 0;
 294 }
 295
 296 static struct btrfs_caching_control *
 297 get_caching_control(struct btrfs_block_group_cache *cache)
 298 {
 299         struct btrfs_caching_control *ctl;
 300
 301         spin_lock(&cache->lock);
 302         if (!cache->caching_ctl) {
 303                 spin_unlock(&cache->lock);
 304                 return NULL;
 305         }
 306
 307         ctl = cache->caching_ctl;
 308         refcount_inc(&ctl->count);
 309         spin_unlock(&cache->lock);
 310         return ctl;
 311 }
 312
 313 static void put_caching_control(struct btrfs_caching_control *ctl)
 314 {
 315         if (refcount_dec_and_test(&ctl->count))
 316                 kfree(ctl);
 317 }
 318
 319 #ifdef CONFIG_BTRFS_DEBUG
 320 static void fragment_free_space(struct btrfs_block_group_cache *block_group)
 321 {
 322         struct btrfs_fs_info *fs_info = block_group->fs_info;
 323         u64 start = block_group->key.objectid;
 324         u64 len = block_group->key.offset;
 325         u64 chunk = block_group->flags & BTRFS_BLOCK_GROUP_METADATA ?
 326                 fs_info->nodesize : fs_info->sectorsize;
 327         u64 step = chunk << 1;
 328
 329         while (len > chunk) {
 330                 btrfs_remove_free_space(block_group, start, chunk);
 331                 start += step;
 332                 if (len < step)
 333                         len = 0;
 334                 else
 335                         len -= step;
 336         }
 337 }
 338 #endif
 339
 340 /*
 341  * this is only called by cache_block_group, since we could have freed extents
 342  * we need to check the pinned_extents for any extents that can't be used yet
 343  * since their free space will be released as soon as the transaction commits.
 344  */
 345 u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 346                        struct btrfs_fs_info *info, u64 start, u64 end)
 347 {
 348         u64 extent_start, extent_end, size, total_added = 0;
 349         int ret;
 350
 351         while (start < end) {
 352                 ret = find_first_extent_bit(info->pinned_extents, start,
 353                                             &extent_start, &extent_end,
 354                                             EXTENT_DIRTY | EXTENT_UPTODATE,
 355                                             NULL);
 356                 if (ret)
 357                         break;
 358
 359                 if (extent_start <= start) {
 360                         start = extent_end + 1;
 361                 } else if (extent_start > start && extent_start < end) {
 362                         size = extent_start - start;
 363                         total_added += size;
 364                         ret = btrfs_add_free_space(block_group, start,
 365                                                    size);
 366                         BUG_ON(ret); /* -ENOMEM or logic error */
 367                         start = extent_end + 1;
 368                 } else {
 369                         break;
 370                 }
 371         }
 372
 373         if (start < end) {
 374                 size = end - start;
 375                 total_added += size;
 376                 ret = btrfs_add_free_space(block_group, start, size);
 377                 BUG_ON(ret); /* -ENOMEM or logic error */
 378         }
 379
 380         return total_added;
 381 }
 382
 383 static int load_extent_tree_free(struct btrfs_caching_control *caching_ctl)
 384 {
 385         struct btrfs_block_group_cache *block_group = caching_ctl->block_group;
 386         struct btrfs_fs_info *fs_info = block_group->fs_info;
 387         struct btrfs_root *extent_root = fs_info->extent_root;
 388         struct btrfs_path *path;
 389         struct extent_buffer *leaf;
 390         struct btrfs_key key;
 391         u64 total_found = 0;
 392         u64 last = 0;
 393         u32 nritems;
 394         int ret;
 395         bool wakeup = true;
 396
 397         path = btrfs_alloc_path();
 398         if (!path)
 399                 return -ENOMEM;
 400
 401         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 402
 403 #ifdef CONFIG_BTRFS_DEBUG
 404         /*
 405          * If we're fragmenting we don't want to make anybody think we can
 406          * allocate from this block group until we've had a chance to fragment
 407          * the free space.
 408          */
 409         if (btrfs_should_fragment_free_space(block_group))
 410                 wakeup = false;
 411 #endif
 412         /*
 413          * We don't want to deadlock with somebody trying to allocate a new
 414          * extent for the extent root while also trying to search the extent
 415          * root to add free space.  So we skip locking and search the commit
 416          * root, since its read-only
 417          */
 418         path->skip_locking = 1;
 419         path->search_commit_root = 1;
 420         path->reada = READA_FORWARD;
 421
 422         key.objectid = last;
 423         key.offset = 0;
 424         key.type = BTRFS_EXTENT_ITEM_KEY;
 425
 426 next:
 427         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 428         if (ret < 0)
 429                 goto out;
 430
 431         leaf = path->nodes[0];
 432         nritems = btrfs_header_nritems(leaf);
 433
 434         while (1) {
 435                 if (btrfs_fs_closing(fs_info) > 1) {
 436                         last = (u64)-1;
 437                         break;
 438                 }
 439
 440                 if (path->slots[0] < nritems) {
 441                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 442                 } else {
 443                         ret = find_next_key(path, 0, &key);
 444                         if (ret)
 445                                 break;
 446
 447                         if (need_resched() ||
 448                             rwsem_is_contended(&fs_info->commit_root_sem)) {
 449                                 if (wakeup)
 450                                         caching_ctl->progress = last;
 451                                 btrfs_release_path(path);
 452                                 up_read(&fs_info->commit_root_sem);
 453                                 mutex_unlock(&caching_ctl->mutex);
 454                                 cond_resched();
 455                                 mutex_lock(&caching_ctl->mutex);
 456                                 down_read(&fs_info->commit_root_sem);
 457                                 goto next;
 458                         }
 459
 460                         ret = btrfs_next_leaf(extent_root, path);
 461                         if (ret < 0)
 462                                 goto out;
 463                         if (ret)
 464                                 break;
 465                         leaf = path->nodes[0];
 466                         nritems = btrfs_header_nritems(leaf);
 467                         continue;
 468                 }
 469
 470                 if (key.objectid < last) {
 471                         key.objectid = last;
 472                         key.offset = 0;
 473                         key.type = BTRFS_EXTENT_ITEM_KEY;
 474
 475                         if (wakeup)
 476                                 caching_ctl->progress = last;
 477                         btrfs_release_path(path);
 478                         goto next;
 479                 }
 480
 481                 if (key.objectid < block_group->key.objectid) {
 482                         path->slots[0]++;
 483                         continue;
 484                 }
 485
 486                 if (key.objectid >= block_group->key.objectid +
 487                     block_group->key.offset)
 488                         break;
 489
 490                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 491                     key.type == BTRFS_METADATA_ITEM_KEY) {
 492                         total_found += add_new_free_space(block_group,
 493                                                           fs_info, last,
 494                                                           key.objectid);
 495                         if (key.type == BTRFS_METADATA_ITEM_KEY)
 496                                 last = key.objectid +
 497                                         fs_info->nodesize;
 498                         else
 499                                 last = key.objectid + key.offset;
 500
 501                         if (total_found > CACHING_CTL_WAKE_UP) {
 502                                 total_found = 0;
 503                                 if (wakeup)
 504                                         wake_up(&caching_ctl->wait);
 505                         }
 506                 }
 507                 path->slots[0]++;
 508         }
 509         ret = 0;
 510
 511         total_found += add_new_free_space(block_group, fs_info, last,
 512                                           block_group->key.objectid +
 513                                           block_group->key.offset);
 514         caching_ctl->progress = (u64)-1;
 515
 516 out:
 517         btrfs_free_path(path);
 518         return ret;
 519 }
 520
 521 static noinline void caching_thread(struct btrfs_work *work)
 522 {
 523         struct btrfs_block_group_cache *block_group;
 524         struct btrfs_fs_info *fs_info;
 525         struct btrfs_caching_control *caching_ctl;
 526         int ret;
 527
 528         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 529         block_group = caching_ctl->block_group;
 530         fs_info = block_group->fs_info;
 531
 532         mutex_lock(&caching_ctl->mutex);
 533         down_read(&fs_info->commit_root_sem);
 534
 535         if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
 536                 ret = load_free_space_tree(caching_ctl);
 537         else
 538                 ret = load_extent_tree_free(caching_ctl);
 539
 540         spin_lock(&block_group->lock);
 541         block_group->caching_ctl = NULL;
 542         block_group->cached = ret ? BTRFS_CACHE_ERROR : BTRFS_CACHE_FINISHED;
 543         spin_unlock(&block_group->lock);
 544
 545 #ifdef CONFIG_BTRFS_DEBUG
 546         if (btrfs_should_fragment_free_space(block_group)) {
 547                 u64 bytes_used;
 548
 549                 spin_lock(&block_group->space_info->lock);
 550                 spin_lock(&block_group->lock);
 551                 bytes_used = block_group->key.offset -
 552                         btrfs_block_group_used(&block_group->item);
 553                 block_group->space_info->bytes_used += bytes_used >> 1;
 554                 spin_unlock(&block_group->lock);
 555                 spin_unlock(&block_group->space_info->lock);
 556                 fragment_free_space(block_group);
 557         }
 558 #endif
 559
 560         caching_ctl->progress = (u64)-1;
 561
 562         up_read(&fs_info->commit_root_sem);
 563         free_excluded_extents(fs_info, block_group);
 564         mutex_unlock(&caching_ctl->mutex);
 565
 566         wake_up(&caching_ctl->wait);
 567
 568         put_caching_control(caching_ctl);
 569         btrfs_put_block_group(block_group);
 570 }
 571
 572 static int cache_block_group(struct btrfs_block_group_cache *cache,
 573                              int load_cache_only)
 574 {
 575         DEFINE_WAIT(wait);
 576         struct btrfs_fs_info *fs_info = cache->fs_info;
 577         struct btrfs_caching_control *caching_ctl;
 578         int ret = 0;
 579
 580         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 581         if (!caching_ctl)
 582                 return -ENOMEM;
 583
 584         INIT_LIST_HEAD(&caching_ctl->list);
 585         mutex_init(&caching_ctl->mutex);
 586         init_waitqueue_head(&caching_ctl->wait);
 587         caching_ctl->block_group = cache;
 588         caching_ctl->progress = cache->key.objectid;
 589         refcount_set(&caching_ctl->count, 1);
 590         btrfs_init_work(&caching_ctl->work, btrfs_cache_helper,
 591                         caching_thread, NULL, NULL);
 592
 593         spin_lock(&cache->lock);
 594         /*
 595          * This should be a rare occasion, but this could happen I think in the
 596          * case where one thread starts to load the space cache info, and then
 597          * some other thread starts a transaction commit which tries to do an
 598          * allocation while the other thread is still loading the space cache
 599          * info.  The previous loop should have kept us from choosing this block
 600          * group, but if we've moved to the state where we will wait on caching
 601          * block groups we need to first check if we're doing a fast load here,
 602          * so we can wait for it to finish, otherwise we could end up allocating
 603          * from a block group who's cache gets evicted for one reason or
 604          * another.
 605          */
 606         while (cache->cached == BTRFS_CACHE_FAST) {
 607                 struct btrfs_caching_control *ctl;
 608
 609                 ctl = cache->caching_ctl;
 610                 refcount_inc(&ctl->count);
 611                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 612                 spin_unlock(&cache->lock);
 613
 614                 schedule();
 615
 616                 finish_wait(&ctl->wait, &wait);
 617                 put_caching_control(ctl);
 618                 spin_lock(&cache->lock);
 619         }
 620
 621         if (cache->cached != BTRFS_CACHE_NO) {
 622                 spin_unlock(&cache->lock);
 623                 kfree(caching_ctl);
 624                 return 0;
 625         }
 626         WARN_ON(cache->caching_ctl);
 627         cache->caching_ctl = caching_ctl;
 628         cache->cached = BTRFS_CACHE_FAST;
 629         spin_unlock(&cache->lock);
 630
 631         if (btrfs_test_opt(fs_info, SPACE_CACHE)) {
 632                 mutex_lock(&caching_ctl->mutex);
 633                 ret = load_free_space_cache(fs_info, cache);
 634
 635                 spin_lock(&cache->lock);
 636                 if (ret == 1) {
 637                         cache->caching_ctl = NULL;
 638                         cache->cached = BTRFS_CACHE_FINISHED;
 639                         cache->last_byte_to_unpin = (u64)-1;
 640                         caching_ctl->progress = (u64)-1;
 641                 } else {
 642                         if (load_cache_only) {
 643                                 cache->caching_ctl = NULL;
 644                                 cache->cached = BTRFS_CACHE_NO;
 645                         } else {
 646                                 cache->cached = BTRFS_CACHE_STARTED;
 647                                 cache->has_caching_ctl = 1;
 648                         }
 649                 }
 650                 spin_unlock(&cache->lock);
 651 #ifdef CONFIG_BTRFS_DEBUG
 652                 if (ret == 1 &&
 653                     btrfs_should_fragment_free_space(cache)) {
 654                         u64 bytes_used;
 655
 656                         spin_lock(&cache->space_info->lock);
 657                         spin_lock(&cache->lock);
 658                         bytes_used = cache->key.offset -
 659                                 btrfs_block_group_used(&cache->item);
 660                         cache->space_info->bytes_used += bytes_used >> 1;
 661                         spin_unlock(&cache->lock);
 662                         spin_unlock(&cache->space_info->lock);
 663                         fragment_free_space(cache);
 664                 }
 665 #endif
 666                 mutex_unlock(&caching_ctl->mutex);
 667
 668                 wake_up(&caching_ctl->wait);
 669                 if (ret == 1) {
 670                         put_caching_control(caching_ctl);
 671                         free_excluded_extents(fs_info, cache);
 672                         return 0;
 673                 }
 674         } else {
 675                 /*
 676                  * We're either using the free space tree or no caching at all.
 677                  * Set cached to the appropriate value and wakeup any waiters.
 678                  */
 679                 spin_lock(&cache->lock);
 680                 if (load_cache_only) {
 681                         cache->caching_ctl = NULL;
 682                         cache->cached = BTRFS_CACHE_NO;
 683                 } else {
 684                         cache->cached = BTRFS_CACHE_STARTED;
 685                         cache->has_caching_ctl = 1;
 686                 }
 687                 spin_unlock(&cache->lock);
 688                 wake_up(&caching_ctl->wait);
 689         }
 690
 691         if (load_cache_only) {
 692                 put_caching_control(caching_ctl);
 693                 return 0;
 694         }
 695
 696         down_write(&fs_info->commit_root_sem);
 697         refcount_inc(&caching_ctl->count);
 698         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 699         up_write(&fs_info->commit_root_sem);
 700
 701         btrfs_get_block_group(cache);
 702
 703         btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
 704
 705         return ret;
 706 }
 707
 708 /*
 709  * return the block group that starts at or after bytenr
 710  */
 711 static struct btrfs_block_group_cache *
 712 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 713 {
 714         return block_group_cache_tree_search(info, bytenr, 0);
 715 }
 716
 717 /*
 718  * return the block group that contains the given bytenr
 719  */
 720 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 721                                                  struct btrfs_fs_info *info,
 722                                                  u64 bytenr)
 723 {
 724         return block_group_cache_tree_search(info, bytenr, 1);
 725 }
 726
 727 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 728                                                   u64 flags)
 729 {
 730         struct list_head *head = &info->space_info;
 731         struct btrfs_space_info *found;
 732
 733         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 734
 735         rcu_read_lock();
 736         list_for_each_entry_rcu(found, head, list) {
 737                 if (found->flags & flags) {
 738                         rcu_read_unlock();
 739                         return found;
 740                 }
 741         }
 742         rcu_read_unlock();
 743         return NULL;
 744 }
 745
 746 static void add_pinned_bytes(struct btrfs_fs_info *fs_info, s64 num_bytes,
 747                              u64 owner, u64 root_objectid)
 748 {
 749         struct btrfs_space_info *space_info;
 750         u64 flags;
 751
 752         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
 753                 if (root_objectid == BTRFS_CHUNK_TREE_OBJECTID)
 754                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
 755                 else
 756                         flags = BTRFS_BLOCK_GROUP_METADATA;
 757         } else {
 758                 flags = BTRFS_BLOCK_GROUP_DATA;
 759         }
 760
 761         space_info = __find_space_info(fs_info, flags);
 762         ASSERT(space_info);
 763         percpu_counter_add(&space_info->total_bytes_pinned, num_bytes);
 764 }
 765
 766 /*
 767  * after adding space to the filesystem, we need to clear the full flags
 768  * on all the space infos.
 769  */
 770 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 771 {
 772         struct list_head *head = &info->space_info;
 773         struct btrfs_space_info *found;
 774
 775         rcu_read_lock();
 776         list_for_each_entry_rcu(found, head, list)
 777                 found->full = 0;
 778         rcu_read_unlock();
 779 }
 780
 781 /* simple helper to search for an existing data extent at a given offset */
 782 int btrfs_lookup_data_extent(struct btrfs_fs_info *fs_info, u64 start, u64 len)
 783 {
 784         int ret;
 785         struct btrfs_key key;
 786         struct btrfs_path *path;
 787
 788         path = btrfs_alloc_path();
 789         if (!path)
 790                 return -ENOMEM;
 791
 792         key.objectid = start;
 793         key.offset = len;
 794         key.type = BTRFS_EXTENT_ITEM_KEY;
 795         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
 796         btrfs_free_path(path);
 797         return ret;
 798 }
 799
 800 /*
 801  * helper function to lookup reference count and flags of a tree block.
 802  *
 803  * the head node for delayed ref is used to store the sum of all the
 804  * reference count modifications queued up in the rbtree. the head
 805  * node may also store the extent flags to set. This way you can check
 806  * to see what the reference count and extent flags would be if all of
 807  * the delayed refs are not processed.
 808  */
 809 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 810                              struct btrfs_fs_info *fs_info, u64 bytenr,
 811                              u64 offset, int metadata, u64 *refs, u64 *flags)
 812 {
 813         struct btrfs_delayed_ref_head *head;
 814         struct btrfs_delayed_ref_root *delayed_refs;
 815         struct btrfs_path *path;
 816         struct btrfs_extent_item *ei;
 817         struct extent_buffer *leaf;
 818         struct btrfs_key key;
 819         u32 item_size;
 820         u64 num_refs;
 821         u64 extent_flags;
 822         int ret;
 823
 824         /*
 825          * If we don't have skinny metadata, don't bother doing anything
 826          * different
 827          */
 828         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA)) {
 829                 offset = fs_info->nodesize;
 830                 metadata = 0;
 831         }
 832
 833         path = btrfs_alloc_path();
 834         if (!path)
 835                 return -ENOMEM;
 836
 837         if (!trans) {
 838                 path->skip_locking = 1;
 839                 path->search_commit_root = 1;
 840         }
 841
 842 search_again:
 843         key.objectid = bytenr;
 844         key.offset = offset;
 845         if (metadata)
 846                 key.type = BTRFS_METADATA_ITEM_KEY;
 847         else
 848                 key.type = BTRFS_EXTENT_ITEM_KEY;
 849
 850         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 0);
 851         if (ret < 0)
 852                 goto out_free;
 853
 854         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 855                 if (path->slots[0]) {
 856                         path->slots[0]--;
 857                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 858                                               path->slots[0]);
 859                         if (key.objectid == bytenr &&
 860                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 861                             key.offset == fs_info->nodesize)
 862                                 ret = 0;
 863                 }
 864         }
 865
 866         if (ret == 0) {
 867                 leaf = path->nodes[0];
 868                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 869                 if (item_size >= sizeof(*ei)) {
 870                         ei = btrfs_item_ptr(leaf, path->slots[0],
 871                                             struct btrfs_extent_item);
 872                         num_refs = btrfs_extent_refs(leaf, ei);
 873                         extent_flags = btrfs_extent_flags(leaf, ei);
 874                 } else {
 875 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 876                         struct btrfs_extent_item_v0 *ei0;
 877                         BUG_ON(item_size != sizeof(*ei0));
 878                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 879                                              struct btrfs_extent_item_v0);
 880                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
 881                         /* FIXME: this isn't correct for data */
 882                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 883 #else
 884                         BUG();
 885 #endif
 886                 }
 887                 BUG_ON(num_refs == 0);
 888         } else {
 889                 num_refs = 0;
 890                 extent_flags = 0;
 891                 ret = 0;
 892         }
 893
 894         if (!trans)
 895                 goto out;
 896
 897         delayed_refs = &trans->transaction->delayed_refs;
 898         spin_lock(&delayed_refs->lock);
 899         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
 900         if (head) {
 901                 if (!mutex_trylock(&head->mutex)) {
 902                         refcount_inc(&head->refs);
 903                         spin_unlock(&delayed_refs->lock);
 904
 905                         btrfs_release_path(path);
 906
 907                         /*
 908                          * Mutex was contended, block until it's released and try
 909                          * again
 910                          */
 911                         mutex_lock(&head->mutex);
 912                         mutex_unlock(&head->mutex);
 913                         btrfs_put_delayed_ref_head(head);
 914                         goto search_again;
 915                 }
 916                 spin_lock(&head->lock);
 917                 if (head->extent_op && head->extent_op->update_flags)
 918                         extent_flags |= head->extent_op->flags_to_set;
 919                 else
 920                         BUG_ON(num_refs == 0);
 921
 922                 num_refs += head->ref_mod;
 923                 spin_unlock(&head->lock);
 924                 mutex_unlock(&head->mutex);
 925         }
 926         spin_unlock(&delayed_refs->lock);
 927 out:
 928         WARN_ON(num_refs == 0);
 929         if (refs)
 930                 *refs = num_refs;
 931         if (flags)
 932                 *flags = extent_flags;
 933 out_free:
 934         btrfs_free_path(path);
 935         return ret;
 936 }
 937
 938 /*
 939  * Back reference rules.  Back refs have three main goals:
 940  *
 941  * 1) differentiate between all holders of references to an extent so that
 942  *    when a reference is dropped we can make sure it was a valid reference
 943  *    before freeing the extent.
 944  *
 945  * 2) Provide enough information to quickly find the holders of an extent
 946  *    if we notice a given block is corrupted or bad.
 947  *
 948  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 949  *    maintenance.  This is actually the same as #2, but with a slightly
 950  *    different use case.
 951  *
 952  * There are two kinds of back refs. The implicit back refs is optimized
 953  * for pointers in non-shared tree blocks. For a given pointer in a block,
 954  * back refs of this kind provide information about the block's owner tree
 955  * and the pointer's key. These information allow us to find the block by
 956  * b-tree searching. The full back refs is for pointers in tree blocks not
 957  * referenced by their owner trees. The location of tree block is recorded
 958  * in the back refs. Actually the full back refs is generic, and can be
 959  * used in all cases the implicit back refs is used. The major shortcoming
 960  * of the full back refs is its overhead. Every time a tree block gets
 961  * COWed, we have to update back refs entry for all pointers in it.
 962  *
 963  * For a newly allocated tree block, we use implicit back refs for
 964  * pointers in it. This means most tree related operations only involve
 965  * implicit back refs. For a tree block created in old transaction, the
 966  * only way to drop a reference to it is COW it. So we can detect the
 967  * event that tree block loses its owner tree's reference and do the
 968  * back refs conversion.
 969  *
 970  * When a tree block is COWed through a tree, there are four cases:
 971  *
 972  * The reference count of the block is one and the tree is the block's
 973  * owner tree. Nothing to do in this case.
 974  *
 975  * The reference count of the block is one and the tree is not the
 976  * block's owner tree. In this case, full back refs is used for pointers
 977  * in the block. Remove these full back refs, add implicit back refs for
 978  * every pointers in the new block.
 979  *
 980  * The reference count of the block is greater than one and the tree is
 981  * the block's owner tree. In this case, implicit back refs is used for
 982  * pointers in the block. Add full back refs for every pointers in the
 983  * block, increase lower level extents' reference counts. The original
 984  * implicit back refs are entailed to the new block.
 985  *
 986  * The reference count of the block is greater than one and the tree is
 987  * not the block's owner tree. Add implicit back refs for every pointer in
 988  * the new block, increase lower level extents' reference count.
 989  *
 990  * Back Reference Key composing:
 991  *
 992  * The key objectid corresponds to the first byte in the extent,
 993  * The key type is used to differentiate between types of back refs.
 994  * There are different meanings of the key offset for different types
 995  * of back refs.
 996  *
 997  * File extents can be referenced by:
 998  *
 999  * - multiple snapshots, subvolumes, or different generations in one subvol
1000  * - different files inside a single subvolume
1001  * - different offsets inside a file (bookend extents in file.c)
1002  *
1003  * The extent ref structure for the implicit back refs has fields for:
1004  *
1005  * - Objectid of the subvolume root
1006  * - objectid of the file holding the reference
1007  * - original offset in the file
1008  * - how many bookend extents
1009  *
1010  * The key offset for the implicit back refs is hash of the first
1011  * three fields.
1012  *
1013  * The extent ref structure for the full back refs has field for:
1014  *
1015  * - number of pointers in the tree leaf
1016  *
1017  * The key offset for the implicit back refs is the first byte of
1018  * the tree leaf
1019  *
1020  * When a file extent is allocated, The implicit back refs is used.
1021  * the fields are filled in:
1022  *
1023  *     (root_key.objectid, inode objectid, offset in file, 1)
1024  *
1025  * When a file extent is removed file truncation, we find the
1026  * corresponding implicit back refs and check the following fields:
1027  *
1028  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
1029  *
1030  * Btree extents can be referenced by:
1031  *
1032  * - Different subvolumes
1033  *
1034  * Both the implicit back refs and the full back refs for tree blocks
1035  * only consist of key. The key offset for the implicit back refs is
1036  * objectid of block's owner tree. The key offset for the full back refs
1037  * is the first byte of parent block.
1038  *
1039  * When implicit back refs is used, information about the lowest key and
1040  * level of the tree block are required. These information are stored in
1041  * tree block info structure.
1042  */
1043
1044 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1045 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
1046                                   struct btrfs_fs_info *fs_info,
1047                                   struct btrfs_path *path,
1048                                   u64 owner, u32 extra_size)
1049 {
1050         struct btrfs_root *root = fs_info->extent_root;
1051         struct btrfs_extent_item *item;
1052         struct btrfs_extent_item_v0 *ei0;
1053         struct btrfs_extent_ref_v0 *ref0;
1054         struct btrfs_tree_block_info *bi;
1055         struct extent_buffer *leaf;
1056         struct btrfs_key key;
1057         struct btrfs_key found_key;
1058         u32 new_size = sizeof(*item);
1059         u64 refs;
1060         int ret;
1061
1062         leaf = path->nodes[0];
1063         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
1064
1065         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1066         ei0 = btrfs_item_ptr(leaf, path->slots[0],
1067                              struct btrfs_extent_item_v0);
1068         refs = btrfs_extent_refs_v0(leaf, ei0);
1069
1070         if (owner == (u64)-1) {
1071                 while (1) {
1072                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1073                                 ret = btrfs_next_leaf(root, path);
1074                                 if (ret < 0)
1075                                         return ret;
1076                                 BUG_ON(ret > 0); /* Corruption */
1077                                 leaf = path->nodes[0];
1078                         }
1079                         btrfs_item_key_to_cpu(leaf, &found_key,
1080                                               path->slots[0]);
1081                         BUG_ON(key.objectid != found_key.objectid);
1082                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
1083                                 path->slots[0]++;
1084                                 continue;
1085                         }
1086                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1087                                               struct btrfs_extent_ref_v0);
1088                         owner = btrfs_ref_objectid_v0(leaf, ref0);
1089                         break;
1090                 }
1091         }
1092         btrfs_release_path(path);
1093
1094         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1095                 new_size += sizeof(*bi);
1096
1097         new_size -= sizeof(*ei0);
1098         ret = btrfs_search_slot(trans, root, &key, path,
1099                                 new_size + extra_size, 1);
1100         if (ret < 0)
1101                 return ret;
1102         BUG_ON(ret); /* Corruption */
1103
1104         btrfs_extend_item(fs_info, path, new_size);
1105
1106         leaf = path->nodes[0];
1107         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1108         btrfs_set_extent_refs(leaf, item, refs);
1109         /* FIXME: get real generation */
1110         btrfs_set_extent_generation(leaf, item, 0);
1111         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1112                 btrfs_set_extent_flags(leaf, item,
1113                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1114                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1115                 bi = (struct btrfs_tree_block_info *)(item + 1);
1116                 /* FIXME: get first key of the block */
1117                 memzero_extent_buffer(leaf, (unsigned long)bi, sizeof(*bi));
1118                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1119         } else {
1120                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1121         }
1122         btrfs_mark_buffer_dirty(leaf);
1123         return 0;
1124 }
1125 #endif
1126
1127 /*
1128  * is_data == BTRFS_REF_TYPE_BLOCK, tree block type is required,
1129  * is_data == BTRFS_REF_TYPE_DATA, data type is requried,
1130  * is_data == BTRFS_REF_TYPE_ANY, either type is OK.
1131  */
1132 int btrfs_get_extent_inline_ref_type(const struct extent_buffer *eb,
1133                                      struct btrfs_extent_inline_ref *iref,
1134                                      enum btrfs_inline_ref_type is_data)
1135 {
1136         int type = btrfs_extent_inline_ref_type(eb, iref);
1137         u64 offset = btrfs_extent_inline_ref_offset(eb, iref);
1138
1139         if (type == BTRFS_TREE_BLOCK_REF_KEY ||
1140             type == BTRFS_SHARED_BLOCK_REF_KEY ||
1141             type == BTRFS_SHARED_DATA_REF_KEY ||
1142             type == BTRFS_EXTENT_DATA_REF_KEY) {
1143                 if (is_data == BTRFS_REF_TYPE_BLOCK) {
1144                         if (type == BTRFS_TREE_BLOCK_REF_KEY)
1145                                 return type;
1146                         if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1147                                 ASSERT(eb->fs_info);
1148                                 /*
1149                                  * Every shared one has parent tree
1150                                  * block, which must be aligned to
1151                                  * nodesize.
1152                                  */
1153                                 if (offset &&
1154                                     IS_ALIGNED(offset, eb->fs_info->nodesize))
1155                                         return type;
1156                         }
1157                 } else if (is_data == BTRFS_REF_TYPE_DATA) {
1158                         if (type == BTRFS_EXTENT_DATA_REF_KEY)
1159                                 return type;
1160                         if (type == BTRFS_SHARED_DATA_REF_KEY) {
1161                                 ASSERT(eb->fs_info);
1162                                 /*
1163                                  * Every shared one has parent tree
1164                                  * block, which must be aligned to
1165                                  * nodesize.
1166                                  */
1167                                 if (offset &&
1168                                     IS_ALIGNED(offset, eb->fs_info->nodesize))
1169                                         return type;
1170                         }
1171                 } else {
1172                         ASSERT(is_data == BTRFS_REF_TYPE_ANY);
1173                         return type;
1174                 }
1175         }
1176
1177         btrfs_print_leaf((struct extent_buffer *)eb);
1178         btrfs_err(eb->fs_info, "eb %llu invalid extent inline ref type %d",
1179                   eb->start, type);
1180         WARN_ON(1);
1181
1182         return BTRFS_REF_TYPE_INVALID;
1183 }
1184
1185 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1186 {
1187         u32 high_crc = ~(u32)0;
1188         u32 low_crc = ~(u32)0;
1189         __le64 lenum;
1190
1191         lenum = cpu_to_le64(root_objectid);
1192         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
1193         lenum = cpu_to_le64(owner);
1194         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1195         lenum = cpu_to_le64(offset);
1196         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1197
1198         return ((u64)high_crc << 31) ^ (u64)low_crc;
1199 }
1200
1201 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1202                                      struct btrfs_extent_data_ref *ref)
1203 {
1204         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1205                                     btrfs_extent_data_ref_objectid(leaf, ref),
1206                                     btrfs_extent_data_ref_offset(leaf, ref));
1207 }
1208
1209 static int match_extent_data_ref(struct extent_buffer *leaf,
1210                                  struct btrfs_extent_data_ref *ref,
1211                                  u64 root_objectid, u64 owner, u64 offset)
1212 {
1213         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1214             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1215             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1216                 return 0;
1217         return 1;
1218 }
1219
1220 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1221                                            struct btrfs_fs_info *fs_info,
1222                                            struct btrfs_path *path,
1223                                            u64 bytenr, u64 parent,
1224                                            u64 root_objectid,
1225                                            u64 owner, u64 offset)
1226 {
1227         struct btrfs_root *root = fs_info->extent_root;
1228         struct btrfs_key key;
1229         struct btrfs_extent_data_ref *ref;
1230         struct extent_buffer *leaf;
1231         u32 nritems;
1232         int ret;
1233         int recow;
1234         int err = -ENOENT;
1235
1236         key.objectid = bytenr;
1237         if (parent) {
1238                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1239                 key.offset = parent;
1240         } else {
1241                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1242                 key.offset = hash_extent_data_ref(root_objectid,
1243                                                   owner, offset);
1244         }
1245 again:
1246         recow = 0;
1247         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1248         if (ret < 0) {
1249                 err = ret;
1250                 goto fail;
1251         }
1252
1253         if (parent) {
1254                 if (!ret)
1255                         return 0;
1256 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1257                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1258                 btrfs_release_path(path);
1259                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1260                 if (ret < 0) {
1261                         err = ret;
1262                         goto fail;
1263                 }
1264                 if (!ret)
1265                         return 0;
1266 #endif
1267                 goto fail;
1268         }
1269
1270         leaf = path->nodes[0];
1271         nritems = btrfs_header_nritems(leaf);
1272         while (1) {
1273                 if (path->slots[0] >= nritems) {
1274                         ret = btrfs_next_leaf(root, path);
1275                         if (ret < 0)
1276                                 err = ret;
1277                         if (ret)
1278                                 goto fail;
1279
1280                         leaf = path->nodes[0];
1281                         nritems = btrfs_header_nritems(leaf);
1282                         recow = 1;
1283                 }
1284
1285                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1286                 if (key.objectid != bytenr ||
1287                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1288                         goto fail;
1289
1290                 ref = btrfs_item_ptr(leaf, path->slots[0],
1291                                      struct btrfs_extent_data_ref);
1292
1293                 if (match_extent_data_ref(leaf, ref, root_objectid,
1294                                           owner, offset)) {
1295                         if (recow) {
1296                                 btrfs_release_path(path);
1297                                 goto again;
1298                         }
1299                         err = 0;
1300                         break;
1301                 }
1302                 path->slots[0]++;
1303         }
1304 fail:
1305         return err;
1306 }
1307
1308 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1309                                            struct btrfs_fs_info *fs_info,
1310                                            struct btrfs_path *path,
1311                                            u64 bytenr, u64 parent,
1312                                            u64 root_objectid, u64 owner,
1313                                            u64 offset, int refs_to_add)
1314 {
1315         struct btrfs_root *root = fs_info->extent_root;
1316         struct btrfs_key key;
1317         struct extent_buffer *leaf;
1318         u32 size;
1319         u32 num_refs;
1320         int ret;
1321
1322         key.objectid = bytenr;
1323         if (parent) {
1324                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1325                 key.offset = parent;
1326                 size = sizeof(struct btrfs_shared_data_ref);
1327         } else {
1328                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1329                 key.offset = hash_extent_data_ref(root_objectid,
1330                                                   owner, offset);
1331                 size = sizeof(struct btrfs_extent_data_ref);
1332         }
1333
1334         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1335         if (ret && ret != -EEXIST)
1336                 goto fail;
1337
1338         leaf = path->nodes[0];
1339         if (parent) {
1340                 struct btrfs_shared_data_ref *ref;
1341                 ref = btrfs_item_ptr(leaf, path->slots[0],
1342                                      struct btrfs_shared_data_ref);
1343                 if (ret == 0) {
1344                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1345                 } else {
1346                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1347                         num_refs += refs_to_add;
1348                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1349                 }
1350         } else {
1351                 struct btrfs_extent_data_ref *ref;
1352                 while (ret == -EEXIST) {
1353                         ref = btrfs_item_ptr(leaf, path->slots[0],
1354                                              struct btrfs_extent_data_ref);
1355                         if (match_extent_data_ref(leaf, ref, root_objectid,
1356                                                   owner, offset))
1357                                 break;
1358                         btrfs_release_path(path);
1359                         key.offset++;
1360                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1361                                                       size);
1362                         if (ret && ret != -EEXIST)
1363                                 goto fail;
1364
1365                         leaf = path->nodes[0];
1366                 }
1367                 ref = btrfs_item_ptr(leaf, path->slots[0],
1368                                      struct btrfs_extent_data_ref);
1369                 if (ret == 0) {
1370                         btrfs_set_extent_data_ref_root(leaf, ref,
1371                                                        root_objectid);
1372                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1373                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1374                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1375                 } else {
1376                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1377                         num_refs += refs_to_add;
1378                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1379                 }
1380         }
1381         btrfs_mark_buffer_dirty(leaf);
1382         ret = 0;
1383 fail:
1384         btrfs_release_path(path);
1385         return ret;
1386 }
1387
1388 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1389                                            struct btrfs_fs_info *fs_info,
1390                                            struct btrfs_path *path,
1391                                            int refs_to_drop, int *last_ref)
1392 {
1393         struct btrfs_key key;
1394         struct btrfs_extent_data_ref *ref1 = NULL;
1395         struct btrfs_shared_data_ref *ref2 = NULL;
1396         struct extent_buffer *leaf;
1397         u32 num_refs = 0;
1398         int ret = 0;
1399
1400         leaf = path->nodes[0];
1401         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1402
1403         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1404                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1405                                       struct btrfs_extent_data_ref);
1406                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1407         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1408                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1409                                       struct btrfs_shared_data_ref);
1410                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1411 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1412         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1413                 struct btrfs_extent_ref_v0 *ref0;
1414                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1415                                       struct btrfs_extent_ref_v0);
1416                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1417 #endif
1418         } else {
1419                 BUG();
1420         }
1421
1422         BUG_ON(num_refs < refs_to_drop);
1423         num_refs -= refs_to_drop;
1424
1425         if (num_refs == 0) {
1426                 ret = btrfs_del_item(trans, fs_info->extent_root, path);
1427                 *last_ref = 1;
1428         } else {
1429                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1430                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1431                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1432                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1433 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1434                 else {
1435                         struct btrfs_extent_ref_v0 *ref0;
1436                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1437                                         struct btrfs_extent_ref_v0);
1438                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1439                 }
1440 #endif
1441                 btrfs_mark_buffer_dirty(leaf);
1442         }
1443         return ret;
1444 }
1445
1446 static noinline u32 extent_data_ref_count(struct btrfs_path *path,
1447                                           struct btrfs_extent_inline_ref *iref)
1448 {
1449         struct btrfs_key key;
1450         struct extent_buffer *leaf;
1451         struct btrfs_extent_data_ref *ref1;
1452         struct btrfs_shared_data_ref *ref2;
1453         u32 num_refs = 0;
1454         int type;
1455
1456         leaf = path->nodes[0];
1457         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1458         if (iref) {
1459                 /*
1460                  * If type is invalid, we should have bailed out earlier than
1461                  * this call.
1462                  */
1463                 type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
1464                 ASSERT(type != BTRFS_REF_TYPE_INVALID);
1465                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1466                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1467                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1468                 } else {
1469                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1470                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1471                 }
1472         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1473                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1474                                       struct btrfs_extent_data_ref);
1475                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1476         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1477                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1478                                       struct btrfs_shared_data_ref);
1479                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1480 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1481         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1482                 struct btrfs_extent_ref_v0 *ref0;
1483                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1484                                       struct btrfs_extent_ref_v0);
1485                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1486 #endif
1487         } else {
1488                 WARN_ON(1);
1489         }
1490         return num_refs;
1491 }
1492
1493 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1494                                           struct btrfs_fs_info *fs_info,
1495                                           struct btrfs_path *path,
1496                                           u64 bytenr, u64 parent,
1497                                           u64 root_objectid)
1498 {
1499         struct btrfs_root *root = fs_info->extent_root;
1500         struct btrfs_key key;
1501         int ret;
1502
1503         key.objectid = bytenr;
1504         if (parent) {
1505                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1506                 key.offset = parent;
1507         } else {
1508                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1509                 key.offset = root_objectid;
1510         }
1511
1512         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1513         if (ret > 0)
1514                 ret = -ENOENT;
1515 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1516         if (ret == -ENOENT && parent) {
1517                 btrfs_release_path(path);
1518                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1519                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1520                 if (ret > 0)
1521                         ret = -ENOENT;
1522         }
1523 #endif
1524         return ret;
1525 }
1526
1527 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1528                                           struct btrfs_fs_info *fs_info,
1529                                           struct btrfs_path *path,
1530                                           u64 bytenr, u64 parent,
1531                                           u64 root_objectid)
1532 {
1533         struct btrfs_key key;
1534         int ret;
1535
1536         key.objectid = bytenr;
1537         if (parent) {
1538                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1539                 key.offset = parent;
1540         } else {
1541                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1542                 key.offset = root_objectid;
1543         }
1544
1545         ret = btrfs_insert_empty_item(trans, fs_info->extent_root,
1546                                       path, &key, 0);
1547         btrfs_release_path(path);
1548         return ret;
1549 }
1550
1551 static inline int extent_ref_type(u64 parent, u64 owner)
1552 {
1553         int type;
1554         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1555                 if (parent > 0)
1556                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1557                 else
1558                         type = BTRFS_TREE_BLOCK_REF_KEY;
1559         } else {
1560                 if (parent > 0)
1561                         type = BTRFS_SHARED_DATA_REF_KEY;
1562                 else
1563                         type = BTRFS_EXTENT_DATA_REF_KEY;
1564         }
1565         return type;
1566 }
1567
1568 static int find_next_key(struct btrfs_path *path, int level,
1569                          struct btrfs_key *key)
1570
1571 {
1572         for (; level < BTRFS_MAX_LEVEL; level++) {
1573                 if (!path->nodes[level])
1574                         break;
1575                 if (path->slots[level] + 1 >=
1576                     btrfs_header_nritems(path->nodes[level]))
1577                         continue;
1578                 if (level == 0)
1579                         btrfs_item_key_to_cpu(path->nodes[level], key,
1580                                               path->slots[level] + 1);
1581                 else
1582                         btrfs_node_key_to_cpu(path->nodes[level], key,
1583                                               path->slots[level] + 1);
1584                 return 0;
1585         }
1586         return 1;
1587 }
1588
1589 /*
1590  * look for inline back ref. if back ref is found, *ref_ret is set
1591  * to the address of inline back ref, and 0 is returned.
1592  *
1593  * if back ref isn't found, *ref_ret is set to the address where it
1594  * should be inserted, and -ENOENT is returned.
1595  *
1596  * if insert is true and there are too many inline back refs, the path
1597  * points to the extent item, and -EAGAIN is returned.
1598  *
1599  * NOTE: inline back refs are ordered in the same way that back ref
1600  *       items in the tree are ordered.
1601  */
1602 static noinline_for_stack
1603 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1604                                  struct btrfs_fs_info *fs_info,
1605                                  struct btrfs_path *path,
1606                                  struct btrfs_extent_inline_ref **ref_ret,
1607                                  u64 bytenr, u64 num_bytes,
1608                                  u64 parent, u64 root_objectid,
1609                                  u64 owner, u64 offset, int insert)
1610 {
1611         struct btrfs_root *root = fs_info->extent_root;
1612         struct btrfs_key key;
1613         struct extent_buffer *leaf;
1614         struct btrfs_extent_item *ei;
1615         struct btrfs_extent_inline_ref *iref;
1616         u64 flags;
1617         u64 item_size;
1618         unsigned long ptr;
1619         unsigned long end;
1620         int extra_size;
1621         int type;
1622         int want;
1623         int ret;
1624         int err = 0;
1625         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
1626         int needed;
1627
1628         key.objectid = bytenr;
1629         key.type = BTRFS_EXTENT_ITEM_KEY;
1630         key.offset = num_bytes;
1631
1632         want = extent_ref_type(parent, owner);
1633         if (insert) {
1634                 extra_size = btrfs_extent_inline_ref_size(want);
1635                 path->keep_locks = 1;
1636         } else
1637                 extra_size = -1;
1638
1639         /*
1640          * Owner is our parent level, so we can just add one to get the level
1641          * for the block we are interested in.
1642          */
1643         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
1644                 key.type = BTRFS_METADATA_ITEM_KEY;
1645                 key.offset = owner;
1646         }
1647
1648 again:
1649         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1650         if (ret < 0) {
1651                 err = ret;
1652                 goto out;
1653         }
1654
1655         /*
1656          * We may be a newly converted file system which still has the old fat
1657          * extent entries for metadata, so try and see if we have one of those.
1658          */
1659         if (ret > 0 && skinny_metadata) {
1660                 skinny_metadata = false;
1661                 if (path->slots[0]) {
1662                         path->slots[0]--;
1663                         btrfs_item_key_to_cpu(path->nodes[0], &key,
1664                                               path->slots[0]);
1665                         if (key.objectid == bytenr &&
1666                             key.type == BTRFS_EXTENT_ITEM_KEY &&
1667                             key.offset == num_bytes)
1668                                 ret = 0;
1669                 }
1670                 if (ret) {
1671                         key.objectid = bytenr;
1672                         key.type = BTRFS_EXTENT_ITEM_KEY;
1673                         key.offset = num_bytes;
1674                         btrfs_release_path(path);
1675                         goto again;
1676                 }
1677         }
1678
1679         if (ret && !insert) {
1680                 err = -ENOENT;
1681                 goto out;
1682         } else if (WARN_ON(ret)) {
1683                 err = -EIO;
1684                 goto out;
1685         }
1686
1687         leaf = path->nodes[0];
1688         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1689 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1690         if (item_size < sizeof(*ei)) {
1691                 if (!insert) {
1692                         err = -ENOENT;
1693                         goto out;
1694                 }
1695                 ret = convert_extent_item_v0(trans, fs_info, path, owner,
1696                                              extra_size);
1697                 if (ret < 0) {
1698                         err = ret;
1699                         goto out;
1700                 }
1701                 leaf = path->nodes[0];
1702                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1703         }
1704 #endif
1705         BUG_ON(item_size < sizeof(*ei));
1706
1707         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1708         flags = btrfs_extent_flags(leaf, ei);
1709
1710         ptr = (unsigned long)(ei + 1);
1711         end = (unsigned long)ei + item_size;
1712
1713         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
1714                 ptr += sizeof(struct btrfs_tree_block_info);
1715                 BUG_ON(ptr > end);
1716         }
1717
1718         if (owner >= BTRFS_FIRST_FREE_OBJECTID)
1719                 needed = BTRFS_REF_TYPE_DATA;
1720         else
1721                 needed = BTRFS_REF_TYPE_BLOCK;
1722
1723         err = -ENOENT;
1724         while (1) {
1725                 if (ptr >= end) {
1726                         WARN_ON(ptr > end);
1727                         break;
1728                 }
1729                 iref = (struct btrfs_extent_inline_ref *)ptr;
1730                 type = btrfs_get_extent_inline_ref_type(leaf, iref, needed);
1731                 if (type == BTRFS_REF_TYPE_INVALID) {
1732                         err = -EINVAL;
1733                         goto out;
1734                 }
1735
1736                 if (want < type)
1737                         break;
1738                 if (want > type) {
1739                         ptr += btrfs_extent_inline_ref_size(type);
1740                         continue;
1741                 }
1742
1743                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1744                         struct btrfs_extent_data_ref *dref;
1745                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1746                         if (match_extent_data_ref(leaf, dref, root_objectid,
1747                                                   owner, offset)) {
1748                                 err = 0;
1749                                 break;
1750                         }
1751                         if (hash_extent_data_ref_item(leaf, dref) <
1752                             hash_extent_data_ref(root_objectid, owner, offset))
1753                                 break;
1754                 } else {
1755                         u64 ref_offset;
1756                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1757                         if (parent > 0) {
1758                                 if (parent == ref_offset) {
1759                                         err = 0;
1760                                         break;
1761                                 }
1762                                 if (ref_offset < parent)
1763                                         break;
1764                         } else {
1765                                 if (root_objectid == ref_offset) {
1766                                         err = 0;
1767                                         break;
1768                                 }
1769                                 if (ref_offset < root_objectid)
1770                                         break;
1771                         }
1772                 }
1773                 ptr += btrfs_extent_inline_ref_size(type);
1774         }
1775         if (err == -ENOENT && insert) {
1776                 if (item_size + extra_size >=
1777                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1778                         err = -EAGAIN;
1779                         goto out;
1780                 }
1781                 /*
1782                  * To add new inline back ref, we have to make sure
1783                  * there is no corresponding back ref item.
1784                  * For simplicity, we just do not add new inline back
1785                  * ref if there is any kind of item for this block
1786                  */
1787                 if (find_next_key(path, 0, &key) == 0 &&
1788                     key.objectid == bytenr &&
1789                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1790                         err = -EAGAIN;
1791                         goto out;
1792                 }
1793         }
1794         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1795 out:
1796         if (insert) {
1797                 path->keep_locks = 0;
1798                 btrfs_unlock_up_safe(path, 1);
1799         }
1800         return err;
1801 }
1802
1803 /*
1804  * helper to add new inline back ref
1805  */
1806 static noinline_for_stack
1807 void setup_inline_extent_backref(struct btrfs_fs_info *fs_info,
1808                                  struct btrfs_path *path,
1809                                  struct btrfs_extent_inline_ref *iref,
1810                                  u64 parent, u64 root_objectid,
1811                                  u64 owner, u64 offset, int refs_to_add,
1812                                  struct btrfs_delayed_extent_op *extent_op)
1813 {
1814         struct extent_buffer *leaf;
1815         struct btrfs_extent_item *ei;
1816         unsigned long ptr;
1817         unsigned long end;
1818         unsigned long item_offset;
1819         u64 refs;
1820         int size;
1821         int type;
1822
1823         leaf = path->nodes[0];
1824         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1825         item_offset = (unsigned long)iref - (unsigned long)ei;
1826
1827         type = extent_ref_type(parent, owner);
1828         size = btrfs_extent_inline_ref_size(type);
1829
1830         btrfs_extend_item(fs_info, path, size);
1831
1832         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1833         refs = btrfs_extent_refs(leaf, ei);
1834         refs += refs_to_add;
1835         btrfs_set_extent_refs(leaf, ei, refs);
1836         if (extent_op)
1837                 __run_delayed_extent_op(extent_op, leaf, ei);
1838
1839         ptr = (unsigned long)ei + item_offset;
1840         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1841         if (ptr < end - size)
1842                 memmove_extent_buffer(leaf, ptr + size, ptr,
1843                                       end - size - ptr);
1844
1845         iref = (struct btrfs_extent_inline_ref *)ptr;
1846         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1847         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1848                 struct btrfs_extent_data_ref *dref;
1849                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1850                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1851                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1852                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1853                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1854         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1855                 struct btrfs_shared_data_ref *sref;
1856                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1857                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1858                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1859         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1860                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1861         } else {
1862                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1863         }
1864         btrfs_mark_buffer_dirty(leaf);
1865 }
1866
1867 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1868                                  struct btrfs_fs_info *fs_info,
1869                                  struct btrfs_path *path,
1870                                  struct btrfs_extent_inline_ref **ref_ret,
1871                                  u64 bytenr, u64 num_bytes, u64 parent,
1872                                  u64 root_objectid, u64 owner, u64 offset)
1873 {
1874         int ret;
1875
1876         ret = lookup_inline_extent_backref(trans, fs_info, path, ref_ret,
1877                                            bytenr, num_bytes, parent,
1878                                            root_objectid, owner, offset, 0);
1879         if (ret != -ENOENT)
1880                 return ret;
1881
1882         btrfs_release_path(path);
1883         *ref_ret = NULL;
1884
1885         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1886                 ret = lookup_tree_block_ref(trans, fs_info, path, bytenr,
1887                                             parent, root_objectid);
1888         } else {
1889                 ret = lookup_extent_data_ref(trans, fs_info, path, bytenr,
1890                                              parent, root_objectid, owner,
1891                                              offset);
1892         }
1893         return ret;
1894 }
1895
1896 /*
1897  * helper to update/remove inline back ref
1898  */
1899 static noinline_for_stack
1900 void update_inline_extent_backref(struct btrfs_fs_info *fs_info,
1901                                   struct btrfs_path *path,
1902                                   struct btrfs_extent_inline_ref *iref,
1903                                   int refs_to_mod,
1904                                   struct btrfs_delayed_extent_op *extent_op,
1905                                   int *last_ref)
1906 {
1907         struct extent_buffer *leaf;
1908         struct btrfs_extent_item *ei;
1909         struct btrfs_extent_data_ref *dref = NULL;
1910         struct btrfs_shared_data_ref *sref = NULL;
1911         unsigned long ptr;
1912         unsigned long end;
1913         u32 item_size;
1914         int size;
1915         int type;
1916         u64 refs;
1917
1918         leaf = path->nodes[0];
1919         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1920         refs = btrfs_extent_refs(leaf, ei);
1921         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1922         refs += refs_to_mod;
1923         btrfs_set_extent_refs(leaf, ei, refs);
1924         if (extent_op)
1925                 __run_delayed_extent_op(extent_op, leaf, ei);
1926
1927         /*
1928          * If type is invalid, we should have bailed out after
1929          * lookup_inline_extent_backref().
1930          */
1931         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_ANY);
1932         ASSERT(type != BTRFS_REF_TYPE_INVALID);
1933
1934         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1935                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1936                 refs = btrfs_extent_data_ref_count(leaf, dref);
1937         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1938                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1939                 refs = btrfs_shared_data_ref_count(leaf, sref);
1940         } else {
1941                 refs = 1;
1942                 BUG_ON(refs_to_mod != -1);
1943         }
1944
1945         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1946         refs += refs_to_mod;
1947
1948         if (refs > 0) {
1949                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1950                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1951                 else
1952                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1953         } else {
1954                 *last_ref = 1;
1955                 size =  btrfs_extent_inline_ref_size(type);
1956                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1957                 ptr = (unsigned long)iref;
1958                 end = (unsigned long)ei + item_size;
1959                 if (ptr + size < end)
1960                         memmove_extent_buffer(leaf, ptr, ptr + size,
1961                                               end - ptr - size);
1962                 item_size -= size;
1963                 btrfs_truncate_item(fs_info, path, item_size, 1);
1964         }
1965         btrfs_mark_buffer_dirty(leaf);
1966 }
1967
1968 static noinline_for_stack
1969 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1970                                  struct btrfs_fs_info *fs_info,
1971                                  struct btrfs_path *path,
1972                                  u64 bytenr, u64 num_bytes, u64 parent,
1973                                  u64 root_objectid, u64 owner,
1974                                  u64 offset, int refs_to_add,
1975                                  struct btrfs_delayed_extent_op *extent_op)
1976 {
1977         struct btrfs_extent_inline_ref *iref;
1978         int ret;
1979
1980         ret = lookup_inline_extent_backref(trans, fs_info, path, &iref,
1981                                            bytenr, num_bytes, parent,
1982                                            root_objectid, owner, offset, 1);
1983         if (ret == 0) {
1984                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1985                 update_inline_extent_backref(fs_info, path, iref,
1986                                              refs_to_add, extent_op, NULL);
1987         } else if (ret == -ENOENT) {
1988                 setup_inline_extent_backref(fs_info, path, iref, parent,
1989                                             root_objectid, owner, offset,
1990                                             refs_to_add, extent_op);
1991                 ret = 0;
1992         }
1993         return ret;
1994 }
1995
1996 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1997                                  struct btrfs_fs_info *fs_info,
1998                                  struct btrfs_path *path,
1999                                  u64 bytenr, u64 parent, u64 root_objectid,
2000                                  u64 owner, u64 offset, int refs_to_add)
2001 {
2002         int ret;
2003         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
2004                 BUG_ON(refs_to_add != 1);
2005                 ret = insert_tree_block_ref(trans, fs_info, path, bytenr,
2006                                             parent, root_objectid);
2007         } else {
2008                 ret = insert_extent_data_ref(trans, fs_info, path, bytenr,
2009                                              parent, root_objectid,
2010                                              owner, offset, refs_to_add);
2011         }
2012         return ret;
2013 }
2014
2015 static int remove_extent_backref(struct btrfs_trans_handle *trans,
2016                                  struct btrfs_fs_info *fs_info,
2017                                  struct btrfs_path *path,
2018                                  struct btrfs_extent_inline_ref *iref,
2019                                  int refs_to_drop, int is_data, int *last_ref)
2020 {
2021         int ret = 0;
2022
2023         BUG_ON(!is_data && refs_to_drop != 1);
2024         if (iref) {
2025                 update_inline_extent_backref(fs_info, path, iref,
2026                                              -refs_to_drop, NULL, last_ref);
2027         } else if (is_data) {
2028                 ret = remove_extent_data_ref(trans, fs_info, path, refs_to_drop,
2029                                              last_ref);
2030         } else {
2031                 *last_ref = 1;
2032                 ret = btrfs_del_item(trans, fs_info->extent_root, path);
2033         }
2034         return ret;
2035 }
2036
2037 #define in_range(b, first, len)        ((b) >= (first) && (b) < (first) + (len))
2038 static int btrfs_issue_discard(struct block_device *bdev, u64 start, u64 len,
2039                                u64 *discarded_bytes)
2040 {
2041         int j, ret = 0;
2042         u64 bytes_left, end;
2043         u64 aligned_start = ALIGN(start, 1 << 9);
2044
2045         if (WARN_ON(start != aligned_start)) {
2046                 len -= aligned_start - start;
2047                 len = round_down(len, 1 << 9);
2048                 start = aligned_start;
2049         }
2050
2051         *discarded_bytes = 0;
2052
2053         if (!len)
2054                 return 0;
2055
2056         end = start + len;
2057         bytes_left = len;
2058
2059         /* Skip any superblocks on this device. */
2060         for (j = 0; j < BTRFS_SUPER_MIRROR_MAX; j++) {
2061                 u64 sb_start = btrfs_sb_offset(j);
2062                 u64 sb_end = sb_start + BTRFS_SUPER_INFO_SIZE;
2063                 u64 size = sb_start - start;
2064
2065                 if (!in_range(sb_start, start, bytes_left) &&
2066                     !in_range(sb_end, start, bytes_left) &&
2067                     !in_range(start, sb_start, BTRFS_SUPER_INFO_SIZE))
2068                         continue;
2069
2070                 /*
2071                  * Superblock spans beginning of range.  Adjust start and
2072                  * try again.
2073                  */
2074                 if (sb_start <= start) {
2075                         start += sb_end - start;
2076                         if (start > end) {
2077                                 bytes_left = 0;
2078                                 break;
2079                         }
2080                         bytes_left = end - start;
2081                         continue;
2082                 }
2083
2084                 if (size) {
2085                         ret = blkdev_issue_discard(bdev, start >> 9, size >> 9,
2086                                                    GFP_NOFS, 0);
2087                         if (!ret)
2088                                 *discarded_bytes += size;
2089                         else if (ret != -EOPNOTSUPP)
2090                                 return ret;
2091                 }
2092
2093                 start = sb_end;
2094                 if (start > end) {
2095                         bytes_left = 0;
2096                         break;
2097                 }
2098                 bytes_left = end - start;
2099         }
2100
2101         if (bytes_left) {
2102                 ret = blkdev_issue_discard(bdev, start >> 9, bytes_left >> 9,
2103                                            GFP_NOFS, 0);
2104                 if (!ret)
2105                         *discarded_bytes += bytes_left;
2106         }
2107         return ret;
2108 }
2109
2110 int btrfs_discard_extent(struct btrfs_fs_info *fs_info, u64 bytenr,
2111                          u64 num_bytes, u64 *actual_bytes)
2112 {
2113         int ret;
2114         u64 discarded_bytes = 0;
2115         struct btrfs_bio *bbio = NULL;
2116
2117
2118         /*
2119          * Avoid races with device replace and make sure our bbio has devices
2120          * associated to its stripes that don't go away while we are discarding.
2121          */
2122         btrfs_bio_counter_inc_blocked(fs_info);
2123         /* Tell the block device(s) that the sectors can be discarded */
2124         ret = btrfs_map_block(fs_info, BTRFS_MAP_DISCARD, bytenr, &num_bytes,
2125                               &bbio, 0);
2126         /* Error condition is -ENOMEM */
2127         if (!ret) {
2128                 struct btrfs_bio_stripe *stripe = bbio->stripes;
2129                 int i;
2130
2131
2132                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
2133                         u64 bytes;
2134                         struct request_queue *req_q;
2135
2136                         if (!stripe->dev->bdev) {
2137                                 ASSERT(btrfs_test_opt(fs_info, DEGRADED));
2138                                 continue;
2139                         }
2140                         req_q = bdev_get_queue(stripe->dev->bdev);
2141                         if (!blk_queue_discard(req_q))
2142                                 continue;
2143
2144                         ret = btrfs_issue_discard(stripe->dev->bdev,
2145                                                   stripe->physical,
2146                                                   stripe->length,
2147                                                   &bytes);
2148                         if (!ret)
2149                                 discarded_bytes += bytes;
2150                         else if (ret != -EOPNOTSUPP)
2151                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
2152
2153                         /*
2154                          * Just in case we get back EOPNOTSUPP for some reason,
2155                          * just ignore the return value so we don't screw up
2156                          * people calling discard_extent.
2157                          */
2158                         ret = 0;
2159                 }
2160                 btrfs_put_bbio(bbio);
2161         }
2162         btrfs_bio_counter_dec(fs_info);
2163
2164         if (actual_bytes)
2165                 *actual_bytes = discarded_bytes;
2166
2167
2168         if (ret == -EOPNOTSUPP)
2169                 ret = 0;
2170         return ret;
2171 }
2172
2173 /* Can return -ENOMEM */
2174 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2175                          struct btrfs_root *root,
2176                          u64 bytenr, u64 num_bytes, u64 parent,
2177                          u64 root_objectid, u64 owner, u64 offset)
2178 {
2179         struct btrfs_fs_info *fs_info = root->fs_info;
2180         int old_ref_mod, new_ref_mod;
2181         int ret;
2182
2183         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
2184                root_objectid == BTRFS_TREE_LOG_OBJECTID);
2185
2186         btrfs_ref_tree_mod(root, bytenr, num_bytes, parent, root_objectid,
2187                            owner, offset, BTRFS_ADD_DELAYED_REF);
2188
2189         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
2190                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
2191                                                  num_bytes, parent,
2192                                                  root_objectid, (int)owner,
2193                                                  BTRFS_ADD_DELAYED_REF, NULL,
2194                                                  &old_ref_mod, &new_ref_mod);
2195         } else {
2196                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
2197                                                  num_bytes, parent,
2198                                                  root_objectid, owner, offset,
2199                                                  0, BTRFS_ADD_DELAYED_REF,
2200                                                  &old_ref_mod, &new_ref_mod);
2201         }
2202
2203         if (ret == 0 && old_ref_mod < 0 && new_ref_mod >= 0)
2204                 add_pinned_bytes(fs_info, -num_bytes, owner, root_objectid);
2205
2206         return ret;
2207 }
2208
2209 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2210                                   struct btrfs_fs_info *fs_info,
2211                                   struct btrfs_delayed_ref_node *node,
2212                                   u64 parent, u64 root_objectid,
2213                                   u64 owner, u64 offset, int refs_to_add,
2214                                   struct btrfs_delayed_extent_op *extent_op)
2215 {
2216         struct btrfs_path *path;
2217         struct extent_buffer *leaf;
2218         struct btrfs_extent_item *item;
2219         struct btrfs_key key;
2220         u64 bytenr = node->bytenr;
2221         u64 num_bytes = node->num_bytes;
2222         u64 refs;
2223         int ret;
2224
2225         path = btrfs_alloc_path();
2226         if (!path)
2227                 return -ENOMEM;
2228
2229         path->reada = READA_FORWARD;
2230         path->leave_spinning = 1;
2231         /* this will setup the path even if it fails to insert the back ref */
2232         ret = insert_inline_extent_backref(trans, fs_info, path, bytenr,
2233                                            num_bytes, parent, root_objectid,
2234                                            owner, offset,
2235                                            refs_to_add, extent_op);
2236         if ((ret < 0 && ret != -EAGAIN) || !ret)
2237                 goto out;
2238
2239         /*
2240          * Ok we had -EAGAIN which means we didn't have space to insert and
2241          * inline extent ref, so just update the reference count and add a
2242          * normal backref.
2243          */
2244         leaf = path->nodes[0];
2245         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2246         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2247         refs = btrfs_extent_refs(leaf, item);
2248         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
2249         if (extent_op)
2250                 __run_delayed_extent_op(extent_op, leaf, item);
2251
2252         btrfs_mark_buffer_dirty(leaf);
2253         btrfs_release_path(path);
2254
2255         path->reada = READA_FORWARD;
2256         path->leave_spinning = 1;
2257         /* now insert the actual backref */
2258         ret = insert_extent_backref(trans, fs_info, path, bytenr, parent,
2259                                     root_objectid, owner, offset, refs_to_add);
2260         if (ret)
2261                 btrfs_abort_transaction(trans, ret);
2262 out:
2263         btrfs_free_path(path);
2264         return ret;
2265 }
2266
2267 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
2268                                 struct btrfs_fs_info *fs_info,
2269                                 struct btrfs_delayed_ref_node *node,
2270                                 struct btrfs_delayed_extent_op *extent_op,
2271                                 int insert_reserved)
2272 {
2273         int ret = 0;
2274         struct btrfs_delayed_data_ref *ref;
2275         struct btrfs_key ins;
2276         u64 parent = 0;
2277         u64 ref_root = 0;
2278         u64 flags = 0;
2279
2280         ins.objectid = node->bytenr;
2281         ins.offset = node->num_bytes;
2282         ins.type = BTRFS_EXTENT_ITEM_KEY;
2283
2284         ref = btrfs_delayed_node_to_data_ref(node);
2285         trace_run_delayed_data_ref(fs_info, node, ref, node->action);
2286
2287         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2288                 parent = ref->parent;
2289         ref_root = ref->root;
2290
2291         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2292                 if (extent_op)
2293                         flags |= extent_op->flags_to_set;
2294                 ret = alloc_reserved_file_extent(trans, fs_info,
2295                                                  parent, ref_root, flags,
2296                                                  ref->objectid, ref->offset,
2297                                                  &ins, node->ref_mod);
2298         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2299                 ret = __btrfs_inc_extent_ref(trans, fs_info, node, parent,
2300                                              ref_root, ref->objectid,
2301                                              ref->offset, node->ref_mod,
2302                                              extent_op);
2303         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2304                 ret = __btrfs_free_extent(trans, fs_info, node, parent,
2305                                           ref_root, ref->objectid,
2306                                           ref->offset, node->ref_mod,
2307                                           extent_op);
2308         } else {
2309                 BUG();
2310         }
2311         return ret;
2312 }
2313
2314 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2315                                     struct extent_buffer *leaf,
2316                                     struct btrfs_extent_item *ei)
2317 {
2318         u64 flags = btrfs_extent_flags(leaf, ei);
2319         if (extent_op->update_flags) {
2320                 flags |= extent_op->flags_to_set;
2321                 btrfs_set_extent_flags(leaf, ei, flags);
2322         }
2323
2324         if (extent_op->update_key) {
2325                 struct btrfs_tree_block_info *bi;
2326                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2327                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2328                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2329         }
2330 }
2331
2332 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2333                                  struct btrfs_fs_info *fs_info,
2334                                  struct btrfs_delayed_ref_head *head,
2335                                  struct btrfs_delayed_extent_op *extent_op)
2336 {
2337         struct btrfs_key key;
2338         struct btrfs_path *path;
2339         struct btrfs_extent_item *ei;
2340         struct extent_buffer *leaf;
2341         u32 item_size;
2342         int ret;
2343         int err = 0;
2344         int metadata = !extent_op->is_data;
2345
2346         if (trans->aborted)
2347                 return 0;
2348
2349         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2350                 metadata = 0;
2351
2352         path = btrfs_alloc_path();
2353         if (!path)
2354                 return -ENOMEM;
2355
2356         key.objectid = head->bytenr;
2357
2358         if (metadata) {
2359                 key.type = BTRFS_METADATA_ITEM_KEY;
2360                 key.offset = extent_op->level;
2361         } else {
2362                 key.type = BTRFS_EXTENT_ITEM_KEY;
2363                 key.offset = head->num_bytes;
2364         }
2365
2366 again:
2367         path->reada = READA_FORWARD;
2368         path->leave_spinning = 1;
2369         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 1);
2370         if (ret < 0) {
2371                 err = ret;
2372                 goto out;
2373         }
2374         if (ret > 0) {
2375                 if (metadata) {
2376                         if (path->slots[0] > 0) {
2377                                 path->slots[0]--;
2378                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
2379                                                       path->slots[0]);
2380                                 if (key.objectid == head->bytenr &&
2381                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
2382                                     key.offset == head->num_bytes)
2383                                         ret = 0;
2384                         }
2385                         if (ret > 0) {
2386                                 btrfs_release_path(path);
2387                                 metadata = 0;
2388
2389                                 key.objectid = head->bytenr;
2390                                 key.offset = head->num_bytes;
2391                                 key.type = BTRFS_EXTENT_ITEM_KEY;
2392                                 goto again;
2393                         }
2394                 } else {
2395                         err = -EIO;
2396                         goto out;
2397                 }
2398         }
2399
2400         leaf = path->nodes[0];
2401         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2402 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2403         if (item_size < sizeof(*ei)) {
2404                 ret = convert_extent_item_v0(trans, fs_info, path, (u64)-1, 0);
2405                 if (ret < 0) {
2406                         err = ret;
2407                         goto out;
2408                 }
2409                 leaf = path->nodes[0];
2410                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2411         }
2412 #endif
2413         BUG_ON(item_size < sizeof(*ei));
2414         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2415         __run_delayed_extent_op(extent_op, leaf, ei);
2416
2417         btrfs_mark_buffer_dirty(leaf);
2418 out:
2419         btrfs_free_path(path);
2420         return err;
2421 }
2422
2423 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2424                                 struct btrfs_fs_info *fs_info,
2425                                 struct btrfs_delayed_ref_node *node,
2426                                 struct btrfs_delayed_extent_op *extent_op,
2427                                 int insert_reserved)
2428 {
2429         int ret = 0;
2430         struct btrfs_delayed_tree_ref *ref;
2431         struct btrfs_key ins;
2432         u64 parent = 0;
2433         u64 ref_root = 0;
2434         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
2435
2436         ref = btrfs_delayed_node_to_tree_ref(node);
2437         trace_run_delayed_tree_ref(fs_info, node, ref, node->action);
2438
2439         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2440                 parent = ref->parent;
2441         ref_root = ref->root;
2442
2443         ins.objectid = node->bytenr;
2444         if (skinny_metadata) {
2445                 ins.offset = ref->level;
2446                 ins.type = BTRFS_METADATA_ITEM_KEY;
2447         } else {
2448                 ins.offset = node->num_bytes;
2449                 ins.type = BTRFS_EXTENT_ITEM_KEY;
2450         }
2451
2452         if (node->ref_mod != 1) {
2453                 btrfs_err(fs_info,
2454         "btree block(%llu) has %d references rather than 1: action %d ref_root %llu parent %llu",
2455                           node->bytenr, node->ref_mod, node->action, ref_root,
2456                           parent);
2457                 return -EIO;
2458         }
2459         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2460                 BUG_ON(!extent_op || !extent_op->update_flags);
2461                 ret = alloc_reserved_tree_block(trans, fs_info,
2462                                                 parent, ref_root,
2463                                                 extent_op->flags_to_set,
2464                                                 &extent_op->key,
2465                                                 ref->level, &ins);
2466         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2467                 ret = __btrfs_inc_extent_ref(trans, fs_info, node,
2468                                              parent, ref_root,
2469                                              ref->level, 0, 1,
2470                                              extent_op);
2471         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2472                 ret = __btrfs_free_extent(trans, fs_info, node,
2473                                           parent, ref_root,
2474                                           ref->level, 0, 1, extent_op);
2475         } else {
2476                 BUG();
2477         }
2478         return ret;
2479 }
2480
2481 /* helper function to actually process a single delayed ref entry */
2482 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2483                                struct btrfs_fs_info *fs_info,
2484                                struct btrfs_delayed_ref_node *node,
2485                                struct btrfs_delayed_extent_op *extent_op,
2486                                int insert_reserved)
2487 {
2488         int ret = 0;
2489
2490         if (trans->aborted) {
2491                 if (insert_reserved)
2492                         btrfs_pin_extent(fs_info, node->bytenr,
2493                                          node->num_bytes, 1);
2494                 return 0;
2495         }
2496
2497         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2498             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2499                 ret = run_delayed_tree_ref(trans, fs_info, node, extent_op,
2500                                            insert_reserved);
2501         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2502                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2503                 ret = run_delayed_data_ref(trans, fs_info, node, extent_op,
2504                                            insert_reserved);
2505         else
2506                 BUG();
2507         return ret;
2508 }
2509
2510 static inline struct btrfs_delayed_ref_node *
2511 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2512 {
2513         struct btrfs_delayed_ref_node *ref;
2514
2515         if (RB_EMPTY_ROOT(&head->ref_tree))
2516                 return NULL;
2517
2518         /*
2519          * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
2520          * This is to prevent a ref count from going down to zero, which deletes
2521          * the extent item from the extent tree, when there still are references
2522          * to add, which would fail because they would not find the extent item.
2523          */
2524         if (!list_empty(&head->ref_add_list))
2525                 return list_first_entry(&head->ref_add_list,
2526                                 struct btrfs_delayed_ref_node, add_list);
2527
2528         ref = rb_entry(rb_first(&head->ref_tree),
2529                        struct btrfs_delayed_ref_node, ref_node);
2530         ASSERT(list_empty(&ref->add_list));
2531         return ref;
2532 }
2533
2534 static void unselect_delayed_ref_head(struct btrfs_delayed_ref_root *delayed_refs,
2535                                       struct btrfs_delayed_ref_head *head)
2536 {
2537         spin_lock(&delayed_refs->lock);
2538         head->processing = 0;
2539         delayed_refs->num_heads_ready++;
2540         spin_unlock(&delayed_refs->lock);
2541         btrfs_delayed_ref_unlock(head);
2542 }
2543
2544 static int cleanup_extent_op(struct btrfs_trans_handle *trans,
2545                              struct btrfs_fs_info *fs_info,
2546                              struct btrfs_delayed_ref_head *head)
2547 {
2548         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
2549         int ret;
2550
2551         if (!extent_op)
2552                 return 0;
2553         head->extent_op = NULL;
2554         if (head->must_insert_reserved) {
2555                 btrfs_free_delayed_extent_op(extent_op);
2556                 return 0;
2557         }
2558         spin_unlock(&head->lock);
2559         ret = run_delayed_extent_op(trans, fs_info, head, extent_op);
2560         btrfs_free_delayed_extent_op(extent_op);
2561         return ret ? ret : 1;
2562 }
2563
2564 static int cleanup_ref_head(struct btrfs_trans_handle *trans,
2565                             struct btrfs_fs_info *fs_info,
2566                             struct btrfs_delayed_ref_head *head)
2567 {
2568         struct btrfs_delayed_ref_root *delayed_refs;
2569         int ret;
2570
2571         delayed_refs = &trans->transaction->delayed_refs;
2572
2573         ret = cleanup_extent_op(trans, fs_info, head);
2574         if (ret < 0) {
2575                 unselect_delayed_ref_head(delayed_refs, head);
2576                 btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
2577                 return ret;
2578         } else if (ret) {
2579                 return ret;
2580         }
2581
2582         /*
2583          * Need to drop our head ref lock and re-acquire the delayed ref lock
2584          * and then re-check to make sure nobody got added.
2585          */
2586         spin_unlock(&head->lock);
2587         spin_lock(&delayed_refs->lock);
2588         spin_lock(&head->lock);
2589         if (!RB_EMPTY_ROOT(&head->ref_tree) || head->extent_op) {
2590                 spin_unlock(&head->lock);
2591                 spin_unlock(&delayed_refs->lock);
2592                 return 1;
2593         }
2594         delayed_refs->num_heads--;
2595         rb_erase(&head->href_node, &delayed_refs->href_root);
2596         RB_CLEAR_NODE(&head->href_node);
2597         spin_unlock(&delayed_refs->lock);
2598         spin_unlock(&head->lock);
2599         atomic_dec(&delayed_refs->num_entries);
2600
2601         trace_run_delayed_ref_head(fs_info, head, 0);
2602
2603         if (head->total_ref_mod < 0) {
2604                 struct btrfs_block_group_cache *cache;
2605
2606                 cache = btrfs_lookup_block_group(fs_info, head->bytenr);
2607                 ASSERT(cache);
2608                 percpu_counter_add(&cache->space_info->total_bytes_pinned,
2609                                    -head->num_bytes);
2610                 btrfs_put_block_group(cache);
2611
2612                 if (head->is_data) {
2613                         spin_lock(&delayed_refs->lock);
2614                         delayed_refs->pending_csums -= head->num_bytes;
2615                         spin_unlock(&delayed_refs->lock);
2616                 }
2617         }
2618
2619         if (head->must_insert_reserved) {
2620                 btrfs_pin_extent(fs_info, head->bytenr,
2621                                  head->num_bytes, 1);
2622                 if (head->is_data) {
2623                         ret = btrfs_del_csums(trans, fs_info, head->bytenr,
2624                                               head->num_bytes);
2625                 }
2626         }
2627
2628         /* Also free its reserved qgroup space */
2629         btrfs_qgroup_free_delayed_ref(fs_info, head->qgroup_ref_root,
2630                                       head->qgroup_reserved);
2631         btrfs_delayed_ref_unlock(head);
2632         btrfs_put_delayed_ref_head(head);
2633         return 0;
2634 }
2635
2636 /*
2637  * Returns 0 on success or if called with an already aborted transaction.
2638  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2639  */
2640 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2641                                              unsigned long nr)
2642 {
2643         struct btrfs_fs_info *fs_info = trans->fs_info;
2644         struct btrfs_delayed_ref_root *delayed_refs;
2645         struct btrfs_delayed_ref_node *ref;
2646         struct btrfs_delayed_ref_head *locked_ref = NULL;
2647         struct btrfs_delayed_extent_op *extent_op;
2648         ktime_t start = ktime_get();
2649         int ret;
2650         unsigned long count = 0;
2651         unsigned long actual_count = 0;
2652         int must_insert_reserved = 0;
2653
2654         delayed_refs = &trans->transaction->delayed_refs;
2655         while (1) {
2656                 if (!locked_ref) {
2657                         if (count >= nr)
2658                                 break;
2659
2660                         spin_lock(&delayed_refs->lock);
2661                         locked_ref = btrfs_select_ref_head(trans);
2662                         if (!locked_ref) {
2663                                 spin_unlock(&delayed_refs->lock);
2664                                 break;
2665                         }
2666
2667                         /* grab the lock that says we are going to process
2668                          * all the refs for this head */
2669                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2670                         spin_unlock(&delayed_refs->lock);
2671                         /*
2672                          * we may have dropped the spin lock to get the head
2673                          * mutex lock, and that might have given someone else
2674                          * time to free the head.  If that's true, it has been
2675                          * removed from our list and we can move on.
2676                          */
2677                         if (ret == -EAGAIN) {
2678                                 locked_ref = NULL;
2679                                 count++;
2680                                 continue;
2681                         }
2682                 }
2683
2684                 /*
2685                  * We need to try and merge add/drops of the same ref since we
2686                  * can run into issues with relocate dropping the implicit ref
2687                  * and then it being added back again before the drop can
2688                  * finish.  If we merged anything we need to re-loop so we can
2689                  * get a good ref.
2690                  * Or we can get node references of the same type that weren't
2691                  * merged when created due to bumps in the tree mod seq, and
2692                  * we need to merge them to prevent adding an inline extent
2693                  * backref before dropping it (triggering a BUG_ON at
2694                  * insert_inline_extent_backref()).
2695                  */
2696                 spin_lock(&locked_ref->lock);
2697                 btrfs_merge_delayed_refs(trans, fs_info, delayed_refs,
2698                                          locked_ref);
2699
2700                 /*
2701                  * locked_ref is the head node, so we have to go one
2702                  * node back for any delayed ref updates
2703                  */
2704                 ref = select_delayed_ref(locked_ref);
2705
2706                 if (ref && ref->seq &&
2707                     btrfs_check_delayed_seq(fs_info, delayed_refs, ref->seq)) {
2708                         spin_unlock(&locked_ref->lock);
2709                         unselect_delayed_ref_head(delayed_refs, locked_ref);
2710                         locked_ref = NULL;
2711                         cond_resched();
2712                         count++;
2713                         continue;
2714                 }
2715
2716                 /*
2717                  * We're done processing refs in this ref_head, clean everything
2718                  * up and move on to the next ref_head.
2719                  */
2720                 if (!ref) {
2721                         ret = cleanup_ref_head(trans, fs_info, locked_ref);
2722                         if (ret > 0 ) {
2723                                 /* We dropped our lock, we need to loop. */
2724                                 ret = 0;
2725                                 continue;
2726                         } else if (ret) {
2727                                 return ret;
2728                         }
2729                         locked_ref = NULL;
2730                         count++;
2731                         continue;
2732                 }
2733
2734                 actual_count++;
2735                 ref->in_tree = 0;
2736                 rb_erase(&ref->ref_node, &locked_ref->ref_tree);
2737                 RB_CLEAR_NODE(&ref->ref_node);
2738                 if (!list_empty(&ref->add_list))
2739                         list_del(&ref->add_list);
2740                 /*
2741                  * When we play the delayed ref, also correct the ref_mod on
2742                  * head
2743                  */
2744                 switch (ref->action) {
2745                 case BTRFS_ADD_DELAYED_REF:
2746                 case BTRFS_ADD_DELAYED_EXTENT:
2747                         locked_ref->ref_mod -= ref->ref_mod;
2748                         break;
2749                 case BTRFS_DROP_DELAYED_REF:
2750                         locked_ref->ref_mod += ref->ref_mod;
2751                         break;
2752                 default:
2753                         WARN_ON(1);
2754                 }
2755                 atomic_dec(&delayed_refs->num_entries);
2756
2757                 /*
2758                  * Record the must-insert_reserved flag before we drop the spin
2759                  * lock.
2760                  */
2761                 must_insert_reserved = locked_ref->must_insert_reserved;
2762                 locked_ref->must_insert_reserved = 0;
2763
2764                 extent_op = locked_ref->extent_op;
2765                 locked_ref->extent_op = NULL;
2766                 spin_unlock(&locked_ref->lock);
2767
2768                 ret = run_one_delayed_ref(trans, fs_info, ref, extent_op,
2769                                           must_insert_reserved);
2770
2771                 btrfs_free_delayed_extent_op(extent_op);
2772                 if (ret) {
2773                         unselect_delayed_ref_head(delayed_refs, locked_ref);
2774                         btrfs_put_delayed_ref(ref);
2775                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d",
2776                                     ret);
2777                         return ret;
2778                 }
2779
2780                 btrfs_put_delayed_ref(ref);
2781                 count++;
2782                 cond_resched();
2783         }
2784
2785         /*
2786          * We don't want to include ref heads since we can have empty ref heads
2787          * and those will drastically skew our runtime down since we just do
2788          * accounting, no actual extent tree updates.
2789          */
2790         if (actual_count > 0) {
2791                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2792                 u64 avg;
2793
2794                 /*
2795                  * We weigh the current average higher than our current runtime
2796                  * to avoid large swings in the average.
2797                  */
2798                 spin_lock(&delayed_refs->lock);
2799                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2800                 fs_info->avg_delayed_ref_runtime = avg >> 2;    /* div by 4 */
2801                 spin_unlock(&delayed_refs->lock);
2802         }
2803         return 0;
2804 }
2805
2806 #ifdef SCRAMBLE_DELAYED_REFS
2807 /*
2808  * Normally delayed refs get processed in ascending bytenr order. This
2809  * correlates in most cases to the order added. To expose dependencies on this
2810  * order, we start to process the tree in the middle instead of the beginning
2811  */
2812 static u64 find_middle(struct rb_root *root)
2813 {
2814         struct rb_node *n = root->rb_node;
2815         struct btrfs_delayed_ref_node *entry;
2816         int alt = 1;
2817         u64 middle;
2818         u64 first = 0, last = 0;
2819
2820         n = rb_first(root);
2821         if (n) {
2822                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2823                 first = entry->bytenr;
2824         }
2825         n = rb_last(root);
2826         if (n) {
2827                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2828                 last = entry->bytenr;
2829         }
2830         n = root->rb_node;
2831
2832         while (n) {
2833                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2834                 WARN_ON(!entry->in_tree);
2835
2836                 middle = entry->bytenr;
2837
2838                 if (alt)
2839                         n = n->rb_left;
2840                 else
2841                         n = n->rb_right;
2842
2843                 alt = 1 - alt;
2844         }
2845         return middle;
2846 }
2847 #endif
2848
2849 static inline u64 heads_to_leaves(struct btrfs_fs_info *fs_info, u64 heads)
2850 {
2851         u64 num_bytes;
2852
2853         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2854                              sizeof(struct btrfs_extent_inline_ref));
2855         if (!btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2856                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2857
2858         /*
2859          * We don't ever fill up leaves all the way so multiply by 2 just to be
2860          * closer to what we're really going to want to use.
2861          */
2862         return div_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(fs_info));
2863 }
2864
2865 /*
2866  * Takes the number of bytes to be csumm'ed and figures out how many leaves it
2867  * would require to store the csums for that many bytes.
2868  */
2869 u64 btrfs_csum_bytes_to_leaves(struct btrfs_fs_info *fs_info, u64 csum_bytes)
2870 {
2871         u64 csum_size;
2872         u64 num_csums_per_leaf;
2873         u64 num_csums;
2874
2875         csum_size = BTRFS_MAX_ITEM_SIZE(fs_info);
2876         num_csums_per_leaf = div64_u64(csum_size,
2877                         (u64)btrfs_super_csum_size(fs_info->super_copy));
2878         num_csums = div64_u64(csum_bytes, fs_info->sectorsize);
2879         num_csums += num_csums_per_leaf - 1;
2880         num_csums = div64_u64(num_csums, num_csums_per_leaf);
2881         return num_csums;
2882 }
2883
2884 int btrfs_check_space_for_delayed_refs(struct btrfs_trans_handle *trans,
2885                                        struct btrfs_fs_info *fs_info)
2886 {
2887         struct btrfs_block_rsv *global_rsv;
2888         u64 num_heads = trans->transaction->delayed_refs.num_heads_ready;
2889         u64 csum_bytes = trans->transaction->delayed_refs.pending_csums;
2890         unsigned int num_dirty_bgs = trans->transaction->num_dirty_bgs;
2891         u64 num_bytes, num_dirty_bgs_bytes;
2892         int ret = 0;
2893
2894         num_bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
2895         num_heads = heads_to_leaves(fs_info, num_heads);
2896         if (num_heads > 1)
2897                 num_bytes += (num_heads - 1) * fs_info->nodesize;
2898         num_bytes <<= 1;
2899         num_bytes += btrfs_csum_bytes_to_leaves(fs_info, csum_bytes) *
2900                                                         fs_info->nodesize;
2901         num_dirty_bgs_bytes = btrfs_calc_trans_metadata_size(fs_info,
2902                                                              num_dirty_bgs);
2903         global_rsv = &fs_info->global_block_rsv;
2904
2905         /*
2906          * If we can't allocate any more chunks lets make sure we have _lots_ of
2907          * wiggle room since running delayed refs can create more delayed refs.
2908          */
2909         if (global_rsv->space_info->full) {
2910                 num_dirty_bgs_bytes <<= 1;
2911                 num_bytes <<= 1;
2912         }
2913
2914         spin_lock(&global_rsv->lock);
2915         if (global_rsv->reserved <= num_bytes + num_dirty_bgs_bytes)
2916                 ret = 1;
2917         spin_unlock(&global_rsv->lock);
2918         return ret;
2919 }
2920
2921 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans,
2922                                        struct btrfs_fs_info *fs_info)
2923 {
2924         u64 num_entries =
2925                 atomic_read(&trans->transaction->delayed_refs.num_entries);
2926         u64 avg_runtime;
2927         u64 val;
2928
2929         smp_mb();
2930         avg_runtime = fs_info->avg_delayed_ref_runtime;
2931         val = num_entries * avg_runtime;
2932         if (val >= NSEC_PER_SEC)
2933                 return 1;
2934         if (val >= NSEC_PER_SEC / 2)
2935                 return 2;
2936
2937         return btrfs_check_space_for_delayed_refs(trans, fs_info);
2938 }
2939
2940 struct async_delayed_refs {
2941         struct btrfs_root *root;
2942         u64 transid;
2943         int count;
2944         int error;
2945         int sync;
2946         struct completion wait;
2947         struct btrfs_work work;
2948 };
2949
2950 static inline struct async_delayed_refs *
2951 to_async_delayed_refs(struct btrfs_work *work)
2952 {
2953         return container_of(work, struct async_delayed_refs, work);
2954 }
2955
2956 static void delayed_ref_async_start(struct btrfs_work *work)
2957 {
2958         struct async_delayed_refs *async = to_async_delayed_refs(work);
2959         struct btrfs_trans_handle *trans;
2960         struct btrfs_fs_info *fs_info = async->root->fs_info;
2961         int ret;
2962
2963         /* if the commit is already started, we don't need to wait here */
2964         if (btrfs_transaction_blocked(fs_info))
2965                 goto done;
2966
2967         trans = btrfs_join_transaction(async->root);
2968         if (IS_ERR(trans)) {
2969                 async->error = PTR_ERR(trans);
2970                 goto done;
2971         }
2972
2973         /*
2974          * trans->sync means that when we call end_transaction, we won't
2975          * wait on delayed refs
2976          */
2977         trans->sync = true;
2978
2979         /* Don't bother flushing if we got into a different transaction */
2980         if (trans->transid > async->transid)
2981                 goto end;
2982
2983         ret = btrfs_run_delayed_refs(trans, async->count);
2984         if (ret)
2985                 async->error = ret;
2986 end:
2987         ret = btrfs_end_transaction(trans);
2988         if (ret && !async->error)
2989                 async->error = ret;
2990 done:
2991         if (async->sync)
2992                 complete(&async->wait);
2993         else
2994                 kfree(async);
2995 }
2996
2997 int btrfs_async_run_delayed_refs(struct btrfs_fs_info *fs_info,
2998                                  unsigned long count, u64 transid, int wait)
2999 {
3000         struct async_delayed_refs *async;
3001         int ret;
3002
3003         async = kmalloc(sizeof(*async), GFP_NOFS);
3004         if (!async)
3005                 return -ENOMEM;
3006
3007         async->root = fs_info->tree_root;
3008         async->count = count;
3009         async->error = 0;
3010         async->transid = transid;
3011         if (wait)
3012                 async->sync = 1;
3013         else
3014                 async->sync = 0;
3015         init_completion(&async->wait);
3016
3017         btrfs_init_work(&async->work, btrfs_extent_refs_helper,
3018                         delayed_ref_async_start, NULL, NULL);
3019
3020         btrfs_queue_work(fs_info->extent_workers, &async->work);
3021
3022         if (wait) {
3023                 wait_for_completion(&async->wait);
3024                 ret = async->error;
3025                 kfree(async);
3026                 return ret;
3027         }
3028         return 0;
3029 }
3030
3031 /*
3032  * this starts processing the delayed reference count updates and
3033  * extent insertions we have queued up so far.  count can be
3034  * 0, which means to process everything in the tree at the start
3035  * of the run (but not newly added entries), or it can be some target
3036  * number you'd like to process.
3037  *
3038  * Returns 0 on success or if called with an aborted transaction
3039  * Returns <0 on error and aborts the transaction
3040  */
3041 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
3042                            unsigned long count)
3043 {
3044         struct btrfs_fs_info *fs_info = trans->fs_info;
3045         struct rb_node *node;
3046         struct btrfs_delayed_ref_root *delayed_refs;
3047         struct btrfs_delayed_ref_head *head;
3048         int ret;
3049         int run_all = count == (unsigned long)-1;
3050         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
3051
3052         /* We'll clean this up in btrfs_cleanup_transaction */
3053         if (trans->aborted)
3054                 return 0;
3055
3056         if (test_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags))
3057                 return 0;
3058
3059         delayed_refs = &trans->transaction->delayed_refs;
3060         if (count == 0)
3061                 count = atomic_read(&delayed_refs->num_entries) * 2;
3062
3063 again:
3064 #ifdef SCRAMBLE_DELAYED_REFS
3065         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
3066 #endif
3067         trans->can_flush_pending_bgs = false;
3068         ret = __btrfs_run_delayed_refs(trans, count);
3069         if (ret < 0) {
3070                 btrfs_abort_transaction(trans, ret);
3071                 return ret;
3072         }
3073
3074         if (run_all) {
3075                 if (!list_empty(&trans->new_bgs))
3076                         btrfs_create_pending_block_groups(trans);
3077
3078                 spin_lock(&delayed_refs->lock);
3079                 node = rb_first(&delayed_refs->href_root);
3080                 if (!node) {
3081                         spin_unlock(&delayed_refs->lock);
3082                         goto out;
3083                 }
3084                 head = rb_entry(node, struct btrfs_delayed_ref_head,
3085                                 href_node);
3086                 refcount_inc(&head->refs);
3087                 spin_unlock(&delayed_refs->lock);
3088
3089                 /* Mutex was contended, block until it's released and retry. */
3090                 mutex_lock(&head->mutex);
3091                 mutex_unlock(&head->mutex);
3092
3093                 btrfs_put_delayed_ref_head(head);
3094                 cond_resched();
3095                 goto again;
3096         }
3097 out:
3098         trans->can_flush_pending_bgs = can_flush_pending_bgs;
3099         return 0;
3100 }
3101
3102 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
3103                                 struct btrfs_fs_info *fs_info,
3104                                 u64 bytenr, u64 num_bytes, u64 flags,
3105                                 int level, int is_data)
3106 {
3107         struct btrfs_delayed_extent_op *extent_op;
3108         int ret;
3109
3110         extent_op = btrfs_alloc_delayed_extent_op();
3111         if (!extent_op)
3112                 return -ENOMEM;
3113
3114         extent_op->flags_to_set = flags;
3115         extent_op->update_flags = true;
3116         extent_op->update_key = false;
3117         extent_op->is_data = is_data ? true : false;
3118         extent_op->level = level;
3119
3120         ret = btrfs_add_delayed_extent_op(fs_info, trans, bytenr,
3121                                           num_bytes, extent_op);
3122         if (ret)
3123                 btrfs_free_delayed_extent_op(extent_op);
3124         return ret;
3125 }
3126
3127 static noinline int check_delayed_ref(struct btrfs_root *root,
3128                                       struct btrfs_path *path,
3129                                       u64 objectid, u64 offset, u64 bytenr)
3130 {
3131         struct btrfs_delayed_ref_head *head;
3132         struct btrfs_delayed_ref_node *ref;
3133         struct btrfs_delayed_data_ref *data_ref;
3134         struct btrfs_delayed_ref_root *delayed_refs;
3135         struct btrfs_transaction *cur_trans;
3136         struct rb_node *node;
3137         int ret = 0;
3138
3139         cur_trans = root->fs_info->running_transaction;
3140         if (!cur_trans)
3141                 return 0;
3142
3143         delayed_refs = &cur_trans->delayed_refs;
3144         spin_lock(&delayed_refs->lock);
3145         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
3146         if (!head) {
3147                 spin_unlock(&delayed_refs->lock);
3148                 return 0;
3149         }
3150
3151         if (!mutex_trylock(&head->mutex)) {
3152                 refcount_inc(&head->refs);
3153                 spin_unlock(&delayed_refs->lock);
3154
3155                 btrfs_release_path(path);
3156
3157                 /*
3158                  * Mutex was contended, block until it's released and let
3159                  * caller try again
3160                  */
3161                 mutex_lock(&head->mutex);
3162                 mutex_unlock(&head->mutex);
3163                 btrfs_put_delayed_ref_head(head);
3164                 return -EAGAIN;
3165         }
3166         spin_unlock(&delayed_refs->lock);
3167
3168         spin_lock(&head->lock);
3169         /*
3170          * XXX: We should replace this with a proper search function in the
3171          * future.
3172          */
3173         for (node = rb_first(&head->ref_tree); node; node = rb_next(node)) {
3174                 ref = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
3175                 /* If it's a shared ref we know a cross reference exists */
3176                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
3177                         ret = 1;
3178                         break;
3179                 }
3180
3181                 data_ref = btrfs_delayed_node_to_data_ref(ref);
3182
3183                 /*
3184                  * If our ref doesn't match the one we're currently looking at
3185                  * then we have a cross reference.
3186                  */
3187                 if (data_ref->root != root->root_key.objectid ||
3188                     data_ref->objectid != objectid ||
3189                     data_ref->offset != offset) {
3190                         ret = 1;
3191                         break;
3192                 }
3193         }
3194         spin_unlock(&head->lock);
3195         mutex_unlock(&head->mutex);
3196         return ret;
3197 }
3198
3199 static noinline int check_committed_ref(struct btrfs_root *root,
3200                                         struct btrfs_path *path,
3201                                         u64 objectid, u64 offset, u64 bytenr)
3202 {
3203         struct btrfs_fs_info *fs_info = root->fs_info;
3204         struct btrfs_root *extent_root = fs_info->extent_root;
3205         struct extent_buffer *leaf;
3206         struct btrfs_extent_data_ref *ref;
3207         struct btrfs_extent_inline_ref *iref;
3208         struct btrfs_extent_item *ei;
3209         struct btrfs_key key;
3210         u32 item_size;
3211         int type;
3212         int ret;
3213
3214         key.objectid = bytenr;
3215         key.offset = (u64)-1;
3216         key.type = BTRFS_EXTENT_ITEM_KEY;
3217
3218         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
3219         if (ret < 0)
3220                 goto out;
3221         BUG_ON(ret == 0); /* Corruption */
3222
3223         ret = -ENOENT;
3224         if (path->slots[0] == 0)
3225                 goto out;
3226
3227         path->slots[0]--;
3228         leaf = path->nodes[0];
3229         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3230
3231         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
3232                 goto out;
3233
3234         ret = 1;
3235         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
3236 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3237         if (item_size < sizeof(*ei)) {
3238                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3239                 goto out;
3240         }
3241 #endif
3242         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
3243
3244         if (item_size != sizeof(*ei) +
3245             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
3246                 goto out;
3247
3248         if (btrfs_extent_generation(leaf, ei) <=
3249             btrfs_root_last_snapshot(&root->root_item))
3250                 goto out;
3251
3252         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
3253
3254         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
3255         if (type != BTRFS_EXTENT_DATA_REF_KEY)
3256                 goto out;
3257
3258         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
3259         if (btrfs_extent_refs(leaf, ei) !=
3260             btrfs_extent_data_ref_count(leaf, ref) ||
3261             btrfs_extent_data_ref_root(leaf, ref) !=
3262             root->root_key.objectid ||
3263             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
3264             btrfs_extent_data_ref_offset(leaf, ref) != offset)
3265                 goto out;
3266
3267         ret = 0;
3268 out:
3269         return ret;
3270 }
3271
3272 int btrfs_cross_ref_exist(struct btrfs_root *root, u64 objectid, u64 offset,
3273                           u64 bytenr)
3274 {
3275         struct btrfs_path *path;
3276         int ret;
3277         int ret2;
3278
3279         path = btrfs_alloc_path();
3280         if (!path)
3281                 return -ENOENT;
3282
3283         do {
3284                 ret = check_committed_ref(root, path, objectid,
3285                                           offset, bytenr);
3286                 if (ret && ret != -ENOENT)
3287                         goto out;
3288
3289                 ret2 = check_delayed_ref(root, path, objectid,
3290                                          offset, bytenr);
3291         } while (ret2 == -EAGAIN);
3292
3293         if (ret2 && ret2 != -ENOENT) {
3294                 ret = ret2;
3295                 goto out;
3296         }
3297
3298         if (ret != -ENOENT || ret2 != -ENOENT)
3299                 ret = 0;
3300 out:
3301         btrfs_free_path(path);
3302         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
3303                 WARN_ON(ret > 0);
3304         return ret;
3305 }
3306
3307 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
3308                            struct btrfs_root *root,
3309                            struct extent_buffer *buf,
3310                            int full_backref, int inc)
3311 {
3312         struct btrfs_fs_info *fs_info = root->fs_info;
3313         u64 bytenr;
3314         u64 num_bytes;
3315         u64 parent;
3316         u64 ref_root;
3317         u32 nritems;
3318         struct btrfs_key key;
3319         struct btrfs_file_extent_item *fi;
3320         int i;
3321         int level;
3322         int ret = 0;
3323         int (*process_func)(struct btrfs_trans_handle *,
3324                             struct btrfs_root *,
3325                             u64, u64, u64, u64, u64, u64);
3326
3327
3328         if (btrfs_is_testing(fs_info))
3329                 return 0;
3330
3331         ref_root = btrfs_header_owner(buf);
3332         nritems = btrfs_header_nritems(buf);
3333         level = btrfs_header_level(buf);
3334
3335         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state) && level == 0)
3336                 return 0;
3337
3338         if (inc)
3339                 process_func = btrfs_inc_extent_ref;
3340         else
3341                 process_func = btrfs_free_extent;
3342
3343         if (full_backref)
3344                 parent = buf->start;
3345         else
3346                 parent = 0;
3347
3348         for (i = 0; i < nritems; i++) {
3349                 if (level == 0) {
3350                         btrfs_item_key_to_cpu(buf, &key, i);
3351                         if (key.type != BTRFS_EXTENT_DATA_KEY)
3352                                 continue;
3353                         fi = btrfs_item_ptr(buf, i,
3354                                             struct btrfs_file_extent_item);
3355                         if (btrfs_file_extent_type(buf, fi) ==
3356                             BTRFS_FILE_EXTENT_INLINE)
3357                                 continue;
3358                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3359                         if (bytenr == 0)
3360                                 continue;
3361
3362                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3363                         key.offset -= btrfs_file_extent_offset(buf, fi);
3364                         ret = process_func(trans, root, bytenr, num_bytes,
3365                                            parent, ref_root, key.objectid,
3366                                            key.offset);
3367                         if (ret)
3368                                 goto fail;
3369                 } else {
3370                         bytenr = btrfs_node_blockptr(buf, i);
3371                         num_bytes = fs_info->nodesize;
3372                         ret = process_func(trans, root, bytenr, num_bytes,
3373                                            parent, ref_root, level - 1, 0);
3374                         if (ret)
3375                                 goto fail;
3376                 }
3377         }
3378         return 0;
3379 fail:
3380         return ret;
3381 }
3382
3383 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3384                   struct extent_buffer *buf, int full_backref)
3385 {
3386         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
3387 }
3388
3389 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3390                   struct extent_buffer *buf, int full_backref)
3391 {
3392         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
3393 }
3394
3395 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3396                                  struct btrfs_fs_info *fs_info,
3397                                  struct btrfs_path *path,
3398                                  struct btrfs_block_group_cache *cache)
3399 {
3400         int ret;
3401         struct btrfs_root *extent_root = fs_info->extent_root;
3402         unsigned long bi;
3403         struct extent_buffer *leaf;
3404
3405         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3406         if (ret) {
3407                 if (ret > 0)
3408                         ret = -ENOENT;
3409                 goto fail;
3410         }
3411
3412         leaf = path->nodes[0];
3413         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3414         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3415         btrfs_mark_buffer_dirty(leaf);
3416 fail:
3417         btrfs_release_path(path);
3418         return ret;
3419
3420 }
3421
3422 static struct btrfs_block_group_cache *
3423 next_block_group(struct btrfs_fs_info *fs_info,
3424                  struct btrfs_block_group_cache *cache)
3425 {
3426         struct rb_node *node;
3427
3428         spin_lock(&fs_info->block_group_cache_lock);
3429
3430         /* If our block group was removed, we need a full search. */
3431         if (RB_EMPTY_NODE(&cache->cache_node)) {
3432                 const u64 next_bytenr = cache->key.objectid + cache->key.offset;
3433
3434                 spin_unlock(&fs_info->block_group_cache_lock);
3435                 btrfs_put_block_group(cache);
3436                 cache = btrfs_lookup_first_block_group(fs_info, next_bytenr); return cache;
3437         }
3438         node = rb_next(&cache->cache_node);
3439         btrfs_put_block_group(cache);
3440         if (node) {
3441                 cache = rb_entry(node, struct btrfs_block_group_cache,
3442                                  cache_node);
3443                 btrfs_get_block_group(cache);
3444         } else
3445                 cache = NULL;
3446         spin_unlock(&fs_info->block_group_cache_lock);
3447         return cache;
3448 }
3449
3450 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3451                             struct btrfs_trans_handle *trans,
3452                             struct btrfs_path *path)
3453 {
3454         struct btrfs_fs_info *fs_info = block_group->fs_info;
3455         struct btrfs_root *root = fs_info->tree_root;
3456         struct inode *inode = NULL;
3457         struct extent_changeset *data_reserved = NULL;
3458         u64 alloc_hint = 0;
3459         int dcs = BTRFS_DC_ERROR;
3460         u64 num_pages = 0;
3461         int retries = 0;
3462         int ret = 0;
3463
3464         /*
3465          * If this block group is smaller than 100 megs don't bother caching the
3466          * block group.
3467          */
3468         if (block_group->key.offset < (100 * SZ_1M)) {
3469                 spin_lock(&block_group->lock);
3470                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3471                 spin_unlock(&block_group->lock);
3472                 return 0;
3473         }
3474
3475         if (trans->aborted)
3476                 return 0;
3477 again:
3478         inode = lookup_free_space_inode(fs_info, block_group, path);
3479         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3480                 ret = PTR_ERR(inode);
3481                 btrfs_release_path(path);
3482                 goto out;
3483         }
3484
3485         if (IS_ERR(inode)) {
3486                 BUG_ON(retries);
3487                 retries++;
3488
3489                 if (block_group->ro)
3490                         goto out_free;
3491
3492                 ret = create_free_space_inode(fs_info, trans, block_group,
3493                                               path);
3494                 if (ret)
3495                         goto out_free;
3496                 goto again;
3497         }
3498
3499         /*
3500          * We want to set the generation to 0, that way if anything goes wrong
3501          * from here on out we know not to trust this cache when we load up next
3502          * time.
3503          */
3504         BTRFS_I(inode)->generation = 0;
3505         ret = btrfs_update_inode(trans, root, inode);
3506         if (ret) {
3507                 /*
3508                  * So theoretically we could recover from this, simply set the
3509                  * super cache generation to 0 so we know to invalidate the
3510                  * cache, but then we'd have to keep track of the block groups
3511                  * that fail this way so we know we _have_ to reset this cache
3512                  * before the next commit or risk reading stale cache.  So to
3513                  * limit our exposure to horrible edge cases lets just abort the
3514                  * transaction, this only happens in really bad situations
3515                  * anyway.
3516                  */
3517                 btrfs_abort_transaction(trans, ret);
3518                 goto out_put;
3519         }
3520         WARN_ON(ret);
3521
3522         /* We've already setup this transaction, go ahead and exit */
3523         if (block_group->cache_generation == trans->transid &&
3524             i_size_read(inode)) {
3525                 dcs = BTRFS_DC_SETUP;
3526                 goto out_put;
3527         }
3528
3529         if (i_size_read(inode) > 0) {
3530                 ret = btrfs_check_trunc_cache_free_space(fs_info,
3531                                         &fs_info->global_block_rsv);
3532                 if (ret)
3533                         goto out_put;
3534
3535                 ret = btrfs_truncate_free_space_cache(trans, NULL, inode);
3536                 if (ret)
3537                         goto out_put;
3538         }
3539
3540         spin_lock(&block_group->lock);
3541         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3542             !btrfs_test_opt(fs_info, SPACE_CACHE)) {
3543                 /*
3544                  * don't bother trying to write stuff out _if_
3545                  * a) we're not cached,
3546                  * b) we're with nospace_cache mount option,
3547                  * c) we're with v2 space_cache (FREE_SPACE_TREE).
3548                  */
3549                 dcs = BTRFS_DC_WRITTEN;
3550                 spin_unlock(&block_group->lock);
3551                 goto out_put;
3552         }
3553         spin_unlock(&block_group->lock);
3554
3555         /*
3556          * We hit an ENOSPC when setting up the cache in this transaction, just
3557          * skip doing the setup, we've already cleared the cache so we're safe.
3558          */
3559         if (test_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags)) {
3560                 ret = -ENOSPC;
3561                 goto out_put;
3562         }
3563
3564         /*
3565          * Try to preallocate enough space based on how big the block group is.
3566          * Keep in mind this has to include any pinned space which could end up
3567          * taking up quite a bit since it's not folded into the other space
3568          * cache.
3569          */
3570         num_pages = div_u64(block_group->key.offset, SZ_256M);
3571         if (!num_pages)
3572                 num_pages = 1;
3573
3574         num_pages *= 16;
3575         num_pages *= PAGE_SIZE;
3576
3577         ret = btrfs_check_data_free_space(inode, &data_reserved, 0, num_pages);
3578         if (ret)
3579                 goto out_put;
3580
3581         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3582                                               num_pages, num_pages,
3583                                               &alloc_hint);
3584         /*
3585          * Our cache requires contiguous chunks so that we don't modify a bunch
3586          * of metadata or split extents when writing the cache out, which means
3587          * we can enospc if we are heavily fragmented in addition to just normal
3588          * out of space conditions.  So if we hit this just skip setting up any
3589          * other block groups for this transaction, maybe we'll unpin enough
3590          * space the next time around.
3591          */
3592         if (!ret)
3593                 dcs = BTRFS_DC_SETUP;
3594         else if (ret == -ENOSPC)
3595                 set_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags);
3596
3597 out_put:
3598         iput(inode);
3599 out_free:
3600         btrfs_release_path(path);
3601 out:
3602         spin_lock(&block_group->lock);
3603         if (!ret && dcs == BTRFS_DC_SETUP)
3604                 block_group->cache_generation = trans->transid;
3605         block_group->disk_cache_state = dcs;
3606         spin_unlock(&block_group->lock);
3607
3608         extent_changeset_free(data_reserved);
3609         return ret;
3610 }
3611
3612 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans,
3613                             struct btrfs_fs_info *fs_info)
3614 {
3615         struct btrfs_block_group_cache *cache, *tmp;
3616         struct btrfs_transaction *cur_trans = trans->transaction;
3617         struct btrfs_path *path;
3618
3619         if (list_empty(&cur_trans->dirty_bgs) ||
3620             !btrfs_test_opt(fs_info, SPACE_CACHE))
3621                 return 0;
3622
3623         path = btrfs_alloc_path();
3624         if (!path)
3625                 return -ENOMEM;
3626
3627         /* Could add new block groups, use _safe just in case */
3628         list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3629                                  dirty_list) {
3630                 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3631                         cache_save_setup(cache, trans, path);
3632         }
3633
3634         btrfs_free_path(path);
3635         return 0;
3636 }
3637
3638 /*
3639  * transaction commit does final block group cache writeback during a
3640  * critical section where nothing is allowed to change the FS.  This is
3641  * required in order for the cache to actually match the block group,
3642  * but can introduce a lot of latency into the commit.
3643  *
3644  * So, btrfs_start_dirty_block_groups is here to kick off block group
3645  * cache IO.  There's a chance we'll have to redo some of it if the
3646  * block group changes again during the commit, but it greatly reduces
3647  * the commit latency by getting rid of the easy block groups while
3648  * we're still allowing others to join the commit.
3649  */
3650 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans)
3651 {
3652         struct btrfs_fs_info *fs_info = trans->fs_info;
3653         struct btrfs_block_group_cache *cache;
3654         struct btrfs_transaction *cur_trans = trans->transaction;
3655         int ret = 0;
3656         int should_put;
3657         struct btrfs_path *path = NULL;
3658         LIST_HEAD(dirty);
3659         struct list_head *io = &cur_trans->io_bgs;
3660         int num_started = 0;
3661         int loops = 0;
3662
3663         spin_lock(&cur_trans->dirty_bgs_lock);
3664         if (list_empty(&cur_trans->dirty_bgs)) {
3665                 spin_unlock(&cur_trans->dirty_bgs_lock);
3666                 return 0;
3667         }
3668         list_splice_init(&cur_trans->dirty_bgs, &dirty);
3669         spin_unlock(&cur_trans->dirty_bgs_lock);
3670
3671 again:
3672         /*
3673          * make sure all the block groups on our dirty list actually
3674          * exist
3675          */
3676         btrfs_create_pending_block_groups(trans);
3677
3678         if (!path) {
3679                 path = btrfs_alloc_path();
3680                 if (!path)
3681                         return -ENOMEM;
3682         }
3683
3684         /*
3685          * cache_write_mutex is here only to save us from balance or automatic
3686          * removal of empty block groups deleting this block group while we are
3687          * writing out the cache
3688          */
3689         mutex_lock(&trans->transaction->cache_write_mutex);
3690         while (!list_empty(&dirty)) {
3691                 cache = list_first_entry(&dirty,
3692                                          struct btrfs_block_group_cache,
3693                                          dirty_list);
3694                 /*
3695                  * this can happen if something re-dirties a block
3696                  * group that is already under IO.  Just wait for it to
3697                  * finish and then do it all again
3698                  */
3699                 if (!list_empty(&cache->io_list)) {
3700                         list_del_init(&cache->io_list);
3701                         btrfs_wait_cache_io(trans, cache, path);
3702                         btrfs_put_block_group(cache);
3703                 }
3704
3705
3706                 /*
3707                  * btrfs_wait_cache_io uses the cache->dirty_list to decide
3708                  * if it should update the cache_state.  Don't delete
3709                  * until after we wait.
3710                  *
3711                  * Since we're not running in the commit critical section
3712                  * we need the dirty_bgs_lock to protect from update_block_group
3713                  */
3714                 spin_lock(&cur_trans->dirty_bgs_lock);
3715                 list_del_init(&cache->dirty_list);
3716                 spin_unlock(&cur_trans->dirty_bgs_lock);
3717
3718                 should_put = 1;
3719
3720                 cache_save_setup(cache, trans, path);
3721
3722                 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
3723                         cache->io_ctl.inode = NULL;
3724                         ret = btrfs_write_out_cache(fs_info, trans,
3725                                                     cache, path);
3726                         if (ret == 0 && cache->io_ctl.inode) {
3727                                 num_started++;
3728                                 should_put = 0;
3729
3730                                 /*
3731                                  * The cache_write_mutex is protecting the
3732                                  * io_list, also refer to the definition of
3733                                  * btrfs_transaction::io_bgs for more details
3734                                  */
3735                                 list_add_tail(&cache->io_list, io);
3736                         } else {
3737                                 /*
3738                                  * if we failed to write the cache, the
3739                                  * generation will be bad and life goes on
3740                                  */
3741                                 ret = 0;
3742                         }
3743                 }
3744                 if (!ret) {
3745                         ret = write_one_cache_group(trans, fs_info,
3746                                                     path, cache);
3747                         /*
3748                          * Our block group might still be attached to the list
3749                          * of new block groups in the transaction handle of some
3750                          * other task (struct btrfs_trans_handle->new_bgs). This
3751                          * means its block group item isn't yet in the extent
3752                          * tree. If this happens ignore the error, as we will
3753                          * try again later in the critical section of the
3754                          * transaction commit.
3755                          */
3756                         if (ret == -ENOENT) {
3757                                 ret = 0;
3758                                 spin_lock(&cur_trans->dirty_bgs_lock);
3759                                 if (list_empty(&cache->dirty_list)) {
3760                                         list_add_tail(&cache->dirty_list,
3761                                                       &cur_trans->dirty_bgs);
3762                                         btrfs_get_block_group(cache);
3763                                 }
3764                                 spin_unlock(&cur_trans->dirty_bgs_lock);
3765                         } else if (ret) {
3766                                 btrfs_abort_transaction(trans, ret);
3767                         }
3768                 }
3769
3770                 /* if its not on the io list, we need to put the block group */
3771                 if (should_put)
3772                         btrfs_put_block_group(cache);
3773
3774                 if (ret)
3775                         break;
3776
3777                 /*
3778                  * Avoid blocking other tasks for too long. It might even save
3779                  * us from writing caches for block groups that are going to be
3780                  * removed.
3781                  */
3782                 mutex_unlock(&trans->transaction->cache_write_mutex);
3783                 mutex_lock(&trans->transaction->cache_write_mutex);
3784         }
3785         mutex_unlock(&trans->transaction->cache_write_mutex);
3786
3787         /*
3788          * go through delayed refs for all the stuff we've just kicked off
3789          * and then loop back (just once)
3790          */
3791         ret = btrfs_run_delayed_refs(trans, 0);
3792         if (!ret && loops == 0) {
3793                 loops++;
3794                 spin_lock(&cur_trans->dirty_bgs_lock);
3795                 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3796                 /*
3797                  * dirty_bgs_lock protects us from concurrent block group
3798                  * deletes too (not just cache_write_mutex).
3799                  */
3800                 if (!list_empty(&dirty)) {
3801                         spin_unlock(&cur_trans->dirty_bgs_lock);
3802                         goto again;
3803                 }
3804                 spin_unlock(&cur_trans->dirty_bgs_lock);
3805         } else if (ret < 0) {
3806                 btrfs_cleanup_dirty_bgs(cur_trans, fs_info);
3807         }
3808
3809         btrfs_free_path(path);
3810         return ret;
3811 }
3812
3813 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3814                                    struct btrfs_fs_info *fs_info)
3815 {
3816         struct btrfs_block_group_cache *cache;
3817         struct btrfs_transaction *cur_trans = trans->transaction;
3818         int ret = 0;
3819         int should_put;
3820         struct btrfs_path *path;
3821         struct list_head *io = &cur_trans->io_bgs;
3822         int num_started = 0;
3823
3824         path = btrfs_alloc_path();
3825         if (!path)
3826                 return -ENOMEM;
3827
3828         /*
3829          * Even though we are in the critical section of the transaction commit,
3830          * we can still have concurrent tasks adding elements to this
3831          * transaction's list of dirty block groups. These tasks correspond to
3832          * endio free space workers started when writeback finishes for a
3833          * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
3834          * allocate new block groups as a result of COWing nodes of the root
3835          * tree when updating the free space inode. The writeback for the space
3836          * caches is triggered by an earlier call to
3837          * btrfs_start_dirty_block_groups() and iterations of the following
3838          * loop.
3839          * Also we want to do the cache_save_setup first and then run the
3840          * delayed refs to make sure we have the best chance at doing this all
3841          * in one shot.
3842          */
3843         spin_lock(&cur_trans->dirty_bgs_lock);
3844         while (!list_empty(&cur_trans->dirty_bgs)) {
3845                 cache = list_first_entry(&cur_trans->dirty_bgs,
3846                                          struct btrfs_block_group_cache,
3847                                          dirty_list);
3848
3849                 /*
3850                  * this can happen if cache_save_setup re-dirties a block
3851                  * group that is already under IO.  Just wait for it to
3852                  * finish and then do it all again
3853                  */
3854                 if (!list_empty(&cache->io_list)) {
3855                         spin_unlock(&cur_trans->dirty_bgs_lock);
3856                         list_del_init(&cache->io_list);
3857                         btrfs_wait_cache_io(trans, cache, path);
3858                         btrfs_put_block_group(cache);
3859                         spin_lock(&cur_trans->dirty_bgs_lock);
3860                 }
3861
3862                 /*
3863                  * don't remove from the dirty list until after we've waited
3864                  * on any pending IO
3865                  */
3866                 list_del_init(&cache->dirty_list);
3867                 spin_unlock(&cur_trans->dirty_bgs_lock);
3868                 should_put = 1;
3869
3870                 cache_save_setup(cache, trans, path);
3871
3872                 if (!ret)
3873                         ret = btrfs_run_delayed_refs(trans,
3874                                                      (unsigned long) -1);
3875
3876                 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3877                         cache->io_ctl.inode = NULL;
3878                         ret = btrfs_write_out_cache(fs_info, trans,
3879                                                     cache, path);
3880                         if (ret == 0 && cache->io_ctl.inode) {
3881                                 num_started++;
3882                                 should_put = 0;
3883                                 list_add_tail(&cache->io_list, io);
3884                         } else {
3885                                 /*
3886                                  * if we failed to write the cache, the
3887                                  * generation will be bad and life goes on
3888                                  */
3889                                 ret = 0;
3890                         }
3891                 }
3892                 if (!ret) {
3893                         ret = write_one_cache_group(trans, fs_info,
3894                                                     path, cache);
3895                         /*
3896                          * One of the free space endio workers might have
3897                          * created a new block group while updating a free space
3898                          * cache's inode (at inode.c:btrfs_finish_ordered_io())
3899                          * and hasn't released its transaction handle yet, in
3900                          * which case the new block group is still attached to
3901                          * its transaction handle and its creation has not
3902                          * finished yet (no block group item in the extent tree
3903                          * yet, etc). If this is the case, wait for all free
3904                          * space endio workers to finish and retry. This is a
3905                          * a very rare case so no need for a more efficient and
3906                          * complex approach.
3907                          */
3908                         if (ret == -ENOENT) {
3909                                 wait_event(cur_trans->writer_wait,
3910                                    atomic_read(&cur_trans->num_writers) == 1);
3911                                 ret = write_one_cache_group(trans, fs_info,
3912                                                             path, cache);
3913                         }
3914                         if (ret)
3915                                 btrfs_abort_transaction(trans, ret);
3916                 }
3917
3918                 /* if its not on the io list, we need to put the block group */
3919                 if (should_put)
3920                         btrfs_put_block_group(cache);
3921                 spin_lock(&cur_trans->dirty_bgs_lock);
3922         }
3923         spin_unlock(&cur_trans->dirty_bgs_lock);
3924
3925         /*
3926          * Refer to the definition of io_bgs member for details why it's safe
3927          * to use it without any locking
3928          */
3929         while (!list_empty(io)) {
3930                 cache = list_first_entry(io, struct btrfs_block_group_cache,
3931                                          io_list);
3932                 list_del_init(&cache->io_list);
3933                 btrfs_wait_cache_io(trans, cache, path);
3934                 btrfs_put_block_group(cache);
3935         }
3936
3937         btrfs_free_path(path);
3938         return ret;
3939 }
3940
3941 int btrfs_extent_readonly(struct btrfs_fs_info *fs_info, u64 bytenr)
3942 {
3943         struct btrfs_block_group_cache *block_group;
3944         int readonly = 0;
3945
3946         block_group = btrfs_lookup_block_group(fs_info, bytenr);
3947         if (!block_group || block_group->ro)
3948                 readonly = 1;
3949         if (block_group)
3950                 btrfs_put_block_group(block_group);
3951         return readonly;
3952 }
3953
3954 bool btrfs_inc_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3955 {
3956         struct btrfs_block_group_cache *bg;
3957         bool ret = true;
3958
3959         bg = btrfs_lookup_block_group(fs_info, bytenr);
3960         if (!bg)
3961                 return false;
3962
3963         spin_lock(&bg->lock);
3964         if (bg->ro)
3965                 ret = false;
3966         else
3967                 atomic_inc(&bg->nocow_writers);
3968         spin_unlock(&bg->lock);
3969
3970         /* no put on block group, done by btrfs_dec_nocow_writers */
3971         if (!ret)
3972                 btrfs_put_block_group(bg);
3973
3974         return ret;
3975
3976 }
3977
3978 void btrfs_dec_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3979 {
3980         struct btrfs_block_group_cache *bg;
3981
3982         bg = btrfs_lookup_block_group(fs_info, bytenr);
3983         ASSERT(bg);
3984         if (atomic_dec_and_test(&bg->nocow_writers))
3985                 wake_up_var(&bg->nocow_writers);
3986         /*
3987          * Once for our lookup and once for the lookup done by a previous call
3988          * to btrfs_inc_nocow_writers()
3989          */
3990         btrfs_put_block_group(bg);
3991         btrfs_put_block_group(bg);
3992 }
3993
3994 void btrfs_wait_nocow_writers(struct btrfs_block_group_cache *bg)
3995 {
3996         wait_var_event(&bg->nocow_writers, !atomic_read(&bg->nocow_writers));
3997 }
3998
3999 static const char *alloc_name(u64 flags)
4000 {
4001         switch (flags) {
4002         case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA:
4003                 return "mixed";
4004         case BTRFS_BLOCK_GROUP_METADATA:
4005                 return "metadata";
4006         case BTRFS_BLOCK_GROUP_DATA:
4007                 return "data";
4008         case BTRFS_BLOCK_GROUP_SYSTEM:
4009                 return "system";
4010         default:
4011                 WARN_ON(1);
4012                 return "invalid-combination";
4013         };
4014 }
4015
4016 static int create_space_info(struct btrfs_fs_info *info, u64 flags,
4017                              struct btrfs_space_info **new)
4018 {
4019
4020         struct btrfs_space_info *space_info;
4021         int i;
4022         int ret;
4023
4024         space_info = kzalloc(sizeof(*space_info), GFP_NOFS);
4025         if (!space_info)
4026                 return -ENOMEM;
4027
4028         ret = percpu_counter_init(&space_info->total_bytes_pinned, 0,
4029                                  GFP_KERNEL);
4030         if (ret) {
4031                 kfree(space_info);
4032                 return ret;
4033         }
4034
4035         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
4036                 INIT_LIST_HEAD(&space_info->block_groups[i]);
4037         init_rwsem(&space_info->groups_sem);
4038         spin_lock_init(&space_info->lock);
4039         space_info->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
4040         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4041         init_waitqueue_head(&space_info->wait);
4042         INIT_LIST_HEAD(&space_info->ro_bgs);
4043         INIT_LIST_HEAD(&space_info->tickets);
4044         INIT_LIST_HEAD(&space_info->priority_tickets);
4045
4046         ret = kobject_init_and_add(&space_info->kobj, &space_info_ktype,
4047                                     info->space_info_kobj, "%s",
4048                                     alloc_name(space_info->flags));
4049         if (ret) {
4050                 percpu_counter_destroy(&space_info->total_bytes_pinned);
4051                 kfree(space_info);
4052                 return ret;
4053         }
4054
4055         *new = space_info;
4056         list_add_rcu(&space_info->list, &info->space_info);
4057         if (flags & BTRFS_BLOCK_GROUP_DATA)
4058                 info->data_sinfo = space_info;
4059
4060         return ret;
4061 }
4062
4063 static void update_space_info(struct btrfs_fs_info *info, u64 flags,
4064                              u64 total_bytes, u64 bytes_used,
4065                              u64 bytes_readonly,
4066                              struct btrfs_space_info **space_info)
4067 {
4068         struct btrfs_space_info *found;
4069         int factor;
4070
4071         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
4072                      BTRFS_BLOCK_GROUP_RAID10))
4073                 factor = 2;
4074         else
4075                 factor = 1;
4076
4077         found = __find_space_info(info, flags);
4078         ASSERT(found);
4079         spin_lock(&found->lock);
4080         found->total_bytes += total_bytes;
4081         found->disk_total += total_bytes * factor;
4082         found->bytes_used += bytes_used;
4083         found->disk_used += bytes_used * factor;
4084         found->bytes_readonly += bytes_readonly;
4085         if (total_bytes > 0)
4086                 found->full = 0;
4087         space_info_add_new_bytes(info, found, total_bytes -
4088                                  bytes_used - bytes_readonly);
4089         spin_unlock(&found->lock);
4090         *space_info = found;
4091 }
4092
4093 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
4094 {
4095         u64 extra_flags = chunk_to_extended(flags) &
4096                                 BTRFS_EXTENDED_PROFILE_MASK;
4097
4098         write_seqlock(&fs_info->profiles_lock);
4099         if (flags & BTRFS_BLOCK_GROUP_DATA)
4100                 fs_info->avail_data_alloc_bits |= extra_flags;
4101         if (flags & BTRFS_BLOCK_GROUP_METADATA)
4102                 fs_info->avail_metadata_alloc_bits |= extra_flags;
4103         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4104                 fs_info->avail_system_alloc_bits |= extra_flags;
4105         write_sequnlock(&fs_info->profiles_lock);
4106 }
4107
4108 /*
4109  * returns target flags in extended format or 0 if restripe for this
4110  * chunk_type is not in progress
4111  *
4112  * should be called with either volume_mutex or balance_lock held
4113  */
4114 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
4115 {
4116         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
4117         u64 target = 0;
4118
4119         if (!bctl)
4120                 return 0;
4121
4122         if (flags & BTRFS_BLOCK_GROUP_DATA &&
4123             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4124                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
4125         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
4126                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4127                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
4128         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
4129                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4130                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
4131         }
4132
4133         return target;
4134 }
4135
4136 /*
4137  * @flags: available profiles in extended format (see ctree.h)
4138  *
4139  * Returns reduced profile in chunk format.  If profile changing is in
4140  * progress (either running or paused) picks the target profile (if it's
4141  * already available), otherwise falls back to plain reducing.
4142  */
4143 static u64 btrfs_reduce_alloc_profile(struct btrfs_fs_info *fs_info, u64 flags)
4144 {
4145         u64 num_devices = fs_info->fs_devices->rw_devices;
4146         u64 target;
4147         u64 raid_type;
4148         u64 allowed = 0;
4149
4150         /*
4151          * see if restripe for this chunk_type is in progress, if so
4152          * try to reduce to the target profile
4153          */
4154         spin_lock(&fs_info->balance_lock);
4155         target = get_restripe_target(fs_info, flags);
4156         if (target) {
4157                 /* pick target profile only if it's already available */
4158                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
4159                         spin_unlock(&fs_info->balance_lock);
4160                         return extended_to_chunk(target);
4161                 }
4162         }
4163         spin_unlock(&fs_info->balance_lock);
4164
4165         /* First, mask out the RAID levels which aren't possible */
4166         for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
4167                 if (num_devices >= btrfs_raid_array[raid_type].devs_min)
4168                         allowed |= btrfs_raid_group[raid_type];
4169         }
4170         allowed &= flags;
4171
4172         if (allowed & BTRFS_BLOCK_GROUP_RAID6)
4173                 allowed = BTRFS_BLOCK_GROUP_RAID6;
4174         else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
4175                 allowed = BTRFS_BLOCK_GROUP_RAID5;
4176         else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
4177                 allowed = BTRFS_BLOCK_GROUP_RAID10;
4178         else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
4179                 allowed = BTRFS_BLOCK_GROUP_RAID1;
4180         else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
4181                 allowed = BTRFS_BLOCK_GROUP_RAID0;
4182
4183         flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
4184
4185         return extended_to_chunk(flags | allowed);
4186 }
4187
4188 static u64 get_alloc_profile(struct btrfs_fs_info *fs_info, u64 orig_flags)
4189 {
4190         unsigned seq;
4191         u64 flags;
4192
4193         do {
4194                 flags = orig_flags;
4195                 seq = read_seqbegin(&fs_info->profiles_lock);
4196
4197                 if (flags & BTRFS_BLOCK_GROUP_DATA)
4198                         flags |= fs_info->avail_data_alloc_bits;
4199                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4200                         flags |= fs_info->avail_system_alloc_bits;
4201                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
4202                         flags |= fs_info->avail_metadata_alloc_bits;
4203         } while (read_seqretry(&fs_info->profiles_lock, seq));
4204
4205         return btrfs_reduce_alloc_profile(fs_info, flags);
4206 }
4207
4208 static u64 get_alloc_profile_by_root(struct btrfs_root *root, int data)
4209 {
4210         struct btrfs_fs_info *fs_info = root->fs_info;
4211         u64 flags;
4212         u64 ret;
4213
4214         if (data)
4215                 flags = BTRFS_BLOCK_GROUP_DATA;
4216         else if (root == fs_info->chunk_root)
4217                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
4218         else
4219                 flags = BTRFS_BLOCK_GROUP_METADATA;
4220
4221         ret = get_alloc_profile(fs_info, flags);
4222         return ret;
4223 }
4224
4225 u64 btrfs_data_alloc_profile(struct btrfs_fs_info *fs_info)
4226 {
4227         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_DATA);
4228 }
4229
4230 u64 btrfs_metadata_alloc_profile(struct btrfs_fs_info *fs_info)
4231 {
4232         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4233 }
4234
4235 u64 btrfs_system_alloc_profile(struct btrfs_fs_info *fs_info)
4236 {
4237         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4238 }
4239
4240 static u64 btrfs_space_info_used(struct btrfs_space_info *s_info,
4241                                  bool may_use_included)
4242 {
4243         ASSERT(s_info);
4244         return s_info->bytes_used + s_info->bytes_reserved +
4245                 s_info->bytes_pinned + s_info->bytes_readonly +
4246                 (may_use_included ? s_info->bytes_may_use : 0);
4247 }
4248
4249 int btrfs_alloc_data_chunk_ondemand(struct btrfs_inode *inode, u64 bytes)
4250 {
4251         struct btrfs_root *root = inode->root;
4252         struct btrfs_fs_info *fs_info = root->fs_info;
4253         struct btrfs_space_info *data_sinfo = fs_info->data_sinfo;
4254         u64 used;
4255         int ret = 0;
4256         int need_commit = 2;
4257         int have_pinned_space;
4258
4259         /* make sure bytes are sectorsize aligned */
4260         bytes = ALIGN(bytes, fs_info->sectorsize);
4261
4262         if (btrfs_is_free_space_inode(inode)) {
4263                 need_commit = 0;
4264                 ASSERT(current->journal_info);
4265         }
4266
4267 again:
4268         /* make sure we have enough space to handle the data first */
4269         spin_lock(&data_sinfo->lock);
4270         used = btrfs_space_info_used(data_sinfo, true);
4271
4272         if (used + bytes > data_sinfo->total_bytes) {
4273                 struct btrfs_trans_handle *trans;
4274
4275                 /*
4276                  * if we don't have enough free bytes in this space then we need
4277                  * to alloc a new chunk.
4278                  */
4279                 if (!data_sinfo->full) {
4280                         u64 alloc_target;
4281
4282                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
4283                         spin_unlock(&data_sinfo->lock);
4284
4285                         alloc_target = btrfs_data_alloc_profile(fs_info);
4286                         /*
4287                          * It is ugly that we don't call nolock join
4288                          * transaction for the free space inode case here.
4289                          * But it is safe because we only do the data space
4290                          * reservation for the free space cache in the
4291                          * transaction context, the common join transaction
4292                          * just increase the counter of the current transaction
4293                          * handler, doesn't try to acquire the trans_lock of
4294                          * the fs.
4295                          */
4296                         trans = btrfs_join_transaction(root);
4297                         if (IS_ERR(trans))
4298                                 return PTR_ERR(trans);
4299
4300                         ret = do_chunk_alloc(trans, fs_info, alloc_target,
4301                                              CHUNK_ALLOC_NO_FORCE);
4302                         btrfs_end_transaction(trans);
4303                         if (ret < 0) {
4304                                 if (ret != -ENOSPC)
4305                                         return ret;
4306                                 else {
4307                                         have_pinned_space = 1;
4308                                         goto commit_trans;
4309                                 }
4310                         }
4311
4312                         goto again;
4313                 }
4314
4315                 /*
4316                  * If we don't have enough pinned space to deal with this
4317                  * allocation, and no removed chunk in current transaction,
4318                  * don't bother committing the transaction.
4319                  */
4320                 have_pinned_space = percpu_counter_compare(
4321                         &data_sinfo->total_bytes_pinned,
4322                         used + bytes - data_sinfo->total_bytes);
4323                 spin_unlock(&data_sinfo->lock);
4324
4325                 /* commit the current transaction and try again */
4326 commit_trans:
4327                 if (need_commit) {
4328                         need_commit--;
4329
4330                         if (need_commit > 0) {
4331                                 btrfs_start_delalloc_roots(fs_info, 0, -1);
4332                                 btrfs_wait_ordered_roots(fs_info, U64_MAX, 0,
4333                                                          (u64)-1);
4334                         }
4335
4336                         trans = btrfs_join_transaction(root);
4337                         if (IS_ERR(trans))
4338                                 return PTR_ERR(trans);
4339                         if (have_pinned_space >= 0 ||
4340                             test_bit(BTRFS_TRANS_HAVE_FREE_BGS,
4341                                      &trans->transaction->flags) ||
4342                             need_commit > 0) {
4343                                 ret = btrfs_commit_transaction(trans);
4344                                 if (ret)
4345                                         return ret;
4346                                 /*
4347                                  * The cleaner kthread might still be doing iput
4348                                  * operations. Wait for it to finish so that
4349                                  * more space is released.
4350                                  */
4351                                 mutex_lock(&fs_info->cleaner_delayed_iput_mutex);
4352                                 mutex_unlock(&fs_info->cleaner_delayed_iput_mutex);
4353                                 goto again;
4354                         } else {
4355                                 btrfs_end_transaction(trans);
4356                         }
4357                 }
4358
4359                 trace_btrfs_space_reservation(fs_info,
4360                                               "space_info:enospc",
4361                                               data_sinfo->flags, bytes, 1);
4362                 return -ENOSPC;
4363         }
4364         data_sinfo->bytes_may_use += bytes;
4365         trace_btrfs_space_reservation(fs_info, "space_info",
4366                                       data_sinfo->flags, bytes, 1);
4367         spin_unlock(&data_sinfo->lock);
4368
4369         return ret;
4370 }
4371
4372 int btrfs_check_data_free_space(struct inode *inode,
4373                         struct extent_changeset **reserved, u64 start, u64 len)
4374 {
4375         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4376         int ret;
4377
4378         /* align the range */
4379         len = round_up(start + len, fs_info->sectorsize) -
4380               round_down(start, fs_info->sectorsize);
4381         start = round_down(start, fs_info->sectorsize);
4382
4383         ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode), len);
4384         if (ret < 0)
4385                 return ret;
4386
4387         /* Use new btrfs_qgroup_reserve_data to reserve precious data space. */
4388         ret = btrfs_qgroup_reserve_data(inode, reserved, start, len);
4389         if (ret < 0)
4390                 btrfs_free_reserved_data_space_noquota(inode, start, len);
4391         else
4392                 ret = 0;
4393         return ret;
4394 }
4395
4396 /*
4397  * Called if we need to clear a data reservation for this inode
4398  * Normally in a error case.
4399  *
4400  * This one will *NOT* use accurate qgroup reserved space API, just for case
4401  * which we can't sleep and is sure it won't affect qgroup reserved space.
4402  * Like clear_bit_hook().
4403  */
4404 void btrfs_free_reserved_data_space_noquota(struct inode *inode, u64 start,
4405                                             u64 len)
4406 {
4407         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4408         struct btrfs_space_info *data_sinfo;
4409
4410         /* Make sure the range is aligned to sectorsize */
4411         len = round_up(start + len, fs_info->sectorsize) -
4412               round_down(start, fs_info->sectorsize);
4413         start = round_down(start, fs_info->sectorsize);
4414
4415         data_sinfo = fs_info->data_sinfo;
4416         spin_lock(&data_sinfo->lock);
4417         if (WARN_ON(data_sinfo->bytes_may_use < len))
4418                 data_sinfo->bytes_may_use = 0;
4419         else
4420                 data_sinfo->bytes_may_use -= len;
4421         trace_btrfs_space_reservation(fs_info, "space_info",
4422                                       data_sinfo->flags, len, 0);
4423         spin_unlock(&data_sinfo->lock);
4424 }
4425
4426 /*
4427  * Called if we need to clear a data reservation for this inode
4428  * Normally in a error case.
4429  *
4430  * This one will handle the per-inode data rsv map for accurate reserved
4431  * space framework.
4432  */
4433 void btrfs_free_reserved_data_space(struct inode *inode,
4434                         struct extent_changeset *reserved, u64 start, u64 len)
4435 {
4436         struct btrfs_root *root = BTRFS_I(inode)->root;
4437
4438         /* Make sure the range is aligned to sectorsize */
4439         len = round_up(start + len, root->fs_info->sectorsize) -
4440               round_down(start, root->fs_info->sectorsize);
4441         start = round_down(start, root->fs_info->sectorsize);
4442
4443         btrfs_free_reserved_data_space_noquota(inode, start, len);
4444         btrfs_qgroup_free_data(inode, reserved, start, len);
4445 }
4446
4447 static void force_metadata_allocation(struct btrfs_fs_info *info)
4448 {
4449         struct list_head *head = &info->space_info;
4450         struct btrfs_space_info *found;
4451
4452         rcu_read_lock();
4453         list_for_each_entry_rcu(found, head, list) {
4454                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
4455                         found->force_alloc = CHUNK_ALLOC_FORCE;
4456         }
4457         rcu_read_unlock();
4458 }
4459
4460 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
4461 {
4462         return (global->size << 1);
4463 }
4464
4465 static int should_alloc_chunk(struct btrfs_fs_info *fs_info,
4466                               struct btrfs_space_info *sinfo, int force)
4467 {
4468         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4469         u64 bytes_used = btrfs_space_info_used(sinfo, false);
4470         u64 thresh;
4471
4472         if (force == CHUNK_ALLOC_FORCE)
4473                 return 1;
4474
4475         /*
4476          * We need to take into account the global rsv because for all intents
4477          * and purposes it's used space.  Don't worry about locking the
4478          * global_rsv, it doesn't change except when the transaction commits.
4479          */
4480         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
4481                 bytes_used += calc_global_rsv_need_space(global_rsv);
4482
4483         /*
4484          * in limited mode, we want to have some free space up to
4485          * about 1% of the FS size.
4486          */
4487         if (force == CHUNK_ALLOC_LIMITED) {
4488                 thresh = btrfs_super_total_bytes(fs_info->super_copy);
4489                 thresh = max_t(u64, SZ_64M, div_factor_fine(thresh, 1));
4490
4491                 if (sinfo->total_bytes - bytes_used < thresh)
4492                         return 1;
4493         }
4494
4495         if (bytes_used + SZ_2M < div_factor(sinfo->total_bytes, 8))
4496                 return 0;
4497         return 1;
4498 }
4499
4500 static u64 get_profile_num_devs(struct btrfs_fs_info *fs_info, u64 type)
4501 {
4502         u64 num_dev;
4503
4504         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
4505                     BTRFS_BLOCK_GROUP_RAID0 |
4506                     BTRFS_BLOCK_GROUP_RAID5 |
4507                     BTRFS_BLOCK_GROUP_RAID6))
4508                 num_dev = fs_info->fs_devices->rw_devices;
4509         else if (type & BTRFS_BLOCK_GROUP_RAID1)
4510                 num_dev = 2;
4511         else
4512                 num_dev = 1;    /* DUP or single */
4513
4514         return num_dev;
4515 }
4516
4517 /*
4518  * If @is_allocation is true, reserve space in the system space info necessary
4519  * for allocating a chunk, otherwise if it's false, reserve space necessary for
4520  * removing a chunk.
4521  */
4522 void check_system_chunk(struct btrfs_trans_handle *trans,
4523                         struct btrfs_fs_info *fs_info, u64 type)
4524 {
4525         struct btrfs_space_info *info;
4526         u64 left;
4527         u64 thresh;
4528         int ret = 0;
4529         u64 num_devs;
4530
4531         /*
4532          * Needed because we can end up allocating a system chunk and for an
4533          * atomic and race free space reservation in the chunk block reserve.
4534          */
4535         lockdep_assert_held(&fs_info->chunk_mutex);
4536
4537         info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4538         spin_lock(&info->lock);
4539         left = info->total_bytes - btrfs_space_info_used(info, true);
4540         spin_unlock(&info->lock);
4541
4542         num_devs = get_profile_num_devs(fs_info, type);
4543
4544         /* num_devs device items to update and 1 chunk item to add or remove */
4545         thresh = btrfs_calc_trunc_metadata_size(fs_info, num_devs) +
4546                 btrfs_calc_trans_metadata_size(fs_info, 1);
4547
4548         if (left < thresh && btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
4549                 btrfs_info(fs_info, "left=%llu, need=%llu, flags=%llu",
4550                            left, thresh, type);
4551                 dump_space_info(fs_info, info, 0, 0);
4552         }
4553
4554         if (left < thresh) {
4555                 u64 flags = btrfs_system_alloc_profile(fs_info);
4556
4557                 /*
4558                  * Ignore failure to create system chunk. We might end up not
4559                  * needing it, as we might not need to COW all nodes/leafs from
4560                  * the paths we visit in the chunk tree (they were already COWed
4561                  * or created in the current transaction for example).
4562                  */
4563                 ret = btrfs_alloc_chunk(trans, fs_info, flags);
4564         }
4565
4566         if (!ret) {
4567                 ret = btrfs_block_rsv_add(fs_info->chunk_root,
4568                                           &fs_info->chunk_block_rsv,
4569                                           thresh, BTRFS_RESERVE_NO_FLUSH);
4570                 if (!ret)
4571                         trans->chunk_bytes_reserved += thresh;
4572         }
4573 }
4574
4575 /*
4576  * If force is CHUNK_ALLOC_FORCE:
4577  *    - return 1 if it successfully allocates a chunk,
4578  *    - return errors including -ENOSPC otherwise.
4579  * If force is NOT CHUNK_ALLOC_FORCE:
4580  *    - return 0 if it doesn't need to allocate a new chunk,
4581  *    - return 1 if it successfully allocates a chunk,
4582  *    - return errors including -ENOSPC otherwise.
4583  */
4584 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
4585                           struct btrfs_fs_info *fs_info, u64 flags, int force)
4586 {
4587         struct btrfs_space_info *space_info;
4588         int wait_for_alloc = 0;
4589         int ret = 0;
4590
4591         /* Don't re-enter if we're already allocating a chunk */
4592         if (trans->allocating_chunk)
4593                 return -ENOSPC;
4594
4595         space_info = __find_space_info(fs_info, flags);
4596         ASSERT(space_info);
4597
4598 again:
4599         spin_lock(&space_info->lock);
4600         if (force < space_info->force_alloc)
4601                 force = space_info->force_alloc;
4602         if (space_info->full) {
4603                 if (should_alloc_chunk(fs_info, space_info, force))
4604                         ret = -ENOSPC;
4605                 else
4606                         ret = 0;
4607                 spin_unlock(&space_info->lock);
4608                 return ret;
4609         }
4610
4611         if (!should_alloc_chunk(fs_info, space_info, force)) {
4612                 spin_unlock(&space_info->lock);
4613                 return 0;
4614         } else if (space_info->chunk_alloc) {
4615                 wait_for_alloc = 1;
4616         } else {
4617                 space_info->chunk_alloc = 1;
4618         }
4619
4620         spin_unlock(&space_info->lock);
4621
4622         mutex_lock(&fs_info->chunk_mutex);
4623
4624         /*
4625          * The chunk_mutex is held throughout the entirety of a chunk
4626          * allocation, so once we've acquired the chunk_mutex we know that the
4627          * other guy is done and we need to recheck and see if we should
4628          * allocate.
4629          */
4630         if (wait_for_alloc) {
4631                 mutex_unlock(&fs_info->chunk_mutex);
4632                 wait_for_alloc = 0;
4633                 cond_resched();
4634                 goto again;
4635         }
4636
4637         trans->allocating_chunk = true;
4638
4639         /*
4640          * If we have mixed data/metadata chunks we want to make sure we keep
4641          * allocating mixed chunks instead of individual chunks.
4642          */
4643         if (btrfs_mixed_space_info(space_info))
4644                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4645
4646         /*
4647          * if we're doing a data chunk, go ahead and make sure that
4648          * we keep a reasonable number of metadata chunks allocated in the
4649          * FS as well.
4650          */
4651         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4652                 fs_info->data_chunk_allocations++;
4653                 if (!(fs_info->data_chunk_allocations %
4654                       fs_info->metadata_ratio))
4655                         force_metadata_allocation(fs_info);
4656         }
4657
4658         /*
4659          * Check if we have enough space in SYSTEM chunk because we may need
4660          * to update devices.
4661          */
4662         check_system_chunk(trans, fs_info, flags);
4663
4664         ret = btrfs_alloc_chunk(trans, fs_info, flags);
4665         trans->allocating_chunk = false;
4666
4667         spin_lock(&space_info->lock);
4668         if (ret < 0 && ret != -ENOSPC)
4669                 goto out;
4670         if (ret)
4671                 space_info->full = 1;
4672         else
4673                 ret = 1;
4674
4675         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4676 out:
4677         space_info->chunk_alloc = 0;
4678         spin_unlock(&space_info->lock);
4679         mutex_unlock(&fs_info->chunk_mutex);
4680         /*
4681          * When we allocate a new chunk we reserve space in the chunk block
4682          * reserve to make sure we can COW nodes/leafs in the chunk tree or
4683          * add new nodes/leafs to it if we end up needing to do it when
4684          * inserting the chunk item and updating device items as part of the
4685          * second phase of chunk allocation, performed by
4686          * btrfs_finish_chunk_alloc(). So make sure we don't accumulate a
4687          * large number of new block groups to create in our transaction
4688          * handle's new_bgs list to avoid exhausting the chunk block reserve
4689          * in extreme cases - like having a single transaction create many new
4690          * block groups when starting to write out the free space caches of all
4691          * the block groups that were made dirty during the lifetime of the
4692          * transaction.
4693          */
4694         if (trans->can_flush_pending_bgs &&
4695             trans->chunk_bytes_reserved >= (u64)SZ_2M) {
4696                 btrfs_create_pending_block_groups(trans);
4697                 btrfs_trans_release_chunk_metadata(trans);
4698         }
4699         return ret;
4700 }
4701
4702 static int can_overcommit(struct btrfs_fs_info *fs_info,
4703                           struct btrfs_space_info *space_info, u64 bytes,
4704                           enum btrfs_reserve_flush_enum flush,
4705                           bool system_chunk)
4706 {
4707         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4708         u64 profile;
4709         u64 space_size;
4710         u64 avail;
4711         u64 used;
4712
4713         /* Don't overcommit when in mixed mode. */
4714         if (space_info->flags & BTRFS_BLOCK_GROUP_DATA)
4715                 return 0;
4716
4717         if (system_chunk)
4718                 profile = btrfs_system_alloc_profile(fs_info);
4719         else
4720                 profile = btrfs_metadata_alloc_profile(fs_info);
4721
4722         used = btrfs_space_info_used(space_info, false);
4723
4724         /*
4725          * We only want to allow over committing if we have lots of actual space
4726          * free, but if we don't have enough space to handle the global reserve
4727          * space then we could end up having a real enospc problem when trying
4728          * to allocate a chunk or some other such important allocation.
4729          */
4730         spin_lock(&global_rsv->lock);
4731         space_size = calc_global_rsv_need_space(global_rsv);
4732         spin_unlock(&global_rsv->lock);
4733         if (used + space_size >= space_info->total_bytes)
4734                 return 0;
4735
4736         used += space_info->bytes_may_use;
4737
4738         avail = atomic64_read(&fs_info->free_chunk_space);
4739
4740         /*
4741          * If we have dup, raid1 or raid10 then only half of the free
4742          * space is actually useable.  For raid56, the space info used
4743          * doesn't include the parity drive, so we don't have to
4744          * change the math
4745          */
4746         if (profile & (BTRFS_BLOCK_GROUP_DUP |
4747                        BTRFS_BLOCK_GROUP_RAID1 |
4748                        BTRFS_BLOCK_GROUP_RAID10))
4749                 avail >>= 1;
4750
4751         /*
4752          * If we aren't flushing all things, let us overcommit up to
4753          * 1/2th of the space. If we can flush, don't let us overcommit
4754          * too much, let it overcommit up to 1/8 of the space.
4755          */
4756         if (flush == BTRFS_RESERVE_FLUSH_ALL)
4757                 avail >>= 3;
4758         else
4759                 avail >>= 1;
4760
4761         if (used + bytes < space_info->total_bytes + avail)
4762                 return 1;
4763         return 0;
4764 }
4765
4766 static void btrfs_writeback_inodes_sb_nr(struct btrfs_fs_info *fs_info,
4767                                          unsigned long nr_pages, int nr_items)
4768 {
4769         struct super_block *sb = fs_info->sb;
4770
4771         if (down_read_trylock(&sb->s_umount)) {
4772                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
4773                 up_read(&sb->s_umount);
4774         } else {
4775                 /*
4776                  * We needn't worry the filesystem going from r/w to r/o though
4777                  * we don't acquire ->s_umount mutex, because the filesystem
4778                  * should guarantee the delalloc inodes list be empty after
4779                  * the filesystem is readonly(all dirty pages are written to
4780                  * the disk).
4781                  */
4782                 btrfs_start_delalloc_roots(fs_info, 0, nr_items);
4783                 if (!current->journal_info)
4784                         btrfs_wait_ordered_roots(fs_info, nr_items, 0, (u64)-1);
4785         }
4786 }
4787
4788 static inline u64 calc_reclaim_items_nr(struct btrfs_fs_info *fs_info,
4789                                         u64 to_reclaim)
4790 {
4791         u64 bytes;
4792         u64 nr;
4793
4794         bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
4795         nr = div64_u64(to_reclaim, bytes);
4796         if (!nr)
4797                 nr = 1;
4798         return nr;
4799 }
4800
4801 #define EXTENT_SIZE_PER_ITEM    SZ_256K
4802
4803 /*
4804  * shrink metadata reservation for delalloc
4805  */
4806 static void shrink_delalloc(struct btrfs_fs_info *fs_info, u64 to_reclaim,
4807                             u64 orig, bool wait_ordered)
4808 {
4809         struct btrfs_space_info *space_info;
4810         struct btrfs_trans_handle *trans;
4811         u64 delalloc_bytes;
4812         u64 max_reclaim;
4813         u64 items;
4814         long time_left;
4815         unsigned long nr_pages;
4816         int loops;
4817
4818         /* Calc the number of the pages we need flush for space reservation */
4819         items = calc_reclaim_items_nr(fs_info, to_reclaim);
4820         to_reclaim = items * EXTENT_SIZE_PER_ITEM;
4821
4822         trans = (struct btrfs_trans_handle *)current->journal_info;
4823         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4824
4825         delalloc_bytes = percpu_counter_sum_positive(
4826                                                 &fs_info->delalloc_bytes);
4827         if (delalloc_bytes == 0) {
4828                 if (trans)
4829                         return;
4830                 if (wait_ordered)
4831                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
4832                 return;
4833         }
4834
4835         loops = 0;
4836         while (delalloc_bytes && loops < 3) {
4837                 max_reclaim = min(delalloc_bytes, to_reclaim);
4838                 nr_pages = max_reclaim >> PAGE_SHIFT;
4839                 btrfs_writeback_inodes_sb_nr(fs_info, nr_pages, items);
4840                 /*
4841                  * We need to wait for the async pages to actually start before
4842                  * we do anything.
4843                  */
4844                 max_reclaim = atomic_read(&fs_info->async_delalloc_pages);
4845                 if (!max_reclaim)
4846                         goto skip_async;
4847
4848                 if (max_reclaim <= nr_pages)
4849                         max_reclaim = 0;
4850                 else
4851                         max_reclaim -= nr_pages;
4852
4853                 wait_event(fs_info->async_submit_wait,
4854                            atomic_read(&fs_info->async_delalloc_pages) <=
4855                            (int)max_reclaim);
4856 skip_async:
4857                 spin_lock(&space_info->lock);
4858                 if (list_empty(&space_info->tickets) &&
4859                     list_empty(&space_info->priority_tickets)) {
4860                         spin_unlock(&space_info->lock);
4861                         break;
4862                 }
4863                 spin_unlock(&space_info->lock);
4864
4865                 loops++;
4866                 if (wait_ordered && !trans) {
4867                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
4868                 } else {
4869                         time_left = schedule_timeout_killable(1);
4870                         if (time_left)
4871                                 break;
4872                 }
4873                 delalloc_bytes = percpu_counter_sum_positive(
4874                                                 &fs_info->delalloc_bytes);
4875         }
4876 }
4877
4878 struct reserve_ticket {
4879         u64 bytes;
4880         int error;
4881         struct list_head list;
4882         wait_queue_head_t wait;
4883 };
4884
4885 /**
4886  * maybe_commit_transaction - possibly commit the transaction if its ok to
4887  * @root - the root we're allocating for
4888  * @bytes - the number of bytes we want to reserve
4889  * @force - force the commit
4890  *
4891  * This will check to make sure that committing the transaction will actually
4892  * get us somewhere and then commit the transaction if it does.  Otherwise it
4893  * will return -ENOSPC.
4894  */
4895 static int may_commit_transaction(struct btrfs_fs_info *fs_info,
4896                                   struct btrfs_space_info *space_info)
4897 {
4898         struct reserve_ticket *ticket = NULL;
4899         struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_block_rsv;
4900         struct btrfs_trans_handle *trans;
4901         u64 bytes;
4902
4903         trans = (struct btrfs_trans_handle *)current->journal_info;
4904         if (trans)
4905                 return -EAGAIN;
4906
4907         spin_lock(&space_info->lock);
4908         if (!list_empty(&space_info->priority_tickets))
4909                 ticket = list_first_entry(&space_info->priority_tickets,
4910                                           struct reserve_ticket, list);
4911         else if (!list_empty(&space_info->tickets))
4912                 ticket = list_first_entry(&space_info->tickets,
4913                                           struct reserve_ticket, list);
4914         bytes = (ticket) ? ticket->bytes : 0;
4915         spin_unlock(&space_info->lock);
4916
4917         if (!bytes)
4918                 return 0;
4919
4920         /* See if there is enough pinned space to make this reservation */
4921         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4922                                    bytes) >= 0)
4923                 goto commit;
4924
4925         /*
4926          * See if there is some space in the delayed insertion reservation for
4927          * this reservation.
4928          */
4929         if (space_info != delayed_rsv->space_info)
4930                 return -ENOSPC;
4931
4932         spin_lock(&delayed_rsv->lock);
4933         if (delayed_rsv->size > bytes)
4934                 bytes = 0;
4935         else
4936                 bytes -= delayed_rsv->size;
4937         spin_unlock(&delayed_rsv->lock);
4938
4939         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4940                                    bytes) < 0) {
4941                 return -ENOSPC;
4942         }
4943
4944 commit:
4945         trans = btrfs_join_transaction(fs_info->extent_root);
4946         if (IS_ERR(trans))
4947                 return -ENOSPC;
4948
4949         return btrfs_commit_transaction(trans);
4950 }
4951
4952 /*
4953  * Try to flush some data based on policy set by @state. This is only advisory
4954  * and may fail for various reasons. The caller is supposed to examine the
4955  * state of @space_info to detect the outcome.
4956  */
4957 static void flush_space(struct btrfs_fs_info *fs_info,
4958                        struct btrfs_space_info *space_info, u64 num_bytes,
4959                        int state)
4960 {
4961         struct btrfs_root *root = fs_info->extent_root;
4962         struct btrfs_trans_handle *trans;
4963         int nr;
4964         int ret = 0;
4965
4966         switch (state) {
4967         case FLUSH_DELAYED_ITEMS_NR:
4968         case FLUSH_DELAYED_ITEMS:
4969                 if (state == FLUSH_DELAYED_ITEMS_NR)
4970                         nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2;
4971                 else
4972                         nr = -1;
4973
4974                 trans = btrfs_join_transaction(root);
4975                 if (IS_ERR(trans)) {
4976                         ret = PTR_ERR(trans);
4977                         break;
4978                 }
4979                 ret = btrfs_run_delayed_items_nr(trans, nr);
4980                 btrfs_end_transaction(trans);
4981                 break;
4982         case FLUSH_DELALLOC:
4983         case FLUSH_DELALLOC_WAIT:
4984                 shrink_delalloc(fs_info, num_bytes * 2, num_bytes,
4985                                 state == FLUSH_DELALLOC_WAIT);
4986                 break;
4987         case ALLOC_CHUNK:
4988                 trans = btrfs_join_transaction(root);
4989                 if (IS_ERR(trans)) {
4990                         ret = PTR_ERR(trans);
4991                         break;
4992                 }
4993                 ret = do_chunk_alloc(trans, fs_info,
4994                                      btrfs_metadata_alloc_profile(fs_info),
4995                                      CHUNK_ALLOC_NO_FORCE);
4996                 btrfs_end_transaction(trans);
4997                 if (ret > 0 || ret == -ENOSPC)
4998                         ret = 0;
4999                 break;
5000         case COMMIT_TRANS:
5001                 ret = may_commit_transaction(fs_info, space_info);
5002                 break;
5003         default:
5004                 ret = -ENOSPC;
5005                 break;
5006         }
5007
5008         trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state,
5009                                 ret);
5010         return;
5011 }
5012
5013 static inline u64
5014 btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info,
5015                                  struct btrfs_space_info *space_info,
5016                                  bool system_chunk)
5017 {
5018         struct reserve_ticket *ticket;
5019         u64 used;
5020         u64 expected;
5021         u64 to_reclaim = 0;
5022
5023         list_for_each_entry(ticket, &space_info->tickets, list)
5024                 to_reclaim += ticket->bytes;
5025         list_for_each_entry(ticket, &space_info->priority_tickets, list)
5026                 to_reclaim += ticket->bytes;
5027         if (to_reclaim)
5028                 return to_reclaim;
5029
5030         to_reclaim = min_t(u64, num_online_cpus() * SZ_1M, SZ_16M);
5031         if (can_overcommit(fs_info, space_info, to_reclaim,
5032                            BTRFS_RESERVE_FLUSH_ALL, system_chunk))
5033                 return 0;
5034
5035         used = btrfs_space_info_used(space_info, true);
5036
5037         if (can_overcommit(fs_info, space_info, SZ_1M,
5038                            BTRFS_RESERVE_FLUSH_ALL, system_chunk))
5039                 expected = div_factor_fine(space_info->total_bytes, 95);
5040         else
5041                 expected = div_factor_fine(space_info->total_bytes, 90);
5042
5043         if (used > expected)
5044                 to_reclaim = used - expected;
5045         else
5046                 to_reclaim = 0;
5047         to_reclaim = min(to_reclaim, space_info->bytes_may_use +
5048                                      space_info->bytes_reserved);
5049         return to_reclaim;
5050 }
5051
5052 static inline int need_do_async_reclaim(struct btrfs_fs_info *fs_info,
5053                                         struct btrfs_space_info *space_info,
5054                                         u64 used, bool system_chunk)
5055 {
5056         u64 thresh = div_factor_fine(space_info->total_bytes, 98);
5057
5058         /* If we're just plain full then async reclaim just slows us down. */
5059         if ((space_info->bytes_used + space_info->bytes_reserved) >= thresh)
5060                 return 0;
5061
5062         if (!btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5063                                               system_chunk))
5064                 return 0;
5065
5066         return (used >= thresh && !btrfs_fs_closing(fs_info) &&
5067                 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
5068 }
5069
5070 static void wake_all_tickets(struct list_head *head)
5071 {
5072         struct reserve_ticket *ticket;
5073
5074         while (!list_empty(head)) {
5075                 ticket = list_first_entry(head, struct reserve_ticket, list);
5076                 list_del_init(&ticket->list);
5077                 ticket->error = -ENOSPC;
5078                 wake_up(&ticket->wait);
5079         }
5080 }
5081
5082 /*
5083  * This is for normal flushers, we can wait all goddamned day if we want to.  We
5084  * will loop and continuously try to flush as long as we are making progress.
5085  * We count progress as clearing off tickets each time we have to loop.
5086  */
5087 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
5088 {
5089         struct btrfs_fs_info *fs_info;
5090         struct btrfs_space_info *space_info;
5091         u64 to_reclaim;
5092         int flush_state;
5093         int commit_cycles = 0;
5094         u64 last_tickets_id;
5095
5096         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
5097         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5098
5099         spin_lock(&space_info->lock);
5100         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5101                                                       false);
5102         if (!to_reclaim) {
5103                 space_info->flush = 0;
5104                 spin_unlock(&space_info->lock);
5105                 return;
5106         }
5107         last_tickets_id = space_info->tickets_id;
5108         spin_unlock(&space_info->lock);
5109
5110         flush_state = FLUSH_DELAYED_ITEMS_NR;
5111         do {
5112                 flush_space(fs_info, space_info, to_reclaim, flush_state);
5113                 spin_lock(&space_info->lock);
5114                 if (list_empty(&space_info->tickets)) {
5115                         space_info->flush = 0;
5116                         spin_unlock(&space_info->lock);
5117                         return;
5118                 }
5119                 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info,
5120                                                               space_info,
5121                                                               false);
5122                 if (last_tickets_id == space_info->tickets_id) {
5123                         flush_state++;
5124                 } else {
5125                         last_tickets_id = space_info->tickets_id;
5126                         flush_state = FLUSH_DELAYED_ITEMS_NR;
5127                         if (commit_cycles)
5128                                 commit_cycles--;
5129                 }
5130
5131                 if (flush_state > COMMIT_TRANS) {
5132                         commit_cycles++;
5133                         if (commit_cycles > 2) {
5134                                 wake_all_tickets(&space_info->tickets);
5135                                 space_info->flush = 0;
5136                         } else {
5137                                 flush_state = FLUSH_DELAYED_ITEMS_NR;
5138                         }
5139                 }
5140                 spin_unlock(&space_info->lock);
5141         } while (flush_state <= COMMIT_TRANS);
5142 }
5143
5144 void btrfs_init_async_reclaim_work(struct work_struct *work)
5145 {
5146         INIT_WORK(work, btrfs_async_reclaim_metadata_space);
5147 }
5148
5149 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info,
5150                                             struct btrfs_space_info *space_info,
5151                                             struct reserve_ticket *ticket)
5152 {
5153         u64 to_reclaim;
5154         int flush_state = FLUSH_DELAYED_ITEMS_NR;
5155
5156         spin_lock(&space_info->lock);
5157         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5158                                                       false);
5159         if (!to_reclaim) {
5160                 spin_unlock(&space_info->lock);
5161                 return;
5162         }
5163         spin_unlock(&space_info->lock);
5164
5165         do {
5166                 flush_space(fs_info, space_info, to_reclaim, flush_state);
5167                 flush_state++;
5168                 spin_lock(&space_info->lock);
5169                 if (ticket->bytes == 0) {
5170                         spin_unlock(&space_info->lock);
5171                         return;
5172                 }
5173                 spin_unlock(&space_info->lock);
5174
5175                 /*
5176                  * Priority flushers can't wait on delalloc without
5177                  * deadlocking.
5178                  */
5179                 if (flush_state == FLUSH_DELALLOC ||
5180                     flush_state == FLUSH_DELALLOC_WAIT)
5181                         flush_state = ALLOC_CHUNK;
5182         } while (flush_state < COMMIT_TRANS);
5183 }
5184
5185 static int wait_reserve_ticket(struct btrfs_fs_info *fs_info,
5186                                struct btrfs_space_info *space_info,
5187                                struct reserve_ticket *ticket, u64 orig_bytes)
5188
5189 {
5190         DEFINE_WAIT(wait);
5191         int ret = 0;
5192
5193         spin_lock(&space_info->lock);
5194         while (ticket->bytes > 0 && ticket->error == 0) {
5195                 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE);
5196                 if (ret) {
5197                         ret = -EINTR;
5198                         break;
5199                 }
5200                 spin_unlock(&space_info->lock);
5201
5202                 schedule();
5203
5204                 finish_wait(&ticket->wait, &wait);
5205                 spin_lock(&space_info->lock);
5206         }
5207         if (!ret)
5208                 ret = ticket->error;
5209         if (!list_empty(&ticket->list))
5210                 list_del_init(&ticket->list);
5211         if (ticket->bytes && ticket->bytes < orig_bytes) {
5212                 u64 num_bytes = orig_bytes - ticket->bytes;
5213                 space_info->bytes_may_use -= num_bytes;
5214                 trace_btrfs_space_reservation(fs_info, "space_info",
5215                                               space_info->flags, num_bytes, 0);
5216         }
5217         spin_unlock(&space_info->lock);
5218
5219         return ret;
5220 }
5221
5222 /**
5223  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5224  * @root - the root we're allocating for
5225  * @space_info - the space info we want to allocate from
5226  * @orig_bytes - the number of bytes we want
5227  * @flush - whether or not we can flush to make our reservation
5228  *
5229  * This will reserve orig_bytes number of bytes from the space info associated
5230  * with the block_rsv.  If there is not enough space it will make an attempt to
5231  * flush out space to make room.  It will do this by flushing delalloc if
5232  * possible or committing the transaction.  If flush is 0 then no attempts to
5233  * regain reservations will be made and this will fail if there is not enough
5234  * space already.
5235  */
5236 static int __reserve_metadata_bytes(struct btrfs_fs_info *fs_info,
5237                                     struct btrfs_space_info *space_info,
5238                                     u64 orig_bytes,
5239                                     enum btrfs_reserve_flush_enum flush,
5240                                     bool system_chunk)
5241 {
5242         struct reserve_ticket ticket;
5243         u64 used;
5244         int ret = 0;
5245
5246         ASSERT(orig_bytes);
5247         ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL);
5248
5249         spin_lock(&space_info->lock);
5250         ret = -ENOSPC;
5251         used = btrfs_space_info_used(space_info, true);
5252
5253         /*
5254          * If we have enough space then hooray, make our reservation and carry
5255          * on.  If not see if we can overcommit, and if we can, hooray carry on.
5256          * If not things get more complicated.
5257          */
5258         if (used + orig_bytes <= space_info->total_bytes) {
5259                 space_info->bytes_may_use += orig_bytes;
5260                 trace_btrfs_space_reservation(fs_info, "space_info",
5261                                               space_info->flags, orig_bytes, 1);
5262                 ret = 0;
5263         } else if (can_overcommit(fs_info, space_info, orig_bytes, flush,
5264                                   system_chunk)) {
5265                 space_info->bytes_may_use += orig_bytes;
5266                 trace_btrfs_space_reservation(fs_info, "space_info",
5267                                               space_info->flags, orig_bytes, 1);
5268                 ret = 0;
5269         }
5270
5271         /*
5272          * If we couldn't make a reservation then setup our reservation ticket
5273          * and kick the async worker if it's not already running.
5274          *
5275          * If we are a priority flusher then we just need to add our ticket to
5276          * the list and we will do our own flushing further down.
5277          */
5278         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
5279                 ticket.bytes = orig_bytes;
5280                 ticket.error = 0;
5281                 init_waitqueue_head(&ticket.wait);
5282                 if (flush == BTRFS_RESERVE_FLUSH_ALL) {
5283                         list_add_tail(&ticket.list, &space_info->tickets);
5284                         if (!space_info->flush) {
5285                                 space_info->flush = 1;
5286                                 trace_btrfs_trigger_flush(fs_info,
5287                                                           space_info->flags,
5288                                                           orig_bytes, flush,
5289                                                           "enospc");
5290                                 queue_work(system_unbound_wq,
5291                                            &fs_info->async_reclaim_work);
5292                         }
5293                 } else {
5294                         list_add_tail(&ticket.list,
5295                                       &space_info->priority_tickets);
5296                 }
5297         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
5298                 used += orig_bytes;
5299                 /*
5300                  * We will do the space reservation dance during log replay,
5301                  * which means we won't have fs_info->fs_root set, so don't do
5302                  * the async reclaim as we will panic.
5303                  */
5304                 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) &&
5305                     need_do_async_reclaim(fs_info, space_info,
5306                                           used, system_chunk) &&
5307                     !work_busy(&fs_info->async_reclaim_work)) {
5308                         trace_btrfs_trigger_flush(fs_info, space_info->flags,
5309                                                   orig_bytes, flush, "preempt");
5310                         queue_work(system_unbound_wq,
5311                                    &fs_info->async_reclaim_work);
5312                 }
5313         }
5314         spin_unlock(&space_info->lock);
5315         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
5316                 return ret;
5317
5318         if (flush == BTRFS_RESERVE_FLUSH_ALL)
5319                 return wait_reserve_ticket(fs_info, space_info, &ticket,
5320                                            orig_bytes);
5321
5322         ret = 0;
5323         priority_reclaim_metadata_space(fs_info, space_info, &ticket);
5324         spin_lock(&space_info->lock);
5325         if (ticket.bytes) {
5326                 if (ticket.bytes < orig_bytes) {
5327                         u64 num_bytes = orig_bytes - ticket.bytes;
5328                         space_info->bytes_may_use -= num_bytes;
5329                         trace_btrfs_space_reservation(fs_info, "space_info",
5330                                                       space_info->flags,
5331                                                       num_bytes, 0);
5332
5333                 }
5334                 list_del_init(&ticket.list);
5335                 ret = -ENOSPC;
5336         }
5337         spin_unlock(&space_info->lock);
5338         ASSERT(list_empty(&ticket.list));
5339         return ret;
5340 }
5341
5342 /**
5343  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5344  * @root - the root we're allocating for
5345  * @block_rsv - the block_rsv we're allocating for
5346  * @orig_bytes - the number of bytes we want
5347  * @flush - whether or not we can flush to make our reservation
5348  *
5349  * This will reserve orgi_bytes number of bytes from the space info associated
5350  * with the block_rsv.  If there is not enough space it will make an attempt to
5351  * flush out space to make room.  It will do this by flushing delalloc if
5352  * possible or committing the transaction.  If flush is 0 then no attempts to
5353  * regain reservations will be made and this will fail if there is not enough
5354  * space already.
5355  */
5356 static int reserve_metadata_bytes(struct btrfs_root *root,
5357                                   struct btrfs_block_rsv *block_rsv,
5358                                   u64 orig_bytes,
5359                                   enum btrfs_reserve_flush_enum flush)
5360 {
5361         struct btrfs_fs_info *fs_info = root->fs_info;
5362         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5363         int ret;
5364         bool system_chunk = (root == fs_info->chunk_root);
5365
5366         ret = __reserve_metadata_bytes(fs_info, block_rsv->space_info,
5367                                        orig_bytes, flush, system_chunk);
5368         if (ret == -ENOSPC &&
5369             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
5370                 if (block_rsv != global_rsv &&
5371                     !block_rsv_use_bytes(global_rsv, orig_bytes))
5372                         ret = 0;
5373         }
5374         if (ret == -ENOSPC) {
5375                 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
5376                                               block_rsv->space_info->flags,
5377                                               orig_bytes, 1);
5378
5379                 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
5380                         dump_space_info(fs_info, block_rsv->space_info,
5381                                         orig_bytes, 0);
5382         }
5383         return ret;
5384 }
5385
5386 static struct btrfs_block_rsv *get_block_rsv(
5387                                         const struct btrfs_trans_handle *trans,
5388                                         const struct btrfs_root *root)
5389 {
5390         struct btrfs_fs_info *fs_info = root->fs_info;
5391         struct btrfs_block_rsv *block_rsv = NULL;
5392
5393         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
5394             (root == fs_info->csum_root && trans->adding_csums) ||
5395             (root == fs_info->uuid_root))
5396                 block_rsv = trans->block_rsv;
5397
5398         if (!block_rsv)
5399                 block_rsv = root->block_rsv;
5400
5401         if (!block_rsv)
5402                 block_rsv = &fs_info->empty_block_rsv;
5403
5404         return block_rsv;
5405 }
5406
5407 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
5408                                u64 num_bytes)
5409 {
5410         int ret = -ENOSPC;
5411         spin_lock(&block_rsv->lock);
5412         if (block_rsv->reserved >= num_bytes) {
5413                 block_rsv->reserved -= num_bytes;
5414                 if (block_rsv->reserved < block_rsv->size)
5415                         block_rsv->full = 0;
5416                 ret = 0;
5417         }
5418         spin_unlock(&block_rsv->lock);
5419         return ret;
5420 }
5421
5422 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
5423                                 u64 num_bytes, int update_size)
5424 {
5425         spin_lock(&block_rsv->lock);
5426         block_rsv->reserved += num_bytes;
5427         if (update_size)
5428                 block_rsv->size += num_bytes;
5429         else if (block_rsv->reserved >= block_rsv->size)
5430                 block_rsv->full = 1;
5431         spin_unlock(&block_rsv->lock);
5432 }
5433
5434 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
5435                              struct btrfs_block_rsv *dest, u64 num_bytes,
5436                              int min_factor)
5437 {
5438         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5439         u64 min_bytes;
5440
5441         if (global_rsv->space_info != dest->space_info)
5442                 return -ENOSPC;
5443
5444         spin_lock(&global_rsv->lock);
5445         min_bytes = div_factor(global_rsv->size, min_factor);
5446         if (global_rsv->reserved < min_bytes + num_bytes) {
5447                 spin_unlock(&global_rsv->lock);
5448                 return -ENOSPC;
5449         }
5450         global_rsv->reserved -= num_bytes;
5451         if (global_rsv->reserved < global_rsv->size)
5452                 global_rsv->full = 0;
5453         spin_unlock(&global_rsv->lock);
5454
5455         block_rsv_add_bytes(dest, num_bytes, 1);
5456         return 0;
5457 }
5458
5459 /*
5460  * This is for space we already have accounted in space_info->bytes_may_use, so
5461  * basically when we're returning space from block_rsv's.
5462  */
5463 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
5464                                      struct btrfs_space_info *space_info,
5465                                      u64 num_bytes)
5466 {
5467         struct reserve_ticket *ticket;
5468         struct list_head *head;
5469         u64 used;
5470         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH;
5471         bool check_overcommit = false;
5472
5473         spin_lock(&space_info->lock);
5474         head = &space_info->priority_tickets;
5475
5476         /*
5477          * If we are over our limit then we need to check and see if we can
5478          * overcommit, and if we can't then we just need to free up our space
5479          * and not satisfy any requests.
5480          */
5481         used = btrfs_space_info_used(space_info, true);
5482         if (used - num_bytes >= space_info->total_bytes)
5483                 check_overcommit = true;
5484 again:
5485         while (!list_empty(head) && num_bytes) {
5486                 ticket = list_first_entry(head, struct reserve_ticket,
5487                                           list);
5488                 /*
5489                  * We use 0 bytes because this space is already reserved, so
5490                  * adding the ticket space would be a double count.
5491                  */
5492                 if (check_overcommit &&
5493                     !can_overcommit(fs_info, space_info, 0, flush, false))
5494                         break;
5495                 if (num_bytes >= ticket->bytes) {
5496                         list_del_init(&ticket->list);
5497                         num_bytes -= ticket->bytes;
5498                         ticket->bytes = 0;
5499                         space_info->tickets_id++;
5500                         wake_up(&ticket->wait);
5501                 } else {
5502                         ticket->bytes -= num_bytes;
5503                         num_bytes = 0;
5504                 }
5505         }
5506
5507         if (num_bytes && head == &space_info->priority_tickets) {
5508                 head = &space_info->tickets;
5509                 flush = BTRFS_RESERVE_FLUSH_ALL;
5510                 goto again;
5511         }
5512         space_info->bytes_may_use -= num_bytes;
5513         trace_btrfs_space_reservation(fs_info, "space_info",
5514                                       space_info->flags, num_bytes, 0);
5515         spin_unlock(&space_info->lock);
5516 }
5517
5518 /*
5519  * This is for newly allocated space that isn't accounted in
5520  * space_info->bytes_may_use yet.  So if we allocate a chunk or unpin an extent
5521  * we use this helper.
5522  */
5523 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
5524                                      struct btrfs_space_info *space_info,
5525                                      u64 num_bytes)
5526 {
5527         struct reserve_ticket *ticket;
5528         struct list_head *head = &space_info->priority_tickets;
5529
5530 again:
5531         while (!list_empty(head) && num_bytes) {
5532                 ticket = list_first_entry(head, struct reserve_ticket,
5533                                           list);
5534                 if (num_bytes >= ticket->bytes) {
5535                         trace_btrfs_space_reservation(fs_info, "space_info",
5536                                                       space_info->flags,
5537                                                       ticket->bytes, 1);
5538                         list_del_init(&ticket->list);
5539                         num_bytes -= ticket->bytes;
5540                         space_info->bytes_may_use += ticket->bytes;
5541                         ticket->bytes = 0;
5542                         space_info->tickets_id++;
5543                         wake_up(&ticket->wait);
5544                 } else {
5545                         trace_btrfs_space_reservation(fs_info, "space_info",
5546                                                       space_info->flags,
5547                                                       num_bytes, 1);
5548                         space_info->bytes_may_use += num_bytes;
5549                         ticket->bytes -= num_bytes;
5550                         num_bytes = 0;
5551                 }
5552         }
5553
5554         if (num_bytes && head == &space_info->priority_tickets) {
5555                 head = &space_info->tickets;
5556                 goto again;
5557         }
5558 }
5559
5560 static u64 block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
5561                                     struct btrfs_block_rsv *block_rsv,
5562                                     struct btrfs_block_rsv *dest, u64 num_bytes)
5563 {
5564         struct btrfs_space_info *space_info = block_rsv->space_info;
5565         u64 ret;
5566
5567         spin_lock(&block_rsv->lock);
5568         if (num_bytes == (u64)-1)
5569                 num_bytes = block_rsv->size;
5570         block_rsv->size -= num_bytes;
5571         if (block_rsv->reserved >= block_rsv->size) {
5572                 num_bytes = block_rsv->reserved - block_rsv->size;
5573                 block_rsv->reserved = block_rsv->size;
5574                 block_rsv->full = 1;
5575         } else {
5576                 num_bytes = 0;
5577         }
5578         spin_unlock(&block_rsv->lock);
5579
5580         ret = num_bytes;
5581         if (num_bytes > 0) {
5582                 if (dest) {
5583                         spin_lock(&dest->lock);
5584                         if (!dest->full) {
5585                                 u64 bytes_to_add;
5586
5587                                 bytes_to_add = dest->size - dest->reserved;
5588                                 bytes_to_add = min(num_bytes, bytes_to_add);
5589                                 dest->reserved += bytes_to_add;
5590                                 if (dest->reserved >= dest->size)
5591                                         dest->full = 1;
5592                                 num_bytes -= bytes_to_add;
5593                         }
5594                         spin_unlock(&dest->lock);
5595                 }
5596                 if (num_bytes)
5597                         space_info_add_old_bytes(fs_info, space_info,
5598                                                  num_bytes);
5599         }
5600         return ret;
5601 }
5602
5603 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src,
5604                             struct btrfs_block_rsv *dst, u64 num_bytes,
5605                             int update_size)
5606 {
5607         int ret;
5608
5609         ret = block_rsv_use_bytes(src, num_bytes);
5610         if (ret)
5611                 return ret;
5612
5613         block_rsv_add_bytes(dst, num_bytes, update_size);
5614         return 0;
5615 }
5616
5617 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
5618 {
5619         memset(rsv, 0, sizeof(*rsv));
5620         spin_lock_init(&rsv->lock);
5621         rsv->type = type;
5622 }
5623
5624 void btrfs_init_metadata_block_rsv(struct btrfs_fs_info *fs_info,
5625                                    struct btrfs_block_rsv *rsv,
5626                                    unsigned short type)
5627 {
5628         btrfs_init_block_rsv(rsv, type);
5629         rsv->space_info = __find_space_info(fs_info,
5630                                             BTRFS_BLOCK_GROUP_METADATA);
5631 }
5632
5633 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_fs_info *fs_info,
5634                                               unsigned short type)
5635 {
5636         struct btrfs_block_rsv *block_rsv;
5637
5638         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
5639         if (!block_rsv)
5640                 return NULL;
5641
5642         btrfs_init_metadata_block_rsv(fs_info, block_rsv, type);
5643         return block_rsv;
5644 }
5645
5646 void btrfs_free_block_rsv(struct btrfs_fs_info *fs_info,
5647                           struct btrfs_block_rsv *rsv)
5648 {
5649         if (!rsv)
5650                 return;
5651         btrfs_block_rsv_release(fs_info, rsv, (u64)-1);
5652         kfree(rsv);
5653 }
5654
5655 void __btrfs_free_block_rsv(struct btrfs_block_rsv *rsv)
5656 {
5657         kfree(rsv);
5658 }
5659
5660 int btrfs_block_rsv_add(struct btrfs_root *root,
5661                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
5662                         enum btrfs_reserve_flush_enum flush)
5663 {
5664         int ret;
5665
5666         if (num_bytes == 0)
5667                 return 0;
5668
5669         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5670         if (!ret) {
5671                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
5672                 return 0;
5673         }
5674
5675         return ret;
5676 }
5677
5678 int btrfs_block_rsv_check(struct btrfs_block_rsv *block_rsv, int min_factor)
5679 {
5680         u64 num_bytes = 0;
5681         int ret = -ENOSPC;
5682
5683         if (!block_rsv)
5684                 return 0;
5685
5686         spin_lock(&block_rsv->lock);
5687         num_bytes = div_factor(block_rsv->size, min_factor);
5688         if (block_rsv->reserved >= num_bytes)
5689                 ret = 0;
5690         spin_unlock(&block_rsv->lock);
5691
5692         return ret;
5693 }
5694
5695 int btrfs_block_rsv_refill(struct btrfs_root *root,
5696                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
5697                            enum btrfs_reserve_flush_enum flush)
5698 {
5699         u64 num_bytes = 0;
5700         int ret = -ENOSPC;
5701
5702         if (!block_rsv)
5703                 return 0;
5704
5705         spin_lock(&block_rsv->lock);
5706         num_bytes = min_reserved;
5707         if (block_rsv->reserved >= num_bytes)
5708                 ret = 0;
5709         else
5710                 num_bytes -= block_rsv->reserved;
5711         spin_unlock(&block_rsv->lock);
5712
5713         if (!ret)
5714                 return 0;
5715
5716         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5717         if (!ret) {
5718                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
5719                 return 0;
5720         }
5721
5722         return ret;
5723 }
5724
5725 /**
5726  * btrfs_inode_rsv_refill - refill the inode block rsv.
5727  * @inode - the inode we are refilling.
5728  * @flush - the flusing restriction.
5729  *
5730  * Essentially the same as btrfs_block_rsv_refill, except it uses the
5731  * block_rsv->size as the minimum size.  We'll either refill the missing amount
5732  * or return if we already have enough space.  This will also handle the resreve
5733  * tracepoint for the reserved amount.
5734  */
5735 static int btrfs_inode_rsv_refill(struct btrfs_inode *inode,
5736                                   enum btrfs_reserve_flush_enum flush)
5737 {
5738         struct btrfs_root *root = inode->root;
5739         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
5740         u64 num_bytes = 0;
5741         int ret = -ENOSPC;
5742
5743         spin_lock(&block_rsv->lock);
5744         if (block_rsv->reserved < block_rsv->size)
5745                 num_bytes = block_rsv->size - block_rsv->reserved;
5746         spin_unlock(&block_rsv->lock);
5747
5748         if (num_bytes == 0)
5749                 return 0;
5750
5751         ret = btrfs_qgroup_reserve_meta_prealloc(root, num_bytes, true);
5752         if (ret)
5753                 return ret;
5754         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5755         if (!ret) {
5756                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
5757                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5758                                               btrfs_ino(inode), num_bytes, 1);
5759         }
5760         return ret;
5761 }
5762
5763 /**
5764  * btrfs_inode_rsv_release - release any excessive reservation.
5765  * @inode - the inode we need to release from.
5766  * @qgroup_free - free or convert qgroup meta.
5767  *   Unlike normal operation, qgroup meta reservation needs to know if we are
5768  *   freeing qgroup reservation or just converting it into per-trans.  Normally
5769  *   @qgroup_free is true for error handling, and false for normal release.
5770  *
5771  * This is the same as btrfs_block_rsv_release, except that it handles the
5772  * tracepoint for the reservation.
5773  */
5774 static void btrfs_inode_rsv_release(struct btrfs_inode *inode, bool qgroup_free)
5775 {
5776         struct btrfs_fs_info *fs_info = inode->root->fs_info;
5777         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5778         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
5779         u64 released = 0;
5780
5781         /*
5782          * Since we statically set the block_rsv->size we just want to say we
5783          * are releasing 0 bytes, and then we'll just get the reservation over
5784          * the size free'd.
5785          */
5786         released = block_rsv_release_bytes(fs_info, block_rsv, global_rsv, 0);
5787         if (released > 0)
5788                 trace_btrfs_space_reservation(fs_info, "delalloc",
5789                                               btrfs_ino(inode), released, 0);
5790         if (qgroup_free)
5791                 btrfs_qgroup_free_meta_prealloc(inode->root, released);
5792         else
5793                 btrfs_qgroup_convert_reserved_meta(inode->root, released);
5794 }
5795
5796 void btrfs_block_rsv_release(struct btrfs_fs_info *fs_info,
5797                              struct btrfs_block_rsv *block_rsv,
5798                              u64 num_bytes)
5799 {
5800         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5801
5802         if (global_rsv == block_rsv ||
5803             block_rsv->space_info != global_rsv->space_info)
5804                 global_rsv = NULL;
5805         block_rsv_release_bytes(fs_info, block_rsv, global_rsv, num_bytes);
5806 }
5807
5808 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
5809 {
5810         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
5811         struct btrfs_space_info *sinfo = block_rsv->space_info;
5812         u64 num_bytes;
5813
5814         /*
5815          * The global block rsv is based on the size of the extent tree, the
5816          * checksum tree and the root tree.  If the fs is empty we want to set
5817          * it to a minimal amount for safety.
5818          */
5819         num_bytes = btrfs_root_used(&fs_info->extent_root->root_item) +
5820                 btrfs_root_used(&fs_info->csum_root->root_item) +
5821                 btrfs_root_used(&fs_info->tree_root->root_item);
5822         num_bytes = max_t(u64, num_bytes, SZ_16M);
5823
5824         spin_lock(&sinfo->lock);
5825         spin_lock(&block_rsv->lock);
5826
5827         block_rsv->size = min_t(u64, num_bytes, SZ_512M);
5828
5829         if (block_rsv->reserved < block_rsv->size) {
5830                 num_bytes = btrfs_space_info_used(sinfo, true);
5831                 if (sinfo->total_bytes > num_bytes) {
5832                         num_bytes = sinfo->total_bytes - num_bytes;
5833                         num_bytes = min(num_bytes,
5834                                         block_rsv->size - block_rsv->reserved);
5835                         block_rsv->reserved += num_bytes;
5836                         sinfo->bytes_may_use += num_bytes;
5837                         trace_btrfs_space_reservation(fs_info, "space_info",
5838                                                       sinfo->flags, num_bytes,
5839                                                       1);
5840                 }
5841         } else if (block_rsv->reserved > block_rsv->size) {
5842                 num_bytes = block_rsv->reserved - block_rsv->size;
5843                 sinfo->bytes_may_use -= num_bytes;
5844                 trace_btrfs_space_reservation(fs_info, "space_info",
5845                                       sinfo->flags, num_bytes, 0);
5846                 block_rsv->reserved = block_rsv->size;
5847         }
5848
5849         if (block_rsv->reserved == block_rsv->size)
5850                 block_rsv->full = 1;
5851         else
5852                 block_rsv->full = 0;
5853
5854         spin_unlock(&block_rsv->lock);
5855         spin_unlock(&sinfo->lock);
5856 }
5857
5858 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
5859 {
5860         struct btrfs_space_info *space_info;
5861
5862         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
5863         fs_info->chunk_block_rsv.space_info = space_info;
5864
5865         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5866         fs_info->global_block_rsv.space_info = space_info;
5867         fs_info->trans_block_rsv.space_info = space_info;
5868         fs_info->empty_block_rsv.space_info = space_info;
5869         fs_info->delayed_block_rsv.space_info = space_info;
5870
5871         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
5872         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
5873         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
5874         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
5875         if (fs_info->quota_root)
5876                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
5877         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
5878
5879         update_global_block_rsv(fs_info);
5880 }
5881
5882 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
5883 {
5884         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
5885                                 (u64)-1);
5886         WARN_ON(fs_info->trans_block_rsv.size > 0);
5887         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
5888         WARN_ON(fs_info->chunk_block_rsv.size > 0);
5889         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
5890         WARN_ON(fs_info->delayed_block_rsv.size > 0);
5891         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
5892 }
5893
5894
5895 /*
5896  * To be called after all the new block groups attached to the transaction
5897  * handle have been created (btrfs_create_pending_block_groups()).
5898  */
5899 void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans)
5900 {
5901         struct btrfs_fs_info *fs_info = trans->fs_info;
5902
5903         if (!trans->chunk_bytes_reserved)
5904                 return;
5905
5906         WARN_ON_ONCE(!list_empty(&trans->new_bgs));
5907
5908         block_rsv_release_bytes(fs_info, &fs_info->chunk_block_rsv, NULL,
5909                                 trans->chunk_bytes_reserved);
5910         trans->chunk_bytes_reserved = 0;
5911 }
5912
5913 /* Can only return 0 or -ENOSPC */
5914 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
5915                                   struct btrfs_inode *inode)
5916 {
5917         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
5918         struct btrfs_root *root = inode->root;
5919         /*
5920          * We always use trans->block_rsv here as we will have reserved space
5921          * for our orphan when starting the transaction, using get_block_rsv()
5922          * here will sometimes make us choose the wrong block rsv as we could be
5923          * doing a reloc inode for a non refcounted root.
5924          */
5925         struct btrfs_block_rsv *src_rsv = trans->block_rsv;
5926         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
5927
5928         /*
5929          * We need to hold space in order to delete our orphan item once we've
5930          * added it, so this takes the reservation so we can release it later
5931          * when we are truly done with the orphan item.
5932          */
5933         u64 num_bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
5934
5935         trace_btrfs_space_reservation(fs_info, "orphan", btrfs_ino(inode),
5936                         num_bytes, 1);
5937         return btrfs_block_rsv_migrate(src_rsv, dst_rsv, num_bytes, 1);
5938 }
5939
5940 void btrfs_orphan_release_metadata(struct btrfs_inode *inode)
5941 {
5942         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
5943         struct btrfs_root *root = inode->root;
5944         u64 num_bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
5945
5946         trace_btrfs_space_reservation(fs_info, "orphan", btrfs_ino(inode),
5947                         num_bytes, 0);
5948         btrfs_block_rsv_release(fs_info, root->orphan_block_rsv, num_bytes);
5949 }
5950
5951 /*
5952  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
5953  * root: the root of the parent directory
5954  * rsv: block reservation
5955  * items: the number of items that we need do reservation
5956  * qgroup_reserved: used to return the reserved size in qgroup
5957  *
5958  * This function is used to reserve the space for snapshot/subvolume
5959  * creation and deletion. Those operations are different with the
5960  * common file/directory operations, they change two fs/file trees
5961  * and root tree, the number of items that the qgroup reserves is
5962  * different with the free space reservation. So we can not use
5963  * the space reservation mechanism in start_transaction().
5964  */
5965 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
5966                                      struct btrfs_block_rsv *rsv,
5967                                      int items,
5968                                      u64 *qgroup_reserved,
5969                                      bool use_global_rsv)
5970 {
5971         u64 num_bytes;
5972         int ret;
5973         struct btrfs_fs_info *fs_info = root->fs_info;
5974         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5975
5976         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)) {
5977                 /* One for parent inode, two for dir entries */
5978                 num_bytes = 3 * fs_info->nodesize;
5979                 ret = btrfs_qgroup_reserve_meta_prealloc(root, num_bytes, true);
5980                 if (ret)
5981                         return ret;
5982         } else {
5983                 num_bytes = 0;
5984         }
5985
5986         *qgroup_reserved = num_bytes;
5987
5988         num_bytes = btrfs_calc_trans_metadata_size(fs_info, items);
5989         rsv->space_info = __find_space_info(fs_info,
5990                                             BTRFS_BLOCK_GROUP_METADATA);
5991         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
5992                                   BTRFS_RESERVE_FLUSH_ALL);
5993
5994         if (ret == -ENOSPC && use_global_rsv)
5995                 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes, 1);
5996
5997         if (ret && *qgroup_reserved)
5998                 btrfs_qgroup_free_meta_prealloc(root, *qgroup_reserved);
5999
6000         return ret;
6001 }
6002
6003 void btrfs_subvolume_release_metadata(struct btrfs_fs_info *fs_info,
6004                                       struct btrfs_block_rsv *rsv)
6005 {
6006         btrfs_block_rsv_release(fs_info, rsv, (u64)-1);
6007 }
6008
6009 static void btrfs_calculate_inode_block_rsv_size(struct btrfs_fs_info *fs_info,
6010                                                  struct btrfs_inode *inode)
6011 {
6012         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
6013         u64 reserve_size = 0;
6014         u64 csum_leaves;
6015         unsigned outstanding_extents;
6016
6017         lockdep_assert_held(&inode->lock);
6018         outstanding_extents = inode->outstanding_extents;
6019         if (outstanding_extents)
6020                 reserve_size = btrfs_calc_trans_metadata_size(fs_info,
6021                                                 outstanding_extents + 1);
6022         csum_leaves = btrfs_csum_bytes_to_leaves(fs_info,
6023                                                  inode->csum_bytes);
6024         reserve_size += btrfs_calc_trans_metadata_size(fs_info,
6025                                                        csum_leaves);
6026
6027         spin_lock(&block_rsv->lock);
6028         block_rsv->size = reserve_size;
6029         spin_unlock(&block_rsv->lock);
6030 }
6031
6032 int btrfs_delalloc_reserve_metadata(struct btrfs_inode *inode, u64 num_bytes)
6033 {
6034         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
6035         unsigned nr_extents;
6036         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
6037         int ret = 0;
6038         bool delalloc_lock = true;
6039
6040         /* If we are a free space inode we need to not flush since we will be in
6041          * the middle of a transaction commit.  We also don't need the delalloc
6042          * mutex since we won't race with anybody.  We need this mostly to make
6043          * lockdep shut its filthy mouth.
6044          *
6045          * If we have a transaction open (can happen if we call truncate_block
6046          * from truncate), then we need FLUSH_LIMIT so we don't deadlock.
6047          */
6048         if (btrfs_is_free_space_inode(inode)) {
6049                 flush = BTRFS_RESERVE_NO_FLUSH;
6050                 delalloc_lock = false;
6051         } else {
6052                 if (current->journal_info)
6053                         flush = BTRFS_RESERVE_FLUSH_LIMIT;
6054
6055                 if (btrfs_transaction_in_commit(fs_info))
6056                         schedule_timeout(1);
6057         }
6058
6059         if (delalloc_lock)
6060                 mutex_lock(&inode->delalloc_mutex);
6061
6062         num_bytes = ALIGN(num_bytes, fs_info->sectorsize);
6063
6064         /* Add our new extents and calculate the new rsv size. */
6065         spin_lock(&inode->lock);
6066         nr_extents = count_max_extents(num_bytes);
6067         btrfs_mod_outstanding_extents(inode, nr_extents);
6068         inode->csum_bytes += num_bytes;
6069         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6070         spin_unlock(&inode->lock);
6071
6072         ret = btrfs_inode_rsv_refill(inode, flush);
6073         if (unlikely(ret))
6074                 goto out_fail;
6075
6076         if (delalloc_lock)
6077                 mutex_unlock(&inode->delalloc_mutex);
6078         return 0;
6079
6080 out_fail:
6081         spin_lock(&inode->lock);
6082         nr_extents = count_max_extents(num_bytes);
6083         btrfs_mod_outstanding_extents(inode, -nr_extents);
6084         inode->csum_bytes -= num_bytes;
6085         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6086         spin_unlock(&inode->lock);
6087
6088         btrfs_inode_rsv_release(inode, true);
6089         if (delalloc_lock)
6090                 mutex_unlock(&inode->delalloc_mutex);
6091         return ret;
6092 }
6093
6094 /**
6095  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
6096  * @inode: the inode to release the reservation for.
6097  * @num_bytes: the number of bytes we are releasing.
6098  * @qgroup_free: free qgroup reservation or convert it to per-trans reservation
6099  *
6100  * This will release the metadata reservation for an inode.  This can be called
6101  * once we complete IO for a given set of bytes to release their metadata
6102  * reservations, or on error for the same reason.
6103  */
6104 void btrfs_delalloc_release_metadata(struct btrfs_inode *inode, u64 num_bytes,
6105                                      bool qgroup_free)
6106 {
6107         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
6108
6109         num_bytes = ALIGN(num_bytes, fs_info->sectorsize);
6110         spin_lock(&inode->lock);
6111         inode->csum_bytes -= num_bytes;
6112         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6113         spin_unlock(&inode->lock);
6114
6115         if (btrfs_is_testing(fs_info))
6116                 return;
6117
6118         btrfs_inode_rsv_release(inode, qgroup_free);
6119 }
6120
6121 /**
6122  * btrfs_delalloc_release_extents - release our outstanding_extents
6123  * @inode: the inode to balance the reservation for.
6124  * @num_bytes: the number of bytes we originally reserved with
6125  * @qgroup_free: do we need to free qgroup meta reservation or convert them.
6126  *
6127  * When we reserve space we increase outstanding_extents for the extents we may
6128  * add.  Once we've set the range as delalloc or created our ordered extents we
6129  * have outstanding_extents to track the real usage, so we use this to free our
6130  * temporarily tracked outstanding_extents.  This _must_ be used in conjunction
6131  * with btrfs_delalloc_reserve_metadata.
6132  */
6133 void btrfs_delalloc_release_extents(struct btrfs_inode *inode, u64 num_bytes,
6134                                     bool qgroup_free)
6135 {
6136         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
6137         unsigned num_extents;
6138
6139         spin_lock(&inode->lock);
6140         num_extents = count_max_extents(num_bytes);
6141         btrfs_mod_outstanding_extents(inode, -num_extents);
6142         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6143         spin_unlock(&inode->lock);
6144
6145         if (btrfs_is_testing(fs_info))
6146                 return;
6147
6148         btrfs_inode_rsv_release(inode, qgroup_free);
6149 }
6150
6151 /**
6152  * btrfs_delalloc_reserve_space - reserve data and metadata space for
6153  * delalloc
6154  * @inode: inode we're writing to
6155  * @start: start range we are writing to
6156  * @len: how long the range we are writing to
6157  * @reserved: mandatory parameter, record actually reserved qgroup ranges of
6158  *            current reservation.
6159  *
6160  * This will do the following things
6161  *
6162  * o reserve space in data space info for num bytes
6163  *   and reserve precious corresponding qgroup space
6164  *   (Done in check_data_free_space)
6165  *
6166  * o reserve space for metadata space, based on the number of outstanding
6167  *   extents and how much csums will be needed
6168  *   also reserve metadata space in a per root over-reserve method.
6169  * o add to the inodes->delalloc_bytes
6170  * o add it to the fs_info's delalloc inodes list.
6171  *   (Above 3 all done in delalloc_reserve_metadata)
6172  *
6173  * Return 0 for success
6174  * Return <0 for error(-ENOSPC or -EQUOT)
6175  */
6176 int btrfs_delalloc_reserve_space(struct inode *inode,
6177                         struct extent_changeset **reserved, u64 start, u64 len)
6178 {
6179         int ret;
6180
6181         ret = btrfs_check_data_free_space(inode, reserved, start, len);
6182         if (ret < 0)
6183                 return ret;
6184         ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode), len);
6185         if (ret < 0)
6186                 btrfs_free_reserved_data_space(inode, *reserved, start, len);
6187         return ret;
6188 }
6189
6190 /**
6191  * btrfs_delalloc_release_space - release data and metadata space for delalloc
6192  * @inode: inode we're releasing space for
6193  * @start: start position of the space already reserved
6194  * @len: the len of the space already reserved
6195  * @release_bytes: the len of the space we consumed or didn't use
6196  *
6197  * This function will release the metadata space that was not used and will
6198  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
6199  * list if there are no delalloc bytes left.
6200  * Also it will handle the qgroup reserved space.
6201  */
6202 void btrfs_delalloc_release_space(struct inode *inode,
6203                                   struct extent_changeset *reserved,
6204                                   u64 start, u64 len, bool qgroup_free)
6205 {
6206         btrfs_delalloc_release_metadata(BTRFS_I(inode), len, qgroup_free);
6207         btrfs_free_reserved_data_space(inode, reserved, start, len);
6208 }
6209
6210 static int update_block_group(struct btrfs_trans_handle *trans,
6211                               struct btrfs_fs_info *info, u64 bytenr,
6212                               u64 num_bytes, int alloc)
6213 {
6214         struct btrfs_block_group_cache *cache = NULL;
6215         u64 total = num_bytes;
6216         u64 old_val;
6217         u64 byte_in_group;
6218         int factor;
6219
6220         /* block accounting for super block */
6221         spin_lock(&info->delalloc_root_lock);
6222         old_val = btrfs_super_bytes_used(info->super_copy);
6223         if (alloc)
6224                 old_val += num_bytes;
6225         else
6226                 old_val -= num_bytes;
6227         btrfs_set_super_bytes_used(info->super_copy, old_val);
6228         spin_unlock(&info->delalloc_root_lock);
6229
6230         while (total) {
6231                 cache = btrfs_lookup_block_group(info, bytenr);
6232                 if (!cache)
6233                         return -ENOENT;
6234                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
6235                                     BTRFS_BLOCK_GROUP_RAID1 |
6236                                     BTRFS_BLOCK_GROUP_RAID10))
6237                         factor = 2;
6238                 else
6239                         factor = 1;
6240                 /*
6241                  * If this block group has free space cache written out, we
6242                  * need to make sure to load it if we are removing space.  This
6243                  * is because we need the unpinning stage to actually add the
6244                  * space back to the block group, otherwise we will leak space.
6245                  */
6246                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
6247                         cache_block_group(cache, 1);
6248
6249                 byte_in_group = bytenr - cache->key.objectid;
6250                 WARN_ON(byte_in_group > cache->key.offset);
6251
6252                 spin_lock(&cache->space_info->lock);
6253                 spin_lock(&cache->lock);
6254
6255                 if (btrfs_test_opt(info, SPACE_CACHE) &&
6256                     cache->disk_cache_state < BTRFS_DC_CLEAR)
6257                         cache->disk_cache_state = BTRFS_DC_CLEAR;
6258
6259                 old_val = btrfs_block_group_used(&cache->item);
6260                 num_bytes = min(total, cache->key.offset - byte_in_group);
6261                 if (alloc) {
6262                         old_val += num_bytes;
6263                         btrfs_set_block_group_used(&cache->item, old_val);
6264                         cache->reserved -= num_bytes;
6265                         cache->space_info->bytes_reserved -= num_bytes;
6266                         cache->space_info->bytes_used += num_bytes;
6267                         cache->space_info->disk_used += num_bytes * factor;
6268                         spin_unlock(&cache->lock);
6269                         spin_unlock(&cache->space_info->lock);
6270                 } else {
6271                         old_val -= num_bytes;
6272                         btrfs_set_block_group_used(&cache->item, old_val);
6273                         cache->pinned += num_bytes;
6274                         cache->space_info->bytes_pinned += num_bytes;
6275                         cache->space_info->bytes_used -= num_bytes;
6276                         cache->space_info->disk_used -= num_bytes * factor;
6277                         spin_unlock(&cache->lock);
6278                         spin_unlock(&cache->space_info->lock);
6279
6280                         trace_btrfs_space_reservation(info, "pinned",
6281                                                       cache->space_info->flags,
6282                                                       num_bytes, 1);
6283                         percpu_counter_add(&cache->space_info->total_bytes_pinned,
6284                                            num_bytes);
6285                         set_extent_dirty(info->pinned_extents,
6286                                          bytenr, bytenr + num_bytes - 1,
6287                                          GFP_NOFS | __GFP_NOFAIL);
6288                 }
6289
6290                 spin_lock(&trans->transaction->dirty_bgs_lock);
6291                 if (list_empty(&cache->dirty_list)) {
6292                         list_add_tail(&cache->dirty_list,
6293                                       &trans->transaction->dirty_bgs);
6294                                 trans->transaction->num_dirty_bgs++;
6295                         btrfs_get_block_group(cache);
6296                 }
6297                 spin_unlock(&trans->transaction->dirty_bgs_lock);
6298
6299                 /*
6300                  * No longer have used bytes in this block group, queue it for
6301                  * deletion. We do this after adding the block group to the
6302                  * dirty list to avoid races between cleaner kthread and space
6303                  * cache writeout.
6304                  */
6305                 if (!alloc && old_val == 0) {
6306                         spin_lock(&info->unused_bgs_lock);
6307                         if (list_empty(&cache->bg_list)) {
6308                                 btrfs_get_block_group(cache);
6309                                 list_add_tail(&cache->bg_list,
6310                                               &info->unused_bgs);
6311                         }
6312                         spin_unlock(&info->unused_bgs_lock);
6313                 }
6314
6315                 btrfs_put_block_group(cache);
6316                 total -= num_bytes;
6317                 bytenr += num_bytes;
6318         }
6319         return 0;
6320 }
6321
6322 static u64 first_logical_byte(struct btrfs_fs_info *fs_info, u64 search_start)
6323 {
6324         struct btrfs_block_group_cache *cache;
6325         u64 bytenr;
6326
6327         spin_lock(&fs_info->block_group_cache_lock);
6328         bytenr = fs_info->first_logical_byte;
6329         spin_unlock(&fs_info->block_group_cache_lock);
6330
6331         if (bytenr < (u64)-1)
6332                 return bytenr;
6333
6334         cache = btrfs_lookup_first_block_group(fs_info, search_start);
6335         if (!cache)
6336                 return 0;
6337
6338         bytenr = cache->key.objectid;
6339         btrfs_put_block_group(cache);
6340
6341         return bytenr;
6342 }
6343
6344 static int pin_down_extent(struct btrfs_fs_info *fs_info,
6345                            struct btrfs_block_group_cache *cache,
6346                            u64 bytenr, u64 num_bytes, int reserved)
6347 {
6348         spin_lock(&cache->space_info->lock);
6349         spin_lock(&cache->lock);
6350         cache->pinned += num_bytes;
6351         cache->space_info->bytes_pinned += num_bytes;
6352         if (reserved) {
6353                 cache->reserved -= num_bytes;
6354                 cache->space_info->bytes_reserved -= num_bytes;
6355         }
6356         spin_unlock(&cache->lock);
6357         spin_unlock(&cache->space_info->lock);
6358
6359         trace_btrfs_space_reservation(fs_info, "pinned",
6360                                       cache->space_info->flags, num_bytes, 1);
6361         percpu_counter_add(&cache->space_info->total_bytes_pinned, num_bytes);
6362         set_extent_dirty(fs_info->pinned_extents, bytenr,
6363                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
6364         return 0;
6365 }
6366
6367 /*
6368  * this function must be called within transaction
6369  */
6370 int btrfs_pin_extent(struct btrfs_fs_info *fs_info,
6371                      u64 bytenr, u64 num_bytes, int reserved)
6372 {
6373         struct btrfs_block_group_cache *cache;
6374
6375         cache = btrfs_lookup_block_group(fs_info, bytenr);
6376         BUG_ON(!cache); /* Logic error */
6377
6378         pin_down_extent(fs_info, cache, bytenr, num_bytes, reserved);
6379
6380         btrfs_put_block_group(cache);
6381         return 0;
6382 }
6383
6384 /*
6385  * this function must be called within transaction
6386  */
6387 int btrfs_pin_extent_for_log_replay(struct btrfs_fs_info *fs_info,
6388                                     u64 bytenr, u64 num_bytes)
6389 {
6390         struct btrfs_block_group_cache *cache;
6391         int ret;
6392
6393         cache = btrfs_lookup_block_group(fs_info, bytenr);
6394         if (!cache)
6395                 return -EINVAL;
6396
6397         /*
6398          * pull in the free space cache (if any) so that our pin
6399          * removes the free space from the cache.  We have load_only set
6400          * to one because the slow code to read in the free extents does check
6401          * the pinned extents.
6402          */
6403         cache_block_group(cache, 1);
6404
6405         pin_down_extent(fs_info, cache, bytenr, num_bytes, 0);
6406
6407         /* remove us from the free space cache (if we're there at all) */
6408         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
6409         btrfs_put_block_group(cache);
6410         return ret;
6411 }
6412
6413 static int __exclude_logged_extent(struct btrfs_fs_info *fs_info,
6414                                    u64 start, u64 num_bytes)
6415 {
6416         int ret;
6417         struct btrfs_block_group_cache *block_group;
6418         struct btrfs_caching_control *caching_ctl;
6419
6420         block_group = btrfs_lookup_block_group(fs_info, start);
6421         if (!block_group)
6422                 return -EINVAL;
6423
6424         cache_block_group(block_group, 0);
6425         caching_ctl = get_caching_control(block_group);
6426
6427         if (!caching_ctl) {
6428                 /* Logic error */
6429                 BUG_ON(!block_group_cache_done(block_group));
6430                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6431         } else {
6432                 mutex_lock(&caching_ctl->mutex);
6433
6434                 if (start >= caching_ctl->progress) {
6435                         ret = add_excluded_extent(fs_info, start, num_bytes);
6436                 } else if (start + num_bytes <= caching_ctl->progress) {
6437                         ret = btrfs_remove_free_space(block_group,
6438                                                       start, num_bytes);
6439                 } else {
6440                         num_bytes = caching_ctl->progress - start;
6441                         ret = btrfs_remove_free_space(block_group,
6442                                                       start, num_bytes);
6443                         if (ret)
6444                                 goto out_lock;
6445
6446                         num_bytes = (start + num_bytes) -
6447                                 caching_ctl->progress;
6448                         start = caching_ctl->progress;
6449                         ret = add_excluded_extent(fs_info, start, num_bytes);
6450                 }
6451 out_lock:
6452                 mutex_unlock(&caching_ctl->mutex);
6453                 put_caching_control(caching_ctl);
6454         }
6455         btrfs_put_block_group(block_group);
6456         return ret;
6457 }
6458
6459 int btrfs_exclude_logged_extents(struct btrfs_fs_info *fs_info,
6460                                  struct extent_buffer *eb)
6461 {
6462         struct btrfs_file_extent_item *item;
6463         struct btrfs_key key;
6464         int found_type;
6465         int i;
6466
6467         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS))
6468                 return 0;
6469
6470         for (i = 0; i < btrfs_header_nritems(eb); i++) {
6471                 btrfs_item_key_to_cpu(eb, &key, i);
6472                 if (key.type != BTRFS_EXTENT_DATA_KEY)
6473                         continue;
6474                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
6475                 found_type = btrfs_file_extent_type(eb, item);
6476                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
6477                         continue;
6478                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
6479                         continue;
6480                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
6481                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
6482                 __exclude_logged_extent(fs_info, key.objectid, key.offset);
6483         }
6484
6485         return 0;
6486 }
6487
6488 static void
6489 btrfs_inc_block_group_reservations(struct btrfs_block_group_cache *bg)
6490 {
6491         atomic_inc(&bg->reservations);
6492 }
6493
6494 void btrfs_dec_block_group_reservations(struct btrfs_fs_info *fs_info,
6495                                         const u64 start)
6496 {
6497         struct btrfs_block_group_cache *bg;
6498
6499         bg = btrfs_lookup_block_group(fs_info, start);
6500         ASSERT(bg);
6501         if (atomic_dec_and_test(&bg->reservations))
6502                 wake_up_var(&bg->reservations);
6503         btrfs_put_block_group(bg);
6504 }
6505
6506 void btrfs_wait_block_group_reservations(struct btrfs_block_group_cache *bg)
6507 {
6508         struct btrfs_space_info *space_info = bg->space_info;
6509
6510         ASSERT(bg->ro);
6511
6512         if (!(bg->flags & BTRFS_BLOCK_GROUP_DATA))
6513                 return;
6514
6515         /*
6516          * Our block group is read only but before we set it to read only,
6517          * some task might have had allocated an extent from it already, but it
6518          * has not yet created a respective ordered extent (and added it to a
6519          * root's list of ordered extents).
6520          * Therefore wait for any task currently allocating extents, since the
6521          * block group's reservations counter is incremented while a read lock
6522          * on the groups' semaphore is held and decremented after releasing
6523          * the read access on that semaphore and creating the ordered extent.
6524          */
6525         down_write(&space_info->groups_sem);
6526         up_write(&space_info->groups_sem);
6527
6528         wait_var_event(&bg->reservations, !atomic_read(&bg->reservations));
6529 }
6530
6531 /**
6532  * btrfs_add_reserved_bytes - update the block_group and space info counters
6533  * @cache:      The cache we are manipulating
6534  * @ram_bytes:  The number of bytes of file content, and will be same to
6535  *              @num_bytes except for the compress path.
6536  * @num_bytes:  The number of bytes in question
6537  * @delalloc:   The blocks are allocated for the delalloc write
6538  *
6539  * This is called by the allocator when it reserves space. If this is a
6540  * reservation and the block group has become read only we cannot make the
6541  * reservation and return -EAGAIN, otherwise this function always succeeds.
6542  */
6543 static int btrfs_add_reserved_bytes(struct btrfs_block_group_cache *cache,
6544                                     u64 ram_bytes, u64 num_bytes, int delalloc)
6545 {
6546         struct btrfs_space_info *space_info = cache->space_info;
6547         int ret = 0;
6548
6549         spin_lock(&space_info->lock);
6550         spin_lock(&cache->lock);
6551         if (cache->ro) {
6552                 ret = -EAGAIN;
6553         } else {
6554                 cache->reserved += num_bytes;
6555                 space_info->bytes_reserved += num_bytes;
6556
6557                 trace_btrfs_space_reservation(cache->fs_info,
6558                                 "space_info", space_info->flags,
6559                                 ram_bytes, 0);
6560                 space_info->bytes_may_use -= ram_bytes;
6561                 if (delalloc)
6562                         cache->delalloc_bytes += num_bytes;
6563         }
6564         spin_unlock(&cache->lock);
6565         spin_unlock(&space_info->lock);
6566         return ret;
6567 }
6568
6569 /**
6570  * btrfs_free_reserved_bytes - update the block_group and space info counters
6571  * @cache:      The cache we are manipulating
6572  * @num_bytes:  The number of bytes in question
6573  * @delalloc:   The blocks are allocated for the delalloc write
6574  *
6575  * This is called by somebody who is freeing space that was never actually used
6576  * on disk.  For example if you reserve some space for a new leaf in transaction
6577  * A and before transaction A commits you free that leaf, you call this with
6578  * reserve set to 0 in order to clear the reservation.
6579  */
6580
6581 static int btrfs_free_reserved_bytes(struct btrfs_block_group_cache *cache,
6582                                      u64 num_bytes, int delalloc)
6583 {
6584         struct btrfs_space_info *space_info = cache->space_info;
6585         int ret = 0;
6586
6587         spin_lock(&space_info->lock);
6588         spin_lock(&cache->lock);
6589         if (cache->ro)
6590                 space_info->bytes_readonly += num_bytes;
6591         cache->reserved -= num_bytes;
6592         space_info->bytes_reserved -= num_bytes;
6593
6594         if (delalloc)
6595                 cache->delalloc_bytes -= num_bytes;
6596         spin_unlock(&cache->lock);
6597         spin_unlock(&space_info->lock);
6598         return ret;
6599 }
6600 void btrfs_prepare_extent_commit(struct btrfs_fs_info *fs_info)
6601 {
6602         struct btrfs_caching_control *next;
6603         struct btrfs_caching_control *caching_ctl;
6604         struct btrfs_block_group_cache *cache;
6605
6606         down_write(&fs_info->commit_root_sem);
6607
6608         list_for_each_entry_safe(caching_ctl, next,
6609                                  &fs_info->caching_block_groups, list) {
6610                 cache = caching_ctl->block_group;
6611                 if (block_group_cache_done(cache)) {
6612                         cache->last_byte_to_unpin = (u64)-1;
6613                         list_del_init(&caching_ctl->list);
6614                         put_caching_control(caching_ctl);
6615                 } else {
6616                         cache->last_byte_to_unpin = caching_ctl->progress;
6617                 }
6618         }
6619
6620         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6621                 fs_info->pinned_extents = &fs_info->freed_extents[1];
6622         else
6623                 fs_info->pinned_extents = &fs_info->freed_extents[0];
6624
6625         up_write(&fs_info->commit_root_sem);
6626
6627         update_global_block_rsv(fs_info);
6628 }
6629
6630 /*
6631  * Returns the free cluster for the given space info and sets empty_cluster to
6632  * what it should be based on the mount options.
6633  */
6634 static struct btrfs_free_cluster *
6635 fetch_cluster_info(struct btrfs_fs_info *fs_info,
6636                    struct btrfs_space_info *space_info, u64 *empty_cluster)
6637 {
6638         struct btrfs_free_cluster *ret = NULL;
6639
6640         *empty_cluster = 0;
6641         if (btrfs_mixed_space_info(space_info))
6642                 return ret;
6643
6644         if (space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
6645                 ret = &fs_info->meta_alloc_cluster;
6646                 if (btrfs_test_opt(fs_info, SSD))
6647                         *empty_cluster = SZ_2M;
6648                 else
6649                         *empty_cluster = SZ_64K;
6650         } else if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) &&
6651                    btrfs_test_opt(fs_info, SSD_SPREAD)) {
6652                 *empty_cluster = SZ_2M;
6653                 ret = &fs_info->data_alloc_cluster;
6654         }
6655
6656         return ret;
6657 }
6658
6659 static int unpin_extent_range(struct btrfs_fs_info *fs_info,
6660                               u64 start, u64 end,
6661                               const bool return_free_space)
6662 {
6663         struct btrfs_block_group_cache *cache = NULL;
6664         struct btrfs_space_info *space_info;
6665         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
6666         struct btrfs_free_cluster *cluster = NULL;
6667         u64 len;
6668         u64 total_unpinned = 0;
6669         u64 empty_cluster = 0;
6670         bool readonly;
6671
6672         while (start <= end) {
6673                 readonly = false;
6674                 if (!cache ||
6675                     start >= cache->key.objectid + cache->key.offset) {
6676                         if (cache)
6677                                 btrfs_put_block_group(cache);
6678                         total_unpinned = 0;
6679                         cache = btrfs_lookup_block_group(fs_info, start);
6680                         BUG_ON(!cache); /* Logic error */
6681
6682                         cluster = fetch_cluster_info(fs_info,
6683                                                      cache->space_info,
6684                                                      &empty_cluster);
6685                         empty_cluster <<= 1;
6686                 }
6687
6688                 len = cache->key.objectid + cache->key.offset - start;
6689                 len = min(len, end + 1 - start);
6690
6691                 if (start < cache->last_byte_to_unpin) {
6692                         len = min(len, cache->last_byte_to_unpin - start);
6693                         if (return_free_space)
6694                                 btrfs_add_free_space(cache, start, len);
6695                 }
6696
6697                 start += len;
6698                 total_unpinned += len;
6699                 space_info = cache->space_info;
6700
6701                 /*
6702                  * If this space cluster has been marked as fragmented and we've
6703                  * unpinned enough in this block group to potentially allow a
6704                  * cluster to be created inside of it go ahead and clear the
6705                  * fragmented check.
6706                  */
6707                 if (cluster && cluster->fragmented &&
6708                     total_unpinned > empty_cluster) {
6709                         spin_lock(&cluster->lock);
6710                         cluster->fragmented = 0;
6711                         spin_unlock(&cluster->lock);
6712                 }
6713
6714                 spin_lock(&space_info->lock);
6715                 spin_lock(&cache->lock);
6716                 cache->pinned -= len;
6717                 space_info->bytes_pinned -= len;
6718
6719                 trace_btrfs_space_reservation(fs_info, "pinned",
6720                                               space_info->flags, len, 0);
6721                 space_info->max_extent_size = 0;
6722                 percpu_counter_add(&space_info->total_bytes_pinned, -len);
6723                 if (cache->ro) {
6724                         space_info->bytes_readonly += len;
6725                         readonly = true;
6726                 }
6727                 spin_unlock(&cache->lock);
6728                 if (!readonly && return_free_space &&
6729                     global_rsv->space_info == space_info) {
6730                         u64 to_add = len;
6731
6732                         spin_lock(&global_rsv->lock);
6733                         if (!global_rsv->full) {
6734                                 to_add = min(len, global_rsv->size -
6735                                              global_rsv->reserved);
6736                                 global_rsv->reserved += to_add;
6737                                 space_info->bytes_may_use += to_add;
6738                                 if (global_rsv->reserved >= global_rsv->size)
6739                                         global_rsv->full = 1;
6740                                 trace_btrfs_space_reservation(fs_info,
6741                                                               "space_info",
6742                                                               space_info->flags,
6743                                                               to_add, 1);
6744                                 len -= to_add;
6745                         }
6746                         spin_unlock(&global_rsv->lock);
6747                         /* Add to any tickets we may have */
6748                         if (len)
6749                                 space_info_add_new_bytes(fs_info, space_info,
6750                                                          len);
6751                 }
6752                 spin_unlock(&space_info->lock);
6753         }
6754
6755         if (cache)
6756                 btrfs_put_block_group(cache);
6757         return 0;
6758 }
6759
6760 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans)
6761 {
6762         struct btrfs_fs_info *fs_info = trans->fs_info;
6763         struct btrfs_block_group_cache *block_group, *tmp;
6764         struct list_head *deleted_bgs;
6765         struct extent_io_tree *unpin;
6766         u64 start;
6767         u64 end;
6768         int ret;
6769
6770         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6771                 unpin = &fs_info->freed_extents[1];
6772         else
6773                 unpin = &fs_info->freed_extents[0];
6774
6775         while (!trans->aborted) {
6776                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
6777                 ret = find_first_extent_bit(unpin, 0, &start, &end,
6778                                             EXTENT_DIRTY, NULL);
6779                 if (ret) {
6780                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6781                         break;
6782                 }
6783
6784                 if (btrfs_test_opt(fs_info, DISCARD))
6785                         ret = btrfs_discard_extent(fs_info, start,
6786                                                    end + 1 - start, NULL);
6787
6788                 clear_extent_dirty(unpin, start, end);
6789                 unpin_extent_range(fs_info, start, end, true);
6790                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6791                 cond_resched();
6792         }
6793
6794         /*
6795          * Transaction is finished.  We don't need the lock anymore.  We
6796          * do need to clean up the block groups in case of a transaction
6797          * abort.
6798          */
6799         deleted_bgs = &trans->transaction->deleted_bgs;
6800         list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
6801                 u64 trimmed = 0;
6802
6803                 ret = -EROFS;
6804                 if (!trans->aborted)
6805                         ret = btrfs_discard_extent(fs_info,
6806                                                    block_group->key.objectid,
6807                                                    block_group->key.offset,
6808                                                    &trimmed);
6809
6810                 list_del_init(&block_group->bg_list);
6811                 btrfs_put_block_group_trimming(block_group);
6812                 btrfs_put_block_group(block_group);
6813
6814                 if (ret) {
6815                         const char *errstr = btrfs_decode_error(ret);
6816                         btrfs_warn(fs_info,
6817                            "discard failed while removing blockgroup: errno=%d %s",
6818                                    ret, errstr);
6819                 }
6820         }
6821
6822         return 0;
6823 }
6824
6825 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
6826                                 struct btrfs_fs_info *info,
6827                                 struct btrfs_delayed_ref_node *node, u64 parent,
6828                                 u64 root_objectid, u64 owner_objectid,
6829                                 u64 owner_offset, int refs_to_drop,
6830                                 struct btrfs_delayed_extent_op *extent_op)
6831 {
6832         struct btrfs_key key;
6833         struct btrfs_path *path;
6834         struct btrfs_root *extent_root = info->extent_root;
6835         struct extent_buffer *leaf;
6836         struct btrfs_extent_item *ei;
6837         struct btrfs_extent_inline_ref *iref;
6838         int ret;
6839         int is_data;
6840         int extent_slot = 0;
6841         int found_extent = 0;
6842         int num_to_del = 1;
6843         u32 item_size;
6844         u64 refs;
6845         u64 bytenr = node->bytenr;
6846         u64 num_bytes = node->num_bytes;
6847         int last_ref = 0;
6848         bool skinny_metadata = btrfs_fs_incompat(info, SKINNY_METADATA);
6849
6850         path = btrfs_alloc_path();
6851         if (!path)
6852                 return -ENOMEM;
6853
6854         path->reada = READA_FORWARD;
6855         path->leave_spinning = 1;
6856
6857         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
6858         BUG_ON(!is_data && refs_to_drop != 1);
6859
6860         if (is_data)
6861                 skinny_metadata = false;
6862
6863         ret = lookup_extent_backref(trans, info, path, &iref,
6864                                     bytenr, num_bytes, parent,
6865                                     root_objectid, owner_objectid,
6866                                     owner_offset);
6867         if (ret == 0) {
6868                 extent_slot = path->slots[0];
6869                 while (extent_slot >= 0) {
6870                         btrfs_item_key_to_cpu(path->nodes[0], &key,
6871                                               extent_slot);
6872                         if (key.objectid != bytenr)
6873                                 break;
6874                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
6875                             key.offset == num_bytes) {
6876                                 found_extent = 1;
6877                                 break;
6878                         }
6879                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
6880                             key.offset == owner_objectid) {
6881                                 found_extent = 1;
6882                                 break;
6883                         }
6884                         if (path->slots[0] - extent_slot > 5)
6885                                 break;
6886                         extent_slot--;
6887                 }
6888 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6889                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
6890                 if (found_extent && item_size < sizeof(*ei))
6891                         found_extent = 0;
6892 #endif
6893                 if (!found_extent) {
6894                         BUG_ON(iref);
6895                         ret = remove_extent_backref(trans, info, path, NULL,
6896                                                     refs_to_drop,
6897                                                     is_data, &last_ref);
6898                         if (ret) {
6899                                 btrfs_abort_transaction(trans, ret);
6900                                 goto out;
6901                         }
6902                         btrfs_release_path(path);
6903                         path->leave_spinning = 1;
6904
6905                         key.objectid = bytenr;
6906                         key.type = BTRFS_EXTENT_ITEM_KEY;
6907                         key.offset = num_bytes;
6908
6909                         if (!is_data && skinny_metadata) {
6910                                 key.type = BTRFS_METADATA_ITEM_KEY;
6911                                 key.offset = owner_objectid;
6912                         }
6913
6914                         ret = btrfs_search_slot(trans, extent_root,
6915                                                 &key, path, -1, 1);
6916                         if (ret > 0 && skinny_metadata && path->slots[0]) {
6917                                 /*
6918                                  * Couldn't find our skinny metadata item,
6919                                  * see if we have ye olde extent item.
6920                                  */
6921                                 path->slots[0]--;
6922                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
6923                                                       path->slots[0]);
6924                                 if (key.objectid == bytenr &&
6925                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
6926                                     key.offset == num_bytes)
6927                                         ret = 0;
6928                         }
6929
6930                         if (ret > 0 && skinny_metadata) {
6931                                 skinny_metadata = false;
6932                                 key.objectid = bytenr;
6933                                 key.type = BTRFS_EXTENT_ITEM_KEY;
6934                                 key.offset = num_bytes;
6935                                 btrfs_release_path(path);
6936                                 ret = btrfs_search_slot(trans, extent_root,
6937                                                         &key, path, -1, 1);
6938                         }
6939
6940                         if (ret) {
6941                                 btrfs_err(info,
6942                                           "umm, got %d back from search, was looking for %llu",
6943                                           ret, bytenr);
6944                                 if (ret > 0)
6945                                         btrfs_print_leaf(path->nodes[0]);
6946                         }
6947                         if (ret < 0) {
6948                                 btrfs_abort_transaction(trans, ret);
6949                                 goto out;
6950                         }
6951                         extent_slot = path->slots[0];
6952                 }
6953         } else if (WARN_ON(ret == -ENOENT)) {
6954                 btrfs_print_leaf(path->nodes[0]);
6955                 btrfs_err(info,
6956                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
6957                         bytenr, parent, root_objectid, owner_objectid,
6958                         owner_offset);
6959                 btrfs_abort_transaction(trans, ret);
6960                 goto out;
6961         } else {
6962                 btrfs_abort_transaction(trans, ret);
6963                 goto out;
6964         }
6965
6966         leaf = path->nodes[0];
6967         item_size = btrfs_item_size_nr(leaf, extent_slot);
6968 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6969         if (item_size < sizeof(*ei)) {
6970                 BUG_ON(found_extent || extent_slot != path->slots[0]);
6971                 ret = convert_extent_item_v0(trans, info, path, owner_objectid,
6972                                              0);
6973                 if (ret < 0) {
6974                         btrfs_abort_transaction(trans, ret);
6975                         goto out;
6976                 }
6977
6978                 btrfs_release_path(path);
6979                 path->leave_spinning = 1;
6980
6981                 key.objectid = bytenr;
6982                 key.type = BTRFS_EXTENT_ITEM_KEY;
6983                 key.offset = num_bytes;
6984
6985                 ret = btrfs_search_slot(trans, extent_root, &key, path,
6986                                         -1, 1);
6987                 if (ret) {
6988                         btrfs_err(info,
6989                                   "umm, got %d back from search, was looking for %llu",
6990                                 ret, bytenr);
6991                         btrfs_print_leaf(path->nodes[0]);
6992                 }
6993                 if (ret < 0) {
6994                         btrfs_abort_transaction(trans, ret);
6995                         goto out;
6996                 }
6997
6998                 extent_slot = path->slots[0];
6999                 leaf = path->nodes[0];
7000                 item_size = btrfs_item_size_nr(leaf, extent_slot);
7001         }
7002 #endif
7003         BUG_ON(item_size < sizeof(*ei));
7004         ei = btrfs_item_ptr(leaf, extent_slot,
7005                             struct btrfs_extent_item);
7006         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
7007             key.type == BTRFS_EXTENT_ITEM_KEY) {
7008                 struct btrfs_tree_block_info *bi;
7009                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
7010                 bi = (struct btrfs_tree_block_info *)(ei + 1);
7011                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
7012         }
7013
7014         refs = btrfs_extent_refs(leaf, ei);
7015         if (refs < refs_to_drop) {
7016                 btrfs_err(info,
7017                           "trying to drop %d refs but we only have %Lu for bytenr %Lu",
7018                           refs_to_drop, refs, bytenr);
7019                 ret = -EINVAL;
7020                 btrfs_abort_transaction(trans, ret);
7021                 goto out;
7022         }
7023         refs -= refs_to_drop;
7024
7025         if (refs > 0) {
7026                 if (extent_op)
7027                         __run_delayed_extent_op(extent_op, leaf, ei);
7028                 /*
7029                  * In the case of inline back ref, reference count will
7030                  * be updated by remove_extent_backref
7031                  */
7032                 if (iref) {
7033                         BUG_ON(!found_extent);
7034                 } else {
7035                         btrfs_set_extent_refs(leaf, ei, refs);
7036                         btrfs_mark_buffer_dirty(leaf);
7037                 }
7038                 if (found_extent) {
7039                         ret = remove_extent_backref(trans, info, path,
7040                                                     iref, refs_to_drop,
7041                                                     is_data, &last_ref);
7042                         if (ret) {
7043                                 btrfs_abort_transaction(trans, ret);
7044                                 goto out;
7045                         }
7046                 }
7047         } else {
7048                 if (found_extent) {
7049                         BUG_ON(is_data && refs_to_drop !=
7050                                extent_data_ref_count(path, iref));
7051                         if (iref) {
7052                                 BUG_ON(path->slots[0] != extent_slot);
7053                         } else {
7054                                 BUG_ON(path->slots[0] != extent_slot + 1);
7055                                 path->slots[0] = extent_slot;
7056                                 num_to_del = 2;
7057                         }
7058                 }
7059
7060                 last_ref = 1;
7061                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
7062                                       num_to_del);
7063                 if (ret) {
7064                         btrfs_abort_transaction(trans, ret);
7065                         goto out;
7066                 }
7067                 btrfs_release_path(path);
7068
7069                 if (is_data) {
7070                         ret = btrfs_del_csums(trans, info, bytenr, num_bytes);
7071                         if (ret) {
7072                                 btrfs_abort_transaction(trans, ret);
7073                                 goto out;
7074                         }
7075                 }
7076
7077                 ret = add_to_free_space_tree(trans, info, bytenr, num_bytes);
7078                 if (ret) {
7079                         btrfs_abort_transaction(trans, ret);
7080                         goto out;
7081                 }
7082
7083                 ret = update_block_group(trans, info, bytenr, num_bytes, 0);
7084                 if (ret) {
7085                         btrfs_abort_transaction(trans, ret);
7086                         goto out;
7087                 }
7088         }
7089         btrfs_release_path(path);
7090
7091 out:
7092         btrfs_free_path(path);
7093         return ret;
7094 }
7095
7096 /*
7097  * when we free an block, it is possible (and likely) that we free the last
7098  * delayed ref for that extent as well.  This searches the delayed ref tree for
7099  * a given extent, and if there are no other delayed refs to be processed, it
7100  * removes it from the tree.
7101  */
7102 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
7103                                       u64 bytenr)
7104 {
7105         struct btrfs_delayed_ref_head *head;
7106         struct btrfs_delayed_ref_root *delayed_refs;
7107         int ret = 0;
7108
7109         delayed_refs = &trans->transaction->delayed_refs;
7110         spin_lock(&delayed_refs->lock);
7111         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
7112         if (!head)
7113                 goto out_delayed_unlock;
7114
7115         spin_lock(&head->lock);
7116         if (!RB_EMPTY_ROOT(&head->ref_tree))
7117                 goto out;
7118
7119         if (head->extent_op) {
7120                 if (!head->must_insert_reserved)
7121                         goto out;
7122                 btrfs_free_delayed_extent_op(head->extent_op);
7123                 head->extent_op = NULL;
7124         }
7125
7126         /*
7127          * waiting for the lock here would deadlock.  If someone else has it
7128          * locked they are already in the process of dropping it anyway
7129          */
7130         if (!mutex_trylock(&head->mutex))
7131                 goto out;
7132
7133         /*
7134          * at this point we have a head with no other entries.  Go
7135          * ahead and process it.
7136          */
7137         rb_erase(&head->href_node, &delayed_refs->href_root);
7138         RB_CLEAR_NODE(&head->href_node);
7139         atomic_dec(&delayed_refs->num_entries);
7140
7141         /*
7142          * we don't take a ref on the node because we're removing it from the
7143          * tree, so we just steal the ref the tree was holding.
7144          */
7145         delayed_refs->num_heads--;
7146         if (head->processing == 0)
7147                 delayed_refs->num_heads_ready--;
7148         head->processing = 0;
7149         spin_unlock(&head->lock);
7150         spin_unlock(&delayed_refs->lock);
7151
7152         BUG_ON(head->extent_op);
7153         if (head->must_insert_reserved)
7154                 ret = 1;
7155
7156         mutex_unlock(&head->mutex);
7157         btrfs_put_delayed_ref_head(head);
7158         return ret;
7159 out:
7160         spin_unlock(&head->lock);
7161
7162 out_delayed_unlock:
7163         spin_unlock(&delayed_refs->lock);
7164         return 0;
7165 }
7166
7167 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
7168                            struct btrfs_root *root,
7169                            struct extent_buffer *buf,
7170                            u64 parent, int last_ref)
7171 {
7172         struct btrfs_fs_info *fs_info = root->fs_info;
7173         int pin = 1;
7174         int ret;
7175
7176         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7177                 int old_ref_mod, new_ref_mod;
7178
7179                 btrfs_ref_tree_mod(root, buf->start, buf->len, parent,
7180                                    root->root_key.objectid,
7181                                    btrfs_header_level(buf), 0,
7182                                    BTRFS_DROP_DELAYED_REF);
7183                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, buf->start,
7184                                                  buf->len, parent,
7185                                                  root->root_key.objectid,
7186                                                  btrfs_header_level(buf),
7187                                                  BTRFS_DROP_DELAYED_REF, NULL,
7188                                                  &old_ref_mod, &new_ref_mod);
7189                 BUG_ON(ret); /* -ENOMEM */
7190                 pin = old_ref_mod >= 0 && new_ref_mod < 0;
7191         }
7192
7193         if (last_ref && btrfs_header_generation(buf) == trans->transid) {
7194                 struct btrfs_block_group_cache *cache;
7195
7196                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7197                         ret = check_ref_cleanup(trans, buf->start);
7198                         if (!ret)
7199                                 goto out;
7200                 }
7201
7202                 pin = 0;
7203                 cache = btrfs_lookup_block_group(fs_info, buf->start);
7204
7205                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
7206                         pin_down_extent(fs_info, cache, buf->start,
7207                                         buf->len, 1);
7208                         btrfs_put_block_group(cache);
7209                         goto out;
7210                 }
7211
7212                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
7213
7214                 btrfs_add_free_space(cache, buf->start, buf->len);
7215                 btrfs_free_reserved_bytes(cache, buf->len, 0);
7216                 btrfs_put_block_group(cache);
7217                 trace_btrfs_reserved_extent_free(fs_info, buf->start, buf->len);
7218         }
7219 out:
7220         if (pin)
7221                 add_pinned_bytes(fs_info, buf->len, btrfs_header_level(buf),
7222                                  root->root_key.objectid);
7223
7224         if (last_ref) {
7225                 /*
7226                  * Deleting the buffer, clear the corrupt flag since it doesn't
7227                  * matter anymore.
7228                  */
7229                 clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
7230         }
7231 }
7232
7233 /* Can return -ENOMEM */
7234 int btrfs_free_extent(struct btrfs_trans_handle *trans,
7235                       struct btrfs_root *root,
7236                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
7237                       u64 owner, u64 offset)
7238 {
7239         struct btrfs_fs_info *fs_info = root->fs_info;
7240         int old_ref_mod, new_ref_mod;
7241         int ret;
7242
7243         if (btrfs_is_testing(fs_info))
7244                 return 0;
7245
7246         if (root_objectid != BTRFS_TREE_LOG_OBJECTID)
7247                 btrfs_ref_tree_mod(root, bytenr, num_bytes, parent,
7248                                    root_objectid, owner, offset,
7249                                    BTRFS_DROP_DELAYED_REF);
7250
7251         /*
7252          * tree log blocks never actually go into the extent allocation
7253          * tree, just update pinning info and exit early.
7254          */
7255         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
7256                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
7257                 /* unlocks the pinned mutex */
7258                 btrfs_pin_extent(fs_info, bytenr, num_bytes, 1);
7259                 old_ref_mod = new_ref_mod = 0;
7260                 ret = 0;
7261         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
7262                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
7263                                                  num_bytes, parent,
7264                                                  root_objectid, (int)owner,
7265                                                  BTRFS_DROP_DELAYED_REF, NULL,
7266                                                  &old_ref_mod, &new_ref_mod);
7267         } else {
7268                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
7269                                                  num_bytes, parent,
7270                                                  root_objectid, owner, offset,
7271                                                  0, BTRFS_DROP_DELAYED_REF,
7272                                                  &old_ref_mod, &new_ref_mod);
7273         }
7274
7275         if (ret == 0 && old_ref_mod >= 0 && new_ref_mod < 0)
7276                 add_pinned_bytes(fs_info, num_bytes, owner, root_objectid);
7277
7278         return ret;
7279 }
7280
7281 /*
7282  * when we wait for progress in the block group caching, its because
7283  * our allocation attempt failed at least once.  So, we must sleep
7284  * and let some progress happen before we try again.
7285  *
7286  * This function will sleep at least once waiting for new free space to
7287  * show up, and then it will check the block group free space numbers
7288  * for our min num_bytes.  Another option is to have it go ahead
7289  * and look in the rbtree for a free extent of a given size, but this
7290  * is a good start.
7291  *
7292  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
7293  * any of the information in this block group.
7294  */
7295 static noinline void
7296 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
7297                                 u64 num_bytes)
7298 {
7299         struct btrfs_caching_control *caching_ctl;
7300
7301         caching_ctl = get_caching_control(cache);
7302         if (!caching_ctl)
7303                 return;
7304
7305         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
7306                    (cache->free_space_ctl->free_space >= num_bytes));
7307
7308         put_caching_control(caching_ctl);
7309 }
7310
7311 static noinline int
7312 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
7313 {
7314         struct btrfs_caching_control *caching_ctl;
7315         int ret = 0;
7316
7317         caching_ctl = get_caching_control(cache);
7318         if (!caching_ctl)
7319                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
7320
7321         wait_event(caching_ctl->wait, block_group_cache_done(cache));
7322         if (cache->cached == BTRFS_CACHE_ERROR)
7323                 ret = -EIO;
7324         put_caching_control(caching_ctl);
7325         return ret;
7326 }
7327
7328 static const char *btrfs_raid_type_names[BTRFS_NR_RAID_TYPES] = {
7329         [BTRFS_RAID_RAID10]     = "raid10",
7330         [BTRFS_RAID_RAID1]      = "raid1",
7331         [BTRFS_RAID_DUP]        = "dup",
7332         [BTRFS_RAID_RAID0]      = "raid0",
7333         [BTRFS_RAID_SINGLE]     = "single",
7334         [BTRFS_RAID_RAID5]      = "raid5",
7335         [BTRFS_RAID_RAID6]      = "raid6",
7336 };
7337
7338 static const char *get_raid_name(enum btrfs_raid_types type)
7339 {
7340         if (type >= BTRFS_NR_RAID_TYPES)
7341                 return NULL;
7342
7343         return btrfs_raid_type_names[type];
7344 }
7345
7346 enum btrfs_loop_type {
7347         LOOP_CACHING_NOWAIT = 0,
7348         LOOP_CACHING_WAIT = 1,
7349         LOOP_ALLOC_CHUNK = 2,
7350         LOOP_NO_EMPTY_SIZE = 3,
7351 };
7352
7353 static inline void
7354 btrfs_lock_block_group(struct btrfs_block_group_cache *cache,
7355                        int delalloc)
7356 {
7357         if (delalloc)
7358                 down_read(&cache->data_rwsem);
7359 }
7360
7361 static inline void
7362 btrfs_grab_block_group(struct btrfs_block_group_cache *cache,
7363                        int delalloc)
7364 {
7365         btrfs_get_block_group(cache);
7366         if (delalloc)
7367                 down_read(&cache->data_rwsem);
7368 }
7369
7370 static struct btrfs_block_group_cache *
7371 btrfs_lock_cluster(struct btrfs_block_group_cache *block_group,
7372                    struct btrfs_free_cluster *cluster,
7373                    int delalloc)
7374 {
7375         struct btrfs_block_group_cache *used_bg = NULL;
7376
7377         spin_lock(&cluster->refill_lock);
7378         while (1) {
7379                 used_bg = cluster->block_group;
7380                 if (!used_bg)
7381                         return NULL;
7382
7383                 if (used_bg == block_group)
7384                         return used_bg;
7385
7386                 btrfs_get_block_group(used_bg);
7387
7388                 if (!delalloc)
7389                         return used_bg;
7390
7391                 if (down_read_trylock(&used_bg->data_rwsem))
7392                         return used_bg;
7393
7394                 spin_unlock(&cluster->refill_lock);
7395
7396                 /* We should only have one-level nested. */
7397                 down_read_nested(&used_bg->data_rwsem, SINGLE_DEPTH_NESTING);
7398
7399                 spin_lock(&cluster->refill_lock);
7400                 if (used_bg == cluster->block_group)
7401                         return used_bg;
7402
7403                 up_read(&used_bg->data_rwsem);
7404                 btrfs_put_block_group(used_bg);
7405         }
7406 }
7407
7408 static inline void
7409 btrfs_release_block_group(struct btrfs_block_group_cache *cache,
7410                          int delalloc)
7411 {
7412         if (delalloc)
7413                 up_read(&cache->data_rwsem);
7414         btrfs_put_block_group(cache);
7415 }
7416
7417 /*
7418  * walks the btree of allocated extents and find a hole of a given size.
7419  * The key ins is changed to record the hole:
7420  * ins->objectid == start position
7421  * ins->flags = BTRFS_EXTENT_ITEM_KEY
7422  * ins->offset == the size of the hole.
7423  * Any available blocks before search_start are skipped.
7424  *
7425  * If there is no suitable free space, we will record the max size of
7426  * the free space extent currently.
7427  */
7428 static noinline int find_free_extent(struct btrfs_fs_info *fs_info,
7429                                 u64 ram_bytes, u64 num_bytes, u64 empty_size,
7430                                 u64 hint_byte, struct btrfs_key *ins,
7431                                 u64 flags, int delalloc)
7432 {
7433         int ret = 0;
7434         struct btrfs_root *root = fs_info->extent_root;
7435         struct btrfs_free_cluster *last_ptr = NULL;
7436         struct btrfs_block_group_cache *block_group = NULL;
7437         u64 search_start = 0;
7438         u64 max_extent_size = 0;
7439         u64 empty_cluster = 0;
7440         struct btrfs_space_info *space_info;
7441         int loop = 0;
7442         int index = btrfs_bg_flags_to_raid_index(flags);
7443         bool failed_cluster_refill = false;
7444         bool failed_alloc = false;
7445         bool use_cluster = true;
7446         bool have_caching_bg = false;
7447         bool orig_have_caching_bg = false;
7448         bool full_search = false;
7449
7450         WARN_ON(num_bytes < fs_info->sectorsize);
7451         ins->type = BTRFS_EXTENT_ITEM_KEY;
7452         ins->objectid = 0;
7453         ins->offset = 0;
7454
7455         trace_find_free_extent(fs_info, num_bytes, empty_size, flags);
7456
7457         space_info = __find_space_info(fs_info, flags);
7458         if (!space_info) {
7459                 btrfs_err(fs_info, "No space info for %llu", flags);
7460                 return -ENOSPC;
7461         }
7462
7463         /*
7464          * If our free space is heavily fragmented we may not be able to make
7465          * big contiguous allocations, so instead of doing the expensive search
7466          * for free space, simply return ENOSPC with our max_extent_size so we
7467          * can go ahead and search for a more manageable chunk.
7468          *
7469          * If our max_extent_size is large enough for our allocation simply
7470          * disable clustering since we will likely not be able to find enough
7471          * space to create a cluster and induce latency trying.
7472          */
7473         if (unlikely(space_info->max_extent_size)) {
7474                 spin_lock(&space_info->lock);
7475                 if (space_info->max_extent_size &&
7476                     num_bytes > space_info->max_extent_size) {
7477                         ins->offset = space_info->max_extent_size;
7478                         spin_unlock(&space_info->lock);
7479                         return -ENOSPC;
7480                 } else if (space_info->max_extent_size) {
7481                         use_cluster = false;
7482                 }
7483                 spin_unlock(&space_info->lock);
7484         }
7485
7486         last_ptr = fetch_cluster_info(fs_info, space_info, &empty_cluster);
7487         if (last_ptr) {
7488                 spin_lock(&last_ptr->lock);
7489                 if (last_ptr->block_group)
7490                         hint_byte = last_ptr->window_start;
7491                 if (last_ptr->fragmented) {
7492                         /*
7493                          * We still set window_start so we can keep track of the
7494                          * last place we found an allocation to try and save
7495                          * some time.
7496                          */
7497                         hint_byte = last_ptr->window_start;
7498                         use_cluster = false;
7499                 }
7500                 spin_unlock(&last_ptr->lock);
7501         }
7502
7503         search_start = max(search_start, first_logical_byte(fs_info, 0));
7504         search_start = max(search_start, hint_byte);
7505         if (search_start == hint_byte) {
7506                 block_group = btrfs_lookup_block_group(fs_info, search_start);
7507                 /*
7508                  * we don't want to use the block group if it doesn't match our
7509                  * allocation bits, or if its not cached.
7510                  *
7511                  * However if we are re-searching with an ideal block group
7512                  * picked out then we don't care that the block group is cached.
7513                  */
7514                 if (block_group && block_group_bits(block_group, flags) &&
7515                     block_group->cached != BTRFS_CACHE_NO) {
7516                         down_read(&space_info->groups_sem);
7517                         if (list_empty(&block_group->list) ||
7518                             block_group->ro) {
7519                                 /*
7520                                  * someone is removing this block group,
7521                                  * we can't jump into the have_block_group
7522                                  * target because our list pointers are not
7523                                  * valid
7524                                  */
7525                                 btrfs_put_block_group(block_group);
7526                                 up_read(&space_info->groups_sem);
7527                         } else {
7528                                 index = btrfs_bg_flags_to_raid_index(
7529                                                 block_group->flags);
7530                                 btrfs_lock_block_group(block_group, delalloc);
7531                                 goto have_block_group;
7532                         }
7533                 } else if (block_group) {
7534                         btrfs_put_block_group(block_group);
7535                 }
7536         }
7537 search:
7538         have_caching_bg = false;
7539         if (index == 0 || index == btrfs_bg_flags_to_raid_index(flags))
7540                 full_search = true;
7541         down_read(&space_info->groups_sem);
7542         list_for_each_entry(block_group, &space_info->block_groups[index],
7543                             list) {
7544                 u64 offset;
7545                 int cached;
7546
7547                 /* If the block group is read-only, we can skip it entirely. */
7548                 if (unlikely(block_group->ro))
7549                         continue;
7550
7551                 btrfs_grab_block_group(block_group, delalloc);
7552                 search_start = block_group->key.objectid;
7553
7554                 /*
7555                  * this can happen if we end up cycling through all the
7556                  * raid types, but we want to make sure we only allocate
7557                  * for the proper type.
7558                  */
7559                 if (!block_group_bits(block_group, flags)) {
7560                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
7561                                 BTRFS_BLOCK_GROUP_RAID1 |
7562                                 BTRFS_BLOCK_GROUP_RAID5 |
7563                                 BTRFS_BLOCK_GROUP_RAID6 |
7564                                 BTRFS_BLOCK_GROUP_RAID10;
7565
7566                         /*
7567                          * if they asked for extra copies and this block group
7568                          * doesn't provide them, bail.  This does allow us to
7569                          * fill raid0 from raid1.
7570                          */
7571                         if ((flags & extra) && !(block_group->flags & extra))
7572                                 goto loop;
7573                 }
7574
7575 have_block_group:
7576                 cached = block_group_cache_done(block_group);
7577                 if (unlikely(!cached)) {
7578                         have_caching_bg = true;
7579                         ret = cache_block_group(block_group, 0);
7580                         BUG_ON(ret < 0);
7581                         ret = 0;
7582                 }
7583
7584                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
7585                         goto loop;
7586
7587                 /*
7588                  * Ok we want to try and use the cluster allocator, so
7589                  * lets look there
7590                  */
7591                 if (last_ptr && use_cluster) {
7592                         struct btrfs_block_group_cache *used_block_group;
7593                         unsigned long aligned_cluster;
7594                         /*
7595                          * the refill lock keeps out other
7596                          * people trying to start a new cluster
7597                          */
7598                         used_block_group = btrfs_lock_cluster(block_group,
7599                                                               last_ptr,
7600                                                               delalloc);
7601                         if (!used_block_group)
7602                                 goto refill_cluster;
7603
7604                         if (used_block_group != block_group &&
7605                             (used_block_group->ro ||
7606                              !block_group_bits(used_block_group, flags)))
7607                                 goto release_cluster;
7608
7609                         offset = btrfs_alloc_from_cluster(used_block_group,
7610                                                 last_ptr,
7611                                                 num_bytes,
7612                                                 used_block_group->key.objectid,
7613                                                 &max_extent_size);
7614                         if (offset) {
7615                                 /* we have a block, we're done */
7616                                 spin_unlock(&last_ptr->refill_lock);
7617                                 trace_btrfs_reserve_extent_cluster(fs_info,
7618                                                 used_block_group,
7619                                                 search_start, num_bytes);
7620                                 if (used_block_group != block_group) {
7621                                         btrfs_release_block_group(block_group,
7622                                                                   delalloc);
7623                                         block_group = used_block_group;
7624                                 }
7625                                 goto checks;
7626                         }
7627
7628                         WARN_ON(last_ptr->block_group != used_block_group);
7629 release_cluster:
7630                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
7631                          * set up a new clusters, so lets just skip it
7632                          * and let the allocator find whatever block
7633                          * it can find.  If we reach this point, we
7634                          * will have tried the cluster allocator
7635                          * plenty of times and not have found
7636                          * anything, so we are likely way too
7637                          * fragmented for the clustering stuff to find
7638                          * anything.
7639                          *
7640                          * However, if the cluster is taken from the
7641                          * current block group, release the cluster
7642                          * first, so that we stand a better chance of
7643                          * succeeding in the unclustered
7644                          * allocation.  */
7645                         if (loop >= LOOP_NO_EMPTY_SIZE &&
7646                             used_block_group != block_group) {
7647                                 spin_unlock(&last_ptr->refill_lock);
7648                                 btrfs_release_block_group(used_block_group,
7649                                                           delalloc);
7650                                 goto unclustered_alloc;
7651                         }
7652
7653                         /*
7654                          * this cluster didn't work out, free it and
7655                          * start over
7656                          */
7657                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7658
7659                         if (used_block_group != block_group)
7660                                 btrfs_release_block_group(used_block_group,
7661                                                           delalloc);
7662 refill_cluster:
7663                         if (loop >= LOOP_NO_EMPTY_SIZE) {
7664                                 spin_unlock(&last_ptr->refill_lock);
7665                                 goto unclustered_alloc;
7666                         }
7667
7668                         aligned_cluster = max_t(unsigned long,
7669                                                 empty_cluster + empty_size,
7670                                               block_group->full_stripe_len);
7671
7672                         /* allocate a cluster in this block group */
7673                         ret = btrfs_find_space_cluster(fs_info, block_group,
7674                                                        last_ptr, search_start,
7675                                                        num_bytes,
7676                                                        aligned_cluster);
7677                         if (ret == 0) {
7678                                 /*
7679                                  * now pull our allocation out of this
7680                                  * cluster
7681                                  */
7682                                 offset = btrfs_alloc_from_cluster(block_group,
7683                                                         last_ptr,
7684                                                         num_bytes,
7685                                                         search_start,
7686                                                         &max_extent_size);
7687                                 if (offset) {
7688                                         /* we found one, proceed */
7689                                         spin_unlock(&last_ptr->refill_lock);
7690                                         trace_btrfs_reserve_extent_cluster(fs_info,
7691                                                 block_group, search_start,
7692                                                 num_bytes);
7693                                         goto checks;
7694                                 }
7695                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
7696                                    && !failed_cluster_refill) {
7697                                 spin_unlock(&last_ptr->refill_lock);
7698
7699                                 failed_cluster_refill = true;
7700                                 wait_block_group_cache_progress(block_group,
7701                                        num_bytes + empty_cluster + empty_size);
7702                                 goto have_block_group;
7703                         }
7704
7705                         /*
7706                          * at this point we either didn't find a cluster
7707                          * or we weren't able to allocate a block from our
7708                          * cluster.  Free the cluster we've been trying
7709                          * to use, and go to the next block group
7710                          */
7711                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7712                         spin_unlock(&last_ptr->refill_lock);
7713                         goto loop;
7714                 }
7715
7716 unclustered_alloc:
7717                 /*
7718                  * We are doing an unclustered alloc, set the fragmented flag so
7719                  * we don't bother trying to setup a cluster again until we get
7720                  * more space.
7721                  */
7722                 if (unlikely(last_ptr)) {
7723                         spin_lock(&last_ptr->lock);
7724                         last_ptr->fragmented = 1;
7725                         spin_unlock(&last_ptr->lock);
7726                 }
7727                 if (cached) {
7728                         struct btrfs_free_space_ctl *ctl =
7729                                 block_group->free_space_ctl;
7730
7731                         spin_lock(&ctl->tree_lock);
7732                         if (ctl->free_space <
7733                             num_bytes + empty_cluster + empty_size) {
7734                                 if (ctl->free_space > max_extent_size)
7735                                         max_extent_size = ctl->free_space;
7736                                 spin_unlock(&ctl->tree_lock);
7737                                 goto loop;
7738                         }
7739                         spin_unlock(&ctl->tree_lock);
7740                 }
7741
7742                 offset = btrfs_find_space_for_alloc(block_group, search_start,
7743                                                     num_bytes, empty_size,
7744                                                     &max_extent_size);
7745                 /*
7746                  * If we didn't find a chunk, and we haven't failed on this
7747                  * block group before, and this block group is in the middle of
7748                  * caching and we are ok with waiting, then go ahead and wait
7749                  * for progress to be made, and set failed_alloc to true.
7750                  *
7751                  * If failed_alloc is true then we've already waited on this
7752                  * block group once and should move on to the next block group.
7753                  */
7754                 if (!offset && !failed_alloc && !cached &&
7755                     loop > LOOP_CACHING_NOWAIT) {
7756                         wait_block_group_cache_progress(block_group,
7757                                                 num_bytes + empty_size);
7758                         failed_alloc = true;
7759                         goto have_block_group;
7760                 } else if (!offset) {
7761                         goto loop;
7762                 }
7763 checks:
7764                 search_start = ALIGN(offset, fs_info->stripesize);
7765
7766                 /* move on to the next group */
7767                 if (search_start + num_bytes >
7768                     block_group->key.objectid + block_group->key.offset) {
7769                         btrfs_add_free_space(block_group, offset, num_bytes);
7770                         goto loop;
7771                 }
7772
7773                 if (offset < search_start)
7774                         btrfs_add_free_space(block_group, offset,
7775                                              search_start - offset);
7776                 BUG_ON(offset > search_start);
7777
7778                 ret = btrfs_add_reserved_bytes(block_group, ram_bytes,
7779                                 num_bytes, delalloc);
7780                 if (ret == -EAGAIN) {
7781                         btrfs_add_free_space(block_group, offset, num_bytes);
7782                         goto loop;
7783                 }
7784                 btrfs_inc_block_group_reservations(block_group);
7785
7786                 /* we are all good, lets return */
7787                 ins->objectid = search_start;
7788                 ins->offset = num_bytes;
7789
7790                 trace_btrfs_reserve_extent(fs_info, block_group,
7791                                            search_start, num_bytes);
7792                 btrfs_release_block_group(block_group, delalloc);
7793                 break;
7794 loop:
7795                 failed_cluster_refill = false;
7796                 failed_alloc = false;
7797                 BUG_ON(btrfs_bg_flags_to_raid_index(block_group->flags) !=
7798                        index);
7799                 btrfs_release_block_group(block_group, delalloc);
7800                 cond_resched();
7801         }
7802         up_read(&space_info->groups_sem);
7803
7804         if ((loop == LOOP_CACHING_NOWAIT) && have_caching_bg
7805                 && !orig_have_caching_bg)
7806                 orig_have_caching_bg = true;
7807
7808         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
7809                 goto search;
7810
7811         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
7812                 goto search;
7813
7814         /*
7815          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
7816          *                      caching kthreads as we move along
7817          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
7818          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
7819          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
7820          *                      again
7821          */
7822         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
7823                 index = 0;
7824                 if (loop == LOOP_CACHING_NOWAIT) {
7825                         /*
7826                          * We want to skip the LOOP_CACHING_WAIT step if we
7827                          * don't have any uncached bgs and we've already done a
7828                          * full search through.
7829                          */
7830                         if (orig_have_caching_bg || !full_search)
7831                                 loop = LOOP_CACHING_WAIT;
7832                         else
7833                                 loop = LOOP_ALLOC_CHUNK;
7834                 } else {
7835                         loop++;
7836                 }
7837
7838                 if (loop == LOOP_ALLOC_CHUNK) {
7839                         struct btrfs_trans_handle *trans;
7840                         int exist = 0;
7841
7842                         trans = current->journal_info;
7843                         if (trans)
7844                                 exist = 1;
7845                         else
7846                                 trans = btrfs_join_transaction(root);
7847
7848                         if (IS_ERR(trans)) {
7849                                 ret = PTR_ERR(trans);
7850                                 goto out;
7851                         }
7852
7853                         ret = do_chunk_alloc(trans, fs_info, flags,
7854                                              CHUNK_ALLOC_FORCE);
7855
7856                         /*
7857                          * If we can't allocate a new chunk we've already looped
7858                          * through at least once, move on to the NO_EMPTY_SIZE
7859                          * case.
7860                          */
7861                         if (ret == -ENOSPC)
7862                                 loop = LOOP_NO_EMPTY_SIZE;
7863
7864                         /*
7865                          * Do not bail out on ENOSPC since we
7866                          * can do more things.
7867                          */
7868                         if (ret < 0 && ret != -ENOSPC)
7869                                 btrfs_abort_transaction(trans, ret);
7870                         else
7871                                 ret = 0;
7872                         if (!exist)
7873                                 btrfs_end_transaction(trans);
7874                         if (ret)
7875                                 goto out;
7876                 }
7877
7878                 if (loop == LOOP_NO_EMPTY_SIZE) {
7879                         /*
7880                          * Don't loop again if we already have no empty_size and
7881                          * no empty_cluster.
7882                          */
7883                         if (empty_size == 0 &&
7884                             empty_cluster == 0) {
7885                                 ret = -ENOSPC;
7886                                 goto out;
7887                         }
7888                         empty_size = 0;
7889                         empty_cluster = 0;
7890                 }
7891
7892                 goto search;
7893         } else if (!ins->objectid) {
7894                 ret = -ENOSPC;
7895         } else if (ins->objectid) {
7896                 if (!use_cluster && last_ptr) {
7897                         spin_lock(&last_ptr->lock);
7898                         last_ptr->window_start = ins->objectid;
7899                         spin_unlock(&last_ptr->lock);
7900                 }
7901                 ret = 0;
7902         }
7903 out:
7904         if (ret == -ENOSPC) {
7905                 spin_lock(&space_info->lock);
7906                 space_info->max_extent_size = max_extent_size;
7907                 spin_unlock(&space_info->lock);
7908                 ins->offset = max_extent_size;
7909         }
7910         return ret;
7911 }
7912
7913 static void dump_space_info(struct btrfs_fs_info *fs_info,
7914                             struct btrfs_space_info *info, u64 bytes,
7915                             int dump_block_groups)
7916 {
7917         struct btrfs_block_group_cache *cache;
7918         int index = 0;
7919
7920         spin_lock(&info->lock);
7921         btrfs_info(fs_info, "space_info %llu has %llu free, is %sfull",
7922                    info->flags,
7923                    info->total_bytes - btrfs_space_info_used(info, true),
7924                    info->full ? "" : "not ");
7925         btrfs_info(fs_info,
7926                 "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu",
7927                 info->total_bytes, info->bytes_used, info->bytes_pinned,
7928                 info->bytes_reserved, info->bytes_may_use,
7929                 info->bytes_readonly);
7930         spin_unlock(&info->lock);
7931
7932         if (!dump_block_groups)
7933                 return;
7934
7935         down_read(&info->groups_sem);
7936 again:
7937         list_for_each_entry(cache, &info->block_groups[index], list) {
7938                 spin_lock(&cache->lock);
7939                 btrfs_info(fs_info,
7940                         "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s",
7941                         cache->key.objectid, cache->key.offset,
7942                         btrfs_block_group_used(&cache->item), cache->pinned,
7943                         cache->reserved, cache->ro ? "[readonly]" : "");
7944                 btrfs_dump_free_space(cache, bytes);
7945                 spin_unlock(&cache->lock);
7946         }
7947         if (++index < BTRFS_NR_RAID_TYPES)
7948                 goto again;
7949         up_read(&info->groups_sem);
7950 }
7951
7952 /*
7953  * btrfs_reserve_extent - entry point to the extent allocator. Tries to find a
7954  *                        hole that is at least as big as @num_bytes.
7955  *
7956  * @root           -    The root that will contain this extent
7957  *
7958  * @ram_bytes      -    The amount of space in ram that @num_bytes take. This
7959  *                      is used for accounting purposes. This value differs
7960  *                      from @num_bytes only in the case of compressed extents.
7961  *
7962  * @num_bytes      -    Number of bytes to allocate on-disk.
7963  *
7964  * @min_alloc_size -    Indicates the minimum amount of space that the
7965  *                      allocator should try to satisfy. In some cases
7966  *                      @num_bytes may be larger than what is required and if
7967  *                      the filesystem is fragmented then allocation fails.
7968  *                      However, the presence of @min_alloc_size gives a
7969  *                      chance to try and satisfy the smaller allocation.
7970  *
7971  * @empty_size     -    A hint that you plan on doing more COW. This is the
7972  *                      size in bytes the allocator should try to find free
7973  *                      next to the block it returns.  This is just a hint and
7974  *                      may be ignored by the allocator.
7975  *
7976  * @hint_byte      -    Hint to the allocator to start searching above the byte
7977  *                      address passed. It might be ignored.
7978  *
7979  * @ins            -    This key is modified to record the found hole. It will
7980  *                      have the following values:
7981  *                      ins->objectid == start position
7982  *                      ins->flags = BTRFS_EXTENT_ITEM_KEY
7983  *                      ins->offset == the size of the hole.
7984  *
7985  * @is_data        -    Boolean flag indicating whether an extent is
7986  *                      allocated for data (true) or metadata (false)
7987  *
7988  * @delalloc       -    Boolean flag indicating whether this allocation is for
7989  *                      delalloc or not. If 'true' data_rwsem of block groups
7990  *                      is going to be acquired.
7991  *
7992  *
7993  * Returns 0 when an allocation succeeded or < 0 when an error occurred. In
7994  * case -ENOSPC is returned then @ins->offset will contain the size of the
7995  * largest available hole the allocator managed to find.
7996  */
7997 int btrfs_reserve_extent(struct btrfs_root *root, u64 ram_bytes,
7998                          u64 num_bytes, u64 min_alloc_size,
7999                          u64 empty_size, u64 hint_byte,
8000                          struct btrfs_key *ins, int is_data, int delalloc)
8001 {
8002         struct btrfs_fs_info *fs_info = root->fs_info;
8003         bool final_tried = num_bytes == min_alloc_size;
8004         u64 flags;
8005         int ret;
8006
8007         flags = get_alloc_profile_by_root(root, is_data);
8008 again:
8009         WARN_ON(num_bytes < fs_info->sectorsize);
8010         ret = find_free_extent(fs_info, ram_bytes, num_bytes, empty_size,
8011                                hint_byte, ins, flags, delalloc);
8012         if (!ret && !is_data) {
8013                 btrfs_dec_block_group_reservations(fs_info, ins->objectid);
8014         } else if (ret == -ENOSPC) {
8015                 if (!final_tried && ins->offset) {
8016                         num_bytes = min(num_bytes >> 1, ins->offset);
8017                         num_bytes = round_down(num_bytes,
8018                                                fs_info->sectorsize);
8019                         num_bytes = max(num_bytes, min_alloc_size);
8020                         ram_bytes = num_bytes;
8021                         if (num_bytes == min_alloc_size)
8022                                 final_tried = true;
8023                         goto again;
8024                 } else if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
8025                         struct btrfs_space_info *sinfo;
8026
8027                         sinfo = __find_space_info(fs_info, flags);
8028                         btrfs_err(fs_info,
8029                                   "allocation failed flags %llu, wanted %llu",
8030                                   flags, num_bytes);
8031                         if (sinfo)
8032                                 dump_space_info(fs_info, sinfo, num_bytes, 1);
8033                 }
8034         }
8035
8036         return ret;
8037 }
8038
8039 static int __btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
8040                                         u64 start, u64 len,
8041                                         int pin, int delalloc)
8042 {
8043         struct btrfs_block_group_cache *cache;
8044         int ret = 0;
8045
8046         cache = btrfs_lookup_block_group(fs_info, start);
8047         if (!cache) {
8048                 btrfs_err(fs_info, "Unable to find block group for %llu",
8049                           start);
8050                 return -ENOSPC;
8051         }
8052
8053         if (pin)
8054                 pin_down_extent(fs_info, cache, start, len, 1);
8055         else {
8056                 if (btrfs_test_opt(fs_info, DISCARD))
8057                         ret = btrfs_discard_extent(fs_info, start, len, NULL);
8058                 btrfs_add_free_space(cache, start, len);
8059                 btrfs_free_reserved_bytes(cache, len, delalloc);
8060                 trace_btrfs_reserved_extent_free(fs_info, start, len);
8061         }
8062
8063         btrfs_put_block_group(cache);
8064         return ret;
8065 }
8066
8067 int btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
8068                                u64 start, u64 len, int delalloc)
8069 {
8070         return __btrfs_free_reserved_extent(fs_info, start, len, 0, delalloc);
8071 }
8072
8073 int btrfs_free_and_pin_reserved_extent(struct btrfs_fs_info *fs_info,
8074                                        u64 start, u64 len)
8075 {
8076         return __btrfs_free_reserved_extent(fs_info, start, len, 1, 0);
8077 }
8078
8079 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8080                                       struct btrfs_fs_info *fs_info,
8081                                       u64 parent, u64 root_objectid,
8082                                       u64 flags, u64 owner, u64 offset,
8083                                       struct btrfs_key *ins, int ref_mod)
8084 {
8085         int ret;
8086         struct btrfs_extent_item *extent_item;
8087         struct btrfs_extent_inline_ref *iref;
8088         struct btrfs_path *path;
8089         struct extent_buffer *leaf;
8090         int type;
8091         u32 size;
8092
8093         if (parent > 0)
8094                 type = BTRFS_SHARED_DATA_REF_KEY;
8095         else
8096                 type = BTRFS_EXTENT_DATA_REF_KEY;
8097
8098         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
8099
8100         path = btrfs_alloc_path();
8101         if (!path)
8102                 return -ENOMEM;
8103
8104         path->leave_spinning = 1;
8105         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
8106                                       ins, size);
8107         if (ret) {
8108                 btrfs_free_path(path);
8109                 return ret;
8110         }
8111
8112         leaf = path->nodes[0];
8113         extent_item = btrfs_item_ptr(leaf, path->slots[0],
8114                                      struct btrfs_extent_item);
8115         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
8116         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8117         btrfs_set_extent_flags(leaf, extent_item,
8118                                flags | BTRFS_EXTENT_FLAG_DATA);
8119
8120         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8121         btrfs_set_extent_inline_ref_type(leaf, iref, type);
8122         if (parent > 0) {
8123                 struct btrfs_shared_data_ref *ref;
8124                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
8125                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
8126                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
8127         } else {
8128                 struct btrfs_extent_data_ref *ref;
8129                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
8130                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
8131                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
8132                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
8133                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
8134         }
8135
8136         btrfs_mark_buffer_dirty(path->nodes[0]);
8137         btrfs_free_path(path);
8138
8139         ret = remove_from_free_space_tree(trans, fs_info, ins->objectid,
8140                                           ins->offset);
8141         if (ret)
8142                 return ret;
8143
8144         ret = update_block_group(trans, fs_info, ins->objectid, ins->offset, 1);
8145         if (ret) { /* -ENOENT, logic error */
8146                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8147                         ins->objectid, ins->offset);
8148                 BUG();
8149         }
8150         trace_btrfs_reserved_extent_alloc(fs_info, ins->objectid, ins->offset);
8151         return ret;
8152 }
8153
8154 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
8155                                      struct btrfs_fs_info *fs_info,
8156                                      u64 parent, u64 root_objectid,
8157                                      u64 flags, struct btrfs_disk_key *key,
8158                                      int level, struct btrfs_key *ins)
8159 {
8160         int ret;
8161         struct btrfs_extent_item *extent_item;
8162         struct btrfs_tree_block_info *block_info;
8163         struct btrfs_extent_inline_ref *iref;
8164         struct btrfs_path *path;
8165         struct extent_buffer *leaf;
8166         u32 size = sizeof(*extent_item) + sizeof(*iref);
8167         u64 num_bytes = ins->offset;
8168         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
8169
8170         if (!skinny_metadata)
8171                 size += sizeof(*block_info);
8172
8173         path = btrfs_alloc_path();
8174         if (!path) {
8175                 btrfs_free_and_pin_reserved_extent(fs_info, ins->objectid,
8176                                                    fs_info->nodesize);
8177                 return -ENOMEM;
8178         }
8179
8180         path->leave_spinning = 1;
8181         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
8182                                       ins, size);
8183         if (ret) {
8184                 btrfs_free_path(path);
8185                 btrfs_free_and_pin_reserved_extent(fs_info, ins->objectid,
8186                                                    fs_info->nodesize);
8187                 return ret;
8188         }
8189
8190         leaf = path->nodes[0];
8191         extent_item = btrfs_item_ptr(leaf, path->slots[0],
8192                                      struct btrfs_extent_item);
8193         btrfs_set_extent_refs(leaf, extent_item, 1);
8194         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8195         btrfs_set_extent_flags(leaf, extent_item,
8196                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
8197
8198         if (skinny_metadata) {
8199                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8200                 num_bytes = fs_info->nodesize;
8201         } else {
8202                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
8203                 btrfs_set_tree_block_key(leaf, block_info, key);
8204                 btrfs_set_tree_block_level(leaf, block_info, level);
8205                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
8206         }
8207
8208         if (parent > 0) {
8209                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
8210                 btrfs_set_extent_inline_ref_type(leaf, iref,
8211                                                  BTRFS_SHARED_BLOCK_REF_KEY);
8212                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
8213         } else {
8214                 btrfs_set_extent_inline_ref_type(leaf, iref,
8215                                                  BTRFS_TREE_BLOCK_REF_KEY);
8216                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
8217         }
8218
8219         btrfs_mark_buffer_dirty(leaf);
8220         btrfs_free_path(path);
8221
8222         ret = remove_from_free_space_tree(trans, fs_info, ins->objectid,
8223                                           num_bytes);
8224         if (ret)
8225                 return ret;
8226
8227         ret = update_block_group(trans, fs_info, ins->objectid,
8228                                  fs_info->nodesize, 1);
8229         if (ret) { /* -ENOENT, logic error */
8230                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8231                         ins->objectid, ins->offset);
8232                 BUG();
8233         }
8234
8235         trace_btrfs_reserved_extent_alloc(fs_info, ins->objectid,
8236                                           fs_info->nodesize);
8237         return ret;
8238 }
8239
8240 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8241                                      struct btrfs_root *root, u64 owner,
8242                                      u64 offset, u64 ram_bytes,
8243                                      struct btrfs_key *ins)
8244 {
8245         struct btrfs_fs_info *fs_info = root->fs_info;
8246         int ret;
8247
8248         BUG_ON(root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
8249
8250         btrfs_ref_tree_mod(root, ins->objectid, ins->offset, 0,
8251                            root->root_key.objectid, owner, offset,
8252                            BTRFS_ADD_DELAYED_EXTENT);
8253
8254         ret = btrfs_add_delayed_data_ref(fs_info, trans, ins->objectid,
8255                                          ins->offset, 0,
8256                                          root->root_key.objectid, owner,
8257                                          offset, ram_bytes,
8258                                          BTRFS_ADD_DELAYED_EXTENT, NULL, NULL);
8259         return ret;
8260 }
8261
8262 /*
8263  * this is used by the tree logging recovery code.  It records that
8264  * an extent has been allocated and makes sure to clear the free
8265  * space cache bits as well
8266  */
8267 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
8268                                    struct btrfs_fs_info *fs_info,
8269                                    u64 root_objectid, u64 owner, u64 offset,
8270                                    struct btrfs_key *ins)
8271 {
8272         int ret;
8273         struct btrfs_block_group_cache *block_group;
8274         struct btrfs_space_info *space_info;
8275
8276         /*
8277          * Mixed block groups will exclude before processing the log so we only
8278          * need to do the exclude dance if this fs isn't mixed.
8279          */
8280         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS)) {
8281                 ret = __exclude_logged_extent(fs_info, ins->objectid,
8282                                               ins->offset);
8283                 if (ret)
8284                         return ret;
8285         }
8286
8287         block_group = btrfs_lookup_block_group(fs_info, ins->objectid);
8288         if (!block_group)
8289                 return -EINVAL;
8290
8291         space_info = block_group->space_info;
8292         spin_lock(&space_info->lock);
8293         spin_lock(&block_group->lock);
8294         space_info->bytes_reserved += ins->offset;
8295         block_group->reserved += ins->offset;
8296         spin_unlock(&block_group->lock);
8297         spin_unlock(&space_info->lock);
8298
8299         ret = alloc_reserved_file_extent(trans, fs_info, 0, root_objectid,
8300                                          0, owner, offset, ins, 1);
8301         btrfs_put_block_group(block_group);
8302         return ret;
8303 }
8304
8305 static struct extent_buffer *
8306 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
8307                       u64 bytenr, int level)
8308 {
8309         struct btrfs_fs_info *fs_info = root->fs_info;
8310         struct extent_buffer *buf;
8311
8312         buf = btrfs_find_create_tree_block(fs_info, bytenr);
8313         if (IS_ERR(buf))
8314                 return buf;
8315
8316         btrfs_set_header_generation(buf, trans->transid);
8317         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
8318         btrfs_tree_lock(buf);
8319         clean_tree_block(fs_info, buf);
8320         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
8321
8322         btrfs_set_lock_blocking(buf);
8323         set_extent_buffer_uptodate(buf);
8324
8325         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
8326                 buf->log_index = root->log_transid % 2;
8327                 /*
8328                  * we allow two log transactions at a time, use different
8329                  * EXENT bit to differentiate dirty pages.
8330                  */
8331                 if (buf->log_index == 0)
8332                         set_extent_dirty(&root->dirty_log_pages, buf->start,
8333                                         buf->start + buf->len - 1, GFP_NOFS);
8334                 else
8335                         set_extent_new(&root->dirty_log_pages, buf->start,
8336                                         buf->start + buf->len - 1);
8337         } else {
8338                 buf->log_index = -1;
8339                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
8340                          buf->start + buf->len - 1, GFP_NOFS);
8341         }
8342         trans->dirty = true;
8343         /* this returns a buffer locked for blocking */
8344         return buf;
8345 }
8346
8347 static struct btrfs_block_rsv *
8348 use_block_rsv(struct btrfs_trans_handle *trans,
8349               struct btrfs_root *root, u32 blocksize)
8350 {
8351         struct btrfs_fs_info *fs_info = root->fs_info;
8352         struct btrfs_block_rsv *block_rsv;
8353         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
8354         int ret;
8355         bool global_updated = false;
8356
8357         block_rsv = get_block_rsv(trans, root);
8358
8359         if (unlikely(block_rsv->size == 0))
8360                 goto try_reserve;
8361 again:
8362         ret = block_rsv_use_bytes(block_rsv, blocksize);
8363         if (!ret)
8364                 return block_rsv;
8365
8366         if (block_rsv->failfast)
8367                 return ERR_PTR(ret);
8368
8369         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
8370                 global_updated = true;
8371                 update_global_block_rsv(fs_info);
8372                 goto again;
8373         }
8374
8375         if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
8376                 static DEFINE_RATELIMIT_STATE(_rs,
8377                                 DEFAULT_RATELIMIT_INTERVAL * 10,
8378                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
8379                 if (__ratelimit(&_rs))
8380                         WARN(1, KERN_DEBUG
8381                                 "BTRFS: block rsv returned %d\n", ret);
8382         }
8383 try_reserve:
8384         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
8385                                      BTRFS_RESERVE_NO_FLUSH);
8386         if (!ret)
8387                 return block_rsv;
8388         /*
8389          * If we couldn't reserve metadata bytes try and use some from
8390          * the global reserve if its space type is the same as the global
8391          * reservation.
8392          */
8393         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
8394             block_rsv->space_info == global_rsv->space_info) {
8395                 ret = block_rsv_use_bytes(global_rsv, blocksize);
8396                 if (!ret)
8397                         return global_rsv;
8398         }
8399         return ERR_PTR(ret);
8400 }
8401
8402 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
8403                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
8404 {
8405         block_rsv_add_bytes(block_rsv, blocksize, 0);
8406         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
8407 }
8408
8409 /*
8410  * finds a free extent and does all the dirty work required for allocation
8411  * returns the tree buffer or an ERR_PTR on error.
8412  */
8413 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
8414                                              struct btrfs_root *root,
8415                                              u64 parent, u64 root_objectid,
8416                                              const struct btrfs_disk_key *key,
8417                                              int level, u64 hint,
8418                                              u64 empty_size)
8419 {
8420         struct btrfs_fs_info *fs_info = root->fs_info;
8421         struct btrfs_key ins;
8422         struct btrfs_block_rsv *block_rsv;
8423         struct extent_buffer *buf;
8424         struct btrfs_delayed_extent_op *extent_op;
8425         u64 flags = 0;
8426         int ret;
8427         u32 blocksize = fs_info->nodesize;
8428         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
8429
8430 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
8431         if (btrfs_is_testing(fs_info)) {
8432                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
8433                                             level);
8434                 if (!IS_ERR(buf))
8435                         root->alloc_bytenr += blocksize;
8436                 return buf;
8437         }
8438 #endif
8439
8440         block_rsv = use_block_rsv(trans, root, blocksize);
8441         if (IS_ERR(block_rsv))
8442                 return ERR_CAST(block_rsv);
8443
8444         ret = btrfs_reserve_extent(root, blocksize, blocksize, blocksize,
8445                                    empty_size, hint, &ins, 0, 0);
8446         if (ret)
8447                 goto out_unuse;
8448
8449         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level);
8450         if (IS_ERR(buf)) {
8451                 ret = PTR_ERR(buf);
8452                 goto out_free_reserved;
8453         }
8454
8455         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
8456                 if (parent == 0)
8457                         parent = ins.objectid;
8458                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
8459         } else
8460                 BUG_ON(parent > 0);
8461
8462         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
8463                 extent_op = btrfs_alloc_delayed_extent_op();
8464                 if (!extent_op) {
8465                         ret = -ENOMEM;
8466                         goto out_free_buf;
8467                 }
8468                 if (key)
8469                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
8470                 else
8471                         memset(&extent_op->key, 0, sizeof(extent_op->key));
8472                 extent_op->flags_to_set = flags;
8473                 extent_op->update_key = skinny_metadata ? false : true;
8474                 extent_op->update_flags = true;
8475                 extent_op->is_data = false;
8476                 extent_op->level = level;
8477
8478                 btrfs_ref_tree_mod(root, ins.objectid, ins.offset, parent,
8479                                    root_objectid, level, 0,
8480                                    BTRFS_ADD_DELAYED_EXTENT);
8481                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, ins.objectid,
8482                                                  ins.offset, parent,
8483                                                  root_objectid, level,
8484                                                  BTRFS_ADD_DELAYED_EXTENT,
8485                                                  extent_op, NULL, NULL);
8486                 if (ret)
8487                         goto out_free_delayed;
8488         }
8489         return buf;
8490
8491 out_free_delayed:
8492         btrfs_free_delayed_extent_op(extent_op);
8493 out_free_buf:
8494         free_extent_buffer(buf);
8495 out_free_reserved:
8496         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 0);
8497 out_unuse:
8498         unuse_block_rsv(fs_info, block_rsv, blocksize);
8499         return ERR_PTR(ret);
8500 }
8501
8502 struct walk_control {
8503         u64 refs[BTRFS_MAX_LEVEL];
8504         u64 flags[BTRFS_MAX_LEVEL];
8505         struct btrfs_key update_progress;
8506         int stage;
8507         int level;
8508         int shared_level;
8509         int update_ref;
8510         int keep_locks;
8511         int reada_slot;
8512         int reada_count;
8513         int for_reloc;
8514 };
8515
8516 #define DROP_REFERENCE  1
8517 #define UPDATE_BACKREF  2
8518
8519 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
8520                                      struct btrfs_root *root,
8521                                      struct walk_control *wc,
8522                                      struct btrfs_path *path)
8523 {
8524         struct btrfs_fs_info *fs_info = root->fs_info;
8525         u64 bytenr;
8526         u64 generation;
8527         u64 refs;
8528         u64 flags;
8529         u32 nritems;
8530         struct btrfs_key key;
8531         struct extent_buffer *eb;
8532         int ret;
8533         int slot;
8534         int nread = 0;
8535
8536         if (path->slots[wc->level] < wc->reada_slot) {
8537                 wc->reada_count = wc->reada_count * 2 / 3;
8538                 wc->reada_count = max(wc->reada_count, 2);
8539         } else {
8540                 wc->reada_count = wc->reada_count * 3 / 2;
8541                 wc->reada_count = min_t(int, wc->reada_count,
8542                                         BTRFS_NODEPTRS_PER_BLOCK(fs_info));
8543         }
8544
8545         eb = path->nodes[wc->level];
8546         nritems = btrfs_header_nritems(eb);
8547
8548         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
8549                 if (nread >= wc->reada_count)
8550                         break;
8551
8552                 cond_resched();
8553                 bytenr = btrfs_node_blockptr(eb, slot);
8554                 generation = btrfs_node_ptr_generation(eb, slot);
8555
8556                 if (slot == path->slots[wc->level])
8557                         goto reada;
8558
8559                 if (wc->stage == UPDATE_BACKREF &&
8560                     generation <= root->root_key.offset)
8561                         continue;
8562
8563                 /* We don't lock the tree block, it's OK to be racy here */
8564                 ret = btrfs_lookup_extent_info(trans, fs_info, bytenr,
8565                                                wc->level - 1, 1, &refs,
8566                                                &flags);
8567                 /* We don't care about errors in readahead. */
8568                 if (ret < 0)
8569                         continue;
8570                 BUG_ON(refs == 0);
8571
8572                 if (wc->stage == DROP_REFERENCE) {
8573                         if (refs == 1)
8574                                 goto reada;
8575
8576                         if (wc->level == 1 &&
8577                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8578                                 continue;
8579                         if (!wc->update_ref ||
8580                             generation <= root->root_key.offset)
8581                                 continue;
8582                         btrfs_node_key_to_cpu(eb, &key, slot);
8583                         ret = btrfs_comp_cpu_keys(&key,
8584                                                   &wc->update_progress);
8585                         if (ret < 0)
8586                                 continue;
8587                 } else {
8588                         if (wc->level == 1 &&
8589                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8590                                 continue;
8591                 }
8592 reada:
8593                 readahead_tree_block(fs_info, bytenr);
8594                 nread++;
8595         }
8596         wc->reada_slot = slot;
8597 }
8598
8599 /*
8600  * helper to process tree block while walking down the tree.
8601  *
8602  * when wc->stage == UPDATE_BACKREF, this function updates
8603  * back refs for pointers in the block.
8604  *
8605  * NOTE: return value 1 means we should stop walking down.
8606  */
8607 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
8608                                    struct btrfs_root *root,
8609                                    struct btrfs_path *path,
8610                                    struct walk_control *wc, int lookup_info)
8611 {
8612         struct btrfs_fs_info *fs_info = root->fs_info;
8613         int level = wc->level;
8614         struct extent_buffer *eb = path->nodes[level];
8615         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8616         int ret;
8617
8618         if (wc->stage == UPDATE_BACKREF &&
8619             btrfs_header_owner(eb) != root->root_key.objectid)
8620                 return 1;
8621
8622         /*
8623          * when reference count of tree block is 1, it won't increase
8624          * again. once full backref flag is set, we never clear it.
8625          */
8626         if (lookup_info &&
8627             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
8628              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
8629                 BUG_ON(!path->locks[level]);
8630                 ret = btrfs_lookup_extent_info(trans, fs_info,
8631                                                eb->start, level, 1,
8632                                                &wc->refs[level],
8633                                                &wc->flags[level]);
8634                 BUG_ON(ret == -ENOMEM);
8635                 if (ret)
8636                         return ret;
8637                 BUG_ON(wc->refs[level] == 0);
8638         }
8639
8640         if (wc->stage == DROP_REFERENCE) {
8641                 if (wc->refs[level] > 1)
8642                         return 1;
8643
8644                 if (path->locks[level] && !wc->keep_locks) {
8645                         btrfs_tree_unlock_rw(eb, path->locks[level]);
8646                         path->locks[level] = 0;
8647                 }
8648                 return 0;
8649         }
8650
8651         /* wc->stage == UPDATE_BACKREF */
8652         if (!(wc->flags[level] & flag)) {
8653                 BUG_ON(!path->locks[level]);
8654                 ret = btrfs_inc_ref(trans, root, eb, 1);
8655                 BUG_ON(ret); /* -ENOMEM */
8656                 ret = btrfs_dec_ref(trans, root, eb, 0);
8657                 BUG_ON(ret); /* -ENOMEM */
8658                 ret = btrfs_set_disk_extent_flags(trans, fs_info, eb->start,
8659                                                   eb->len, flag,
8660                                                   btrfs_header_level(eb), 0);
8661                 BUG_ON(ret); /* -ENOMEM */
8662                 wc->flags[level] |= flag;
8663         }
8664
8665         /*
8666          * the block is shared by multiple trees, so it's not good to
8667          * keep the tree lock
8668          */
8669         if (path->locks[level] && level > 0) {
8670                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8671                 path->locks[level] = 0;
8672         }
8673         return 0;
8674 }
8675
8676 /*
8677  * helper to process tree block pointer.
8678  *
8679  * when wc->stage == DROP_REFERENCE, this function checks
8680  * reference count of the block pointed to. if the block
8681  * is shared and we need update back refs for the subtree
8682  * rooted at the block, this function changes wc->stage to
8683  * UPDATE_BACKREF. if the block is shared and there is no
8684  * need to update back, this function drops the reference
8685  * to the block.
8686  *
8687  * NOTE: return value 1 means we should stop walking down.
8688  */
8689 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
8690                                  struct btrfs_root *root,
8691                                  struct btrfs_path *path,
8692                                  struct walk_control *wc, int *lookup_info)
8693 {
8694         struct btrfs_fs_info *fs_info = root->fs_info;
8695         u64 bytenr;
8696         u64 generation;
8697         u64 parent;
8698         u32 blocksize;
8699         struct btrfs_key key;
8700         struct btrfs_key first_key;
8701         struct extent_buffer *next;
8702         int level = wc->level;
8703         int reada = 0;
8704         int ret = 0;
8705         bool need_account = false;
8706
8707         generation = btrfs_node_ptr_generation(path->nodes[level],
8708                                                path->slots[level]);
8709         /*
8710          * if the lower level block was created before the snapshot
8711          * was created, we know there is no need to update back refs
8712          * for the subtree
8713          */
8714         if (wc->stage == UPDATE_BACKREF &&
8715             generation <= root->root_key.offset) {
8716                 *lookup_info = 1;
8717                 return 1;
8718         }
8719
8720         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
8721         btrfs_node_key_to_cpu(path->nodes[level], &first_key,
8722                               path->slots[level]);
8723         blocksize = fs_info->nodesize;
8724
8725         next = find_extent_buffer(fs_info, bytenr);
8726         if (!next) {
8727                 next = btrfs_find_create_tree_block(fs_info, bytenr);
8728                 if (IS_ERR(next))
8729                         return PTR_ERR(next);
8730
8731                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
8732                                                level - 1);
8733                 reada = 1;
8734         }
8735         btrfs_tree_lock(next);
8736         btrfs_set_lock_blocking(next);
8737
8738         ret = btrfs_lookup_extent_info(trans, fs_info, bytenr, level - 1, 1,
8739                                        &wc->refs[level - 1],
8740                                        &wc->flags[level - 1]);
8741         if (ret < 0)
8742                 goto out_unlock;
8743
8744         if (unlikely(wc->refs[level - 1] == 0)) {
8745                 btrfs_err(fs_info, "Missing references.");
8746                 ret = -EIO;
8747                 goto out_unlock;
8748         }
8749         *lookup_info = 0;
8750
8751         if (wc->stage == DROP_REFERENCE) {
8752                 if (wc->refs[level - 1] > 1) {
8753                         need_account = true;
8754                         if (level == 1 &&
8755                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8756                                 goto skip;
8757
8758                         if (!wc->update_ref ||
8759                             generation <= root->root_key.offset)
8760                                 goto skip;
8761
8762                         btrfs_node_key_to_cpu(path->nodes[level], &key,
8763                                               path->slots[level]);
8764                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
8765                         if (ret < 0)
8766                                 goto skip;
8767
8768                         wc->stage = UPDATE_BACKREF;
8769                         wc->shared_level = level - 1;
8770                 }
8771         } else {
8772                 if (level == 1 &&
8773                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8774                         goto skip;
8775         }
8776
8777         if (!btrfs_buffer_uptodate(next, generation, 0)) {
8778                 btrfs_tree_unlock(next);
8779                 free_extent_buffer(next);
8780                 next = NULL;
8781                 *lookup_info = 1;
8782         }
8783
8784         if (!next) {
8785                 if (reada && level == 1)
8786                         reada_walk_down(trans, root, wc, path);
8787                 next = read_tree_block(fs_info, bytenr, generation, level - 1,
8788                                        &first_key);
8789                 if (IS_ERR(next)) {
8790                         return PTR_ERR(next);
8791                 } else if (!extent_buffer_uptodate(next)) {
8792                         free_extent_buffer(next);
8793                         return -EIO;
8794                 }
8795                 btrfs_tree_lock(next);
8796                 btrfs_set_lock_blocking(next);
8797         }
8798
8799         level--;
8800         ASSERT(level == btrfs_header_level(next));
8801         if (level != btrfs_header_level(next)) {
8802                 btrfs_err(root->fs_info, "mismatched level");
8803                 ret = -EIO;
8804                 goto out_unlock;
8805         }
8806         path->nodes[level] = next;
8807         path->slots[level] = 0;
8808         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8809         wc->level = level;
8810         if (wc->level == 1)
8811                 wc->reada_slot = 0;
8812         return 0;
8813 skip:
8814         wc->refs[level - 1] = 0;
8815         wc->flags[level - 1] = 0;
8816         if (wc->stage == DROP_REFERENCE) {
8817                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
8818                         parent = path->nodes[level]->start;
8819                 } else {
8820                         ASSERT(root->root_key.objectid ==
8821                                btrfs_header_owner(path->nodes[level]));
8822                         if (root->root_key.objectid !=
8823                             btrfs_header_owner(path->nodes[level])) {
8824                                 btrfs_err(root->fs_info,
8825                                                 "mismatched block owner");
8826                                 ret = -EIO;
8827                                 goto out_unlock;
8828                         }
8829                         parent = 0;
8830                 }
8831
8832                 if (need_account) {
8833                         ret = btrfs_qgroup_trace_subtree(trans, root, next,
8834                                                          generation, level - 1);
8835                         if (ret) {
8836                                 btrfs_err_rl(fs_info,
8837                                              "Error %d accounting shared subtree. Quota is out of sync, rescan required.",
8838                                              ret);
8839                         }
8840                 }
8841                 ret = btrfs_free_extent(trans, root, bytenr, blocksize,
8842                                         parent, root->root_key.objectid,
8843                                         level - 1, 0);
8844                 if (ret)
8845                         goto out_unlock;
8846         }
8847
8848         *lookup_info = 1;
8849         ret = 1;
8850
8851 out_unlock:
8852         btrfs_tree_unlock(next);
8853         free_extent_buffer(next);
8854
8855         return ret;
8856 }
8857
8858 /*
8859  * helper to process tree block while walking up the tree.
8860  *
8861  * when wc->stage == DROP_REFERENCE, this function drops
8862  * reference count on the block.
8863  *
8864  * when wc->stage == UPDATE_BACKREF, this function changes
8865  * wc->stage back to DROP_REFERENCE if we changed wc->stage
8866  * to UPDATE_BACKREF previously while processing the block.
8867  *
8868  * NOTE: return value 1 means we should stop walking up.
8869  */
8870 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
8871                                  struct btrfs_root *root,
8872                                  struct btrfs_path *path,
8873                                  struct walk_control *wc)
8874 {
8875         struct btrfs_fs_info *fs_info = root->fs_info;
8876         int ret;
8877         int level = wc->level;
8878         struct extent_buffer *eb = path->nodes[level];
8879         u64 parent = 0;
8880
8881         if (wc->stage == UPDATE_BACKREF) {
8882                 BUG_ON(wc->shared_level < level);
8883                 if (level < wc->shared_level)
8884                         goto out;
8885
8886                 ret = find_next_key(path, level + 1, &wc->update_progress);
8887                 if (ret > 0)
8888                         wc->update_ref = 0;
8889
8890                 wc->stage = DROP_REFERENCE;
8891                 wc->shared_level = -1;
8892                 path->slots[level] = 0;
8893
8894                 /*
8895                  * check reference count again if the block isn't locked.
8896                  * we should start walking down the tree again if reference
8897                  * count is one.
8898                  */
8899                 if (!path->locks[level]) {
8900                         BUG_ON(level == 0);
8901                         btrfs_tree_lock(eb);
8902                         btrfs_set_lock_blocking(eb);
8903                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8904
8905                         ret = btrfs_lookup_extent_info(trans, fs_info,
8906                                                        eb->start, level, 1,
8907                                                        &wc->refs[level],
8908                                                        &wc->flags[level]);
8909                         if (ret < 0) {
8910                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8911                                 path->locks[level] = 0;
8912                                 return ret;
8913                         }
8914                         BUG_ON(wc->refs[level] == 0);
8915                         if (wc->refs[level] == 1) {
8916                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8917                                 path->locks[level] = 0;
8918                                 return 1;
8919                         }
8920                 }
8921         }
8922
8923         /* wc->stage == DROP_REFERENCE */
8924         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
8925
8926         if (wc->refs[level] == 1) {
8927                 if (level == 0) {
8928                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8929                                 ret = btrfs_dec_ref(trans, root, eb, 1);
8930                         else
8931                                 ret = btrfs_dec_ref(trans, root, eb, 0);
8932                         BUG_ON(ret); /* -ENOMEM */
8933                         ret = btrfs_qgroup_trace_leaf_items(trans, fs_info, eb);
8934                         if (ret) {
8935                                 btrfs_err_rl(fs_info,
8936                                              "error %d accounting leaf items. Quota is out of sync, rescan required.",
8937                                              ret);
8938                         }
8939                 }
8940                 /* make block locked assertion in clean_tree_block happy */
8941                 if (!path->locks[level] &&
8942                     btrfs_header_generation(eb) == trans->transid) {
8943                         btrfs_tree_lock(eb);
8944                         btrfs_set_lock_blocking(eb);
8945                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8946                 }
8947                 clean_tree_block(fs_info, eb);
8948         }
8949
8950         if (eb == root->node) {
8951                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8952                         parent = eb->start;
8953                 else
8954                         BUG_ON(root->root_key.objectid !=
8955                                btrfs_header_owner(eb));
8956         } else {
8957                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8958                         parent = path->nodes[level + 1]->start;
8959                 else
8960                         BUG_ON(root->root_key.objectid !=
8961                                btrfs_header_owner(path->nodes[level + 1]));
8962         }
8963
8964         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
8965 out:
8966         wc->refs[level] = 0;
8967         wc->flags[level] = 0;
8968         return 0;
8969 }
8970
8971 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
8972                                    struct btrfs_root *root,
8973                                    struct btrfs_path *path,
8974                                    struct walk_control *wc)
8975 {
8976         int level = wc->level;
8977         int lookup_info = 1;
8978         int ret;
8979
8980         while (level >= 0) {
8981                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
8982                 if (ret > 0)
8983                         break;
8984
8985                 if (level == 0)
8986                         break;
8987
8988                 if (path->slots[level] >=
8989                     btrfs_header_nritems(path->nodes[level]))
8990                         break;
8991
8992                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
8993                 if (ret > 0) {
8994                         path->slots[level]++;
8995                         continue;
8996                 } else if (ret < 0)
8997                         return ret;
8998                 level = wc->level;
8999         }
9000         return 0;
9001 }
9002
9003 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
9004                                  struct btrfs_root *root,
9005                                  struct btrfs_path *path,
9006                                  struct walk_control *wc, int max_level)
9007 {
9008         int level = wc->level;
9009         int ret;
9010
9011         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
9012         while (level < max_level && path->nodes[level]) {
9013                 wc->level = level;
9014                 if (path->slots[level] + 1 <
9015                     btrfs_header_nritems(path->nodes[level])) {
9016                         path->slots[level]++;
9017                         return 0;
9018                 } else {
9019                         ret = walk_up_proc(trans, root, path, wc);
9020                         if (ret > 0)
9021                                 return 0;
9022
9023                         if (path->locks[level]) {
9024                                 btrfs_tree_unlock_rw(path->nodes[level],
9025                                                      path->locks[level]);
9026                                 path->locks[level] = 0;
9027                         }
9028                         free_extent_buffer(path->nodes[level]);
9029                         path->nodes[level] = NULL;
9030                         level++;
9031                 }
9032         }
9033         return 1;
9034 }
9035
9036 /*
9037  * drop a subvolume tree.
9038  *
9039  * this function traverses the tree freeing any blocks that only
9040  * referenced by the tree.
9041  *
9042  * when a shared tree block is found. this function decreases its
9043  * reference count by one. if update_ref is true, this function
9044  * also make sure backrefs for the shared block and all lower level
9045  * blocks are properly updated.
9046  *
9047  * If called with for_reloc == 0, may exit early with -EAGAIN
9048  */
9049 int btrfs_drop_snapshot(struct btrfs_root *root,
9050                          struct btrfs_block_rsv *block_rsv, int update_ref,
9051                          int for_reloc)
9052 {
9053         struct btrfs_fs_info *fs_info = root->fs_info;
9054         struct btrfs_path *path;
9055         struct btrfs_trans_handle *trans;
9056         struct btrfs_root *tree_root = fs_info->tree_root;
9057         struct btrfs_root_item *root_item = &root->root_item;
9058         struct walk_control *wc;
9059         struct btrfs_key key;
9060         int err = 0;
9061         int ret;
9062         int level;
9063         bool root_dropped = false;
9064
9065         btrfs_debug(fs_info, "Drop subvolume %llu", root->objectid);
9066
9067         path = btrfs_alloc_path();
9068         if (!path) {
9069                 err = -ENOMEM;
9070                 goto out;
9071         }
9072
9073         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9074         if (!wc) {
9075                 btrfs_free_path(path);
9076                 err = -ENOMEM;
9077                 goto out;
9078         }
9079
9080         trans = btrfs_start_transaction(tree_root, 0);
9081         if (IS_ERR(trans)) {
9082                 err = PTR_ERR(trans);
9083                 goto out_free;
9084         }
9085
9086         if (block_rsv)
9087                 trans->block_rsv = block_rsv;
9088
9089         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
9090                 level = btrfs_header_level(root->node);
9091                 path->nodes[level] = btrfs_lock_root_node(root);
9092                 btrfs_set_lock_blocking(path->nodes[level]);
9093                 path->slots[level] = 0;
9094                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9095                 memset(&wc->update_progress, 0,
9096                        sizeof(wc->update_progress));
9097         } else {
9098                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
9099                 memcpy(&wc->update_progress, &key,
9100                        sizeof(wc->update_progress));
9101
9102                 level = root_item->drop_level;
9103                 BUG_ON(level == 0);
9104                 path->lowest_level = level;
9105                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
9106                 path->lowest_level = 0;
9107                 if (ret < 0) {
9108                         err = ret;
9109                         goto out_end_trans;
9110                 }
9111                 WARN_ON(ret > 0);
9112
9113                 /*
9114                  * unlock our path, this is safe because only this
9115                  * function is allowed to delete this snapshot
9116                  */
9117                 btrfs_unlock_up_safe(path, 0);
9118
9119                 level = btrfs_header_level(root->node);
9120                 while (1) {
9121                         btrfs_tree_lock(path->nodes[level]);
9122                         btrfs_set_lock_blocking(path->nodes[level]);
9123                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9124
9125                         ret = btrfs_lookup_extent_info(trans, fs_info,
9126                                                 path->nodes[level]->start,
9127                                                 level, 1, &wc->refs[level],
9128                                                 &wc->flags[level]);
9129                         if (ret < 0) {
9130                                 err = ret;
9131                                 goto out_end_trans;
9132                         }
9133                         BUG_ON(wc->refs[level] == 0);
9134
9135                         if (level == root_item->drop_level)
9136                                 break;
9137
9138                         btrfs_tree_unlock(path->nodes[level]);
9139                         path->locks[level] = 0;
9140                         WARN_ON(wc->refs[level] != 1);
9141                         level--;
9142                 }
9143         }
9144
9145         wc->level = level;
9146         wc->shared_level = -1;
9147         wc->stage = DROP_REFERENCE;
9148         wc->update_ref = update_ref;
9149         wc->keep_locks = 0;
9150         wc->for_reloc = for_reloc;
9151         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
9152
9153         while (1) {
9154
9155                 ret = walk_down_tree(trans, root, path, wc);
9156                 if (ret < 0) {
9157                         err = ret;
9158                         break;
9159                 }
9160
9161                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
9162                 if (ret < 0) {
9163                         err = ret;
9164                         break;
9165                 }
9166
9167                 if (ret > 0) {
9168                         BUG_ON(wc->stage != DROP_REFERENCE);
9169                         break;
9170                 }
9171
9172                 if (wc->stage == DROP_REFERENCE) {
9173                         level = wc->level;
9174                         btrfs_node_key(path->nodes[level],
9175                                        &root_item->drop_progress,
9176                                        path->slots[level]);
9177                         root_item->drop_level = level;
9178                 }
9179
9180                 BUG_ON(wc->level == 0);
9181                 if (btrfs_should_end_transaction(trans) ||
9182                     (!for_reloc && btrfs_need_cleaner_sleep(fs_info))) {
9183                         ret = btrfs_update_root(trans, tree_root,
9184                                                 &root->root_key,
9185                                                 root_item);
9186                         if (ret) {
9187                                 btrfs_abort_transaction(trans, ret);
9188                                 err = ret;
9189                                 goto out_end_trans;
9190                         }
9191
9192                         btrfs_end_transaction_throttle(trans);
9193                         if (!for_reloc && btrfs_need_cleaner_sleep(fs_info)) {
9194                                 btrfs_debug(fs_info,
9195                                             "drop snapshot early exit");
9196                                 err = -EAGAIN;
9197                                 goto out_free;
9198                         }
9199
9200                         trans = btrfs_start_transaction(tree_root, 0);
9201                         if (IS_ERR(trans)) {
9202                                 err = PTR_ERR(trans);
9203                                 goto out_free;
9204                         }
9205                         if (block_rsv)
9206                                 trans->block_rsv = block_rsv;
9207                 }
9208         }
9209         btrfs_release_path(path);
9210         if (err)
9211                 goto out_end_trans;
9212
9213         ret = btrfs_del_root(trans, fs_info, &root->root_key);
9214         if (ret) {
9215                 btrfs_abort_transaction(trans, ret);
9216                 err = ret;
9217                 goto out_end_trans;
9218         }
9219
9220         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
9221                 ret = btrfs_find_root(tree_root, &root->root_key, path,
9222                                       NULL, NULL);
9223                 if (ret < 0) {
9224                         btrfs_abort_transaction(trans, ret);
9225                         err = ret;
9226                         goto out_end_trans;
9227                 } else if (ret > 0) {
9228                         /* if we fail to delete the orphan item this time
9229                          * around, it'll get picked up the next time.
9230                          *
9231                          * The most common failure here is just -ENOENT.
9232                          */
9233                         btrfs_del_orphan_item(trans, tree_root,
9234                                               root->root_key.objectid);
9235                 }
9236         }
9237
9238         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
9239                 btrfs_add_dropped_root(trans, root);
9240         } else {
9241                 free_extent_buffer(root->node);
9242                 free_extent_buffer(root->commit_root);
9243                 btrfs_put_fs_root(root);
9244         }
9245         root_dropped = true;
9246 out_end_trans:
9247         btrfs_end_transaction_throttle(trans);
9248 out_free:
9249         kfree(wc);
9250         btrfs_free_path(path);
9251 out:
9252         /*
9253          * So if we need to stop dropping the snapshot for whatever reason we
9254          * need to make sure to add it back to the dead root list so that we
9255          * keep trying to do the work later.  This also cleans up roots if we
9256          * don't have it in the radix (like when we recover after a power fail
9257          * or unmount) so we don't leak memory.
9258          */
9259         if (!for_reloc && !root_dropped)
9260                 btrfs_add_dead_root(root);
9261         if (err && err != -EAGAIN)
9262                 btrfs_handle_fs_error(fs_info, err, NULL);
9263         return err;
9264 }
9265
9266 /*
9267  * drop subtree rooted at tree block 'node'.
9268  *
9269  * NOTE: this function will unlock and release tree block 'node'
9270  * only used by relocation code
9271  */
9272 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
9273                         struct btrfs_root *root,
9274                         struct extent_buffer *node,
9275                         struct extent_buffer *parent)
9276 {
9277         struct btrfs_fs_info *fs_info = root->fs_info;
9278         struct btrfs_path *path;
9279         struct walk_control *wc;
9280         int level;
9281         int parent_level;
9282         int ret = 0;
9283         int wret;
9284
9285         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
9286
9287         path = btrfs_alloc_path();
9288         if (!path)
9289                 return -ENOMEM;
9290
9291         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9292         if (!wc) {
9293                 btrfs_free_path(path);
9294                 return -ENOMEM;
9295         }
9296
9297         btrfs_assert_tree_locked(parent);
9298         parent_level = btrfs_header_level(parent);
9299         extent_buffer_get(parent);
9300         path->nodes[parent_level] = parent;
9301         path->slots[parent_level] = btrfs_header_nritems(parent);
9302
9303         btrfs_assert_tree_locked(node);
9304         level = btrfs_header_level(node);
9305         path->nodes[level] = node;
9306         path->slots[level] = 0;
9307         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9308
9309         wc->refs[parent_level] = 1;
9310         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
9311         wc->level = level;
9312         wc->shared_level = -1;
9313         wc->stage = DROP_REFERENCE;
9314         wc->update_ref = 0;
9315         wc->keep_locks = 1;
9316         wc->for_reloc = 1;
9317         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
9318
9319         while (1) {
9320                 wret = walk_down_tree(trans, root, path, wc);
9321                 if (wret < 0) {
9322                         ret = wret;
9323                         break;
9324                 }
9325
9326                 wret = walk_up_tree(trans, root, path, wc, parent_level);
9327                 if (wret < 0)
9328                         ret = wret;
9329                 if (wret != 0)
9330                         break;
9331         }
9332
9333         kfree(wc);
9334         btrfs_free_path(path);
9335         return ret;
9336 }
9337
9338 static u64 update_block_group_flags(struct btrfs_fs_info *fs_info, u64 flags)
9339 {
9340         u64 num_devices;
9341         u64 stripped;
9342
9343         /*
9344          * if restripe for this chunk_type is on pick target profile and
9345          * return, otherwise do the usual balance
9346          */
9347         stripped = get_restripe_target(fs_info, flags);
9348         if (stripped)
9349                 return extended_to_chunk(stripped);
9350
9351         num_devices = fs_info->fs_devices->rw_devices;
9352
9353         stripped = BTRFS_BLOCK_GROUP_RAID0 |
9354                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
9355                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
9356
9357         if (num_devices == 1) {
9358                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9359                 stripped = flags & ~stripped;
9360
9361                 /* turn raid0 into single device chunks */
9362                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
9363                         return stripped;
9364
9365                 /* turn mirroring into duplication */
9366                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
9367                              BTRFS_BLOCK_GROUP_RAID10))
9368                         return stripped | BTRFS_BLOCK_GROUP_DUP;
9369         } else {
9370                 /* they already had raid on here, just return */
9371                 if (flags & stripped)
9372                         return flags;
9373
9374                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9375                 stripped = flags & ~stripped;
9376
9377                 /* switch duplicated blocks with raid1 */
9378                 if (flags & BTRFS_BLOCK_GROUP_DUP)
9379                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
9380
9381                 /* this is drive concat, leave it alone */
9382         }
9383
9384         return flags;
9385 }
9386
9387 static int inc_block_group_ro(struct btrfs_block_group_cache *cache, int force)
9388 {
9389         struct btrfs_space_info *sinfo = cache->space_info;
9390         u64 num_bytes;
9391         u64 min_allocable_bytes;
9392         int ret = -ENOSPC;
9393
9394         /*
9395          * We need some metadata space and system metadata space for
9396          * allocating chunks in some corner cases until we force to set
9397          * it to be readonly.
9398          */
9399         if ((sinfo->flags &
9400              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
9401             !force)
9402                 min_allocable_bytes = SZ_1M;
9403         else
9404                 min_allocable_bytes = 0;
9405
9406         spin_lock(&sinfo->lock);
9407         spin_lock(&cache->lock);
9408
9409         if (cache->ro) {
9410                 cache->ro++;
9411                 ret = 0;
9412                 goto out;
9413         }
9414
9415         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
9416                     cache->bytes_super - btrfs_block_group_used(&cache->item);
9417
9418         if (btrfs_space_info_used(sinfo, true) + num_bytes +
9419             min_allocable_bytes <= sinfo->total_bytes) {
9420                 sinfo->bytes_readonly += num_bytes;
9421                 cache->ro++;
9422                 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
9423                 ret = 0;
9424         }
9425 out:
9426         spin_unlock(&cache->lock);
9427         spin_unlock(&sinfo->lock);
9428         return ret;
9429 }
9430
9431 int btrfs_inc_block_group_ro(struct btrfs_fs_info *fs_info,
9432                              struct btrfs_block_group_cache *cache)
9433
9434 {
9435         struct btrfs_trans_handle *trans;
9436         u64 alloc_flags;
9437         int ret;
9438
9439 again:
9440         trans = btrfs_join_transaction(fs_info->extent_root);
9441         if (IS_ERR(trans))
9442                 return PTR_ERR(trans);
9443
9444         /*
9445          * we're not allowed to set block groups readonly after the dirty
9446          * block groups cache has started writing.  If it already started,
9447          * back off and let this transaction commit
9448          */
9449         mutex_lock(&fs_info->ro_block_group_mutex);
9450         if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
9451                 u64 transid = trans->transid;
9452
9453                 mutex_unlock(&fs_info->ro_block_group_mutex);
9454                 btrfs_end_transaction(trans);
9455
9456                 ret = btrfs_wait_for_commit(fs_info, transid);
9457                 if (ret)
9458                         return ret;
9459                 goto again;
9460         }
9461
9462         /*
9463          * if we are changing raid levels, try to allocate a corresponding
9464          * block group with the new raid level.
9465          */
9466         alloc_flags = update_block_group_flags(fs_info, cache->flags);
9467         if (alloc_flags != cache->flags) {
9468                 ret = do_chunk_alloc(trans, fs_info, alloc_flags,
9469                                      CHUNK_ALLOC_FORCE);
9470                 /*
9471                  * ENOSPC is allowed here, we may have enough space
9472                  * already allocated at the new raid level to
9473                  * carry on
9474                  */
9475                 if (ret == -ENOSPC)
9476                         ret = 0;
9477                 if (ret < 0)
9478                         goto out;
9479         }
9480
9481         ret = inc_block_group_ro(cache, 0);
9482         if (!ret)
9483                 goto out;
9484         alloc_flags = get_alloc_profile(fs_info, cache->space_info->flags);
9485         ret = do_chunk_alloc(trans, fs_info, alloc_flags,
9486                              CHUNK_ALLOC_FORCE);
9487         if (ret < 0)
9488                 goto out;
9489         ret = inc_block_group_ro(cache, 0);
9490 out:
9491         if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
9492                 alloc_flags = update_block_group_flags(fs_info, cache->flags);
9493                 mutex_lock(&fs_info->chunk_mutex);
9494                 check_system_chunk(trans, fs_info, alloc_flags);
9495                 mutex_unlock(&fs_info->chunk_mutex);
9496         }
9497         mutex_unlock(&fs_info->ro_block_group_mutex);
9498
9499         btrfs_end_transaction(trans);
9500         return ret;
9501 }
9502
9503 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
9504                             struct btrfs_fs_info *fs_info, u64 type)
9505 {
9506         u64 alloc_flags = get_alloc_profile(fs_info, type);
9507
9508         return do_chunk_alloc(trans, fs_info, alloc_flags, CHUNK_ALLOC_FORCE);
9509 }
9510
9511 /*
9512  * helper to account the unused space of all the readonly block group in the
9513  * space_info. takes mirrors into account.
9514  */
9515 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
9516 {
9517         struct btrfs_block_group_cache *block_group;
9518         u64 free_bytes = 0;
9519         int factor;
9520
9521         /* It's df, we don't care if it's racy */
9522         if (list_empty(&sinfo->ro_bgs))
9523                 return 0;
9524
9525         spin_lock(&sinfo->lock);
9526         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
9527                 spin_lock(&block_group->lock);
9528
9529                 if (!block_group->ro) {
9530                         spin_unlock(&block_group->lock);
9531                         continue;
9532                 }
9533
9534                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
9535                                           BTRFS_BLOCK_GROUP_RAID10 |
9536                                           BTRFS_BLOCK_GROUP_DUP))
9537                         factor = 2;
9538                 else
9539                         factor = 1;
9540
9541                 free_bytes += (block_group->key.offset -
9542                                btrfs_block_group_used(&block_group->item)) *
9543                                factor;
9544
9545                 spin_unlock(&block_group->lock);
9546         }
9547         spin_unlock(&sinfo->lock);
9548
9549         return free_bytes;
9550 }
9551
9552 void btrfs_dec_block_group_ro(struct btrfs_block_group_cache *cache)
9553 {
9554         struct btrfs_space_info *sinfo = cache->space_info;
9555         u64 num_bytes;
9556
9557         BUG_ON(!cache->ro);
9558
9559         spin_lock(&sinfo->lock);
9560         spin_lock(&cache->lock);
9561         if (!--cache->ro) {
9562                 num_bytes = cache->key.offset - cache->reserved -
9563                             cache->pinned - cache->bytes_super -
9564                             btrfs_block_group_used(&cache->item);
9565                 sinfo->bytes_readonly -= num_bytes;
9566                 list_del_init(&cache->ro_list);
9567         }
9568         spin_unlock(&cache->lock);
9569         spin_unlock(&sinfo->lock);
9570 }
9571
9572 /*
9573  * checks to see if its even possible to relocate this block group.
9574  *
9575  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
9576  * ok to go ahead and try.
9577  */
9578 int btrfs_can_relocate(struct btrfs_fs_info *fs_info, u64 bytenr)
9579 {
9580         struct btrfs_root *root = fs_info->extent_root;
9581         struct btrfs_block_group_cache *block_group;
9582         struct btrfs_space_info *space_info;
9583         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
9584         struct btrfs_device *device;
9585         struct btrfs_trans_handle *trans;
9586         u64 min_free;
9587         u64 dev_min = 1;
9588         u64 dev_nr = 0;
9589         u64 target;
9590         int debug;
9591         int index;
9592         int full = 0;
9593         int ret = 0;
9594
9595         debug = btrfs_test_opt(fs_info, ENOSPC_DEBUG);
9596
9597         block_group = btrfs_lookup_block_group(fs_info, bytenr);
9598
9599         /* odd, couldn't find the block group, leave it alone */
9600         if (!block_group) {
9601                 if (debug)
9602                         btrfs_warn(fs_info,
9603                                    "can't find block group for bytenr %llu",
9604                                    bytenr);
9605                 return -1;
9606         }
9607
9608         min_free = btrfs_block_group_used(&block_group->item);
9609
9610         /* no bytes used, we're good */
9611         if (!min_free)
9612                 goto out;
9613
9614         space_info = block_group->space_info;
9615         spin_lock(&space_info->lock);
9616
9617         full = space_info->full;
9618
9619         /*
9620          * if this is the last block group we have in this space, we can't
9621          * relocate it unless we're able to allocate a new chunk below.
9622          *
9623          * Otherwise, we need to make sure we have room in the space to handle
9624          * all of the extents from this block group.  If we can, we're good
9625          */
9626         if ((space_info->total_bytes != block_group->key.offset) &&
9627             (btrfs_space_info_used(space_info, false) + min_free <
9628              space_info->total_bytes)) {
9629                 spin_unlock(&space_info->lock);
9630                 goto out;
9631         }
9632         spin_unlock(&space_info->lock);
9633
9634         /*
9635          * ok we don't have enough space, but maybe we have free space on our
9636          * devices to allocate new chunks for relocation, so loop through our
9637          * alloc devices and guess if we have enough space.  if this block
9638          * group is going to be restriped, run checks against the target
9639          * profile instead of the current one.
9640          */
9641         ret = -1;
9642
9643         /*
9644          * index:
9645          *      0: raid10
9646          *      1: raid1
9647          *      2: dup
9648          *      3: raid0
9649          *      4: single
9650          */
9651         target = get_restripe_target(fs_info, block_group->flags);
9652         if (target) {
9653                 index = btrfs_bg_flags_to_raid_index(extended_to_chunk(target));
9654         } else {
9655                 /*
9656                  * this is just a balance, so if we were marked as full
9657                  * we know there is no space for a new chunk
9658                  */
9659                 if (full) {
9660                         if (debug)
9661                                 btrfs_warn(fs_info,
9662                                            "no space to alloc new chunk for block group %llu",
9663                                            block_group->key.objectid);
9664                         goto out;
9665                 }
9666
9667                 index = btrfs_bg_flags_to_raid_index(block_group->flags);
9668         }
9669
9670         if (index == BTRFS_RAID_RAID10) {
9671                 dev_min = 4;
9672                 /* Divide by 2 */
9673                 min_free >>= 1;
9674         } else if (index == BTRFS_RAID_RAID1) {
9675                 dev_min = 2;
9676         } else if (index == BTRFS_RAID_DUP) {
9677                 /* Multiply by 2 */
9678                 min_free <<= 1;
9679         } else if (index == BTRFS_RAID_RAID0) {
9680                 dev_min = fs_devices->rw_devices;
9681                 min_free = div64_u64(min_free, dev_min);
9682         }
9683
9684         /* We need to do this so that we can look at pending chunks */
9685         trans = btrfs_join_transaction(root);
9686         if (IS_ERR(trans)) {
9687                 ret = PTR_ERR(trans);
9688                 goto out;
9689         }
9690
9691         mutex_lock(&fs_info->chunk_mutex);
9692         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
9693                 u64 dev_offset;
9694
9695                 /*
9696                  * check to make sure we can actually find a chunk with enough
9697                  * space to fit our block group in.
9698                  */
9699                 if (device->total_bytes > device->bytes_used + min_free &&
9700                     !test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state)) {
9701                         ret = find_free_dev_extent(trans, device, min_free,
9702                                                    &dev_offset, NULL);
9703                         if (!ret)
9704                                 dev_nr++;
9705
9706                         if (dev_nr >= dev_min)
9707                                 break;
9708
9709                         ret = -1;
9710                 }
9711         }
9712         if (debug && ret == -1)
9713                 btrfs_warn(fs_info,
9714                            "no space to allocate a new chunk for block group %llu",
9715                            block_group->key.objectid);
9716         mutex_unlock(&fs_info->chunk_mutex);
9717         btrfs_end_transaction(trans);
9718 out:
9719         btrfs_put_block_group(block_group);
9720         return ret;
9721 }
9722
9723 static int find_first_block_group(struct btrfs_fs_info *fs_info,
9724                                   struct btrfs_path *path,
9725                                   struct btrfs_key *key)
9726 {
9727         struct btrfs_root *root = fs_info->extent_root;
9728         int ret = 0;
9729         struct btrfs_key found_key;
9730         struct extent_buffer *leaf;
9731         int slot;
9732
9733         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
9734         if (ret < 0)
9735                 goto out;
9736
9737         while (1) {
9738                 slot = path->slots[0];
9739                 leaf = path->nodes[0];
9740                 if (slot >= btrfs_header_nritems(leaf)) {
9741                         ret = btrfs_next_leaf(root, path);
9742                         if (ret == 0)
9743                                 continue;
9744                         if (ret < 0)
9745                                 goto out;
9746                         break;
9747                 }
9748                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
9749
9750                 if (found_key.objectid >= key->objectid &&
9751                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
9752                         struct extent_map_tree *em_tree;
9753                         struct extent_map *em;
9754
9755                         em_tree = &root->fs_info->mapping_tree.map_tree;
9756                         read_lock(&em_tree->lock);
9757                         em = lookup_extent_mapping(em_tree, found_key.objectid,
9758                                                    found_key.offset);
9759                         read_unlock(&em_tree->lock);
9760                         if (!em) {
9761                                 btrfs_err(fs_info,
9762                         "logical %llu len %llu found bg but no related chunk",
9763                                           found_key.objectid, found_key.offset);
9764                                 ret = -ENOENT;
9765                         } else {
9766                                 ret = 0;
9767                         }
9768                         free_extent_map(em);
9769                         goto out;
9770                 }
9771                 path->slots[0]++;
9772         }
9773 out:
9774         return ret;
9775 }
9776
9777 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
9778 {
9779         struct btrfs_block_group_cache *block_group;
9780         u64 last = 0;
9781
9782         while (1) {
9783                 struct inode *inode;
9784
9785                 block_group = btrfs_lookup_first_block_group(info, last);
9786                 while (block_group) {
9787                         spin_lock(&block_group->lock);
9788                         if (block_group->iref)
9789                                 break;
9790                         spin_unlock(&block_group->lock);
9791                         block_group = next_block_group(info, block_group);
9792                 }
9793                 if (!block_group) {
9794                         if (last == 0)
9795                                 break;
9796                         last = 0;
9797                         continue;
9798                 }
9799
9800                 inode = block_group->inode;
9801                 block_group->iref = 0;
9802                 block_group->inode = NULL;
9803                 spin_unlock(&block_group->lock);
9804                 ASSERT(block_group->io_ctl.inode == NULL);
9805                 iput(inode);
9806                 last = block_group->key.objectid + block_group->key.offset;
9807                 btrfs_put_block_group(block_group);
9808         }
9809 }
9810
9811 /*
9812  * Must be called only after stopping all workers, since we could have block
9813  * group caching kthreads running, and therefore they could race with us if we
9814  * freed the block groups before stopping them.
9815  */
9816 int btrfs_free_block_groups(struct btrfs_fs_info *info)
9817 {
9818         struct btrfs_block_group_cache *block_group;
9819         struct btrfs_space_info *space_info;
9820         struct btrfs_caching_control *caching_ctl;
9821         struct rb_node *n;
9822
9823         down_write(&info->commit_root_sem);
9824         while (!list_empty(&info->caching_block_groups)) {
9825                 caching_ctl = list_entry(info->caching_block_groups.next,
9826                                          struct btrfs_caching_control, list);
9827                 list_del(&caching_ctl->list);
9828                 put_caching_control(caching_ctl);
9829         }
9830         up_write(&info->commit_root_sem);
9831
9832         spin_lock(&info->unused_bgs_lock);
9833         while (!list_empty(&info->unused_bgs)) {
9834                 block_group = list_first_entry(&info->unused_bgs,
9835                                                struct btrfs_block_group_cache,
9836                                                bg_list);
9837                 list_del_init(&block_group->bg_list);
9838                 btrfs_put_block_group(block_group);
9839         }
9840         spin_unlock(&info->unused_bgs_lock);
9841
9842         spin_lock(&info->block_group_cache_lock);
9843         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
9844                 block_group = rb_entry(n, struct btrfs_block_group_cache,
9845                                        cache_node);
9846                 rb_erase(&block_group->cache_node,
9847                          &info->block_group_cache_tree);
9848                 RB_CLEAR_NODE(&block_group->cache_node);
9849                 spin_unlock(&info->block_group_cache_lock);
9850
9851                 down_write(&block_group->space_info->groups_sem);
9852                 list_del(&block_group->list);
9853                 up_write(&block_group->space_info->groups_sem);
9854
9855                 /*
9856                  * We haven't cached this block group, which means we could
9857                  * possibly have excluded extents on this block group.
9858                  */
9859                 if (block_group->cached == BTRFS_CACHE_NO ||
9860                     block_group->cached == BTRFS_CACHE_ERROR)
9861                         free_excluded_extents(info, block_group);
9862
9863                 btrfs_remove_free_space_cache(block_group);
9864                 ASSERT(block_group->cached != BTRFS_CACHE_STARTED);
9865                 ASSERT(list_empty(&block_group->dirty_list));
9866                 ASSERT(list_empty(&block_group->io_list));
9867                 ASSERT(list_empty(&block_group->bg_list));
9868                 ASSERT(atomic_read(&block_group->count) == 1);
9869                 btrfs_put_block_group(block_group);
9870
9871                 spin_lock(&info->block_group_cache_lock);
9872         }
9873         spin_unlock(&info->block_group_cache_lock);
9874
9875         /* now that all the block groups are freed, go through and
9876          * free all the space_info structs.  This is only called during
9877          * the final stages of unmount, and so we know nobody is
9878          * using them.  We call synchronize_rcu() once before we start,
9879          * just to be on the safe side.
9880          */
9881         synchronize_rcu();
9882
9883         release_global_block_rsv(info);
9884
9885         while (!list_empty(&info->space_info)) {
9886                 int i;
9887
9888                 space_info = list_entry(info->space_info.next,
9889                                         struct btrfs_space_info,
9890                                         list);
9891
9892                 /*
9893                  * Do not hide this behind enospc_debug, this is actually
9894                  * important and indicates a real bug if this happens.
9895                  */
9896                 if (WARN_ON(space_info->bytes_pinned > 0 ||
9897                             space_info->bytes_reserved > 0 ||
9898                             space_info->bytes_may_use > 0))
9899                         dump_space_info(info, space_info, 0, 0);
9900                 list_del(&space_info->list);
9901                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
9902                         struct kobject *kobj;
9903                         kobj = space_info->block_group_kobjs[i];
9904                         space_info->block_group_kobjs[i] = NULL;
9905                         if (kobj) {
9906                                 kobject_del(kobj);
9907                                 kobject_put(kobj);
9908                         }
9909                 }
9910                 kobject_del(&space_info->kobj);
9911                 kobject_put(&space_info->kobj);
9912         }
9913         return 0;
9914 }
9915
9916 /* link_block_group will queue up kobjects to add when we're reclaim-safe */
9917 void btrfs_add_raid_kobjects(struct btrfs_fs_info *fs_info)
9918 {
9919         struct btrfs_space_info *space_info;
9920         struct raid_kobject *rkobj;
9921         LIST_HEAD(list);
9922         int index;
9923         int ret = 0;
9924
9925         spin_lock(&fs_info->pending_raid_kobjs_lock);
9926         list_splice_init(&fs_info->pending_raid_kobjs, &list);
9927         spin_unlock(&fs_info->pending_raid_kobjs_lock);
9928
9929         list_for_each_entry(rkobj, &list, list) {
9930                 space_info = __find_space_info(fs_info, rkobj->flags);
9931                 index = btrfs_bg_flags_to_raid_index(rkobj->flags);
9932
9933                 ret = kobject_add(&rkobj->kobj, &space_info->kobj,
9934                                   "%s", get_raid_name(index));
9935                 if (ret) {
9936                         kobject_put(&rkobj->kobj);
9937                         break;
9938                 }
9939         }
9940         if (ret)
9941                 btrfs_warn(fs_info,
9942                            "failed to add kobject for block cache, ignoring");
9943 }
9944
9945 static void link_block_group(struct btrfs_block_group_cache *cache)
9946 {
9947         struct btrfs_space_info *space_info = cache->space_info;
9948         struct btrfs_fs_info *fs_info = cache->fs_info;
9949         int index = btrfs_bg_flags_to_raid_index(cache->flags);
9950         bool first = false;
9951
9952         down_write(&space_info->groups_sem);
9953         if (list_empty(&space_info->block_groups[index]))
9954                 first = true;
9955         list_add_tail(&cache->list, &space_info->block_groups[index]);
9956         up_write(&space_info->groups_sem);
9957
9958         if (first) {
9959                 struct raid_kobject *rkobj = kzalloc(sizeof(*rkobj), GFP_NOFS);
9960                 if (!rkobj) {
9961                         btrfs_warn(cache->fs_info,
9962                                 "couldn't alloc memory for raid level kobject");
9963                         return;
9964                 }
9965                 rkobj->flags = cache->flags;
9966                 kobject_init(&rkobj->kobj, &btrfs_raid_ktype);
9967
9968                 spin_lock(&fs_info->pending_raid_kobjs_lock);
9969                 list_add_tail(&rkobj->list, &fs_info->pending_raid_kobjs);
9970                 spin_unlock(&fs_info->pending_raid_kobjs_lock);
9971                 space_info->block_group_kobjs[index] = &rkobj->kobj;
9972         }
9973 }
9974
9975 static struct btrfs_block_group_cache *
9976 btrfs_create_block_group_cache(struct btrfs_fs_info *fs_info,
9977                                u64 start, u64 size)
9978 {
9979         struct btrfs_block_group_cache *cache;
9980
9981         cache = kzalloc(sizeof(*cache), GFP_NOFS);
9982         if (!cache)
9983                 return NULL;
9984
9985         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
9986                                         GFP_NOFS);
9987         if (!cache->free_space_ctl) {
9988                 kfree(cache);
9989                 return NULL;
9990         }
9991
9992         cache->key.objectid = start;
9993         cache->key.offset = size;
9994         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
9995
9996         cache->fs_info = fs_info;
9997         cache->full_stripe_len = btrfs_full_stripe_len(fs_info, start);
9998         set_free_space_tree_thresholds(cache);
9999
10000         atomic_set(&cache->count, 1);
10001         spin_lock_init(&cache->lock);
10002         init_rwsem(&cache->data_rwsem);
10003         INIT_LIST_HEAD(&cache->list);
10004         INIT_LIST_HEAD(&cache->cluster_list);
10005         INIT_LIST_HEAD(&cache->bg_list);
10006         INIT_LIST_HEAD(&cache->ro_list);
10007         INIT_LIST_HEAD(&cache->dirty_list);
10008         INIT_LIST_HEAD(&cache->io_list);
10009         btrfs_init_free_space_ctl(cache);
10010         atomic_set(&cache->trimming, 0);
10011         mutex_init(&cache->free_space_lock);
10012         btrfs_init_full_stripe_locks_tree(&cache->full_stripe_locks_root);
10013
10014         return cache;
10015 }
10016
10017 int btrfs_read_block_groups(struct btrfs_fs_info *info)
10018 {
10019         struct btrfs_path *path;
10020         int ret;
10021         struct btrfs_block_group_cache *cache;
10022         struct btrfs_space_info *space_info;
10023         struct btrfs_key key;
10024         struct btrfs_key found_key;
10025         struct extent_buffer *leaf;
10026         int need_clear = 0;
10027         u64 cache_gen;
10028         u64 feature;
10029         int mixed;
10030
10031         feature = btrfs_super_incompat_flags(info->super_copy);
10032         mixed = !!(feature & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS);
10033
10034         key.objectid = 0;
10035         key.offset = 0;
10036         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
10037         path = btrfs_alloc_path();
10038         if (!path)
10039                 return -ENOMEM;
10040         path->reada = READA_FORWARD;
10041
10042         cache_gen = btrfs_super_cache_generation(info->super_copy);
10043         if (btrfs_test_opt(info, SPACE_CACHE) &&
10044             btrfs_super_generation(info->super_copy) != cache_gen)
10045                 need_clear = 1;
10046         if (btrfs_test_opt(info, CLEAR_CACHE))
10047                 need_clear = 1;
10048
10049         while (1) {
10050                 ret = find_first_block_group(info, path, &key);
10051                 if (ret > 0)
10052                         break;
10053                 if (ret != 0)
10054                         goto error;
10055
10056                 leaf = path->nodes[0];
10057                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
10058
10059                 cache = btrfs_create_block_group_cache(info, found_key.objectid,
10060                                                        found_key.offset);
10061                 if (!cache) {
10062                         ret = -ENOMEM;
10063                         goto error;
10064                 }
10065
10066                 if (need_clear) {
10067                         /*
10068                          * When we mount with old space cache, we need to
10069                          * set BTRFS_DC_CLEAR and set dirty flag.
10070                          *
10071                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
10072                          *    truncate the old free space cache inode and
10073                          *    setup a new one.
10074                          * b) Setting 'dirty flag' makes sure that we flush
10075                          *    the new space cache info onto disk.
10076                          */
10077                         if (btrfs_test_opt(info, SPACE_CACHE))
10078                                 cache->disk_cache_state = BTRFS_DC_CLEAR;
10079                 }
10080
10081                 read_extent_buffer(leaf, &cache->item,
10082                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
10083                                    sizeof(cache->item));
10084                 cache->flags = btrfs_block_group_flags(&cache->item);
10085                 if (!mixed &&
10086                     ((cache->flags & BTRFS_BLOCK_GROUP_METADATA) &&
10087                     (cache->flags & BTRFS_BLOCK_GROUP_DATA))) {
10088                         btrfs_err(info,
10089 "bg %llu is a mixed block group but filesystem hasn't enabled mixed block groups",
10090                                   cache->key.objectid);
10091                         ret = -EINVAL;
10092                         goto error;
10093                 }
10094
10095                 key.objectid = found_key.objectid + found_key.offset;
10096                 btrfs_release_path(path);
10097
10098                 /*
10099                  * We need to exclude the super stripes now so that the space
10100                  * info has super bytes accounted for, otherwise we'll think
10101                  * we have more space than we actually do.
10102                  */
10103                 ret = exclude_super_stripes(info, cache);
10104                 if (ret) {
10105                         /*
10106                          * We may have excluded something, so call this just in
10107                          * case.
10108                          */
10109                         free_excluded_extents(info, cache);
10110                         btrfs_put_block_group(cache);
10111                         goto error;
10112                 }
10113
10114                 /*
10115                  * check for two cases, either we are full, and therefore
10116                  * don't need to bother with the caching work since we won't
10117                  * find any space, or we are empty, and we can just add all
10118                  * the space in and be done with it.  This saves us _alot_ of
10119                  * time, particularly in the full case.
10120                  */
10121                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
10122                         cache->last_byte_to_unpin = (u64)-1;
10123                         cache->cached = BTRFS_CACHE_FINISHED;
10124                         free_excluded_extents(info, cache);
10125                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10126                         cache->last_byte_to_unpin = (u64)-1;
10127                         cache->cached = BTRFS_CACHE_FINISHED;
10128                         add_new_free_space(cache, info,
10129                                            found_key.objectid,
10130                                            found_key.objectid +
10131                                            found_key.offset);
10132                         free_excluded_extents(info, cache);
10133                 }
10134
10135                 ret = btrfs_add_block_group_cache(info, cache);
10136                 if (ret) {
10137                         btrfs_remove_free_space_cache(cache);
10138                         btrfs_put_block_group(cache);
10139                         goto error;
10140                 }
10141
10142                 trace_btrfs_add_block_group(info, cache, 0);
10143                 update_space_info(info, cache->flags, found_key.offset,
10144                                   btrfs_block_group_used(&cache->item),
10145                                   cache->bytes_super, &space_info);
10146
10147                 cache->space_info = space_info;
10148
10149                 link_block_group(cache);
10150
10151                 set_avail_alloc_bits(info, cache->flags);
10152                 if (btrfs_chunk_readonly(info, cache->key.objectid)) {
10153                         inc_block_group_ro(cache, 1);
10154                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10155                         spin_lock(&info->unused_bgs_lock);
10156                         /* Should always be true but just in case. */
10157                         if (list_empty(&cache->bg_list)) {
10158                                 btrfs_get_block_group(cache);
10159                                 list_add_tail(&cache->bg_list,
10160                                               &info->unused_bgs);
10161                         }
10162                         spin_unlock(&info->unused_bgs_lock);
10163                 }
10164         }
10165
10166         list_for_each_entry_rcu(space_info, &info->space_info, list) {
10167                 if (!(get_alloc_profile(info, space_info->flags) &
10168                       (BTRFS_BLOCK_GROUP_RAID10 |
10169                        BTRFS_BLOCK_GROUP_RAID1 |
10170                        BTRFS_BLOCK_GROUP_RAID5 |
10171                        BTRFS_BLOCK_GROUP_RAID6 |
10172                        BTRFS_BLOCK_GROUP_DUP)))
10173                         continue;
10174                 /*
10175                  * avoid allocating from un-mirrored block group if there are
10176                  * mirrored block groups.
10177                  */
10178                 list_for_each_entry(cache,
10179                                 &space_info->block_groups[BTRFS_RAID_RAID0],
10180                                 list)
10181                         inc_block_group_ro(cache, 1);
10182                 list_for_each_entry(cache,
10183                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
10184                                 list)
10185                         inc_block_group_ro(cache, 1);
10186         }
10187
10188         btrfs_add_raid_kobjects(info);
10189         init_global_block_rsv(info);
10190         ret = 0;
10191 error:
10192         btrfs_free_path(path);
10193         return ret;
10194 }
10195
10196 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans)
10197 {
10198         struct btrfs_fs_info *fs_info = trans->fs_info;
10199         struct btrfs_block_group_cache *block_group, *tmp;
10200         struct btrfs_root *extent_root = fs_info->extent_root;
10201         struct btrfs_block_group_item item;
10202         struct btrfs_key key;
10203         int ret = 0;
10204         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
10205
10206         trans->can_flush_pending_bgs = false;
10207         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) {
10208                 if (ret)
10209                         goto next;
10210
10211                 spin_lock(&block_group->lock);
10212                 memcpy(&item, &block_group->item, sizeof(item));
10213                 memcpy(&key, &block_group->key, sizeof(key));
10214                 spin_unlock(&block_group->lock);
10215
10216                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
10217                                         sizeof(item));
10218                 if (ret)
10219                         btrfs_abort_transaction(trans, ret);
10220                 ret = btrfs_finish_chunk_alloc(trans, fs_info, key.objectid,
10221                                                key.offset);
10222                 if (ret)
10223                         btrfs_abort_transaction(trans, ret);
10224                 add_block_group_free_space(trans, fs_info, block_group);
10225                 /* already aborted the transaction if it failed. */
10226 next:
10227                 list_del_init(&block_group->bg_list);
10228         }
10229         trans->can_flush_pending_bgs = can_flush_pending_bgs;
10230 }
10231
10232 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
10233                            struct btrfs_fs_info *fs_info, u64 bytes_used,
10234                            u64 type, u64 chunk_offset, u64 size)
10235 {
10236         struct btrfs_block_group_cache *cache;
10237         int ret;
10238
10239         btrfs_set_log_full_commit(fs_info, trans);
10240
10241         cache = btrfs_create_block_group_cache(fs_info, chunk_offset, size);
10242         if (!cache)
10243                 return -ENOMEM;
10244
10245         btrfs_set_block_group_used(&cache->item, bytes_used);
10246         btrfs_set_block_group_chunk_objectid(&cache->item,
10247                                              BTRFS_FIRST_CHUNK_TREE_OBJECTID);
10248         btrfs_set_block_group_flags(&cache->item, type);
10249
10250         cache->flags = type;
10251         cache->last_byte_to_unpin = (u64)-1;
10252         cache->cached = BTRFS_CACHE_FINISHED;
10253         cache->needs_free_space = 1;
10254         ret = exclude_super_stripes(fs_info, cache);
10255         if (ret) {
10256                 /*
10257                  * We may have excluded something, so call this just in
10258                  * case.
10259                  */
10260                 free_excluded_extents(fs_info, cache);
10261                 btrfs_put_block_group(cache);
10262                 return ret;
10263         }
10264
10265         add_new_free_space(cache, fs_info, chunk_offset, chunk_offset + size);
10266
10267         free_excluded_extents(fs_info, cache);
10268
10269 #ifdef CONFIG_BTRFS_DEBUG
10270         if (btrfs_should_fragment_free_space(cache)) {
10271                 u64 new_bytes_used = size - bytes_used;
10272
10273                 bytes_used += new_bytes_used >> 1;
10274                 fragment_free_space(cache);
10275         }
10276 #endif
10277         /*
10278          * Ensure the corresponding space_info object is created and
10279          * assigned to our block group. We want our bg to be added to the rbtree
10280          * with its ->space_info set.
10281          */
10282         cache->space_info = __find_space_info(fs_info, cache->flags);
10283         ASSERT(cache->space_info);
10284
10285         ret = btrfs_add_block_group_cache(fs_info, cache);
10286         if (ret) {
10287                 btrfs_remove_free_space_cache(cache);
10288                 btrfs_put_block_group(cache);
10289                 return ret;
10290         }
10291
10292         /*
10293          * Now that our block group has its ->space_info set and is inserted in
10294          * the rbtree, update the space info's counters.
10295          */
10296         trace_btrfs_add_block_group(fs_info, cache, 1);
10297         update_space_info(fs_info, cache->flags, size, bytes_used,
10298                                 cache->bytes_super, &cache->space_info);
10299         update_global_block_rsv(fs_info);
10300
10301         link_block_group(cache);
10302
10303         list_add_tail(&cache->bg_list, &trans->new_bgs);
10304
10305         set_avail_alloc_bits(fs_info, type);
10306         return 0;
10307 }
10308
10309 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
10310 {
10311         u64 extra_flags = chunk_to_extended(flags) &
10312                                 BTRFS_EXTENDED_PROFILE_MASK;
10313
10314         write_seqlock(&fs_info->profiles_lock);
10315         if (flags & BTRFS_BLOCK_GROUP_DATA)
10316                 fs_info->avail_data_alloc_bits &= ~extra_flags;
10317         if (flags & BTRFS_BLOCK_GROUP_METADATA)
10318                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
10319         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
10320                 fs_info->avail_system_alloc_bits &= ~extra_flags;
10321         write_sequnlock(&fs_info->profiles_lock);
10322 }
10323
10324 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
10325                              struct btrfs_fs_info *fs_info, u64 group_start,
10326                              struct extent_map *em)
10327 {
10328         struct btrfs_root *root = fs_info->extent_root;
10329         struct btrfs_path *path;
10330         struct btrfs_block_group_cache *block_group;
10331         struct btrfs_free_cluster *cluster;
10332         struct btrfs_root *tree_root = fs_info->tree_root;
10333         struct btrfs_key key;
10334         struct inode *inode;
10335         struct kobject *kobj = NULL;
10336         int ret;
10337         int index;
10338         int factor;
10339         struct btrfs_caching_control *caching_ctl = NULL;
10340         bool remove_em;
10341
10342         block_group = btrfs_lookup_block_group(fs_info, group_start);
10343         BUG_ON(!block_group);
10344         BUG_ON(!block_group->ro);
10345
10346         /*
10347          * Free the reserved super bytes from this block group before
10348          * remove it.
10349          */
10350         free_excluded_extents(fs_info, block_group);
10351         btrfs_free_ref_tree_range(fs_info, block_group->key.objectid,
10352                                   block_group->key.offset);
10353
10354         memcpy(&key, &block_group->key, sizeof(key));
10355         index = btrfs_bg_flags_to_raid_index(block_group->flags);
10356         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
10357                                   BTRFS_BLOCK_GROUP_RAID1 |
10358                                   BTRFS_BLOCK_GROUP_RAID10))
10359                 factor = 2;
10360         else
10361                 factor = 1;
10362
10363         /* make sure this block group isn't part of an allocation cluster */
10364         cluster = &fs_info->data_alloc_cluster;
10365         spin_lock(&cluster->refill_lock);
10366         btrfs_return_cluster_to_free_space(block_group, cluster);
10367         spin_unlock(&cluster->refill_lock);
10368
10369         /*
10370          * make sure this block group isn't part of a metadata
10371          * allocation cluster
10372          */
10373         cluster = &fs_info->meta_alloc_cluster;
10374         spin_lock(&cluster->refill_lock);
10375         btrfs_return_cluster_to_free_space(block_group, cluster);
10376         spin_unlock(&cluster->refill_lock);
10377
10378         path = btrfs_alloc_path();
10379         if (!path) {
10380                 ret = -ENOMEM;
10381                 goto out;
10382         }
10383
10384         /*
10385          * get the inode first so any iput calls done for the io_list
10386          * aren't the final iput (no unlinks allowed now)
10387          */
10388         inode = lookup_free_space_inode(fs_info, block_group, path);
10389
10390         mutex_lock(&trans->transaction->cache_write_mutex);
10391         /*
10392          * make sure our free spache cache IO is done before remove the
10393          * free space inode
10394          */
10395         spin_lock(&trans->transaction->dirty_bgs_lock);
10396         if (!list_empty(&block_group->io_list)) {
10397                 list_del_init(&block_group->io_list);
10398
10399                 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
10400
10401                 spin_unlock(&trans->transaction->dirty_bgs_lock);
10402                 btrfs_wait_cache_io(trans, block_group, path);
10403                 btrfs_put_block_group(block_group);
10404                 spin_lock(&trans->transaction->dirty_bgs_lock);
10405         }
10406
10407         if (!list_empty(&block_group->dirty_list)) {
10408                 list_del_init(&block_group->dirty_list);
10409                 btrfs_put_block_group(block_group);
10410         }
10411         spin_unlock(&trans->transaction->dirty_bgs_lock);
10412         mutex_unlock(&trans->transaction->cache_write_mutex);
10413
10414         if (!IS_ERR(inode)) {
10415                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
10416                 if (ret) {
10417                         btrfs_add_delayed_iput(inode);
10418                         goto out;
10419                 }
10420                 clear_nlink(inode);
10421                 /* One for the block groups ref */
10422                 spin_lock(&block_group->lock);
10423                 if (block_group->iref) {
10424                         block_group->iref = 0;
10425                         block_group->inode = NULL;
10426                         spin_unlock(&block_group->lock);
10427                         iput(inode);
10428                 } else {
10429                         spin_unlock(&block_group->lock);
10430                 }
10431                 /* One for our lookup ref */
10432                 btrfs_add_delayed_iput(inode);
10433         }
10434
10435         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
10436         key.offset = block_group->key.objectid;
10437         key.type = 0;
10438
10439         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
10440         if (ret < 0)
10441                 goto out;
10442         if (ret > 0)
10443                 btrfs_release_path(path);
10444         if (ret == 0) {
10445                 ret = btrfs_del_item(trans, tree_root, path);
10446                 if (ret)
10447                         goto out;
10448                 btrfs_release_path(path);
10449         }
10450
10451         spin_lock(&fs_info->block_group_cache_lock);
10452         rb_erase(&block_group->cache_node,
10453                  &fs_info->block_group_cache_tree);
10454         RB_CLEAR_NODE(&block_group->cache_node);
10455
10456         if (fs_info->first_logical_byte == block_group->key.objectid)
10457                 fs_info->first_logical_byte = (u64)-1;
10458         spin_unlock(&fs_info->block_group_cache_lock);
10459
10460         down_write(&block_group->space_info->groups_sem);
10461         /*
10462          * we must use list_del_init so people can check to see if they
10463          * are still on the list after taking the semaphore
10464          */
10465         list_del_init(&block_group->list);
10466         if (list_empty(&block_group->space_info->block_groups[index])) {
10467                 kobj = block_group->space_info->block_group_kobjs[index];
10468                 block_group->space_info->block_group_kobjs[index] = NULL;
10469                 clear_avail_alloc_bits(fs_info, block_group->flags);
10470         }
10471         up_write(&block_group->space_info->groups_sem);
10472         if (kobj) {
10473                 kobject_del(kobj);
10474                 kobject_put(kobj);
10475         }
10476
10477         if (block_group->has_caching_ctl)
10478                 caching_ctl = get_caching_control(block_group);
10479         if (block_group->cached == BTRFS_CACHE_STARTED)
10480                 wait_block_group_cache_done(block_group);
10481         if (block_group->has_caching_ctl) {
10482                 down_write(&fs_info->commit_root_sem);
10483                 if (!caching_ctl) {
10484                         struct btrfs_caching_control *ctl;
10485
10486                         list_for_each_entry(ctl,
10487                                     &fs_info->caching_block_groups, list)
10488                                 if (ctl->block_group == block_group) {
10489                                         caching_ctl = ctl;
10490                                         refcount_inc(&caching_ctl->count);
10491                                         break;
10492                                 }
10493                 }
10494                 if (caching_ctl)
10495                         list_del_init(&caching_ctl->list);
10496                 up_write(&fs_info->commit_root_sem);
10497                 if (caching_ctl) {
10498                         /* Once for the caching bgs list and once for us. */
10499                         put_caching_control(caching_ctl);
10500                         put_caching_control(caching_ctl);
10501                 }
10502         }
10503
10504         spin_lock(&trans->transaction->dirty_bgs_lock);
10505         if (!list_empty(&block_group->dirty_list)) {
10506                 WARN_ON(1);
10507         }
10508         if (!list_empty(&block_group->io_list)) {
10509                 WARN_ON(1);
10510         }
10511         spin_unlock(&trans->transaction->dirty_bgs_lock);
10512         btrfs_remove_free_space_cache(block_group);
10513
10514         spin_lock(&block_group->space_info->lock);
10515         list_del_init(&block_group->ro_list);
10516
10517         if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
10518                 WARN_ON(block_group->space_info->total_bytes
10519                         < block_group->key.offset);
10520                 WARN_ON(block_group->space_info->bytes_readonly
10521                         < block_group->key.offset);
10522                 WARN_ON(block_group->space_info->disk_total
10523                         < block_group->key.offset * factor);
10524         }
10525         block_group->space_info->total_bytes -= block_group->key.offset;
10526         block_group->space_info->bytes_readonly -= block_group->key.offset;
10527         block_group->space_info->disk_total -= block_group->key.offset * factor;
10528
10529         spin_unlock(&block_group->space_info->lock);
10530
10531         memcpy(&key, &block_group->key, sizeof(key));
10532
10533         mutex_lock(&fs_info->chunk_mutex);
10534         if (!list_empty(&em->list)) {
10535                 /* We're in the transaction->pending_chunks list. */
10536                 free_extent_map(em);
10537         }
10538         spin_lock(&block_group->lock);
10539         block_group->removed = 1;
10540         /*
10541          * At this point trimming can't start on this block group, because we
10542          * removed the block group from the tree fs_info->block_group_cache_tree
10543          * so no one can't find it anymore and even if someone already got this
10544          * block group before we removed it from the rbtree, they have already
10545          * incremented block_group->trimming - if they didn't, they won't find
10546          * any free space entries because we already removed them all when we
10547          * called btrfs_remove_free_space_cache().
10548          *
10549          * And we must not remove the extent map from the fs_info->mapping_tree
10550          * to prevent the same logical address range and physical device space
10551          * ranges from being reused for a new block group. This is because our
10552          * fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
10553          * completely transactionless, so while it is trimming a range the
10554          * currently running transaction might finish and a new one start,
10555          * allowing for new block groups to be created that can reuse the same
10556          * physical device locations unless we take this special care.
10557          *
10558          * There may also be an implicit trim operation if the file system
10559          * is mounted with -odiscard. The same protections must remain
10560          * in place until the extents have been discarded completely when
10561          * the transaction commit has completed.
10562          */
10563         remove_em = (atomic_read(&block_group->trimming) == 0);
10564         /*
10565          * Make sure a trimmer task always sees the em in the pinned_chunks list
10566          * if it sees block_group->removed == 1 (needs to lock block_group->lock
10567          * before checking block_group->removed).
10568          */
10569         if (!remove_em) {
10570                 /*
10571                  * Our em might be in trans->transaction->pending_chunks which
10572                  * is protected by fs_info->chunk_mutex ([lock|unlock]_chunks),
10573                  * and so is the fs_info->pinned_chunks list.
10574                  *
10575                  * So at this point we must be holding the chunk_mutex to avoid
10576                  * any races with chunk allocation (more specifically at
10577                  * volumes.c:contains_pending_extent()), to ensure it always
10578                  * sees the em, either in the pending_chunks list or in the
10579                  * pinned_chunks list.
10580                  */
10581                 list_move_tail(&em->list, &fs_info->pinned_chunks);
10582         }
10583         spin_unlock(&block_group->lock);
10584
10585         if (remove_em) {
10586                 struct extent_map_tree *em_tree;
10587
10588                 em_tree = &fs_info->mapping_tree.map_tree;
10589                 write_lock(&em_tree->lock);
10590                 /*
10591                  * The em might be in the pending_chunks list, so make sure the
10592                  * chunk mutex is locked, since remove_extent_mapping() will
10593                  * delete us from that list.
10594                  */
10595                 remove_extent_mapping(em_tree, em);
10596                 write_unlock(&em_tree->lock);
10597                 /* once for the tree */
10598                 free_extent_map(em);
10599         }
10600
10601         mutex_unlock(&fs_info->chunk_mutex);
10602
10603         ret = remove_block_group_free_space(trans, fs_info, block_group);
10604         if (ret)
10605                 goto out;
10606
10607         btrfs_put_block_group(block_group);
10608         btrfs_put_block_group(block_group);
10609
10610         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
10611         if (ret > 0)
10612                 ret = -EIO;
10613         if (ret < 0)
10614                 goto out;
10615
10616         ret = btrfs_del_item(trans, root, path);
10617 out:
10618         btrfs_free_path(path);
10619         return ret;
10620 }
10621
10622 struct btrfs_trans_handle *
10623 btrfs_start_trans_remove_block_group(struct btrfs_fs_info *fs_info,
10624                                      const u64 chunk_offset)
10625 {
10626         struct extent_map_tree *em_tree = &fs_info->mapping_tree.map_tree;
10627         struct extent_map *em;
10628         struct map_lookup *map;
10629         unsigned int num_items;
10630
10631         read_lock(&em_tree->lock);
10632         em = lookup_extent_mapping(em_tree, chunk_offset, 1);
10633         read_unlock(&em_tree->lock);
10634         ASSERT(em && em->start == chunk_offset);
10635
10636         /*
10637          * We need to reserve 3 + N units from the metadata space info in order
10638          * to remove a block group (done at btrfs_remove_chunk() and at
10639          * btrfs_remove_block_group()), which are used for:
10640          *
10641          * 1 unit for adding the free space inode's orphan (located in the tree
10642          * of tree roots).
10643          * 1 unit for deleting the block group item (located in the extent
10644          * tree).
10645          * 1 unit for deleting the free space item (located in tree of tree
10646          * roots).
10647          * N units for deleting N device extent items corresponding to each
10648          * stripe (located in the device tree).
10649          *
10650          * In order to remove a block group we also need to reserve units in the
10651          * system space info in order to update the chunk tree (update one or
10652          * more device items and remove one chunk item), but this is done at
10653          * btrfs_remove_chunk() through a call to check_system_chunk().
10654          */
10655         map = em->map_lookup;
10656         num_items = 3 + map->num_stripes;
10657         free_extent_map(em);
10658
10659         return btrfs_start_transaction_fallback_global_rsv(fs_info->extent_root,
10660                                                            num_items, 1);
10661 }
10662
10663 /*
10664  * Process the unused_bgs list and remove any that don't have any allocated
10665  * space inside of them.
10666  */
10667 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
10668 {
10669         struct btrfs_block_group_cache *block_group;
10670         struct btrfs_space_info *space_info;
10671         struct btrfs_trans_handle *trans;
10672         int ret = 0;
10673
10674         if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
10675                 return;
10676
10677         spin_lock(&fs_info->unused_bgs_lock);
10678         while (!list_empty(&fs_info->unused_bgs)) {
10679                 u64 start, end;
10680                 int trimming;
10681
10682                 block_group = list_first_entry(&fs_info->unused_bgs,
10683                                                struct btrfs_block_group_cache,
10684                                                bg_list);
10685                 list_del_init(&block_group->bg_list);
10686
10687                 space_info = block_group->space_info;
10688
10689                 if (ret || btrfs_mixed_space_info(space_info)) {
10690                         btrfs_put_block_group(block_group);
10691                         continue;
10692                 }
10693                 spin_unlock(&fs_info->unused_bgs_lock);
10694
10695                 mutex_lock(&fs_info->delete_unused_bgs_mutex);
10696
10697                 /* Don't want to race with allocators so take the groups_sem */
10698                 down_write(&space_info->groups_sem);
10699                 spin_lock(&block_group->lock);
10700                 if (block_group->reserved ||
10701                     btrfs_block_group_used(&block_group->item) ||
10702                     block_group->ro ||
10703                     list_is_singular(&block_group->list)) {
10704                         /*
10705                          * We want to bail if we made new allocations or have
10706                          * outstanding allocations in this block group.  We do
10707                          * the ro check in case balance is currently acting on
10708                          * this block group.
10709                          */
10710                         spin_unlock(&block_group->lock);
10711                         up_write(&space_info->groups_sem);
10712                         goto next;
10713                 }
10714                 spin_unlock(&block_group->lock);
10715
10716                 /* We don't want to force the issue, only flip if it's ok. */
10717                 ret = inc_block_group_ro(block_group, 0);
10718                 up_write(&space_info->groups_sem);
10719                 if (ret < 0) {
10720                         ret = 0;
10721                         goto next;
10722                 }
10723
10724                 /*
10725                  * Want to do this before we do anything else so we can recover
10726                  * properly if we fail to join the transaction.
10727                  */
10728                 trans = btrfs_start_trans_remove_block_group(fs_info,
10729                                                      block_group->key.objectid);
10730                 if (IS_ERR(trans)) {
10731                         btrfs_dec_block_group_ro(block_group);
10732                         ret = PTR_ERR(trans);
10733                         goto next;
10734                 }
10735
10736                 /*
10737                  * We could have pending pinned extents for this block group,
10738                  * just delete them, we don't care about them anymore.
10739                  */
10740                 start = block_group->key.objectid;
10741                 end = start + block_group->key.offset - 1;
10742                 /*
10743                  * Hold the unused_bg_unpin_mutex lock to avoid racing with
10744                  * btrfs_finish_extent_commit(). If we are at transaction N,
10745                  * another task might be running finish_extent_commit() for the
10746                  * previous transaction N - 1, and have seen a range belonging
10747                  * to the block group in freed_extents[] before we were able to
10748                  * clear the whole block group range from freed_extents[]. This
10749                  * means that task can lookup for the block group after we
10750                  * unpinned it from freed_extents[] and removed it, leading to
10751                  * a BUG_ON() at btrfs_unpin_extent_range().
10752                  */
10753                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
10754                 ret = clear_extent_bits(&fs_info->freed_extents[0], start, end,
10755                                   EXTENT_DIRTY);
10756                 if (ret) {
10757                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10758                         btrfs_dec_block_group_ro(block_group);
10759                         goto end_trans;
10760                 }
10761                 ret = clear_extent_bits(&fs_info->freed_extents[1], start, end,
10762                                   EXTENT_DIRTY);
10763                 if (ret) {
10764                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10765                         btrfs_dec_block_group_ro(block_group);
10766                         goto end_trans;
10767                 }
10768                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10769
10770                 /* Reset pinned so btrfs_put_block_group doesn't complain */
10771                 spin_lock(&space_info->lock);
10772                 spin_lock(&block_group->lock);
10773
10774                 space_info->bytes_pinned -= block_group->pinned;
10775                 space_info->bytes_readonly += block_group->pinned;
10776                 percpu_counter_add(&space_info->total_bytes_pinned,
10777                                    -block_group->pinned);
10778                 block_group->pinned = 0;
10779
10780                 spin_unlock(&block_group->lock);
10781                 spin_unlock(&space_info->lock);
10782
10783                 /* DISCARD can flip during remount */
10784                 trimming = btrfs_test_opt(fs_info, DISCARD);
10785
10786                 /* Implicit trim during transaction commit. */
10787                 if (trimming)
10788                         btrfs_get_block_group_trimming(block_group);
10789
10790                 /*
10791                  * Btrfs_remove_chunk will abort the transaction if things go
10792                  * horribly wrong.
10793                  */
10794                 ret = btrfs_remove_chunk(trans, fs_info,
10795                                          block_group->key.objectid);
10796
10797                 if (ret) {
10798                         if (trimming)
10799                                 btrfs_put_block_group_trimming(block_group);
10800                         goto end_trans;
10801                 }
10802
10803                 /*
10804                  * If we're not mounted with -odiscard, we can just forget
10805                  * about this block group. Otherwise we'll need to wait
10806                  * until transaction commit to do the actual discard.
10807                  */
10808                 if (trimming) {
10809                         spin_lock(&fs_info->unused_bgs_lock);
10810                         /*
10811                          * A concurrent scrub might have added us to the list
10812                          * fs_info->unused_bgs, so use a list_move operation
10813                          * to add the block group to the deleted_bgs list.
10814                          */
10815                         list_move(&block_group->bg_list,
10816                                   &trans->transaction->deleted_bgs);
10817                         spin_unlock(&fs_info->unused_bgs_lock);
10818                         btrfs_get_block_group(block_group);
10819                 }
10820 end_trans:
10821                 btrfs_end_transaction(trans);
10822 next:
10823                 mutex_unlock(&fs_info->delete_unused_bgs_mutex);
10824                 btrfs_put_block_group(block_group);
10825                 spin_lock(&fs_info->unused_bgs_lock);
10826         }
10827         spin_unlock(&fs_info->unused_bgs_lock);
10828 }
10829
10830 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
10831 {
10832         struct btrfs_space_info *space_info;
10833         struct btrfs_super_block *disk_super;
10834         u64 features;
10835         u64 flags;
10836         int mixed = 0;
10837         int ret;
10838
10839         disk_super = fs_info->super_copy;
10840         if (!btrfs_super_root(disk_super))
10841                 return -EINVAL;
10842
10843         features = btrfs_super_incompat_flags(disk_super);
10844         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
10845                 mixed = 1;
10846
10847         flags = BTRFS_BLOCK_GROUP_SYSTEM;
10848         ret = create_space_info(fs_info, flags, &space_info);
10849         if (ret)
10850                 goto out;
10851
10852         if (mixed) {
10853                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
10854                 ret = create_space_info(fs_info, flags, &space_info);
10855         } else {
10856                 flags = BTRFS_BLOCK_GROUP_METADATA;
10857                 ret = create_space_info(fs_info, flags, &space_info);
10858                 if (ret)
10859                         goto out;
10860
10861                 flags = BTRFS_BLOCK_GROUP_DATA;
10862                 ret = create_space_info(fs_info, flags, &space_info);
10863         }
10864 out:
10865         return ret;
10866 }
10867
10868 int btrfs_error_unpin_extent_range(struct btrfs_fs_info *fs_info,
10869                                    u64 start, u64 end)
10870 {
10871         return unpin_extent_range(fs_info, start, end, false);
10872 }
10873
10874 /*
10875  * It used to be that old block groups would be left around forever.
10876  * Iterating over them would be enough to trim unused space.  Since we
10877  * now automatically remove them, we also need to iterate over unallocated
10878  * space.
10879  *
10880  * We don't want a transaction for this since the discard may take a
10881  * substantial amount of time.  We don't require that a transaction be
10882  * running, but we do need to take a running transaction into account
10883  * to ensure that we're not discarding chunks that were released in
10884  * the current transaction.
10885  *
10886  * Holding the chunks lock will prevent other threads from allocating
10887  * or releasing chunks, but it won't prevent a running transaction
10888  * from committing and releasing the memory that the pending chunks
10889  * list head uses.  For that, we need to take a reference to the
10890  * transaction.
10891  */
10892 static int btrfs_trim_free_extents(struct btrfs_device *device,
10893                                    u64 minlen, u64 *trimmed)
10894 {
10895         u64 start = 0, len = 0;
10896         int ret;
10897
10898         *trimmed = 0;
10899
10900         /* Not writeable = nothing to do. */
10901         if (!test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state))
10902                 return 0;
10903
10904         /* No free space = nothing to do. */
10905         if (device->total_bytes <= device->bytes_used)
10906                 return 0;
10907
10908         ret = 0;
10909
10910         while (1) {
10911                 struct btrfs_fs_info *fs_info = device->fs_info;
10912                 struct btrfs_transaction *trans;
10913                 u64 bytes;
10914
10915                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
10916                 if (ret)
10917                         return ret;
10918
10919                 down_read(&fs_info->commit_root_sem);
10920
10921                 spin_lock(&fs_info->trans_lock);
10922                 trans = fs_info->running_transaction;
10923                 if (trans)
10924                         refcount_inc(&trans->use_count);
10925                 spin_unlock(&fs_info->trans_lock);
10926
10927                 ret = find_free_dev_extent_start(trans, device, minlen, start,
10928                                                  &start, &len);
10929                 if (trans)
10930                         btrfs_put_transaction(trans);
10931
10932                 if (ret) {
10933                         up_read(&fs_info->commit_root_sem);
10934                         mutex_unlock(&fs_info->chunk_mutex);
10935                         if (ret == -ENOSPC)
10936                                 ret = 0;
10937                         break;
10938                 }
10939
10940                 ret = btrfs_issue_discard(device->bdev, start, len, &bytes);
10941                 up_read(&fs_info->commit_root_sem);
10942                 mutex_unlock(&fs_info->chunk_mutex);
10943
10944                 if (ret)
10945                         break;
10946
10947                 start += len;
10948                 *trimmed += bytes;
10949
10950                 if (fatal_signal_pending(current)) {
10951                         ret = -ERESTARTSYS;
10952                         break;
10953                 }
10954
10955                 cond_resched();
10956         }
10957
10958         return ret;
10959 }
10960
10961 int btrfs_trim_fs(struct btrfs_fs_info *fs_info, struct fstrim_range *range)
10962 {
10963         struct btrfs_block_group_cache *cache = NULL;
10964         struct btrfs_device *device;
10965         struct list_head *devices;
10966         u64 group_trimmed;
10967         u64 start;
10968         u64 end;
10969         u64 trimmed = 0;
10970         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
10971         int ret = 0;
10972
10973         /*
10974          * try to trim all FS space, our block group may start from non-zero.
10975          */
10976         if (range->len == total_bytes)
10977                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
10978         else
10979                 cache = btrfs_lookup_block_group(fs_info, range->start);
10980
10981         while (cache) {
10982                 if (cache->key.objectid >= (range->start + range->len)) {
10983                         btrfs_put_block_group(cache);
10984                         break;
10985                 }
10986
10987                 start = max(range->start, cache->key.objectid);
10988                 end = min(range->start + range->len,
10989                                 cache->key.objectid + cache->key.offset);
10990
10991                 if (end - start >= range->minlen) {
10992                         if (!block_group_cache_done(cache)) {
10993                                 ret = cache_block_group(cache, 0);
10994                                 if (ret) {
10995                                         btrfs_put_block_group(cache);
10996                                         break;
10997                                 }
10998                                 ret = wait_block_group_cache_done(cache);
10999                                 if (ret) {
11000                                         btrfs_put_block_group(cache);
11001                                         break;
11002                                 }
11003                         }
11004                         ret = btrfs_trim_block_group(cache,
11005                                                      &group_trimmed,
11006                                                      start,
11007                                                      end,
11008                                                      range->minlen);
11009
11010                         trimmed += group_trimmed;
11011                         if (ret) {
11012                                 btrfs_put_block_group(cache);
11013                                 break;
11014                         }
11015                 }
11016
11017                 cache = next_block_group(fs_info, cache);
11018         }
11019
11020         mutex_lock(&fs_info->fs_devices->device_list_mutex);
11021         devices = &fs_info->fs_devices->alloc_list;
11022         list_for_each_entry(device, devices, dev_alloc_list) {
11023                 ret = btrfs_trim_free_extents(device, range->minlen,
11024                                               &group_trimmed);
11025                 if (ret)
11026                         break;
11027
11028                 trimmed += group_trimmed;
11029         }
11030         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
11031
11032         range->len = trimmed;
11033         return ret;
11034 }
11035
11036 /*
11037  * btrfs_{start,end}_write_no_snapshotting() are similar to
11038  * mnt_{want,drop}_write(), they are used to prevent some tasks from writing
11039  * data into the page cache through nocow before the subvolume is snapshoted,
11040  * but flush the data into disk after the snapshot creation, or to prevent
11041  * operations while snapshotting is ongoing and that cause the snapshot to be
11042  * inconsistent (writes followed by expanding truncates for example).
11043  */
11044 void btrfs_end_write_no_snapshotting(struct btrfs_root *root)
11045 {
11046         percpu_counter_dec(&root->subv_writers->counter);
11047         /*
11048          * Make sure counter is updated before we wake up waiters.
11049          */
11050         smp_mb();
11051         if (waitqueue_active(&root->subv_writers->wait))
11052                 wake_up(&root->subv_writers->wait);
11053 }
11054
11055 int btrfs_start_write_no_snapshotting(struct btrfs_root *root)
11056 {
11057         if (atomic_read(&root->will_be_snapshotted))
11058                 return 0;
11059
11060         percpu_counter_inc(&root->subv_writers->counter);
11061         /*
11062          * Make sure counter is updated before we check for snapshot creation.
11063          */
11064         smp_mb();
11065         if (atomic_read(&root->will_be_snapshotted)) {
11066                 btrfs_end_write_no_snapshotting(root);
11067                 return 0;
11068         }
11069         return 1;
11070 }
11071
11072 void btrfs_wait_for_snapshot_creation(struct btrfs_root *root)
11073 {
11074         while (true) {
11075                 int ret;
11076
11077                 ret = btrfs_start_write_no_snapshotting(root);
11078                 if (ret)
11079                         break;
11080                 wait_var_event(&root->will_be_snapshotted,
11081                                !atomic_read(&root->will_be_snapshotted));
11082         }
11083 }