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 "misc.h"
  20 #include "tree-log.h"
  21 #include "disk-io.h"
  22 #include "print-tree.h"
  23 #include "volumes.h"
  24 #include "raid56.h"
  25 #include "locking.h"
  26 #include "free-space-cache.h"
  27 #include "free-space-tree.h"
  28 #include "sysfs.h"
  29 #include "qgroup.h"
  30 #include "ref-verify.h"
  31 #include "space-info.h"
  32 #include "block-rsv.h"
  33 #include "delalloc-space.h"
  34 #include "block-group.h"
  35 #include "discard.h"
  36 #include "rcu-string.h"
  37
  38 #undef SCRAMBLE_DELAYED_REFS
  39
  40
  41 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  42                                struct btrfs_delayed_ref_node *node, u64 parent,
  43                                u64 root_objectid, u64 owner_objectid,
  44                                u64 owner_offset, int refs_to_drop,
  45                                struct btrfs_delayed_extent_op *extra_op);
  46 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  47                                     struct extent_buffer *leaf,
  48                                     struct btrfs_extent_item *ei);
  49 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  50                                       u64 parent, u64 root_objectid,
  51                                       u64 flags, u64 owner, u64 offset,
  52                                       struct btrfs_key *ins, int ref_mod);
  53 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  54                                      struct btrfs_delayed_ref_node *node,
  55                                      struct btrfs_delayed_extent_op *extent_op);
  56 static int find_next_key(struct btrfs_path *path, int level,
  57                          struct btrfs_key *key);
  58
  59 static int block_group_bits(struct btrfs_block_group *cache, u64 bits)
  60 {
  61         return (cache->flags & bits) == bits;
  62 }
  63
  64 int btrfs_add_excluded_extent(struct btrfs_fs_info *fs_info,
  65                               u64 start, u64 num_bytes)
  66 {
  67         u64 end = start + num_bytes - 1;
  68         set_extent_bits(&fs_info->excluded_extents, start, end,
  69                         EXTENT_UPTODATE);
  70         return 0;
  71 }
  72
  73 void btrfs_free_excluded_extents(struct btrfs_block_group *cache)
  74 {
  75         struct btrfs_fs_info *fs_info = cache->fs_info;
  76         u64 start, end;
  77
  78         start = cache->start;
  79         end = start + cache->length - 1;
  80
  81         clear_extent_bits(&fs_info->excluded_extents, start, end,
  82                           EXTENT_UPTODATE);
  83 }
  84
  85 static u64 generic_ref_to_space_flags(struct btrfs_ref *ref)
  86 {
  87         if (ref->type == BTRFS_REF_METADATA) {
  88                 if (ref->tree_ref.root == BTRFS_CHUNK_TREE_OBJECTID)
  89                         return BTRFS_BLOCK_GROUP_SYSTEM;
  90                 else
  91                         return BTRFS_BLOCK_GROUP_METADATA;
  92         }
  93         return BTRFS_BLOCK_GROUP_DATA;
  94 }
  95
  96 static void add_pinned_bytes(struct btrfs_fs_info *fs_info,
  97                              struct btrfs_ref *ref)
  98 {
  99         struct btrfs_space_info *space_info;
 100         u64 flags = generic_ref_to_space_flags(ref);
 101
 102         space_info = btrfs_find_space_info(fs_info, flags);
 103         ASSERT(space_info);
 104         percpu_counter_add_batch(&space_info->total_bytes_pinned, ref->len,
 105                     BTRFS_TOTAL_BYTES_PINNED_BATCH);
 106 }
 107
 108 static void sub_pinned_bytes(struct btrfs_fs_info *fs_info,
 109                              struct btrfs_ref *ref)
 110 {
 111         struct btrfs_space_info *space_info;
 112         u64 flags = generic_ref_to_space_flags(ref);
 113
 114         space_info = btrfs_find_space_info(fs_info, flags);
 115         ASSERT(space_info);
 116         percpu_counter_add_batch(&space_info->total_bytes_pinned, -ref->len,
 117                     BTRFS_TOTAL_BYTES_PINNED_BATCH);
 118 }
 119
 120 /* simple helper to search for an existing data extent at a given offset */
 121 int btrfs_lookup_data_extent(struct btrfs_fs_info *fs_info, u64 start, u64 len)
 122 {
 123         int ret;
 124         struct btrfs_key key;
 125         struct btrfs_path *path;
 126
 127         path = btrfs_alloc_path();
 128         if (!path)
 129                 return -ENOMEM;
 130
 131         key.objectid = start;
 132         key.offset = len;
 133         key.type = BTRFS_EXTENT_ITEM_KEY;
 134         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
 135         btrfs_free_path(path);
 136         return ret;
 137 }
 138
 139 /*
 140  * helper function to lookup reference count and flags of a tree block.
 141  *
 142  * the head node for delayed ref is used to store the sum of all the
 143  * reference count modifications queued up in the rbtree. the head
 144  * node may also store the extent flags to set. This way you can check
 145  * to see what the reference count and extent flags would be if all of
 146  * the delayed refs are not processed.
 147  */
 148 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 149                              struct btrfs_fs_info *fs_info, u64 bytenr,
 150                              u64 offset, int metadata, u64 *refs, u64 *flags)
 151 {
 152         struct btrfs_delayed_ref_head *head;
 153         struct btrfs_delayed_ref_root *delayed_refs;
 154         struct btrfs_path *path;
 155         struct btrfs_extent_item *ei;
 156         struct extent_buffer *leaf;
 157         struct btrfs_key key;
 158         u32 item_size;
 159         u64 num_refs;
 160         u64 extent_flags;
 161         int ret;
 162
 163         /*
 164          * If we don't have skinny metadata, don't bother doing anything
 165          * different
 166          */
 167         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA)) {
 168                 offset = fs_info->nodesize;
 169                 metadata = 0;
 170         }
 171
 172         path = btrfs_alloc_path();
 173         if (!path)
 174                 return -ENOMEM;
 175
 176         if (!trans) {
 177                 path->skip_locking = 1;
 178                 path->search_commit_root = 1;
 179         }
 180
 181 search_again:
 182         key.objectid = bytenr;
 183         key.offset = offset;
 184         if (metadata)
 185                 key.type = BTRFS_METADATA_ITEM_KEY;
 186         else
 187                 key.type = BTRFS_EXTENT_ITEM_KEY;
 188
 189         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 0);
 190         if (ret < 0)
 191                 goto out_free;
 192
 193         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 194                 if (path->slots[0]) {
 195                         path->slots[0]--;
 196                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 197                                               path->slots[0]);
 198                         if (key.objectid == bytenr &&
 199                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 200                             key.offset == fs_info->nodesize)
 201                                 ret = 0;
 202                 }
 203         }
 204
 205         if (ret == 0) {
 206                 leaf = path->nodes[0];
 207                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 208                 if (item_size >= sizeof(*ei)) {
 209                         ei = btrfs_item_ptr(leaf, path->slots[0],
 210                                             struct btrfs_extent_item);
 211                         num_refs = btrfs_extent_refs(leaf, ei);
 212                         extent_flags = btrfs_extent_flags(leaf, ei);
 213                 } else {
 214                         ret = -EINVAL;
 215                         btrfs_print_v0_err(fs_info);
 216                         if (trans)
 217                                 btrfs_abort_transaction(trans, ret);
 218                         else
 219                                 btrfs_handle_fs_error(fs_info, ret, NULL);
 220
 221                         goto out_free;
 222                 }
 223
 224                 BUG_ON(num_refs == 0);
 225         } else {
 226                 num_refs = 0;
 227                 extent_flags = 0;
 228                 ret = 0;
 229         }
 230
 231         if (!trans)
 232                 goto out;
 233
 234         delayed_refs = &trans->transaction->delayed_refs;
 235         spin_lock(&delayed_refs->lock);
 236         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
 237         if (head) {
 238                 if (!mutex_trylock(&head->mutex)) {
 239                         refcount_inc(&head->refs);
 240                         spin_unlock(&delayed_refs->lock);
 241
 242                         btrfs_release_path(path);
 243
 244                         /*
 245                          * Mutex was contended, block until it's released and try
 246                          * again
 247                          */
 248                         mutex_lock(&head->mutex);
 249                         mutex_unlock(&head->mutex);
 250                         btrfs_put_delayed_ref_head(head);
 251                         goto search_again;
 252                 }
 253                 spin_lock(&head->lock);
 254                 if (head->extent_op && head->extent_op->update_flags)
 255                         extent_flags |= head->extent_op->flags_to_set;
 256                 else
 257                         BUG_ON(num_refs == 0);
 258
 259                 num_refs += head->ref_mod;
 260                 spin_unlock(&head->lock);
 261                 mutex_unlock(&head->mutex);
 262         }
 263         spin_unlock(&delayed_refs->lock);
 264 out:
 265         WARN_ON(num_refs == 0);
 266         if (refs)
 267                 *refs = num_refs;
 268         if (flags)
 269                 *flags = extent_flags;
 270 out_free:
 271         btrfs_free_path(path);
 272         return ret;
 273 }
 274
 275 /*
 276  * Back reference rules.  Back refs have three main goals:
 277  *
 278  * 1) differentiate between all holders of references to an extent so that
 279  *    when a reference is dropped we can make sure it was a valid reference
 280  *    before freeing the extent.
 281  *
 282  * 2) Provide enough information to quickly find the holders of an extent
 283  *    if we notice a given block is corrupted or bad.
 284  *
 285  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 286  *    maintenance.  This is actually the same as #2, but with a slightly
 287  *    different use case.
 288  *
 289  * There are two kinds of back refs. The implicit back refs is optimized
 290  * for pointers in non-shared tree blocks. For a given pointer in a block,
 291  * back refs of this kind provide information about the block's owner tree
 292  * and the pointer's key. These information allow us to find the block by
 293  * b-tree searching. The full back refs is for pointers in tree blocks not
 294  * referenced by their owner trees. The location of tree block is recorded
 295  * in the back refs. Actually the full back refs is generic, and can be
 296  * used in all cases the implicit back refs is used. The major shortcoming
 297  * of the full back refs is its overhead. Every time a tree block gets
 298  * COWed, we have to update back refs entry for all pointers in it.
 299  *
 300  * For a newly allocated tree block, we use implicit back refs for
 301  * pointers in it. This means most tree related operations only involve
 302  * implicit back refs. For a tree block created in old transaction, the
 303  * only way to drop a reference to it is COW it. So we can detect the
 304  * event that tree block loses its owner tree's reference and do the
 305  * back refs conversion.
 306  *
 307  * When a tree block is COWed through a tree, there are four cases:
 308  *
 309  * The reference count of the block is one and the tree is the block's
 310  * owner tree. Nothing to do in this case.
 311  *
 312  * The reference count of the block is one and the tree is not the
 313  * block's owner tree. In this case, full back refs is used for pointers
 314  * in the block. Remove these full back refs, add implicit back refs for
 315  * every pointers in the new block.
 316  *
 317  * The reference count of the block is greater than one and the tree is
 318  * the block's owner tree. In this case, implicit back refs is used for
 319  * pointers in the block. Add full back refs for every pointers in the
 320  * block, increase lower level extents' reference counts. The original
 321  * implicit back refs are entailed to the new block.
 322  *
 323  * The reference count of the block is greater than one and the tree is
 324  * not the block's owner tree. Add implicit back refs for every pointer in
 325  * the new block, increase lower level extents' reference count.
 326  *
 327  * Back Reference Key composing:
 328  *
 329  * The key objectid corresponds to the first byte in the extent,
 330  * The key type is used to differentiate between types of back refs.
 331  * There are different meanings of the key offset for different types
 332  * of back refs.
 333  *
 334  * File extents can be referenced by:
 335  *
 336  * - multiple snapshots, subvolumes, or different generations in one subvol
 337  * - different files inside a single subvolume
 338  * - different offsets inside a file (bookend extents in file.c)
 339  *
 340  * The extent ref structure for the implicit back refs has fields for:
 341  *
 342  * - Objectid of the subvolume root
 343  * - objectid of the file holding the reference
 344  * - original offset in the file
 345  * - how many bookend extents
 346  *
 347  * The key offset for the implicit back refs is hash of the first
 348  * three fields.
 349  *
 350  * The extent ref structure for the full back refs has field for:
 351  *
 352  * - number of pointers in the tree leaf
 353  *
 354  * The key offset for the implicit back refs is the first byte of
 355  * the tree leaf
 356  *
 357  * When a file extent is allocated, The implicit back refs is used.
 358  * the fields are filled in:
 359  *
 360  *     (root_key.objectid, inode objectid, offset in file, 1)
 361  *
 362  * When a file extent is removed file truncation, we find the
 363  * corresponding implicit back refs and check the following fields:
 364  *
 365  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
 366  *
 367  * Btree extents can be referenced by:
 368  *
 369  * - Different subvolumes
 370  *
 371  * Both the implicit back refs and the full back refs for tree blocks
 372  * only consist of key. The key offset for the implicit back refs is
 373  * objectid of block's owner tree. The key offset for the full back refs
 374  * is the first byte of parent block.
 375  *
 376  * When implicit back refs is used, information about the lowest key and
 377  * level of the tree block are required. These information are stored in
 378  * tree block info structure.
 379  */
 380
 381 /*
 382  * is_data == BTRFS_REF_TYPE_BLOCK, tree block type is required,
 383  * is_data == BTRFS_REF_TYPE_DATA, data type is requiried,
 384  * is_data == BTRFS_REF_TYPE_ANY, either type is OK.
 385  */
 386 int btrfs_get_extent_inline_ref_type(const struct extent_buffer *eb,
 387                                      struct btrfs_extent_inline_ref *iref,
 388                                      enum btrfs_inline_ref_type is_data)
 389 {
 390         int type = btrfs_extent_inline_ref_type(eb, iref);
 391         u64 offset = btrfs_extent_inline_ref_offset(eb, iref);
 392
 393         if (type == BTRFS_TREE_BLOCK_REF_KEY ||
 394             type == BTRFS_SHARED_BLOCK_REF_KEY ||
 395             type == BTRFS_SHARED_DATA_REF_KEY ||
 396             type == BTRFS_EXTENT_DATA_REF_KEY) {
 397                 if (is_data == BTRFS_REF_TYPE_BLOCK) {
 398                         if (type == BTRFS_TREE_BLOCK_REF_KEY)
 399                                 return type;
 400                         if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
 401                                 ASSERT(eb->fs_info);
 402                                 /*
 403                                  * Every shared one has parent tree block,
 404                                  * which must be aligned to sector size.
 405                                  */
 406                                 if (offset &&
 407                                     IS_ALIGNED(offset, eb->fs_info->sectorsize))
 408                                         return type;
 409                         }
 410                 } else if (is_data == BTRFS_REF_TYPE_DATA) {
 411                         if (type == BTRFS_EXTENT_DATA_REF_KEY)
 412                                 return type;
 413                         if (type == BTRFS_SHARED_DATA_REF_KEY) {
 414                                 ASSERT(eb->fs_info);
 415                                 /*
 416                                  * Every shared one has parent tree block,
 417                                  * which must be aligned to sector size.
 418                                  */
 419                                 if (offset &&
 420                                     IS_ALIGNED(offset, eb->fs_info->sectorsize))
 421                                         return type;
 422                         }
 423                 } else {
 424                         ASSERT(is_data == BTRFS_REF_TYPE_ANY);
 425                         return type;
 426                 }
 427         }
 428
 429         btrfs_print_leaf((struct extent_buffer *)eb);
 430         btrfs_err(eb->fs_info,
 431                   "eb %llu iref 0x%lx invalid extent inline ref type %d",
 432                   eb->start, (unsigned long)iref, type);
 433         WARN_ON(1);
 434
 435         return BTRFS_REF_TYPE_INVALID;
 436 }
 437
 438 u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
 439 {
 440         u32 high_crc = ~(u32)0;
 441         u32 low_crc = ~(u32)0;
 442         __le64 lenum;
 443
 444         lenum = cpu_to_le64(root_objectid);
 445         high_crc = btrfs_crc32c(high_crc, &lenum, sizeof(lenum));
 446         lenum = cpu_to_le64(owner);
 447         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
 448         lenum = cpu_to_le64(offset);
 449         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
 450
 451         return ((u64)high_crc << 31) ^ (u64)low_crc;
 452 }
 453
 454 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
 455                                      struct btrfs_extent_data_ref *ref)
 456 {
 457         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
 458                                     btrfs_extent_data_ref_objectid(leaf, ref),
 459                                     btrfs_extent_data_ref_offset(leaf, ref));
 460 }
 461
 462 static int match_extent_data_ref(struct extent_buffer *leaf,
 463                                  struct btrfs_extent_data_ref *ref,
 464                                  u64 root_objectid, u64 owner, u64 offset)
 465 {
 466         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
 467             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
 468             btrfs_extent_data_ref_offset(leaf, ref) != offset)
 469                 return 0;
 470         return 1;
 471 }
 472
 473 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
 474                                            struct btrfs_path *path,
 475                                            u64 bytenr, u64 parent,
 476                                            u64 root_objectid,
 477                                            u64 owner, u64 offset)
 478 {
 479         struct btrfs_root *root = trans->fs_info->extent_root;
 480         struct btrfs_key key;
 481         struct btrfs_extent_data_ref *ref;
 482         struct extent_buffer *leaf;
 483         u32 nritems;
 484         int ret;
 485         int recow;
 486         int err = -ENOENT;
 487
 488         key.objectid = bytenr;
 489         if (parent) {
 490                 key.type = BTRFS_SHARED_DATA_REF_KEY;
 491                 key.offset = parent;
 492         } else {
 493                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
 494                 key.offset = hash_extent_data_ref(root_objectid,
 495                                                   owner, offset);
 496         }
 497 again:
 498         recow = 0;
 499         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
 500         if (ret < 0) {
 501                 err = ret;
 502                 goto fail;
 503         }
 504
 505         if (parent) {
 506                 if (!ret)
 507                         return 0;
 508                 goto fail;
 509         }
 510
 511         leaf = path->nodes[0];
 512         nritems = btrfs_header_nritems(leaf);
 513         while (1) {
 514                 if (path->slots[0] >= nritems) {
 515                         ret = btrfs_next_leaf(root, path);
 516                         if (ret < 0)
 517                                 err = ret;
 518                         if (ret)
 519                                 goto fail;
 520
 521                         leaf = path->nodes[0];
 522                         nritems = btrfs_header_nritems(leaf);
 523                         recow = 1;
 524                 }
 525
 526                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 527                 if (key.objectid != bytenr ||
 528                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
 529                         goto fail;
 530
 531                 ref = btrfs_item_ptr(leaf, path->slots[0],
 532                                      struct btrfs_extent_data_ref);
 533
 534                 if (match_extent_data_ref(leaf, ref, root_objectid,
 535                                           owner, offset)) {
 536                         if (recow) {
 537                                 btrfs_release_path(path);
 538                                 goto again;
 539                         }
 540                         err = 0;
 541                         break;
 542                 }
 543                 path->slots[0]++;
 544         }
 545 fail:
 546         return err;
 547 }
 548
 549 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
 550                                            struct btrfs_path *path,
 551                                            u64 bytenr, u64 parent,
 552                                            u64 root_objectid, u64 owner,
 553                                            u64 offset, int refs_to_add)
 554 {
 555         struct btrfs_root *root = trans->fs_info->extent_root;
 556         struct btrfs_key key;
 557         struct extent_buffer *leaf;
 558         u32 size;
 559         u32 num_refs;
 560         int ret;
 561
 562         key.objectid = bytenr;
 563         if (parent) {
 564                 key.type = BTRFS_SHARED_DATA_REF_KEY;
 565                 key.offset = parent;
 566                 size = sizeof(struct btrfs_shared_data_ref);
 567         } else {
 568                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
 569                 key.offset = hash_extent_data_ref(root_objectid,
 570                                                   owner, offset);
 571                 size = sizeof(struct btrfs_extent_data_ref);
 572         }
 573
 574         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
 575         if (ret && ret != -EEXIST)
 576                 goto fail;
 577
 578         leaf = path->nodes[0];
 579         if (parent) {
 580                 struct btrfs_shared_data_ref *ref;
 581                 ref = btrfs_item_ptr(leaf, path->slots[0],
 582                                      struct btrfs_shared_data_ref);
 583                 if (ret == 0) {
 584                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
 585                 } else {
 586                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
 587                         num_refs += refs_to_add;
 588                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
 589                 }
 590         } else {
 591                 struct btrfs_extent_data_ref *ref;
 592                 while (ret == -EEXIST) {
 593                         ref = btrfs_item_ptr(leaf, path->slots[0],
 594                                              struct btrfs_extent_data_ref);
 595                         if (match_extent_data_ref(leaf, ref, root_objectid,
 596                                                   owner, offset))
 597                                 break;
 598                         btrfs_release_path(path);
 599                         key.offset++;
 600                         ret = btrfs_insert_empty_item(trans, root, path, &key,
 601                                                       size);
 602                         if (ret && ret != -EEXIST)
 603                                 goto fail;
 604
 605                         leaf = path->nodes[0];
 606                 }
 607                 ref = btrfs_item_ptr(leaf, path->slots[0],
 608                                      struct btrfs_extent_data_ref);
 609                 if (ret == 0) {
 610                         btrfs_set_extent_data_ref_root(leaf, ref,
 611                                                        root_objectid);
 612                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
 613                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
 614                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
 615                 } else {
 616                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
 617                         num_refs += refs_to_add;
 618                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
 619                 }
 620         }
 621         btrfs_mark_buffer_dirty(leaf);
 622         ret = 0;
 623 fail:
 624         btrfs_release_path(path);
 625         return ret;
 626 }
 627
 628 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
 629                                            struct btrfs_path *path,
 630                                            int refs_to_drop, int *last_ref)
 631 {
 632         struct btrfs_key key;
 633         struct btrfs_extent_data_ref *ref1 = NULL;
 634         struct btrfs_shared_data_ref *ref2 = NULL;
 635         struct extent_buffer *leaf;
 636         u32 num_refs = 0;
 637         int ret = 0;
 638
 639         leaf = path->nodes[0];
 640         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 641
 642         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
 643                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
 644                                       struct btrfs_extent_data_ref);
 645                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
 646         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
 647                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
 648                                       struct btrfs_shared_data_ref);
 649                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
 650         } else if (unlikely(key.type == BTRFS_EXTENT_REF_V0_KEY)) {
 651                 btrfs_print_v0_err(trans->fs_info);
 652                 btrfs_abort_transaction(trans, -EINVAL);
 653                 return -EINVAL;
 654         } else {
 655                 BUG();
 656         }
 657
 658         BUG_ON(num_refs < refs_to_drop);
 659         num_refs -= refs_to_drop;
 660
 661         if (num_refs == 0) {
 662                 ret = btrfs_del_item(trans, trans->fs_info->extent_root, path);
 663                 *last_ref = 1;
 664         } else {
 665                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
 666                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
 667                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
 668                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
 669                 btrfs_mark_buffer_dirty(leaf);
 670         }
 671         return ret;
 672 }
 673
 674 static noinline u32 extent_data_ref_count(struct btrfs_path *path,
 675                                           struct btrfs_extent_inline_ref *iref)
 676 {
 677         struct btrfs_key key;
 678         struct extent_buffer *leaf;
 679         struct btrfs_extent_data_ref *ref1;
 680         struct btrfs_shared_data_ref *ref2;
 681         u32 num_refs = 0;
 682         int type;
 683
 684         leaf = path->nodes[0];
 685         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 686
 687         BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
 688         if (iref) {
 689                 /*
 690                  * If type is invalid, we should have bailed out earlier than
 691                  * this call.
 692                  */
 693                 type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
 694                 ASSERT(type != BTRFS_REF_TYPE_INVALID);
 695                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
 696                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
 697                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
 698                 } else {
 699                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
 700                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
 701                 }
 702         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
 703                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
 704                                       struct btrfs_extent_data_ref);
 705                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
 706         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
 707                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
 708                                       struct btrfs_shared_data_ref);
 709                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
 710         } else {
 711                 WARN_ON(1);
 712         }
 713         return num_refs;
 714 }
 715
 716 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
 717                                           struct btrfs_path *path,
 718                                           u64 bytenr, u64 parent,
 719                                           u64 root_objectid)
 720 {
 721         struct btrfs_root *root = trans->fs_info->extent_root;
 722         struct btrfs_key key;
 723         int ret;
 724
 725         key.objectid = bytenr;
 726         if (parent) {
 727                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
 728                 key.offset = parent;
 729         } else {
 730                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
 731                 key.offset = root_objectid;
 732         }
 733
 734         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
 735         if (ret > 0)
 736                 ret = -ENOENT;
 737         return ret;
 738 }
 739
 740 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
 741                                           struct btrfs_path *path,
 742                                           u64 bytenr, u64 parent,
 743                                           u64 root_objectid)
 744 {
 745         struct btrfs_key key;
 746         int ret;
 747
 748         key.objectid = bytenr;
 749         if (parent) {
 750                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
 751                 key.offset = parent;
 752         } else {
 753                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
 754                 key.offset = root_objectid;
 755         }
 756
 757         ret = btrfs_insert_empty_item(trans, trans->fs_info->extent_root,
 758                                       path, &key, 0);
 759         btrfs_release_path(path);
 760         return ret;
 761 }
 762
 763 static inline int extent_ref_type(u64 parent, u64 owner)
 764 {
 765         int type;
 766         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
 767                 if (parent > 0)
 768                         type = BTRFS_SHARED_BLOCK_REF_KEY;
 769                 else
 770                         type = BTRFS_TREE_BLOCK_REF_KEY;
 771         } else {
 772                 if (parent > 0)
 773                         type = BTRFS_SHARED_DATA_REF_KEY;
 774                 else
 775                         type = BTRFS_EXTENT_DATA_REF_KEY;
 776         }
 777         return type;
 778 }
 779
 780 static int find_next_key(struct btrfs_path *path, int level,
 781                          struct btrfs_key *key)
 782
 783 {
 784         for (; level < BTRFS_MAX_LEVEL; level++) {
 785                 if (!path->nodes[level])
 786                         break;
 787                 if (path->slots[level] + 1 >=
 788                     btrfs_header_nritems(path->nodes[level]))
 789                         continue;
 790                 if (level == 0)
 791                         btrfs_item_key_to_cpu(path->nodes[level], key,
 792                                               path->slots[level] + 1);
 793                 else
 794                         btrfs_node_key_to_cpu(path->nodes[level], key,
 795                                               path->slots[level] + 1);
 796                 return 0;
 797         }
 798         return 1;
 799 }
 800
 801 /*
 802  * look for inline back ref. if back ref is found, *ref_ret is set
 803  * to the address of inline back ref, and 0 is returned.
 804  *
 805  * if back ref isn't found, *ref_ret is set to the address where it
 806  * should be inserted, and -ENOENT is returned.
 807  *
 808  * if insert is true and there are too many inline back refs, the path
 809  * points to the extent item, and -EAGAIN is returned.
 810  *
 811  * NOTE: inline back refs are ordered in the same way that back ref
 812  *       items in the tree are ordered.
 813  */
 814 static noinline_for_stack
 815 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
 816                                  struct btrfs_path *path,
 817                                  struct btrfs_extent_inline_ref **ref_ret,
 818                                  u64 bytenr, u64 num_bytes,
 819                                  u64 parent, u64 root_objectid,
 820                                  u64 owner, u64 offset, int insert)
 821 {
 822         struct btrfs_fs_info *fs_info = trans->fs_info;
 823         struct btrfs_root *root = fs_info->extent_root;
 824         struct btrfs_key key;
 825         struct extent_buffer *leaf;
 826         struct btrfs_extent_item *ei;
 827         struct btrfs_extent_inline_ref *iref;
 828         u64 flags;
 829         u64 item_size;
 830         unsigned long ptr;
 831         unsigned long end;
 832         int extra_size;
 833         int type;
 834         int want;
 835         int ret;
 836         int err = 0;
 837         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
 838         int needed;
 839
 840         key.objectid = bytenr;
 841         key.type = BTRFS_EXTENT_ITEM_KEY;
 842         key.offset = num_bytes;
 843
 844         want = extent_ref_type(parent, owner);
 845         if (insert) {
 846                 extra_size = btrfs_extent_inline_ref_size(want);
 847                 path->keep_locks = 1;
 848         } else
 849                 extra_size = -1;
 850
 851         /*
 852          * Owner is our level, so we can just add one to get the level for the
 853          * block we are interested in.
 854          */
 855         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
 856                 key.type = BTRFS_METADATA_ITEM_KEY;
 857                 key.offset = owner;
 858         }
 859
 860 again:
 861         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
 862         if (ret < 0) {
 863                 err = ret;
 864                 goto out;
 865         }
 866
 867         /*
 868          * We may be a newly converted file system which still has the old fat
 869          * extent entries for metadata, so try and see if we have one of those.
 870          */
 871         if (ret > 0 && skinny_metadata) {
 872                 skinny_metadata = false;
 873                 if (path->slots[0]) {
 874                         path->slots[0]--;
 875                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 876                                               path->slots[0]);
 877                         if (key.objectid == bytenr &&
 878                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 879                             key.offset == num_bytes)
 880                                 ret = 0;
 881                 }
 882                 if (ret) {
 883                         key.objectid = bytenr;
 884                         key.type = BTRFS_EXTENT_ITEM_KEY;
 885                         key.offset = num_bytes;
 886                         btrfs_release_path(path);
 887                         goto again;
 888                 }
 889         }
 890
 891         if (ret && !insert) {
 892                 err = -ENOENT;
 893                 goto out;
 894         } else if (WARN_ON(ret)) {
 895                 err = -EIO;
 896                 goto out;
 897         }
 898
 899         leaf = path->nodes[0];
 900         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 901         if (unlikely(item_size < sizeof(*ei))) {
 902                 err = -EINVAL;
 903                 btrfs_print_v0_err(fs_info);
 904                 btrfs_abort_transaction(trans, err);
 905                 goto out;
 906         }
 907
 908         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
 909         flags = btrfs_extent_flags(leaf, ei);
 910
 911         ptr = (unsigned long)(ei + 1);
 912         end = (unsigned long)ei + item_size;
 913
 914         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
 915                 ptr += sizeof(struct btrfs_tree_block_info);
 916                 BUG_ON(ptr > end);
 917         }
 918
 919         if (owner >= BTRFS_FIRST_FREE_OBJECTID)
 920                 needed = BTRFS_REF_TYPE_DATA;
 921         else
 922                 needed = BTRFS_REF_TYPE_BLOCK;
 923
 924         err = -ENOENT;
 925         while (1) {
 926                 if (ptr >= end) {
 927                         WARN_ON(ptr > end);
 928                         break;
 929                 }
 930                 iref = (struct btrfs_extent_inline_ref *)ptr;
 931                 type = btrfs_get_extent_inline_ref_type(leaf, iref, needed);
 932                 if (type == BTRFS_REF_TYPE_INVALID) {
 933                         err = -EUCLEAN;
 934                         goto out;
 935                 }
 936
 937                 if (want < type)
 938                         break;
 939                 if (want > type) {
 940                         ptr += btrfs_extent_inline_ref_size(type);
 941                         continue;
 942                 }
 943
 944                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
 945                         struct btrfs_extent_data_ref *dref;
 946                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
 947                         if (match_extent_data_ref(leaf, dref, root_objectid,
 948                                                   owner, offset)) {
 949                                 err = 0;
 950                                 break;
 951                         }
 952                         if (hash_extent_data_ref_item(leaf, dref) <
 953                             hash_extent_data_ref(root_objectid, owner, offset))
 954                                 break;
 955                 } else {
 956                         u64 ref_offset;
 957                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
 958                         if (parent > 0) {
 959                                 if (parent == ref_offset) {
 960                                         err = 0;
 961                                         break;
 962                                 }
 963                                 if (ref_offset < parent)
 964                                         break;
 965                         } else {
 966                                 if (root_objectid == ref_offset) {
 967                                         err = 0;
 968                                         break;
 969                                 }
 970                                 if (ref_offset < root_objectid)
 971                                         break;
 972                         }
 973                 }
 974                 ptr += btrfs_extent_inline_ref_size(type);
 975         }
 976         if (err == -ENOENT && insert) {
 977                 if (item_size + extra_size >=
 978                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
 979                         err = -EAGAIN;
 980                         goto out;
 981                 }
 982                 /*
 983                  * To add new inline back ref, we have to make sure
 984                  * there is no corresponding back ref item.
 985                  * For simplicity, we just do not add new inline back
 986                  * ref if there is any kind of item for this block
 987                  */
 988                 if (find_next_key(path, 0, &key) == 0 &&
 989                     key.objectid == bytenr &&
 990                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
 991                         err = -EAGAIN;
 992                         goto out;
 993                 }
 994         }
 995         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
 996 out:
 997         if (insert) {
 998                 path->keep_locks = 0;
 999                 btrfs_unlock_up_safe(path, 1);
1000         }
1001         return err;
1002 }
1003
1004 /*
1005  * helper to add new inline back ref
1006  */
1007 static noinline_for_stack
1008 void setup_inline_extent_backref(struct btrfs_fs_info *fs_info,
1009                                  struct btrfs_path *path,
1010                                  struct btrfs_extent_inline_ref *iref,
1011                                  u64 parent, u64 root_objectid,
1012                                  u64 owner, u64 offset, int refs_to_add,
1013                                  struct btrfs_delayed_extent_op *extent_op)
1014 {
1015         struct extent_buffer *leaf;
1016         struct btrfs_extent_item *ei;
1017         unsigned long ptr;
1018         unsigned long end;
1019         unsigned long item_offset;
1020         u64 refs;
1021         int size;
1022         int type;
1023
1024         leaf = path->nodes[0];
1025         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1026         item_offset = (unsigned long)iref - (unsigned long)ei;
1027
1028         type = extent_ref_type(parent, owner);
1029         size = btrfs_extent_inline_ref_size(type);
1030
1031         btrfs_extend_item(path, size);
1032
1033         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1034         refs = btrfs_extent_refs(leaf, ei);
1035         refs += refs_to_add;
1036         btrfs_set_extent_refs(leaf, ei, refs);
1037         if (extent_op)
1038                 __run_delayed_extent_op(extent_op, leaf, ei);
1039
1040         ptr = (unsigned long)ei + item_offset;
1041         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1042         if (ptr < end - size)
1043                 memmove_extent_buffer(leaf, ptr + size, ptr,
1044                                       end - size - ptr);
1045
1046         iref = (struct btrfs_extent_inline_ref *)ptr;
1047         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1048         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1049                 struct btrfs_extent_data_ref *dref;
1050                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1051                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1052                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1053                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1054                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1055         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1056                 struct btrfs_shared_data_ref *sref;
1057                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1058                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1059                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1060         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1061                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1062         } else {
1063                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1064         }
1065         btrfs_mark_buffer_dirty(leaf);
1066 }
1067
1068 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1069                                  struct btrfs_path *path,
1070                                  struct btrfs_extent_inline_ref **ref_ret,
1071                                  u64 bytenr, u64 num_bytes, u64 parent,
1072                                  u64 root_objectid, u64 owner, u64 offset)
1073 {
1074         int ret;
1075
1076         ret = lookup_inline_extent_backref(trans, path, ref_ret, bytenr,
1077                                            num_bytes, parent, root_objectid,
1078                                            owner, offset, 0);
1079         if (ret != -ENOENT)
1080                 return ret;
1081
1082         btrfs_release_path(path);
1083         *ref_ret = NULL;
1084
1085         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1086                 ret = lookup_tree_block_ref(trans, path, bytenr, parent,
1087                                             root_objectid);
1088         } else {
1089                 ret = lookup_extent_data_ref(trans, path, bytenr, parent,
1090                                              root_objectid, owner, offset);
1091         }
1092         return ret;
1093 }
1094
1095 /*
1096  * helper to update/remove inline back ref
1097  */
1098 static noinline_for_stack
1099 void update_inline_extent_backref(struct btrfs_path *path,
1100                                   struct btrfs_extent_inline_ref *iref,
1101                                   int refs_to_mod,
1102                                   struct btrfs_delayed_extent_op *extent_op,
1103                                   int *last_ref)
1104 {
1105         struct extent_buffer *leaf = path->nodes[0];
1106         struct btrfs_extent_item *ei;
1107         struct btrfs_extent_data_ref *dref = NULL;
1108         struct btrfs_shared_data_ref *sref = NULL;
1109         unsigned long ptr;
1110         unsigned long end;
1111         u32 item_size;
1112         int size;
1113         int type;
1114         u64 refs;
1115
1116         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1117         refs = btrfs_extent_refs(leaf, ei);
1118         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1119         refs += refs_to_mod;
1120         btrfs_set_extent_refs(leaf, ei, refs);
1121         if (extent_op)
1122                 __run_delayed_extent_op(extent_op, leaf, ei);
1123
1124         /*
1125          * If type is invalid, we should have bailed out after
1126          * lookup_inline_extent_backref().
1127          */
1128         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_ANY);
1129         ASSERT(type != BTRFS_REF_TYPE_INVALID);
1130
1131         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1132                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1133                 refs = btrfs_extent_data_ref_count(leaf, dref);
1134         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1135                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1136                 refs = btrfs_shared_data_ref_count(leaf, sref);
1137         } else {
1138                 refs = 1;
1139                 BUG_ON(refs_to_mod != -1);
1140         }
1141
1142         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1143         refs += refs_to_mod;
1144
1145         if (refs > 0) {
1146                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1147                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1148                 else
1149                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1150         } else {
1151                 *last_ref = 1;
1152                 size =  btrfs_extent_inline_ref_size(type);
1153                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1154                 ptr = (unsigned long)iref;
1155                 end = (unsigned long)ei + item_size;
1156                 if (ptr + size < end)
1157                         memmove_extent_buffer(leaf, ptr, ptr + size,
1158                                               end - ptr - size);
1159                 item_size -= size;
1160                 btrfs_truncate_item(path, item_size, 1);
1161         }
1162         btrfs_mark_buffer_dirty(leaf);
1163 }
1164
1165 static noinline_for_stack
1166 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1167                                  struct btrfs_path *path,
1168                                  u64 bytenr, u64 num_bytes, u64 parent,
1169                                  u64 root_objectid, u64 owner,
1170                                  u64 offset, int refs_to_add,
1171                                  struct btrfs_delayed_extent_op *extent_op)
1172 {
1173         struct btrfs_extent_inline_ref *iref;
1174         int ret;
1175
1176         ret = lookup_inline_extent_backref(trans, path, &iref, bytenr,
1177                                            num_bytes, parent, root_objectid,
1178                                            owner, offset, 1);
1179         if (ret == 0) {
1180                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1181                 update_inline_extent_backref(path, iref, refs_to_add,
1182                                              extent_op, NULL);
1183         } else if (ret == -ENOENT) {
1184                 setup_inline_extent_backref(trans->fs_info, path, iref, parent,
1185                                             root_objectid, owner, offset,
1186                                             refs_to_add, extent_op);
1187                 ret = 0;
1188         }
1189         return ret;
1190 }
1191
1192 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1193                                  struct btrfs_path *path,
1194                                  struct btrfs_extent_inline_ref *iref,
1195                                  int refs_to_drop, int is_data, int *last_ref)
1196 {
1197         int ret = 0;
1198
1199         BUG_ON(!is_data && refs_to_drop != 1);
1200         if (iref) {
1201                 update_inline_extent_backref(path, iref, -refs_to_drop, NULL,
1202                                              last_ref);
1203         } else if (is_data) {
1204                 ret = remove_extent_data_ref(trans, path, refs_to_drop,
1205                                              last_ref);
1206         } else {
1207                 *last_ref = 1;
1208                 ret = btrfs_del_item(trans, trans->fs_info->extent_root, path);
1209         }
1210         return ret;
1211 }
1212
1213 static int btrfs_issue_discard(struct block_device *bdev, u64 start, u64 len,
1214                                u64 *discarded_bytes)
1215 {
1216         int j, ret = 0;
1217         u64 bytes_left, end;
1218         u64 aligned_start = ALIGN(start, 1 << 9);
1219
1220         if (WARN_ON(start != aligned_start)) {
1221                 len -= aligned_start - start;
1222                 len = round_down(len, 1 << 9);
1223                 start = aligned_start;
1224         }
1225
1226         *discarded_bytes = 0;
1227
1228         if (!len)
1229                 return 0;
1230
1231         end = start + len;
1232         bytes_left = len;
1233
1234         /* Skip any superblocks on this device. */
1235         for (j = 0; j < BTRFS_SUPER_MIRROR_MAX; j++) {
1236                 u64 sb_start = btrfs_sb_offset(j);
1237                 u64 sb_end = sb_start + BTRFS_SUPER_INFO_SIZE;
1238                 u64 size = sb_start - start;
1239
1240                 if (!in_range(sb_start, start, bytes_left) &&
1241                     !in_range(sb_end, start, bytes_left) &&
1242                     !in_range(start, sb_start, BTRFS_SUPER_INFO_SIZE))
1243                         continue;
1244
1245                 /*
1246                  * Superblock spans beginning of range.  Adjust start and
1247                  * try again.
1248                  */
1249                 if (sb_start <= start) {
1250                         start += sb_end - start;
1251                         if (start > end) {
1252                                 bytes_left = 0;
1253                                 break;
1254                         }
1255                         bytes_left = end - start;
1256                         continue;
1257                 }
1258
1259                 if (size) {
1260                         ret = blkdev_issue_discard(bdev, start >> 9, size >> 9,
1261                                                    GFP_NOFS, 0);
1262                         if (!ret)
1263                                 *discarded_bytes += size;
1264                         else if (ret != -EOPNOTSUPP)
1265                                 return ret;
1266                 }
1267
1268                 start = sb_end;
1269                 if (start > end) {
1270                         bytes_left = 0;
1271                         break;
1272                 }
1273                 bytes_left = end - start;
1274         }
1275
1276         if (bytes_left) {
1277                 ret = blkdev_issue_discard(bdev, start >> 9, bytes_left >> 9,
1278                                            GFP_NOFS, 0);
1279                 if (!ret)
1280                         *discarded_bytes += bytes_left;
1281         }
1282         return ret;
1283 }
1284
1285 int btrfs_discard_extent(struct btrfs_fs_info *fs_info, u64 bytenr,
1286                          u64 num_bytes, u64 *actual_bytes)
1287 {
1288         int ret = 0;
1289         u64 discarded_bytes = 0;
1290         u64 end = bytenr + num_bytes;
1291         u64 cur = bytenr;
1292         struct btrfs_bio *bbio = NULL;
1293
1294
1295         /*
1296          * Avoid races with device replace and make sure our bbio has devices
1297          * associated to its stripes that don't go away while we are discarding.
1298          */
1299         btrfs_bio_counter_inc_blocked(fs_info);
1300         while (cur < end) {
1301                 struct btrfs_bio_stripe *stripe;
1302                 int i;
1303
1304                 num_bytes = end - cur;
1305                 /* Tell the block device(s) that the sectors can be discarded */
1306                 ret = btrfs_map_block(fs_info, BTRFS_MAP_DISCARD, cur,
1307                                       &num_bytes, &bbio, 0);
1308                 /*
1309                  * Error can be -ENOMEM, -ENOENT (no such chunk mapping) or
1310                  * -EOPNOTSUPP. For any such error, @num_bytes is not updated,
1311                  * thus we can't continue anyway.
1312                  */
1313                 if (ret < 0)
1314                         goto out;
1315
1316                 stripe = bbio->stripes;
1317                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
1318                         u64 bytes;
1319                         struct request_queue *req_q;
1320
1321                         if (!stripe->dev->bdev) {
1322                                 ASSERT(btrfs_test_opt(fs_info, DEGRADED));
1323                                 continue;
1324                         }
1325                         req_q = bdev_get_queue(stripe->dev->bdev);
1326                         if (!blk_queue_discard(req_q))
1327                                 continue;
1328
1329                         ret = btrfs_issue_discard(stripe->dev->bdev,
1330                                                   stripe->physical,
1331                                                   stripe->length,
1332                                                   &bytes);
1333                         if (!ret) {
1334                                 discarded_bytes += bytes;
1335                         } else if (ret != -EOPNOTSUPP) {
1336                                 /*
1337                                  * Logic errors or -ENOMEM, or -EIO, but
1338                                  * unlikely to happen.
1339                                  *
1340                                  * And since there are two loops, explicitly
1341                                  * go to out to avoid confusion.
1342                                  */
1343                                 btrfs_put_bbio(bbio);
1344                                 goto out;
1345                         }
1346
1347                         /*
1348                          * Just in case we get back EOPNOTSUPP for some reason,
1349                          * just ignore the return value so we don't screw up
1350                          * people calling discard_extent.
1351                          */
1352                         ret = 0;
1353                 }
1354                 btrfs_put_bbio(bbio);
1355                 cur += num_bytes;
1356         }
1357 out:
1358         btrfs_bio_counter_dec(fs_info);
1359
1360         if (actual_bytes)
1361                 *actual_bytes = discarded_bytes;
1362
1363
1364         if (ret == -EOPNOTSUPP)
1365                 ret = 0;
1366         return ret;
1367 }
1368
1369 /* Can return -ENOMEM */
1370 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1371                          struct btrfs_ref *generic_ref)
1372 {
1373         struct btrfs_fs_info *fs_info = trans->fs_info;
1374         int old_ref_mod, new_ref_mod;
1375         int ret;
1376
1377         ASSERT(generic_ref->type != BTRFS_REF_NOT_SET &&
1378                generic_ref->action);
1379         BUG_ON(generic_ref->type == BTRFS_REF_METADATA &&
1380                generic_ref->tree_ref.root == BTRFS_TREE_LOG_OBJECTID);
1381
1382         if (generic_ref->type == BTRFS_REF_METADATA)
1383                 ret = btrfs_add_delayed_tree_ref(trans, generic_ref,
1384                                 NULL, &old_ref_mod, &new_ref_mod);
1385         else
1386                 ret = btrfs_add_delayed_data_ref(trans, generic_ref, 0,
1387                                                  &old_ref_mod, &new_ref_mod);
1388
1389         btrfs_ref_tree_mod(fs_info, generic_ref);
1390
1391         if (ret == 0 && old_ref_mod < 0 && new_ref_mod >= 0)
1392                 sub_pinned_bytes(fs_info, generic_ref);
1393
1394         return ret;
1395 }
1396
1397 /*
1398  * __btrfs_inc_extent_ref - insert backreference for a given extent
1399  *
1400  * @trans:          Handle of transaction
1401  *
1402  * @node:           The delayed ref node used to get the bytenr/length for
1403  *                  extent whose references are incremented.
1404  *
1405  * @parent:         If this is a shared extent (BTRFS_SHARED_DATA_REF_KEY/
1406  *                  BTRFS_SHARED_BLOCK_REF_KEY) then it holds the logical
1407  *                  bytenr of the parent block. Since new extents are always
1408  *                  created with indirect references, this will only be the case
1409  *                  when relocating a shared extent. In that case, root_objectid
1410  *                  will be BTRFS_TREE_RELOC_OBJECTID. Otheriwse, parent must
1411  *                  be 0
1412  *
1413  * @root_objectid:  The id of the root where this modification has originated,
1414  *                  this can be either one of the well-known metadata trees or
1415  *                  the subvolume id which references this extent.
1416  *
1417  * @owner:          For data extents it is the inode number of the owning file.
1418  *                  For metadata extents this parameter holds the level in the
1419  *                  tree of the extent.
1420  *
1421  * @offset:         For metadata extents the offset is ignored and is currently
1422  *                  always passed as 0. For data extents it is the fileoffset
1423  *                  this extent belongs to.
1424  *
1425  * @refs_to_add     Number of references to add
1426  *
1427  * @extent_op       Pointer to a structure, holding information necessary when
1428  *                  updating a tree block's flags
1429  *
1430  */
1431 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1432                                   struct btrfs_delayed_ref_node *node,
1433                                   u64 parent, u64 root_objectid,
1434                                   u64 owner, u64 offset, int refs_to_add,
1435                                   struct btrfs_delayed_extent_op *extent_op)
1436 {
1437         struct btrfs_path *path;
1438         struct extent_buffer *leaf;
1439         struct btrfs_extent_item *item;
1440         struct btrfs_key key;
1441         u64 bytenr = node->bytenr;
1442         u64 num_bytes = node->num_bytes;
1443         u64 refs;
1444         int ret;
1445
1446         path = btrfs_alloc_path();
1447         if (!path)
1448                 return -ENOMEM;
1449
1450         path->leave_spinning = 1;
1451         /* this will setup the path even if it fails to insert the back ref */
1452         ret = insert_inline_extent_backref(trans, path, bytenr, num_bytes,
1453                                            parent, root_objectid, owner,
1454                                            offset, refs_to_add, extent_op);
1455         if ((ret < 0 && ret != -EAGAIN) || !ret)
1456                 goto out;
1457
1458         /*
1459          * Ok we had -EAGAIN which means we didn't have space to insert and
1460          * inline extent ref, so just update the reference count and add a
1461          * normal backref.
1462          */
1463         leaf = path->nodes[0];
1464         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1465         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1466         refs = btrfs_extent_refs(leaf, item);
1467         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
1468         if (extent_op)
1469                 __run_delayed_extent_op(extent_op, leaf, item);
1470
1471         btrfs_mark_buffer_dirty(leaf);
1472         btrfs_release_path(path);
1473
1474         path->leave_spinning = 1;
1475         /* now insert the actual backref */
1476         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1477                 BUG_ON(refs_to_add != 1);
1478                 ret = insert_tree_block_ref(trans, path, bytenr, parent,
1479                                             root_objectid);
1480         } else {
1481                 ret = insert_extent_data_ref(trans, path, bytenr, parent,
1482                                              root_objectid, owner, offset,
1483                                              refs_to_add);
1484         }
1485         if (ret)
1486                 btrfs_abort_transaction(trans, ret);
1487 out:
1488         btrfs_free_path(path);
1489         return ret;
1490 }
1491
1492 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
1493                                 struct btrfs_delayed_ref_node *node,
1494                                 struct btrfs_delayed_extent_op *extent_op,
1495                                 int insert_reserved)
1496 {
1497         int ret = 0;
1498         struct btrfs_delayed_data_ref *ref;
1499         struct btrfs_key ins;
1500         u64 parent = 0;
1501         u64 ref_root = 0;
1502         u64 flags = 0;
1503
1504         ins.objectid = node->bytenr;
1505         ins.offset = node->num_bytes;
1506         ins.type = BTRFS_EXTENT_ITEM_KEY;
1507
1508         ref = btrfs_delayed_node_to_data_ref(node);
1509         trace_run_delayed_data_ref(trans->fs_info, node, ref, node->action);
1510
1511         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
1512                 parent = ref->parent;
1513         ref_root = ref->root;
1514
1515         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1516                 if (extent_op)
1517                         flags |= extent_op->flags_to_set;
1518                 ret = alloc_reserved_file_extent(trans, parent, ref_root,
1519                                                  flags, ref->objectid,
1520                                                  ref->offset, &ins,
1521                                                  node->ref_mod);
1522         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1523                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
1524                                              ref->objectid, ref->offset,
1525                                              node->ref_mod, extent_op);
1526         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1527                 ret = __btrfs_free_extent(trans, node, parent,
1528                                           ref_root, ref->objectid,
1529                                           ref->offset, node->ref_mod,
1530                                           extent_op);
1531         } else {
1532                 BUG();
1533         }
1534         return ret;
1535 }
1536
1537 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
1538                                     struct extent_buffer *leaf,
1539                                     struct btrfs_extent_item *ei)
1540 {
1541         u64 flags = btrfs_extent_flags(leaf, ei);
1542         if (extent_op->update_flags) {
1543                 flags |= extent_op->flags_to_set;
1544                 btrfs_set_extent_flags(leaf, ei, flags);
1545         }
1546
1547         if (extent_op->update_key) {
1548                 struct btrfs_tree_block_info *bi;
1549                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
1550                 bi = (struct btrfs_tree_block_info *)(ei + 1);
1551                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
1552         }
1553 }
1554
1555 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
1556                                  struct btrfs_delayed_ref_head *head,
1557                                  struct btrfs_delayed_extent_op *extent_op)
1558 {
1559         struct btrfs_fs_info *fs_info = trans->fs_info;
1560         struct btrfs_key key;
1561         struct btrfs_path *path;
1562         struct btrfs_extent_item *ei;
1563         struct extent_buffer *leaf;
1564         u32 item_size;
1565         int ret;
1566         int err = 0;
1567         int metadata = !extent_op->is_data;
1568
1569         if (TRANS_ABORTED(trans))
1570                 return 0;
1571
1572         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1573                 metadata = 0;
1574
1575         path = btrfs_alloc_path();
1576         if (!path)
1577                 return -ENOMEM;
1578
1579         key.objectid = head->bytenr;
1580
1581         if (metadata) {
1582                 key.type = BTRFS_METADATA_ITEM_KEY;
1583                 key.offset = extent_op->level;
1584         } else {
1585                 key.type = BTRFS_EXTENT_ITEM_KEY;
1586                 key.offset = head->num_bytes;
1587         }
1588
1589 again:
1590         path->leave_spinning = 1;
1591         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 1);
1592         if (ret < 0) {
1593                 err = ret;
1594                 goto out;
1595         }
1596         if (ret > 0) {
1597                 if (metadata) {
1598                         if (path->slots[0] > 0) {
1599                                 path->slots[0]--;
1600                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
1601                                                       path->slots[0]);
1602                                 if (key.objectid == head->bytenr &&
1603                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
1604                                     key.offset == head->num_bytes)
1605                                         ret = 0;
1606                         }
1607                         if (ret > 0) {
1608                                 btrfs_release_path(path);
1609                                 metadata = 0;
1610
1611                                 key.objectid = head->bytenr;
1612                                 key.offset = head->num_bytes;
1613                                 key.type = BTRFS_EXTENT_ITEM_KEY;
1614                                 goto again;
1615                         }
1616                 } else {
1617                         err = -EIO;
1618                         goto out;
1619                 }
1620         }
1621
1622         leaf = path->nodes[0];
1623         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1624
1625         if (unlikely(item_size < sizeof(*ei))) {
1626                 err = -EINVAL;
1627                 btrfs_print_v0_err(fs_info);
1628                 btrfs_abort_transaction(trans, err);
1629                 goto out;
1630         }
1631
1632         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1633         __run_delayed_extent_op(extent_op, leaf, ei);
1634
1635         btrfs_mark_buffer_dirty(leaf);
1636 out:
1637         btrfs_free_path(path);
1638         return err;
1639 }
1640
1641 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
1642                                 struct btrfs_delayed_ref_node *node,
1643                                 struct btrfs_delayed_extent_op *extent_op,
1644                                 int insert_reserved)
1645 {
1646         int ret = 0;
1647         struct btrfs_delayed_tree_ref *ref;
1648         u64 parent = 0;
1649         u64 ref_root = 0;
1650
1651         ref = btrfs_delayed_node_to_tree_ref(node);
1652         trace_run_delayed_tree_ref(trans->fs_info, node, ref, node->action);
1653
1654         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
1655                 parent = ref->parent;
1656         ref_root = ref->root;
1657
1658         if (node->ref_mod != 1) {
1659                 btrfs_err(trans->fs_info,
1660         "btree block(%llu) has %d references rather than 1: action %d ref_root %llu parent %llu",
1661                           node->bytenr, node->ref_mod, node->action, ref_root,
1662                           parent);
1663                 return -EIO;
1664         }
1665         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1666                 BUG_ON(!extent_op || !extent_op->update_flags);
1667                 ret = alloc_reserved_tree_block(trans, node, extent_op);
1668         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1669                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
1670                                              ref->level, 0, 1, extent_op);
1671         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1672                 ret = __btrfs_free_extent(trans, node, parent, ref_root,
1673                                           ref->level, 0, 1, extent_op);
1674         } else {
1675                 BUG();
1676         }
1677         return ret;
1678 }
1679
1680 /* helper function to actually process a single delayed ref entry */
1681 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
1682                                struct btrfs_delayed_ref_node *node,
1683                                struct btrfs_delayed_extent_op *extent_op,
1684                                int insert_reserved)
1685 {
1686         int ret = 0;
1687
1688         if (TRANS_ABORTED(trans)) {
1689                 if (insert_reserved)
1690                         btrfs_pin_extent(trans, node->bytenr, node->num_bytes, 1);
1691                 return 0;
1692         }
1693
1694         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
1695             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
1696                 ret = run_delayed_tree_ref(trans, node, extent_op,
1697                                            insert_reserved);
1698         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
1699                  node->type == BTRFS_SHARED_DATA_REF_KEY)
1700                 ret = run_delayed_data_ref(trans, node, extent_op,
1701                                            insert_reserved);
1702         else
1703                 BUG();
1704         if (ret && insert_reserved)
1705                 btrfs_pin_extent(trans, node->bytenr, node->num_bytes, 1);
1706         return ret;
1707 }
1708
1709 static inline struct btrfs_delayed_ref_node *
1710 select_delayed_ref(struct btrfs_delayed_ref_head *head)
1711 {
1712         struct btrfs_delayed_ref_node *ref;
1713
1714         if (RB_EMPTY_ROOT(&head->ref_tree.rb_root))
1715                 return NULL;
1716
1717         /*
1718          * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
1719          * This is to prevent a ref count from going down to zero, which deletes
1720          * the extent item from the extent tree, when there still are references
1721          * to add, which would fail because they would not find the extent item.
1722          */
1723         if (!list_empty(&head->ref_add_list))
1724                 return list_first_entry(&head->ref_add_list,
1725                                 struct btrfs_delayed_ref_node, add_list);
1726
1727         ref = rb_entry(rb_first_cached(&head->ref_tree),
1728                        struct btrfs_delayed_ref_node, ref_node);
1729         ASSERT(list_empty(&ref->add_list));
1730         return ref;
1731 }
1732
1733 static void unselect_delayed_ref_head(struct btrfs_delayed_ref_root *delayed_refs,
1734                                       struct btrfs_delayed_ref_head *head)
1735 {
1736         spin_lock(&delayed_refs->lock);
1737         head->processing = 0;
1738         delayed_refs->num_heads_ready++;
1739         spin_unlock(&delayed_refs->lock);
1740         btrfs_delayed_ref_unlock(head);
1741 }
1742
1743 static struct btrfs_delayed_extent_op *cleanup_extent_op(
1744                                 struct btrfs_delayed_ref_head *head)
1745 {
1746         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
1747
1748         if (!extent_op)
1749                 return NULL;
1750
1751         if (head->must_insert_reserved) {
1752                 head->extent_op = NULL;
1753                 btrfs_free_delayed_extent_op(extent_op);
1754                 return NULL;
1755         }
1756         return extent_op;
1757 }
1758
1759 static int run_and_cleanup_extent_op(struct btrfs_trans_handle *trans,
1760                                      struct btrfs_delayed_ref_head *head)
1761 {
1762         struct btrfs_delayed_extent_op *extent_op;
1763         int ret;
1764
1765         extent_op = cleanup_extent_op(head);
1766         if (!extent_op)
1767                 return 0;
1768         head->extent_op = NULL;
1769         spin_unlock(&head->lock);
1770         ret = run_delayed_extent_op(trans, head, extent_op);
1771         btrfs_free_delayed_extent_op(extent_op);
1772         return ret ? ret : 1;
1773 }
1774
1775 void btrfs_cleanup_ref_head_accounting(struct btrfs_fs_info *fs_info,
1776                                   struct btrfs_delayed_ref_root *delayed_refs,
1777                                   struct btrfs_delayed_ref_head *head)
1778 {
1779         int nr_items = 1;       /* Dropping this ref head update. */
1780
1781         if (head->total_ref_mod < 0) {
1782                 struct btrfs_space_info *space_info;
1783                 u64 flags;
1784
1785                 if (head->is_data)
1786                         flags = BTRFS_BLOCK_GROUP_DATA;
1787                 else if (head->is_system)
1788                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
1789                 else
1790                         flags = BTRFS_BLOCK_GROUP_METADATA;
1791                 space_info = btrfs_find_space_info(fs_info, flags);
1792                 ASSERT(space_info);
1793                 percpu_counter_add_batch(&space_info->total_bytes_pinned,
1794                                    -head->num_bytes,
1795                                    BTRFS_TOTAL_BYTES_PINNED_BATCH);
1796
1797                 /*
1798                  * We had csum deletions accounted for in our delayed refs rsv,
1799                  * we need to drop the csum leaves for this update from our
1800                  * delayed_refs_rsv.
1801                  */
1802                 if (head->is_data) {
1803                         spin_lock(&delayed_refs->lock);
1804                         delayed_refs->pending_csums -= head->num_bytes;
1805                         spin_unlock(&delayed_refs->lock);
1806                         nr_items += btrfs_csum_bytes_to_leaves(fs_info,
1807                                 head->num_bytes);
1808                 }
1809         }
1810
1811         btrfs_delayed_refs_rsv_release(fs_info, nr_items);
1812 }
1813
1814 static int cleanup_ref_head(struct btrfs_trans_handle *trans,
1815                             struct btrfs_delayed_ref_head *head)
1816 {
1817
1818         struct btrfs_fs_info *fs_info = trans->fs_info;
1819         struct btrfs_delayed_ref_root *delayed_refs;
1820         int ret;
1821
1822         delayed_refs = &trans->transaction->delayed_refs;
1823
1824         ret = run_and_cleanup_extent_op(trans, head);
1825         if (ret < 0) {
1826                 unselect_delayed_ref_head(delayed_refs, head);
1827                 btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
1828                 return ret;
1829         } else if (ret) {
1830                 return ret;
1831         }
1832
1833         /*
1834          * Need to drop our head ref lock and re-acquire the delayed ref lock
1835          * and then re-check to make sure nobody got added.
1836          */
1837         spin_unlock(&head->lock);
1838         spin_lock(&delayed_refs->lock);
1839         spin_lock(&head->lock);
1840         if (!RB_EMPTY_ROOT(&head->ref_tree.rb_root) || head->extent_op) {
1841                 spin_unlock(&head->lock);
1842                 spin_unlock(&delayed_refs->lock);
1843                 return 1;
1844         }
1845         btrfs_delete_ref_head(delayed_refs, head);
1846         spin_unlock(&head->lock);
1847         spin_unlock(&delayed_refs->lock);
1848
1849         if (head->must_insert_reserved) {
1850                 btrfs_pin_extent(trans, head->bytenr, head->num_bytes, 1);
1851                 if (head->is_data) {
1852                         ret = btrfs_del_csums(trans, fs_info->csum_root,
1853                                               head->bytenr, head->num_bytes);
1854                 }
1855         }
1856
1857         btrfs_cleanup_ref_head_accounting(fs_info, delayed_refs, head);
1858
1859         trace_run_delayed_ref_head(fs_info, head, 0);
1860         btrfs_delayed_ref_unlock(head);
1861         btrfs_put_delayed_ref_head(head);
1862         return 0;
1863 }
1864
1865 static struct btrfs_delayed_ref_head *btrfs_obtain_ref_head(
1866                                         struct btrfs_trans_handle *trans)
1867 {
1868         struct btrfs_delayed_ref_root *delayed_refs =
1869                 &trans->transaction->delayed_refs;
1870         struct btrfs_delayed_ref_head *head = NULL;
1871         int ret;
1872
1873         spin_lock(&delayed_refs->lock);
1874         head = btrfs_select_ref_head(delayed_refs);
1875         if (!head) {
1876                 spin_unlock(&delayed_refs->lock);
1877                 return head;
1878         }
1879
1880         /*
1881          * Grab the lock that says we are going to process all the refs for
1882          * this head
1883          */
1884         ret = btrfs_delayed_ref_lock(delayed_refs, head);
1885         spin_unlock(&delayed_refs->lock);
1886
1887         /*
1888          * We may have dropped the spin lock to get the head mutex lock, and
1889          * that might have given someone else time to free the head.  If that's
1890          * true, it has been removed from our list and we can move on.
1891          */
1892         if (ret == -EAGAIN)
1893                 head = ERR_PTR(-EAGAIN);
1894
1895         return head;
1896 }
1897
1898 static int btrfs_run_delayed_refs_for_head(struct btrfs_trans_handle *trans,
1899                                     struct btrfs_delayed_ref_head *locked_ref,
1900                                     unsigned long *run_refs)
1901 {
1902         struct btrfs_fs_info *fs_info = trans->fs_info;
1903         struct btrfs_delayed_ref_root *delayed_refs;
1904         struct btrfs_delayed_extent_op *extent_op;
1905         struct btrfs_delayed_ref_node *ref;
1906         int must_insert_reserved = 0;
1907         int ret;
1908
1909         delayed_refs = &trans->transaction->delayed_refs;
1910
1911         lockdep_assert_held(&locked_ref->mutex);
1912         lockdep_assert_held(&locked_ref->lock);
1913
1914         while ((ref = select_delayed_ref(locked_ref))) {
1915                 if (ref->seq &&
1916                     btrfs_check_delayed_seq(fs_info, ref->seq)) {
1917                         spin_unlock(&locked_ref->lock);
1918                         unselect_delayed_ref_head(delayed_refs, locked_ref);
1919                         return -EAGAIN;
1920                 }
1921
1922                 (*run_refs)++;
1923                 ref->in_tree = 0;
1924                 rb_erase_cached(&ref->ref_node, &locked_ref->ref_tree);
1925                 RB_CLEAR_NODE(&ref->ref_node);
1926                 if (!list_empty(&ref->add_list))
1927                         list_del(&ref->add_list);
1928                 /*
1929                  * When we play the delayed ref, also correct the ref_mod on
1930                  * head
1931                  */
1932                 switch (ref->action) {
1933                 case BTRFS_ADD_DELAYED_REF:
1934                 case BTRFS_ADD_DELAYED_EXTENT:
1935                         locked_ref->ref_mod -= ref->ref_mod;
1936                         break;
1937                 case BTRFS_DROP_DELAYED_REF:
1938                         locked_ref->ref_mod += ref->ref_mod;
1939                         break;
1940                 default:
1941                         WARN_ON(1);
1942                 }
1943                 atomic_dec(&delayed_refs->num_entries);
1944
1945                 /*
1946                  * Record the must_insert_reserved flag before we drop the
1947                  * spin lock.
1948                  */
1949                 must_insert_reserved = locked_ref->must_insert_reserved;
1950                 locked_ref->must_insert_reserved = 0;
1951
1952                 extent_op = locked_ref->extent_op;
1953                 locked_ref->extent_op = NULL;
1954                 spin_unlock(&locked_ref->lock);
1955
1956                 ret = run_one_delayed_ref(trans, ref, extent_op,
1957                                           must_insert_reserved);
1958
1959                 btrfs_free_delayed_extent_op(extent_op);
1960                 if (ret) {
1961                         unselect_delayed_ref_head(delayed_refs, locked_ref);
1962                         btrfs_put_delayed_ref(ref);
1963                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d",
1964                                     ret);
1965                         return ret;
1966                 }
1967
1968                 btrfs_put_delayed_ref(ref);
1969                 cond_resched();
1970
1971                 spin_lock(&locked_ref->lock);
1972                 btrfs_merge_delayed_refs(trans, delayed_refs, locked_ref);
1973         }
1974
1975         return 0;
1976 }
1977
1978 /*
1979  * Returns 0 on success or if called with an already aborted transaction.
1980  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
1981  */
1982 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
1983                                              unsigned long nr)
1984 {
1985         struct btrfs_fs_info *fs_info = trans->fs_info;
1986         struct btrfs_delayed_ref_root *delayed_refs;
1987         struct btrfs_delayed_ref_head *locked_ref = NULL;
1988         ktime_t start = ktime_get();
1989         int ret;
1990         unsigned long count = 0;
1991         unsigned long actual_count = 0;
1992
1993         delayed_refs = &trans->transaction->delayed_refs;
1994         do {
1995                 if (!locked_ref) {
1996                         locked_ref = btrfs_obtain_ref_head(trans);
1997                         if (IS_ERR_OR_NULL(locked_ref)) {
1998                                 if (PTR_ERR(locked_ref) == -EAGAIN) {
1999                                         continue;
2000                                 } else {
2001                                         break;
2002                                 }
2003                         }
2004                         count++;
2005                 }
2006                 /*
2007                  * We need to try and merge add/drops of the same ref since we
2008                  * can run into issues with relocate dropping the implicit ref
2009                  * and then it being added back again before the drop can
2010                  * finish.  If we merged anything we need to re-loop so we can
2011                  * get a good ref.
2012                  * Or we can get node references of the same type that weren't
2013                  * merged when created due to bumps in the tree mod seq, and
2014                  * we need to merge them to prevent adding an inline extent
2015                  * backref before dropping it (triggering a BUG_ON at
2016                  * insert_inline_extent_backref()).
2017                  */
2018                 spin_lock(&locked_ref->lock);
2019                 btrfs_merge_delayed_refs(trans, delayed_refs, locked_ref);
2020
2021                 ret = btrfs_run_delayed_refs_for_head(trans, locked_ref,
2022                                                       &actual_count);
2023                 if (ret < 0 && ret != -EAGAIN) {
2024                         /*
2025                          * Error, btrfs_run_delayed_refs_for_head already
2026                          * unlocked everything so just bail out
2027                          */
2028                         return ret;
2029                 } else if (!ret) {
2030                         /*
2031                          * Success, perform the usual cleanup of a processed
2032                          * head
2033                          */
2034                         ret = cleanup_ref_head(trans, locked_ref);
2035                         if (ret > 0 ) {
2036                                 /* We dropped our lock, we need to loop. */
2037                                 ret = 0;
2038                                 continue;
2039                         } else if (ret) {
2040                                 return ret;
2041                         }
2042                 }
2043
2044                 /*
2045                  * Either success case or btrfs_run_delayed_refs_for_head
2046                  * returned -EAGAIN, meaning we need to select another head
2047                  */
2048
2049                 locked_ref = NULL;
2050                 cond_resched();
2051         } while ((nr != -1 && count < nr) || locked_ref);
2052
2053         /*
2054          * We don't want to include ref heads since we can have empty ref heads
2055          * and those will drastically skew our runtime down since we just do
2056          * accounting, no actual extent tree updates.
2057          */
2058         if (actual_count > 0) {
2059                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2060                 u64 avg;
2061
2062                 /*
2063                  * We weigh the current average higher than our current runtime
2064                  * to avoid large swings in the average.
2065                  */
2066                 spin_lock(&delayed_refs->lock);
2067                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2068                 fs_info->avg_delayed_ref_runtime = avg >> 2;    /* div by 4 */
2069                 spin_unlock(&delayed_refs->lock);
2070         }
2071         return 0;
2072 }
2073
2074 #ifdef SCRAMBLE_DELAYED_REFS
2075 /*
2076  * Normally delayed refs get processed in ascending bytenr order. This
2077  * correlates in most cases to the order added. To expose dependencies on this
2078  * order, we start to process the tree in the middle instead of the beginning
2079  */
2080 static u64 find_middle(struct rb_root *root)
2081 {
2082         struct rb_node *n = root->rb_node;
2083         struct btrfs_delayed_ref_node *entry;
2084         int alt = 1;
2085         u64 middle;
2086         u64 first = 0, last = 0;
2087
2088         n = rb_first(root);
2089         if (n) {
2090                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2091                 first = entry->bytenr;
2092         }
2093         n = rb_last(root);
2094         if (n) {
2095                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2096                 last = entry->bytenr;
2097         }
2098         n = root->rb_node;
2099
2100         while (n) {
2101                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2102                 WARN_ON(!entry->in_tree);
2103
2104                 middle = entry->bytenr;
2105
2106                 if (alt)
2107                         n = n->rb_left;
2108                 else
2109                         n = n->rb_right;
2110
2111                 alt = 1 - alt;
2112         }
2113         return middle;
2114 }
2115 #endif
2116
2117 /*
2118  * Takes the number of bytes to be csumm'ed and figures out how many leaves it
2119  * would require to store the csums for that many bytes.
2120  */
2121 u64 btrfs_csum_bytes_to_leaves(struct btrfs_fs_info *fs_info, u64 csum_bytes)
2122 {
2123         u64 csum_size;
2124         u64 num_csums_per_leaf;
2125         u64 num_csums;
2126
2127         csum_size = BTRFS_MAX_ITEM_SIZE(fs_info);
2128         num_csums_per_leaf = div64_u64(csum_size,
2129                         (u64)btrfs_super_csum_size(fs_info->super_copy));
2130         num_csums = div64_u64(csum_bytes, fs_info->sectorsize);
2131         num_csums += num_csums_per_leaf - 1;
2132         num_csums = div64_u64(num_csums, num_csums_per_leaf);
2133         return num_csums;
2134 }
2135
2136 /*
2137  * this starts processing the delayed reference count updates and
2138  * extent insertions we have queued up so far.  count can be
2139  * 0, which means to process everything in the tree at the start
2140  * of the run (but not newly added entries), or it can be some target
2141  * number you'd like to process.
2142  *
2143  * Returns 0 on success or if called with an aborted transaction
2144  * Returns <0 on error and aborts the transaction
2145  */
2146 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2147                            unsigned long count)
2148 {
2149         struct btrfs_fs_info *fs_info = trans->fs_info;
2150         struct rb_node *node;
2151         struct btrfs_delayed_ref_root *delayed_refs;
2152         struct btrfs_delayed_ref_head *head;
2153         int ret;
2154         int run_all = count == (unsigned long)-1;
2155
2156         /* We'll clean this up in btrfs_cleanup_transaction */
2157         if (TRANS_ABORTED(trans))
2158                 return 0;
2159
2160         if (test_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags))
2161                 return 0;
2162
2163         delayed_refs = &trans->transaction->delayed_refs;
2164         if (count == 0)
2165                 count = atomic_read(&delayed_refs->num_entries) * 2;
2166
2167 again:
2168 #ifdef SCRAMBLE_DELAYED_REFS
2169         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2170 #endif
2171         ret = __btrfs_run_delayed_refs(trans, count);
2172         if (ret < 0) {
2173                 btrfs_abort_transaction(trans, ret);
2174                 return ret;
2175         }
2176
2177         if (run_all) {
2178                 btrfs_create_pending_block_groups(trans);
2179
2180                 spin_lock(&delayed_refs->lock);
2181                 node = rb_first_cached(&delayed_refs->href_root);
2182                 if (!node) {
2183                         spin_unlock(&delayed_refs->lock);
2184                         goto out;
2185                 }
2186                 head = rb_entry(node, struct btrfs_delayed_ref_head,
2187                                 href_node);
2188                 refcount_inc(&head->refs);
2189                 spin_unlock(&delayed_refs->lock);
2190
2191                 /* Mutex was contended, block until it's released and retry. */
2192                 mutex_lock(&head->mutex);
2193                 mutex_unlock(&head->mutex);
2194
2195                 btrfs_put_delayed_ref_head(head);
2196                 cond_resched();
2197                 goto again;
2198         }
2199 out:
2200         return 0;
2201 }
2202
2203 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2204                                 struct extent_buffer *eb, u64 flags,
2205                                 int level, int is_data)
2206 {
2207         struct btrfs_delayed_extent_op *extent_op;
2208         int ret;
2209
2210         extent_op = btrfs_alloc_delayed_extent_op();
2211         if (!extent_op)
2212                 return -ENOMEM;
2213
2214         extent_op->flags_to_set = flags;
2215         extent_op->update_flags = true;
2216         extent_op->update_key = false;
2217         extent_op->is_data = is_data ? true : false;
2218         extent_op->level = level;
2219
2220         ret = btrfs_add_delayed_extent_op(trans, eb->start, eb->len, extent_op);
2221         if (ret)
2222                 btrfs_free_delayed_extent_op(extent_op);
2223         return ret;
2224 }
2225
2226 static noinline int check_delayed_ref(struct btrfs_root *root,
2227                                       struct btrfs_path *path,
2228                                       u64 objectid, u64 offset, u64 bytenr)
2229 {
2230         struct btrfs_delayed_ref_head *head;
2231         struct btrfs_delayed_ref_node *ref;
2232         struct btrfs_delayed_data_ref *data_ref;
2233         struct btrfs_delayed_ref_root *delayed_refs;
2234         struct btrfs_transaction *cur_trans;
2235         struct rb_node *node;
2236         int ret = 0;
2237
2238         spin_lock(&root->fs_info->trans_lock);
2239         cur_trans = root->fs_info->running_transaction;
2240         if (cur_trans)
2241                 refcount_inc(&cur_trans->use_count);
2242         spin_unlock(&root->fs_info->trans_lock);
2243         if (!cur_trans)
2244                 return 0;
2245
2246         delayed_refs = &cur_trans->delayed_refs;
2247         spin_lock(&delayed_refs->lock);
2248         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
2249         if (!head) {
2250                 spin_unlock(&delayed_refs->lock);
2251                 btrfs_put_transaction(cur_trans);
2252                 return 0;
2253         }
2254
2255         if (!mutex_trylock(&head->mutex)) {
2256                 refcount_inc(&head->refs);
2257                 spin_unlock(&delayed_refs->lock);
2258
2259                 btrfs_release_path(path);
2260
2261                 /*
2262                  * Mutex was contended, block until it's released and let
2263                  * caller try again
2264                  */
2265                 mutex_lock(&head->mutex);
2266                 mutex_unlock(&head->mutex);
2267                 btrfs_put_delayed_ref_head(head);
2268                 btrfs_put_transaction(cur_trans);
2269                 return -EAGAIN;
2270         }
2271         spin_unlock(&delayed_refs->lock);
2272
2273         spin_lock(&head->lock);
2274         /*
2275          * XXX: We should replace this with a proper search function in the
2276          * future.
2277          */
2278         for (node = rb_first_cached(&head->ref_tree); node;
2279              node = rb_next(node)) {
2280                 ref = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
2281                 /* If it's a shared ref we know a cross reference exists */
2282                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
2283                         ret = 1;
2284                         break;
2285                 }
2286
2287                 data_ref = btrfs_delayed_node_to_data_ref(ref);
2288
2289                 /*
2290                  * If our ref doesn't match the one we're currently looking at
2291                  * then we have a cross reference.
2292                  */
2293                 if (data_ref->root != root->root_key.objectid ||
2294                     data_ref->objectid != objectid ||
2295                     data_ref->offset != offset) {
2296                         ret = 1;
2297                         break;
2298                 }
2299         }
2300         spin_unlock(&head->lock);
2301         mutex_unlock(&head->mutex);
2302         btrfs_put_transaction(cur_trans);
2303         return ret;
2304 }
2305
2306 static noinline int check_committed_ref(struct btrfs_root *root,
2307                                         struct btrfs_path *path,
2308                                         u64 objectid, u64 offset, u64 bytenr,
2309                                         bool strict)
2310 {
2311         struct btrfs_fs_info *fs_info = root->fs_info;
2312         struct btrfs_root *extent_root = fs_info->extent_root;
2313         struct extent_buffer *leaf;
2314         struct btrfs_extent_data_ref *ref;
2315         struct btrfs_extent_inline_ref *iref;
2316         struct btrfs_extent_item *ei;
2317         struct btrfs_key key;
2318         u32 item_size;
2319         int type;
2320         int ret;
2321
2322         key.objectid = bytenr;
2323         key.offset = (u64)-1;
2324         key.type = BTRFS_EXTENT_ITEM_KEY;
2325
2326         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2327         if (ret < 0)
2328                 goto out;
2329         BUG_ON(ret == 0); /* Corruption */
2330
2331         ret = -ENOENT;
2332         if (path->slots[0] == 0)
2333                 goto out;
2334
2335         path->slots[0]--;
2336         leaf = path->nodes[0];
2337         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2338
2339         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2340                 goto out;
2341
2342         ret = 1;
2343         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2344         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2345
2346         /* If extent item has more than 1 inline ref then it's shared */
2347         if (item_size != sizeof(*ei) +
2348             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2349                 goto out;
2350
2351         /*
2352          * If extent created before last snapshot => it's shared unless the
2353          * snapshot has been deleted. Use the heuristic if strict is false.
2354          */
2355         if (!strict &&
2356             (btrfs_extent_generation(leaf, ei) <=
2357              btrfs_root_last_snapshot(&root->root_item)))
2358                 goto out;
2359
2360         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2361
2362         /* If this extent has SHARED_DATA_REF then it's shared */
2363         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
2364         if (type != BTRFS_EXTENT_DATA_REF_KEY)
2365                 goto out;
2366
2367         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2368         if (btrfs_extent_refs(leaf, ei) !=
2369             btrfs_extent_data_ref_count(leaf, ref) ||
2370             btrfs_extent_data_ref_root(leaf, ref) !=
2371             root->root_key.objectid ||
2372             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2373             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2374                 goto out;
2375
2376         ret = 0;
2377 out:
2378         return ret;
2379 }
2380
2381 int btrfs_cross_ref_exist(struct btrfs_root *root, u64 objectid, u64 offset,
2382                           u64 bytenr, bool strict)
2383 {
2384         struct btrfs_path *path;
2385         int ret;
2386
2387         path = btrfs_alloc_path();
2388         if (!path)
2389                 return -ENOMEM;
2390
2391         do {
2392                 ret = check_committed_ref(root, path, objectid,
2393                                           offset, bytenr, strict);
2394                 if (ret && ret != -ENOENT)
2395                         goto out;
2396
2397                 ret = check_delayed_ref(root, path, objectid, offset, bytenr);
2398         } while (ret == -EAGAIN);
2399
2400 out:
2401         btrfs_free_path(path);
2402         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2403                 WARN_ON(ret > 0);
2404         return ret;
2405 }
2406
2407 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2408                            struct btrfs_root *root,
2409                            struct extent_buffer *buf,
2410                            int full_backref, int inc)
2411 {
2412         struct btrfs_fs_info *fs_info = root->fs_info;
2413         u64 bytenr;
2414         u64 num_bytes;
2415         u64 parent;
2416         u64 ref_root;
2417         u32 nritems;
2418         struct btrfs_key key;
2419         struct btrfs_file_extent_item *fi;
2420         struct btrfs_ref generic_ref = { 0 };
2421         bool for_reloc = btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC);
2422         int i;
2423         int action;
2424         int level;
2425         int ret = 0;
2426
2427         if (btrfs_is_testing(fs_info))
2428                 return 0;
2429
2430         ref_root = btrfs_header_owner(buf);
2431         nritems = btrfs_header_nritems(buf);
2432         level = btrfs_header_level(buf);
2433
2434         if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state) && level == 0)
2435                 return 0;
2436
2437         if (full_backref)
2438                 parent = buf->start;
2439         else
2440                 parent = 0;
2441         if (inc)
2442                 action = BTRFS_ADD_DELAYED_REF;
2443         else
2444                 action = BTRFS_DROP_DELAYED_REF;
2445
2446         for (i = 0; i < nritems; i++) {
2447                 if (level == 0) {
2448                         btrfs_item_key_to_cpu(buf, &key, i);
2449                         if (key.type != BTRFS_EXTENT_DATA_KEY)
2450                                 continue;
2451                         fi = btrfs_item_ptr(buf, i,
2452                                             struct btrfs_file_extent_item);
2453                         if (btrfs_file_extent_type(buf, fi) ==
2454                             BTRFS_FILE_EXTENT_INLINE)
2455                                 continue;
2456                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2457                         if (bytenr == 0)
2458                                 continue;
2459
2460                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
2461                         key.offset -= btrfs_file_extent_offset(buf, fi);
2462                         btrfs_init_generic_ref(&generic_ref, action, bytenr,
2463                                                num_bytes, parent);
2464                         generic_ref.real_root = root->root_key.objectid;
2465                         btrfs_init_data_ref(&generic_ref, ref_root, key.objectid,
2466                                             key.offset);
2467                         generic_ref.skip_qgroup = for_reloc;
2468                         if (inc)
2469                                 ret = btrfs_inc_extent_ref(trans, &generic_ref);
2470                         else
2471                                 ret = btrfs_free_extent(trans, &generic_ref);
2472                         if (ret)
2473                                 goto fail;
2474                 } else {
2475                         bytenr = btrfs_node_blockptr(buf, i);
2476                         num_bytes = fs_info->nodesize;
2477                         btrfs_init_generic_ref(&generic_ref, action, bytenr,
2478                                                num_bytes, parent);
2479                         generic_ref.real_root = root->root_key.objectid;
2480                         btrfs_init_tree_ref(&generic_ref, level - 1, ref_root);
2481                         generic_ref.skip_qgroup = for_reloc;
2482                         if (inc)
2483                                 ret = btrfs_inc_extent_ref(trans, &generic_ref);
2484                         else
2485                                 ret = btrfs_free_extent(trans, &generic_ref);
2486                         if (ret)
2487                                 goto fail;
2488                 }
2489         }
2490         return 0;
2491 fail:
2492         return ret;
2493 }
2494
2495 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2496                   struct extent_buffer *buf, int full_backref)
2497 {
2498         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
2499 }
2500
2501 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2502                   struct extent_buffer *buf, int full_backref)
2503 {
2504         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
2505 }
2506
2507 int btrfs_extent_readonly(struct btrfs_fs_info *fs_info, u64 bytenr)
2508 {
2509         struct btrfs_block_group *block_group;
2510         int readonly = 0;
2511
2512         block_group = btrfs_lookup_block_group(fs_info, bytenr);
2513         if (!block_group || block_group->ro)
2514                 readonly = 1;
2515         if (block_group)
2516                 btrfs_put_block_group(block_group);
2517         return readonly;
2518 }
2519
2520 static u64 get_alloc_profile_by_root(struct btrfs_root *root, int data)
2521 {
2522         struct btrfs_fs_info *fs_info = root->fs_info;
2523         u64 flags;
2524         u64 ret;
2525
2526         if (data)
2527                 flags = BTRFS_BLOCK_GROUP_DATA;
2528         else if (root == fs_info->chunk_root)
2529                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
2530         else
2531                 flags = BTRFS_BLOCK_GROUP_METADATA;
2532
2533         ret = btrfs_get_alloc_profile(fs_info, flags);
2534         return ret;
2535 }
2536
2537 static u64 first_logical_byte(struct btrfs_fs_info *fs_info, u64 search_start)
2538 {
2539         struct btrfs_block_group *cache;
2540         u64 bytenr;
2541
2542         spin_lock(&fs_info->block_group_cache_lock);
2543         bytenr = fs_info->first_logical_byte;
2544         spin_unlock(&fs_info->block_group_cache_lock);
2545
2546         if (bytenr < (u64)-1)
2547                 return bytenr;
2548
2549         cache = btrfs_lookup_first_block_group(fs_info, search_start);
2550         if (!cache)
2551                 return 0;
2552
2553         bytenr = cache->start;
2554         btrfs_put_block_group(cache);
2555
2556         return bytenr;
2557 }
2558
2559 static int pin_down_extent(struct btrfs_trans_handle *trans,
2560                            struct btrfs_block_group *cache,
2561                            u64 bytenr, u64 num_bytes, int reserved)
2562 {
2563         struct btrfs_fs_info *fs_info = cache->fs_info;
2564
2565         spin_lock(&cache->space_info->lock);
2566         spin_lock(&cache->lock);
2567         cache->pinned += num_bytes;
2568         btrfs_space_info_update_bytes_pinned(fs_info, cache->space_info,
2569                                              num_bytes);
2570         if (reserved) {
2571                 cache->reserved -= num_bytes;
2572                 cache->space_info->bytes_reserved -= num_bytes;
2573         }
2574         spin_unlock(&cache->lock);
2575         spin_unlock(&cache->space_info->lock);
2576
2577         percpu_counter_add_batch(&cache->space_info->total_bytes_pinned,
2578                     num_bytes, BTRFS_TOTAL_BYTES_PINNED_BATCH);
2579         set_extent_dirty(&trans->transaction->pinned_extents, bytenr,
2580                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
2581         return 0;
2582 }
2583
2584 int btrfs_pin_extent(struct btrfs_trans_handle *trans,
2585                      u64 bytenr, u64 num_bytes, int reserved)
2586 {
2587         struct btrfs_block_group *cache;
2588
2589         cache = btrfs_lookup_block_group(trans->fs_info, bytenr);
2590         BUG_ON(!cache); /* Logic error */
2591
2592         pin_down_extent(trans, cache, bytenr, num_bytes, reserved);
2593
2594         btrfs_put_block_group(cache);
2595         return 0;
2596 }
2597
2598 /*
2599  * this function must be called within transaction
2600  */
2601 int btrfs_pin_extent_for_log_replay(struct btrfs_trans_handle *trans,
2602                                     u64 bytenr, u64 num_bytes)
2603 {
2604         struct btrfs_block_group *cache;
2605         int ret;
2606
2607         btrfs_add_excluded_extent(trans->fs_info, bytenr, num_bytes);
2608
2609         cache = btrfs_lookup_block_group(trans->fs_info, bytenr);
2610         if (!cache)
2611                 return -EINVAL;
2612
2613         /*
2614          * pull in the free space cache (if any) so that our pin
2615          * removes the free space from the cache.  We have load_only set
2616          * to one because the slow code to read in the free extents does check
2617          * the pinned extents.
2618          */
2619         btrfs_cache_block_group(cache, 1);
2620
2621         pin_down_extent(trans, cache, bytenr, num_bytes, 0);
2622
2623         /* remove us from the free space cache (if we're there at all) */
2624         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
2625         btrfs_put_block_group(cache);
2626         return ret;
2627 }
2628
2629 static int __exclude_logged_extent(struct btrfs_fs_info *fs_info,
2630                                    u64 start, u64 num_bytes)
2631 {
2632         int ret;
2633         struct btrfs_block_group *block_group;
2634         struct btrfs_caching_control *caching_ctl;
2635
2636         block_group = btrfs_lookup_block_group(fs_info, start);
2637         if (!block_group)
2638                 return -EINVAL;
2639
2640         btrfs_cache_block_group(block_group, 0);
2641         caching_ctl = btrfs_get_caching_control(block_group);
2642
2643         if (!caching_ctl) {
2644                 /* Logic error */
2645                 BUG_ON(!btrfs_block_group_done(block_group));
2646                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
2647         } else {
2648                 mutex_lock(&caching_ctl->mutex);
2649
2650                 if (start >= caching_ctl->progress) {
2651                         ret = btrfs_add_excluded_extent(fs_info, start,
2652                                                         num_bytes);
2653                 } else if (start + num_bytes <= caching_ctl->progress) {
2654                         ret = btrfs_remove_free_space(block_group,
2655                                                       start, num_bytes);
2656                 } else {
2657                         num_bytes = caching_ctl->progress - start;
2658                         ret = btrfs_remove_free_space(block_group,
2659                                                       start, num_bytes);
2660                         if (ret)
2661                                 goto out_lock;
2662
2663                         num_bytes = (start + num_bytes) -
2664                                 caching_ctl->progress;
2665                         start = caching_ctl->progress;
2666                         ret = btrfs_add_excluded_extent(fs_info, start,
2667                                                         num_bytes);
2668                 }
2669 out_lock:
2670                 mutex_unlock(&caching_ctl->mutex);
2671                 btrfs_put_caching_control(caching_ctl);
2672         }
2673         btrfs_put_block_group(block_group);
2674         return ret;
2675 }
2676
2677 int btrfs_exclude_logged_extents(struct extent_buffer *eb)
2678 {
2679         struct btrfs_fs_info *fs_info = eb->fs_info;
2680         struct btrfs_file_extent_item *item;
2681         struct btrfs_key key;
2682         int found_type;
2683         int i;
2684         int ret = 0;
2685
2686         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS))
2687                 return 0;
2688
2689         for (i = 0; i < btrfs_header_nritems(eb); i++) {
2690                 btrfs_item_key_to_cpu(eb, &key, i);
2691                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2692                         continue;
2693                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
2694                 found_type = btrfs_file_extent_type(eb, item);
2695                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
2696                         continue;
2697                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
2698                         continue;
2699                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
2700                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
2701                 ret = __exclude_logged_extent(fs_info, key.objectid, key.offset);
2702                 if (ret)
2703                         break;
2704         }
2705
2706         return ret;
2707 }
2708
2709 static void
2710 btrfs_inc_block_group_reservations(struct btrfs_block_group *bg)
2711 {
2712         atomic_inc(&bg->reservations);
2713 }
2714
2715 void btrfs_prepare_extent_commit(struct btrfs_fs_info *fs_info)
2716 {
2717         struct btrfs_caching_control *next;
2718         struct btrfs_caching_control *caching_ctl;
2719         struct btrfs_block_group *cache;
2720
2721         down_write(&fs_info->commit_root_sem);
2722
2723         list_for_each_entry_safe(caching_ctl, next,
2724                                  &fs_info->caching_block_groups, list) {
2725                 cache = caching_ctl->block_group;
2726                 if (btrfs_block_group_done(cache)) {
2727                         cache->last_byte_to_unpin = (u64)-1;
2728                         list_del_init(&caching_ctl->list);
2729                         btrfs_put_caching_control(caching_ctl);
2730                 } else {
2731                         cache->last_byte_to_unpin = caching_ctl->progress;
2732                 }
2733         }
2734
2735         up_write(&fs_info->commit_root_sem);
2736
2737         btrfs_update_global_block_rsv(fs_info);
2738 }
2739
2740 /*
2741  * Returns the free cluster for the given space info and sets empty_cluster to
2742  * what it should be based on the mount options.
2743  */
2744 static struct btrfs_free_cluster *
2745 fetch_cluster_info(struct btrfs_fs_info *fs_info,
2746                    struct btrfs_space_info *space_info, u64 *empty_cluster)
2747 {
2748         struct btrfs_free_cluster *ret = NULL;
2749
2750         *empty_cluster = 0;
2751         if (btrfs_mixed_space_info(space_info))
2752                 return ret;
2753
2754         if (space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
2755                 ret = &fs_info->meta_alloc_cluster;
2756                 if (btrfs_test_opt(fs_info, SSD))
2757                         *empty_cluster = SZ_2M;
2758                 else
2759                         *empty_cluster = SZ_64K;
2760         } else if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) &&
2761                    btrfs_test_opt(fs_info, SSD_SPREAD)) {
2762                 *empty_cluster = SZ_2M;
2763                 ret = &fs_info->data_alloc_cluster;
2764         }
2765
2766         return ret;
2767 }
2768
2769 static int unpin_extent_range(struct btrfs_fs_info *fs_info,
2770                               u64 start, u64 end,
2771                               const bool return_free_space)
2772 {
2773         struct btrfs_block_group *cache = NULL;
2774         struct btrfs_space_info *space_info;
2775         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
2776         struct btrfs_free_cluster *cluster = NULL;
2777         u64 len;
2778         u64 total_unpinned = 0;
2779         u64 empty_cluster = 0;
2780         bool readonly;
2781
2782         while (start <= end) {
2783                 readonly = false;
2784                 if (!cache ||
2785                     start >= cache->start + cache->length) {
2786                         if (cache)
2787                                 btrfs_put_block_group(cache);
2788                         total_unpinned = 0;
2789                         cache = btrfs_lookup_block_group(fs_info, start);
2790                         BUG_ON(!cache); /* Logic error */
2791
2792                         cluster = fetch_cluster_info(fs_info,
2793                                                      cache->space_info,
2794                                                      &empty_cluster);
2795                         empty_cluster <<= 1;
2796                 }
2797
2798                 len = cache->start + cache->length - start;
2799                 len = min(len, end + 1 - start);
2800
2801                 if (start < cache->last_byte_to_unpin) {
2802                         len = min(len, cache->last_byte_to_unpin - start);
2803                         if (return_free_space)
2804                                 btrfs_add_free_space(cache, start, len);
2805                 }
2806
2807                 start += len;
2808                 total_unpinned += len;
2809                 space_info = cache->space_info;
2810
2811                 /*
2812                  * If this space cluster has been marked as fragmented and we've
2813                  * unpinned enough in this block group to potentially allow a
2814                  * cluster to be created inside of it go ahead and clear the
2815                  * fragmented check.
2816                  */
2817                 if (cluster && cluster->fragmented &&
2818                     total_unpinned > empty_cluster) {
2819                         spin_lock(&cluster->lock);
2820                         cluster->fragmented = 0;
2821                         spin_unlock(&cluster->lock);
2822                 }
2823
2824                 spin_lock(&space_info->lock);
2825                 spin_lock(&cache->lock);
2826                 cache->pinned -= len;
2827                 btrfs_space_info_update_bytes_pinned(fs_info, space_info, -len);
2828                 space_info->max_extent_size = 0;
2829                 percpu_counter_add_batch(&space_info->total_bytes_pinned,
2830                             -len, BTRFS_TOTAL_BYTES_PINNED_BATCH);
2831                 if (cache->ro) {
2832                         space_info->bytes_readonly += len;
2833                         readonly = true;
2834                 }
2835                 spin_unlock(&cache->lock);
2836                 if (!readonly && return_free_space &&
2837                     global_rsv->space_info == space_info) {
2838                         u64 to_add = len;
2839
2840                         spin_lock(&global_rsv->lock);
2841                         if (!global_rsv->full) {
2842                                 to_add = min(len, global_rsv->size -
2843                                              global_rsv->reserved);
2844                                 global_rsv->reserved += to_add;
2845                                 btrfs_space_info_update_bytes_may_use(fs_info,
2846                                                 space_info, to_add);
2847                                 if (global_rsv->reserved >= global_rsv->size)
2848                                         global_rsv->full = 1;
2849                                 len -= to_add;
2850                         }
2851                         spin_unlock(&global_rsv->lock);
2852                         /* Add to any tickets we may have */
2853                         if (len)
2854                                 btrfs_try_granting_tickets(fs_info,
2855                                                            space_info);
2856                 }
2857                 spin_unlock(&space_info->lock);
2858         }
2859
2860         if (cache)
2861                 btrfs_put_block_group(cache);
2862         return 0;
2863 }
2864
2865 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans)
2866 {
2867         struct btrfs_fs_info *fs_info = trans->fs_info;
2868         struct btrfs_block_group *block_group, *tmp;
2869         struct list_head *deleted_bgs;
2870         struct extent_io_tree *unpin;
2871         u64 start;
2872         u64 end;
2873         int ret;
2874
2875         unpin = &trans->transaction->pinned_extents;
2876
2877         while (!TRANS_ABORTED(trans)) {
2878                 struct extent_state *cached_state = NULL;
2879
2880                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
2881                 ret = find_first_extent_bit(unpin, 0, &start, &end,
2882                                             EXTENT_DIRTY, &cached_state);
2883                 if (ret) {
2884                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
2885                         break;
2886                 }
2887                 if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
2888                         clear_extent_bits(&fs_info->excluded_extents, start,
2889                                           end, EXTENT_UPTODATE);
2890
2891                 if (btrfs_test_opt(fs_info, DISCARD_SYNC))
2892                         ret = btrfs_discard_extent(fs_info, start,
2893                                                    end + 1 - start, NULL);
2894
2895                 clear_extent_dirty(unpin, start, end, &cached_state);
2896                 unpin_extent_range(fs_info, start, end, true);
2897                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
2898                 free_extent_state(cached_state);
2899                 cond_resched();
2900         }
2901
2902         if (btrfs_test_opt(fs_info, DISCARD_ASYNC)) {
2903                 btrfs_discard_calc_delay(&fs_info->discard_ctl);
2904                 btrfs_discard_schedule_work(&fs_info->discard_ctl, true);
2905         }
2906
2907         /*
2908          * Transaction is finished.  We don't need the lock anymore.  We
2909          * do need to clean up the block groups in case of a transaction
2910          * abort.
2911          */
2912         deleted_bgs = &trans->transaction->deleted_bgs;
2913         list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
2914                 u64 trimmed = 0;
2915
2916                 ret = -EROFS;
2917                 if (!TRANS_ABORTED(trans))
2918                         ret = btrfs_discard_extent(fs_info,
2919                                                    block_group->start,
2920                                                    block_group->length,
2921                                                    &trimmed);
2922
2923                 list_del_init(&block_group->bg_list);
2924                 btrfs_unfreeze_block_group(block_group);
2925                 btrfs_put_block_group(block_group);
2926
2927                 if (ret) {
2928                         const char *errstr = btrfs_decode_error(ret);
2929                         btrfs_warn(fs_info,
2930                            "discard failed while removing blockgroup: errno=%d %s",
2931                                    ret, errstr);
2932                 }
2933         }
2934
2935         return 0;
2936 }
2937
2938 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
2939                                struct btrfs_delayed_ref_node *node, u64 parent,
2940                                u64 root_objectid, u64 owner_objectid,
2941                                u64 owner_offset, int refs_to_drop,
2942                                struct btrfs_delayed_extent_op *extent_op)
2943 {
2944         struct btrfs_fs_info *info = trans->fs_info;
2945         struct btrfs_key key;
2946         struct btrfs_path *path;
2947         struct btrfs_root *extent_root = info->extent_root;
2948         struct extent_buffer *leaf;
2949         struct btrfs_extent_item *ei;
2950         struct btrfs_extent_inline_ref *iref;
2951         int ret;
2952         int is_data;
2953         int extent_slot = 0;
2954         int found_extent = 0;
2955         int num_to_del = 1;
2956         u32 item_size;
2957         u64 refs;
2958         u64 bytenr = node->bytenr;
2959         u64 num_bytes = node->num_bytes;
2960         int last_ref = 0;
2961         bool skinny_metadata = btrfs_fs_incompat(info, SKINNY_METADATA);
2962
2963         path = btrfs_alloc_path();
2964         if (!path)
2965                 return -ENOMEM;
2966
2967         path->leave_spinning = 1;
2968
2969         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
2970         BUG_ON(!is_data && refs_to_drop != 1);
2971
2972         if (is_data)
2973                 skinny_metadata = false;
2974
2975         ret = lookup_extent_backref(trans, path, &iref, bytenr, num_bytes,
2976                                     parent, root_objectid, owner_objectid,
2977                                     owner_offset);
2978         if (ret == 0) {
2979                 extent_slot = path->slots[0];
2980                 while (extent_slot >= 0) {
2981                         btrfs_item_key_to_cpu(path->nodes[0], &key,
2982                                               extent_slot);
2983                         if (key.objectid != bytenr)
2984                                 break;
2985                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
2986                             key.offset == num_bytes) {
2987                                 found_extent = 1;
2988                                 break;
2989                         }
2990                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
2991                             key.offset == owner_objectid) {
2992                                 found_extent = 1;
2993                                 break;
2994                         }
2995                         if (path->slots[0] - extent_slot > 5)
2996                                 break;
2997                         extent_slot--;
2998                 }
2999
3000                 if (!found_extent) {
3001                         BUG_ON(iref);
3002                         ret = remove_extent_backref(trans, path, NULL,
3003                                                     refs_to_drop,
3004                                                     is_data, &last_ref);
3005                         if (ret) {
3006                                 btrfs_abort_transaction(trans, ret);
3007                                 goto out;
3008                         }
3009                         btrfs_release_path(path);
3010                         path->leave_spinning = 1;
3011
3012                         key.objectid = bytenr;
3013                         key.type = BTRFS_EXTENT_ITEM_KEY;
3014                         key.offset = num_bytes;
3015
3016                         if (!is_data && skinny_metadata) {
3017                                 key.type = BTRFS_METADATA_ITEM_KEY;
3018                                 key.offset = owner_objectid;
3019                         }
3020
3021                         ret = btrfs_search_slot(trans, extent_root,
3022                                                 &key, path, -1, 1);
3023                         if (ret > 0 && skinny_metadata && path->slots[0]) {
3024                                 /*
3025                                  * Couldn't find our skinny metadata item,
3026                                  * see if we have ye olde extent item.
3027                                  */
3028                                 path->slots[0]--;
3029                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
3030                                                       path->slots[0]);
3031                                 if (key.objectid == bytenr &&
3032                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
3033                                     key.offset == num_bytes)
3034                                         ret = 0;
3035                         }
3036
3037                         if (ret > 0 && skinny_metadata) {
3038                                 skinny_metadata = false;
3039                                 key.objectid = bytenr;
3040                                 key.type = BTRFS_EXTENT_ITEM_KEY;
3041                                 key.offset = num_bytes;
3042                                 btrfs_release_path(path);
3043                                 ret = btrfs_search_slot(trans, extent_root,
3044                                                         &key, path, -1, 1);
3045                         }
3046
3047                         if (ret) {
3048                                 btrfs_err(info,
3049                                           "umm, got %d back from search, was looking for %llu",
3050                                           ret, bytenr);
3051                                 if (ret > 0)
3052                                         btrfs_print_leaf(path->nodes[0]);
3053                         }
3054                         if (ret < 0) {
3055                                 btrfs_abort_transaction(trans, ret);
3056                                 goto out;
3057                         }
3058                         extent_slot = path->slots[0];
3059                 }
3060         } else if (WARN_ON(ret == -ENOENT)) {
3061                 btrfs_print_leaf(path->nodes[0]);
3062                 btrfs_err(info,
3063                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
3064                         bytenr, parent, root_objectid, owner_objectid,
3065                         owner_offset);
3066                 btrfs_abort_transaction(trans, ret);
3067                 goto out;
3068         } else {
3069                 btrfs_abort_transaction(trans, ret);
3070                 goto out;
3071         }
3072
3073         leaf = path->nodes[0];
3074         item_size = btrfs_item_size_nr(leaf, extent_slot);
3075         if (unlikely(item_size < sizeof(*ei))) {
3076                 ret = -EINVAL;
3077                 btrfs_print_v0_err(info);
3078                 btrfs_abort_transaction(trans, ret);
3079                 goto out;
3080         }
3081         ei = btrfs_item_ptr(leaf, extent_slot,
3082                             struct btrfs_extent_item);
3083         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
3084             key.type == BTRFS_EXTENT_ITEM_KEY) {
3085                 struct btrfs_tree_block_info *bi;
3086                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
3087                 bi = (struct btrfs_tree_block_info *)(ei + 1);
3088                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
3089         }
3090
3091         refs = btrfs_extent_refs(leaf, ei);
3092         if (refs < refs_to_drop) {
3093                 btrfs_err(info,
3094                           "trying to drop %d refs but we only have %Lu for bytenr %Lu",
3095                           refs_to_drop, refs, bytenr);
3096                 ret = -EINVAL;
3097                 btrfs_abort_transaction(trans, ret);
3098                 goto out;
3099         }
3100         refs -= refs_to_drop;
3101
3102         if (refs > 0) {
3103                 if (extent_op)
3104                         __run_delayed_extent_op(extent_op, leaf, ei);
3105                 /*
3106                  * In the case of inline back ref, reference count will
3107                  * be updated by remove_extent_backref
3108                  */
3109                 if (iref) {
3110                         BUG_ON(!found_extent);
3111                 } else {
3112                         btrfs_set_extent_refs(leaf, ei, refs);
3113                         btrfs_mark_buffer_dirty(leaf);
3114                 }
3115                 if (found_extent) {
3116                         ret = remove_extent_backref(trans, path, iref,
3117                                                     refs_to_drop, is_data,
3118                                                     &last_ref);
3119                         if (ret) {
3120                                 btrfs_abort_transaction(trans, ret);
3121                                 goto out;
3122                         }
3123                 }
3124         } else {
3125                 if (found_extent) {
3126                         BUG_ON(is_data && refs_to_drop !=
3127                                extent_data_ref_count(path, iref));
3128                         if (iref) {
3129                                 BUG_ON(path->slots[0] != extent_slot);
3130                         } else {
3131                                 BUG_ON(path->slots[0] != extent_slot + 1);
3132                                 path->slots[0] = extent_slot;
3133                                 num_to_del = 2;
3134                         }
3135                 }
3136
3137                 last_ref = 1;
3138                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
3139                                       num_to_del);
3140                 if (ret) {
3141                         btrfs_abort_transaction(trans, ret);
3142                         goto out;
3143                 }
3144                 btrfs_release_path(path);
3145
3146                 if (is_data) {
3147                         ret = btrfs_del_csums(trans, info->csum_root, bytenr,
3148                                               num_bytes);
3149                         if (ret) {
3150                                 btrfs_abort_transaction(trans, ret);
3151                                 goto out;
3152                         }
3153                 }
3154
3155                 ret = add_to_free_space_tree(trans, bytenr, num_bytes);
3156                 if (ret) {
3157                         btrfs_abort_transaction(trans, ret);
3158                         goto out;
3159                 }
3160
3161                 ret = btrfs_update_block_group(trans, bytenr, num_bytes, 0);
3162                 if (ret) {
3163                         btrfs_abort_transaction(trans, ret);
3164                         goto out;
3165                 }
3166         }
3167         btrfs_release_path(path);
3168
3169 out:
3170         btrfs_free_path(path);
3171         return ret;
3172 }
3173
3174 /*
3175  * when we free an block, it is possible (and likely) that we free the last
3176  * delayed ref for that extent as well.  This searches the delayed ref tree for
3177  * a given extent, and if there are no other delayed refs to be processed, it
3178  * removes it from the tree.
3179  */
3180 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
3181                                       u64 bytenr)
3182 {
3183         struct btrfs_delayed_ref_head *head;
3184         struct btrfs_delayed_ref_root *delayed_refs;
3185         int ret = 0;
3186
3187         delayed_refs = &trans->transaction->delayed_refs;
3188         spin_lock(&delayed_refs->lock);
3189         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
3190         if (!head)
3191                 goto out_delayed_unlock;
3192
3193         spin_lock(&head->lock);
3194         if (!RB_EMPTY_ROOT(&head->ref_tree.rb_root))
3195                 goto out;
3196
3197         if (cleanup_extent_op(head) != NULL)
3198                 goto out;
3199
3200         /*
3201          * waiting for the lock here would deadlock.  If someone else has it
3202          * locked they are already in the process of dropping it anyway
3203          */
3204         if (!mutex_trylock(&head->mutex))
3205                 goto out;
3206
3207         btrfs_delete_ref_head(delayed_refs, head);
3208         head->processing = 0;
3209
3210         spin_unlock(&head->lock);
3211         spin_unlock(&delayed_refs->lock);
3212
3213         BUG_ON(head->extent_op);
3214         if (head->must_insert_reserved)
3215                 ret = 1;
3216
3217         btrfs_cleanup_ref_head_accounting(trans->fs_info, delayed_refs, head);
3218         mutex_unlock(&head->mutex);
3219         btrfs_put_delayed_ref_head(head);
3220         return ret;
3221 out:
3222         spin_unlock(&head->lock);
3223
3224 out_delayed_unlock:
3225         spin_unlock(&delayed_refs->lock);
3226         return 0;
3227 }
3228
3229 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
3230                            struct btrfs_root *root,
3231                            struct extent_buffer *buf,
3232                            u64 parent, int last_ref)
3233 {
3234         struct btrfs_fs_info *fs_info = root->fs_info;
3235         struct btrfs_ref generic_ref = { 0 };
3236         int pin = 1;
3237         int ret;
3238
3239         btrfs_init_generic_ref(&generic_ref, BTRFS_DROP_DELAYED_REF,
3240                                buf->start, buf->len, parent);
3241         btrfs_init_tree_ref(&generic_ref, btrfs_header_level(buf),
3242                             root->root_key.objectid);
3243
3244         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
3245                 int old_ref_mod, new_ref_mod;
3246
3247                 btrfs_ref_tree_mod(fs_info, &generic_ref);
3248                 ret = btrfs_add_delayed_tree_ref(trans, &generic_ref, NULL,
3249                                                  &old_ref_mod, &new_ref_mod);
3250                 BUG_ON(ret); /* -ENOMEM */
3251                 pin = old_ref_mod >= 0 && new_ref_mod < 0;
3252         }
3253
3254         if (last_ref && btrfs_header_generation(buf) == trans->transid) {
3255                 struct btrfs_block_group *cache;
3256
3257                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
3258                         ret = check_ref_cleanup(trans, buf->start);
3259                         if (!ret)
3260                                 goto out;
3261                 }
3262
3263                 pin = 0;
3264                 cache = btrfs_lookup_block_group(fs_info, buf->start);
3265
3266                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
3267                         pin_down_extent(trans, cache, buf->start, buf->len, 1);
3268                         btrfs_put_block_group(cache);
3269                         goto out;
3270                 }
3271
3272                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
3273
3274                 btrfs_add_free_space(cache, buf->start, buf->len);
3275                 btrfs_free_reserved_bytes(cache, buf->len, 0);
3276                 btrfs_put_block_group(cache);
3277                 trace_btrfs_reserved_extent_free(fs_info, buf->start, buf->len);
3278         }
3279 out:
3280         if (pin)
3281                 add_pinned_bytes(fs_info, &generic_ref);
3282
3283         if (last_ref) {
3284                 /*
3285                  * Deleting the buffer, clear the corrupt flag since it doesn't
3286                  * matter anymore.
3287                  */
3288                 clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
3289         }
3290 }
3291
3292 /* Can return -ENOMEM */
3293 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_ref *ref)
3294 {
3295         struct btrfs_fs_info *fs_info = trans->fs_info;
3296         int old_ref_mod, new_ref_mod;
3297         int ret;
3298
3299         if (btrfs_is_testing(fs_info))
3300                 return 0;
3301
3302         /*
3303          * tree log blocks never actually go into the extent allocation
3304          * tree, just update pinning info and exit early.
3305          */
3306         if ((ref->type == BTRFS_REF_METADATA &&
3307              ref->tree_ref.root == BTRFS_TREE_LOG_OBJECTID) ||
3308             (ref->type == BTRFS_REF_DATA &&
3309              ref->data_ref.ref_root == BTRFS_TREE_LOG_OBJECTID)) {
3310                 /* unlocks the pinned mutex */
3311                 btrfs_pin_extent(trans, ref->bytenr, ref->len, 1);
3312                 old_ref_mod = new_ref_mod = 0;
3313                 ret = 0;
3314         } else if (ref->type == BTRFS_REF_METADATA) {
3315                 ret = btrfs_add_delayed_tree_ref(trans, ref, NULL,
3316                                                  &old_ref_mod, &new_ref_mod);
3317         } else {
3318                 ret = btrfs_add_delayed_data_ref(trans, ref, 0,
3319                                                  &old_ref_mod, &new_ref_mod);
3320         }
3321
3322         if (!((ref->type == BTRFS_REF_METADATA &&
3323                ref->tree_ref.root == BTRFS_TREE_LOG_OBJECTID) ||
3324               (ref->type == BTRFS_REF_DATA &&
3325                ref->data_ref.ref_root == BTRFS_TREE_LOG_OBJECTID)))
3326                 btrfs_ref_tree_mod(fs_info, ref);
3327
3328         if (ret == 0 && old_ref_mod >= 0 && new_ref_mod < 0)
3329                 add_pinned_bytes(fs_info, ref);
3330
3331         return ret;
3332 }
3333
3334 enum btrfs_loop_type {
3335         LOOP_CACHING_NOWAIT,
3336         LOOP_CACHING_WAIT,
3337         LOOP_ALLOC_CHUNK,
3338         LOOP_NO_EMPTY_SIZE,
3339 };
3340
3341 static inline void
3342 btrfs_lock_block_group(struct btrfs_block_group *cache,
3343                        int delalloc)
3344 {
3345         if (delalloc)
3346                 down_read(&cache->data_rwsem);
3347 }
3348
3349 static inline void btrfs_grab_block_group(struct btrfs_block_group *cache,
3350                        int delalloc)
3351 {
3352         btrfs_get_block_group(cache);
3353         if (delalloc)
3354                 down_read(&cache->data_rwsem);
3355 }
3356
3357 static struct btrfs_block_group *btrfs_lock_cluster(
3358                    struct btrfs_block_group *block_group,
3359                    struct btrfs_free_cluster *cluster,
3360                    int delalloc)
3361         __acquires(&cluster->refill_lock)
3362 {
3363         struct btrfs_block_group *used_bg = NULL;
3364
3365         spin_lock(&cluster->refill_lock);
3366         while (1) {
3367                 used_bg = cluster->block_group;
3368                 if (!used_bg)
3369                         return NULL;
3370
3371                 if (used_bg == block_group)
3372                         return used_bg;
3373
3374                 btrfs_get_block_group(used_bg);
3375
3376                 if (!delalloc)
3377                         return used_bg;
3378
3379                 if (down_read_trylock(&used_bg->data_rwsem))
3380                         return used_bg;
3381
3382                 spin_unlock(&cluster->refill_lock);
3383
3384                 /* We should only have one-level nested. */
3385                 down_read_nested(&used_bg->data_rwsem, SINGLE_DEPTH_NESTING);
3386
3387                 spin_lock(&cluster->refill_lock);
3388                 if (used_bg == cluster->block_group)
3389                         return used_bg;
3390
3391                 up_read(&used_bg->data_rwsem);
3392                 btrfs_put_block_group(used_bg);
3393         }
3394 }
3395
3396 static inline void
3397 btrfs_release_block_group(struct btrfs_block_group *cache,
3398                          int delalloc)
3399 {
3400         if (delalloc)
3401                 up_read(&cache->data_rwsem);
3402         btrfs_put_block_group(cache);
3403 }
3404
3405 enum btrfs_extent_allocation_policy {
3406         BTRFS_EXTENT_ALLOC_CLUSTERED,
3407 };
3408
3409 /*
3410  * Structure used internally for find_free_extent() function.  Wraps needed
3411  * parameters.
3412  */
3413 struct find_free_extent_ctl {
3414         /* Basic allocation info */
3415         u64 num_bytes;
3416         u64 empty_size;
3417         u64 flags;
3418         int delalloc;
3419
3420         /* Where to start the search inside the bg */
3421         u64 search_start;
3422
3423         /* For clustered allocation */
3424         u64 empty_cluster;
3425         struct btrfs_free_cluster *last_ptr;
3426         bool use_cluster;
3427
3428         bool have_caching_bg;
3429         bool orig_have_caching_bg;
3430
3431         /* RAID index, converted from flags */
3432         int index;
3433
3434         /*
3435          * Current loop number, check find_free_extent_update_loop() for details
3436          */
3437         int loop;
3438
3439         /*
3440          * Whether we're refilling a cluster, if true we need to re-search
3441          * current block group but don't try to refill the cluster again.
3442          */
3443         bool retry_clustered;
3444
3445         /*
3446          * Whether we're updating free space cache, if true we need to re-search
3447          * current block group but don't try updating free space cache again.
3448          */
3449         bool retry_unclustered;
3450
3451         /* If current block group is cached */
3452         int cached;
3453
3454         /* Max contiguous hole found */
3455         u64 max_extent_size;
3456
3457         /* Total free space from free space cache, not always contiguous */
3458         u64 total_free_space;
3459
3460         /* Found result */
3461         u64 found_offset;
3462
3463         /* Hint where to start looking for an empty space */
3464         u64 hint_byte;
3465
3466         /* Allocation policy */
3467         enum btrfs_extent_allocation_policy policy;
3468 };
3469
3470
3471 /*
3472  * Helper function for find_free_extent().
3473  *
3474  * Return -ENOENT to inform caller that we need fallback to unclustered mode.
3475  * Return -EAGAIN to inform caller that we need to re-search this block group
3476  * Return >0 to inform caller that we find nothing
3477  * Return 0 means we have found a location and set ffe_ctl->found_offset.
3478  */
3479 static int find_free_extent_clustered(struct btrfs_block_group *bg,
3480                                       struct find_free_extent_ctl *ffe_ctl,
3481                                       struct btrfs_block_group **cluster_bg_ret)
3482 {
3483         struct btrfs_block_group *cluster_bg;
3484         struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
3485         u64 aligned_cluster;
3486         u64 offset;
3487         int ret;
3488
3489         cluster_bg = btrfs_lock_cluster(bg, last_ptr, ffe_ctl->delalloc);
3490         if (!cluster_bg)
3491                 goto refill_cluster;
3492         if (cluster_bg != bg && (cluster_bg->ro ||
3493             !block_group_bits(cluster_bg, ffe_ctl->flags)))
3494                 goto release_cluster;
3495
3496         offset = btrfs_alloc_from_cluster(cluster_bg, last_ptr,
3497                         ffe_ctl->num_bytes, cluster_bg->start,
3498                         &ffe_ctl->max_extent_size);
3499         if (offset) {
3500                 /* We have a block, we're done */
3501                 spin_unlock(&last_ptr->refill_lock);
3502                 trace_btrfs_reserve_extent_cluster(cluster_bg,
3503                                 ffe_ctl->search_start, ffe_ctl->num_bytes);
3504                 *cluster_bg_ret = cluster_bg;
3505                 ffe_ctl->found_offset = offset;
3506                 return 0;
3507         }
3508         WARN_ON(last_ptr->block_group != cluster_bg);
3509
3510 release_cluster:
3511         /*
3512          * If we are on LOOP_NO_EMPTY_SIZE, we can't set up a new clusters, so
3513          * lets just skip it and let the allocator find whatever block it can
3514          * find. If we reach this point, we will have tried the cluster
3515          * allocator plenty of times and not have found anything, so we are
3516          * likely way too fragmented for the clustering stuff to find anything.
3517          *
3518          * However, if the cluster is taken from the current block group,
3519          * release the cluster first, so that we stand a better chance of
3520          * succeeding in the unclustered allocation.
3521          */
3522         if (ffe_ctl->loop >= LOOP_NO_EMPTY_SIZE && cluster_bg != bg) {
3523                 spin_unlock(&last_ptr->refill_lock);
3524                 btrfs_release_block_group(cluster_bg, ffe_ctl->delalloc);
3525                 return -ENOENT;
3526         }
3527
3528         /* This cluster didn't work out, free it and start over */
3529         btrfs_return_cluster_to_free_space(NULL, last_ptr);
3530
3531         if (cluster_bg != bg)
3532                 btrfs_release_block_group(cluster_bg, ffe_ctl->delalloc);
3533
3534 refill_cluster:
3535         if (ffe_ctl->loop >= LOOP_NO_EMPTY_SIZE) {
3536                 spin_unlock(&last_ptr->refill_lock);
3537                 return -ENOENT;
3538         }
3539
3540         aligned_cluster = max_t(u64,
3541                         ffe_ctl->empty_cluster + ffe_ctl->empty_size,
3542                         bg->full_stripe_len);
3543         ret = btrfs_find_space_cluster(bg, last_ptr, ffe_ctl->search_start,
3544                         ffe_ctl->num_bytes, aligned_cluster);
3545         if (ret == 0) {
3546                 /* Now pull our allocation out of this cluster */
3547                 offset = btrfs_alloc_from_cluster(bg, last_ptr,
3548                                 ffe_ctl->num_bytes, ffe_ctl->search_start,
3549                                 &ffe_ctl->max_extent_size);
3550                 if (offset) {
3551                         /* We found one, proceed */
3552                         spin_unlock(&last_ptr->refill_lock);
3553                         trace_btrfs_reserve_extent_cluster(bg,
3554                                         ffe_ctl->search_start,
3555                                         ffe_ctl->num_bytes);
3556                         ffe_ctl->found_offset = offset;
3557                         return 0;
3558                 }
3559         } else if (!ffe_ctl->cached && ffe_ctl->loop > LOOP_CACHING_NOWAIT &&
3560                    !ffe_ctl->retry_clustered) {
3561                 spin_unlock(&last_ptr->refill_lock);
3562
3563                 ffe_ctl->retry_clustered = true;
3564                 btrfs_wait_block_group_cache_progress(bg, ffe_ctl->num_bytes +
3565                                 ffe_ctl->empty_cluster + ffe_ctl->empty_size);
3566                 return -EAGAIN;
3567         }
3568         /*
3569          * At this point we either didn't find a cluster or we weren't able to
3570          * allocate a block from our cluster.  Free the cluster we've been
3571          * trying to use, and go to the next block group.
3572          */
3573         btrfs_return_cluster_to_free_space(NULL, last_ptr);
3574         spin_unlock(&last_ptr->refill_lock);
3575         return 1;
3576 }
3577
3578 /*
3579  * Return >0 to inform caller that we find nothing
3580  * Return 0 when we found an free extent and set ffe_ctrl->found_offset
3581  * Return -EAGAIN to inform caller that we need to re-search this block group
3582  */
3583 static int find_free_extent_unclustered(struct btrfs_block_group *bg,
3584                                         struct find_free_extent_ctl *ffe_ctl)
3585 {
3586         struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
3587         u64 offset;
3588
3589         /*
3590          * We are doing an unclustered allocation, set the fragmented flag so
3591          * we don't bother trying to setup a cluster again until we get more
3592          * space.
3593          */
3594         if (unlikely(last_ptr)) {
3595                 spin_lock(&last_ptr->lock);
3596                 last_ptr->fragmented = 1;
3597                 spin_unlock(&last_ptr->lock);
3598         }
3599         if (ffe_ctl->cached) {
3600                 struct btrfs_free_space_ctl *free_space_ctl;
3601
3602                 free_space_ctl = bg->free_space_ctl;
3603                 spin_lock(&free_space_ctl->tree_lock);
3604                 if (free_space_ctl->free_space <
3605                     ffe_ctl->num_bytes + ffe_ctl->empty_cluster +
3606                     ffe_ctl->empty_size) {
3607                         ffe_ctl->total_free_space = max_t(u64,
3608                                         ffe_ctl->total_free_space,
3609                                         free_space_ctl->free_space);
3610                         spin_unlock(&free_space_ctl->tree_lock);
3611                         return 1;
3612                 }
3613                 spin_unlock(&free_space_ctl->tree_lock);
3614         }
3615
3616         offset = btrfs_find_space_for_alloc(bg, ffe_ctl->search_start,
3617                         ffe_ctl->num_bytes, ffe_ctl->empty_size,
3618                         &ffe_ctl->max_extent_size);
3619
3620         /*
3621          * If we didn't find a chunk, and we haven't failed on this block group
3622          * before, and this block group is in the middle of caching and we are
3623          * ok with waiting, then go ahead and wait for progress to be made, and
3624          * set @retry_unclustered to true.
3625          *
3626          * If @retry_unclustered is true then we've already waited on this
3627          * block group once and should move on to the next block group.
3628          */
3629         if (!offset && !ffe_ctl->retry_unclustered && !ffe_ctl->cached &&
3630             ffe_ctl->loop > LOOP_CACHING_NOWAIT) {
3631                 btrfs_wait_block_group_cache_progress(bg, ffe_ctl->num_bytes +
3632                                                       ffe_ctl->empty_size);
3633                 ffe_ctl->retry_unclustered = true;
3634                 return -EAGAIN;
3635         } else if (!offset) {
3636                 return 1;
3637         }
3638         ffe_ctl->found_offset = offset;
3639         return 0;
3640 }
3641
3642 static int do_allocation_clustered(struct btrfs_block_group *block_group,
3643                                    struct find_free_extent_ctl *ffe_ctl,
3644                                    struct btrfs_block_group **bg_ret)
3645 {
3646         int ret;
3647
3648         /* We want to try and use the cluster allocator, so lets look there */
3649         if (ffe_ctl->last_ptr && ffe_ctl->use_cluster) {
3650                 ret = find_free_extent_clustered(block_group, ffe_ctl, bg_ret);
3651                 if (ret >= 0 || ret == -EAGAIN)
3652                         return ret;
3653                 /* ret == -ENOENT case falls through */
3654         }
3655
3656         return find_free_extent_unclustered(block_group, ffe_ctl);
3657 }
3658
3659 static int do_allocation(struct btrfs_block_group *block_group,
3660                          struct find_free_extent_ctl *ffe_ctl,
3661                          struct btrfs_block_group **bg_ret)
3662 {
3663         switch (ffe_ctl->policy) {
3664         case BTRFS_EXTENT_ALLOC_CLUSTERED:
3665                 return do_allocation_clustered(block_group, ffe_ctl, bg_ret);
3666         default:
3667                 BUG();
3668         }
3669 }
3670
3671 static void release_block_group(struct btrfs_block_group *block_group,
3672                                 struct find_free_extent_ctl *ffe_ctl,
3673                                 int delalloc)
3674 {
3675         switch (ffe_ctl->policy) {
3676         case BTRFS_EXTENT_ALLOC_CLUSTERED:
3677                 ffe_ctl->retry_clustered = false;
3678                 ffe_ctl->retry_unclustered = false;
3679                 break;
3680         default:
3681                 BUG();
3682         }
3683
3684         BUG_ON(btrfs_bg_flags_to_raid_index(block_group->flags) !=
3685                ffe_ctl->index);
3686         btrfs_release_block_group(block_group, delalloc);
3687 }
3688
3689 static void found_extent_clustered(struct find_free_extent_ctl *ffe_ctl,
3690                                    struct btrfs_key *ins)
3691 {
3692         struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
3693
3694         if (!ffe_ctl->use_cluster && last_ptr) {
3695                 spin_lock(&last_ptr->lock);
3696                 last_ptr->window_start = ins->objectid;
3697                 spin_unlock(&last_ptr->lock);
3698         }
3699 }
3700
3701 static void found_extent(struct find_free_extent_ctl *ffe_ctl,
3702                          struct btrfs_key *ins)
3703 {
3704         switch (ffe_ctl->policy) {
3705         case BTRFS_EXTENT_ALLOC_CLUSTERED:
3706                 found_extent_clustered(ffe_ctl, ins);
3707                 break;
3708         default:
3709                 BUG();
3710         }
3711 }
3712
3713 static int chunk_allocation_failed(struct find_free_extent_ctl *ffe_ctl)
3714 {
3715         switch (ffe_ctl->policy) {
3716         case BTRFS_EXTENT_ALLOC_CLUSTERED:
3717                 /*
3718                  * If we can't allocate a new chunk we've already looped through
3719                  * at least once, move on to the NO_EMPTY_SIZE case.
3720                  */
3721                 ffe_ctl->loop = LOOP_NO_EMPTY_SIZE;
3722                 return 0;
3723         default:
3724                 BUG();
3725         }
3726 }
3727
3728 /*
3729  * Return >0 means caller needs to re-search for free extent
3730  * Return 0 means we have the needed free extent.
3731  * Return <0 means we failed to locate any free extent.
3732  */
3733 static int find_free_extent_update_loop(struct btrfs_fs_info *fs_info,
3734                                         struct btrfs_key *ins,
3735                                         struct find_free_extent_ctl *ffe_ctl,
3736                                         bool full_search)
3737 {
3738         struct btrfs_root *root = fs_info->extent_root;
3739         int ret;
3740
3741         if ((ffe_ctl->loop == LOOP_CACHING_NOWAIT) &&
3742             ffe_ctl->have_caching_bg && !ffe_ctl->orig_have_caching_bg)
3743                 ffe_ctl->orig_have_caching_bg = true;
3744
3745         if (!ins->objectid && ffe_ctl->loop >= LOOP_CACHING_WAIT &&
3746             ffe_ctl->have_caching_bg)
3747                 return 1;
3748
3749         if (!ins->objectid && ++(ffe_ctl->index) < BTRFS_NR_RAID_TYPES)
3750                 return 1;
3751
3752         if (ins->objectid) {
3753                 found_extent(ffe_ctl, ins);
3754                 return 0;
3755         }
3756
3757         /*
3758          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
3759          *                      caching kthreads as we move along
3760          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
3761          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
3762          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
3763          *                     again
3764          */
3765         if (ffe_ctl->loop < LOOP_NO_EMPTY_SIZE) {
3766                 ffe_ctl->index = 0;
3767                 if (ffe_ctl->loop == LOOP_CACHING_NOWAIT) {
3768                         /*
3769                          * We want to skip the LOOP_CACHING_WAIT step if we
3770                          * don't have any uncached bgs and we've already done a
3771                          * full search through.
3772                          */
3773                         if (ffe_ctl->orig_have_caching_bg || !full_search)
3774                                 ffe_ctl->loop = LOOP_CACHING_WAIT;
3775                         else
3776                                 ffe_ctl->loop = LOOP_ALLOC_CHUNK;
3777                 } else {
3778                         ffe_ctl->loop++;
3779                 }
3780
3781                 if (ffe_ctl->loop == LOOP_ALLOC_CHUNK) {
3782                         struct btrfs_trans_handle *trans;
3783                         int exist = 0;
3784
3785                         trans = current->journal_info;
3786                         if (trans)
3787                                 exist = 1;
3788                         else
3789                                 trans = btrfs_join_transaction(root);
3790
3791                         if (IS_ERR(trans)) {
3792                                 ret = PTR_ERR(trans);
3793                                 return ret;
3794                         }
3795
3796                         ret = btrfs_chunk_alloc(trans, ffe_ctl->flags,
3797                                                 CHUNK_ALLOC_FORCE);
3798
3799                         /* Do not bail out on ENOSPC since we can do more. */
3800                         if (ret == -ENOSPC)
3801                                 ret = chunk_allocation_failed(ffe_ctl);
3802                         else if (ret < 0)
3803                                 btrfs_abort_transaction(trans, ret);
3804                         else
3805                                 ret = 0;
3806                         if (!exist)
3807                                 btrfs_end_transaction(trans);
3808                         if (ret)
3809                                 return ret;
3810                 }
3811
3812                 if (ffe_ctl->loop == LOOP_NO_EMPTY_SIZE) {
3813                         if (ffe_ctl->policy != BTRFS_EXTENT_ALLOC_CLUSTERED)
3814                                 return -ENOSPC;
3815
3816                         /*
3817                          * Don't loop again if we already have no empty_size and
3818                          * no empty_cluster.
3819                          */
3820                         if (ffe_ctl->empty_size == 0 &&
3821                             ffe_ctl->empty_cluster == 0)
3822                                 return -ENOSPC;
3823                         ffe_ctl->empty_size = 0;
3824                         ffe_ctl->empty_cluster = 0;
3825                 }
3826                 return 1;
3827         }
3828         return -ENOSPC;
3829 }
3830
3831 static int prepare_allocation_clustered(struct btrfs_fs_info *fs_info,
3832                                         struct find_free_extent_ctl *ffe_ctl,
3833                                         struct btrfs_space_info *space_info,
3834                                         struct btrfs_key *ins)
3835 {
3836         /*
3837          * If our free space is heavily fragmented we may not be able to make
3838          * big contiguous allocations, so instead of doing the expensive search
3839          * for free space, simply return ENOSPC with our max_extent_size so we
3840          * can go ahead and search for a more manageable chunk.
3841          *
3842          * If our max_extent_size is large enough for our allocation simply
3843          * disable clustering since we will likely not be able to find enough
3844          * space to create a cluster and induce latency trying.
3845          */
3846         if (space_info->max_extent_size) {
3847                 spin_lock(&space_info->lock);
3848                 if (space_info->max_extent_size &&
3849                     ffe_ctl->num_bytes > space_info->max_extent_size) {
3850                         ins->offset = space_info->max_extent_size;
3851                         spin_unlock(&space_info->lock);
3852                         return -ENOSPC;
3853                 } else if (space_info->max_extent_size) {
3854                         ffe_ctl->use_cluster = false;
3855                 }
3856                 spin_unlock(&space_info->lock);
3857         }
3858
3859         ffe_ctl->last_ptr = fetch_cluster_info(fs_info, space_info,
3860                                                &ffe_ctl->empty_cluster);
3861         if (ffe_ctl->last_ptr) {
3862                 struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
3863
3864                 spin_lock(&last_ptr->lock);
3865                 if (last_ptr->block_group)
3866                         ffe_ctl->hint_byte = last_ptr->window_start;
3867                 if (last_ptr->fragmented) {
3868                         /*
3869                          * We still set window_start so we can keep track of the
3870                          * last place we found an allocation to try and save
3871                          * some time.
3872                          */
3873                         ffe_ctl->hint_byte = last_ptr->window_start;
3874                         ffe_ctl->use_cluster = false;
3875                 }
3876                 spin_unlock(&last_ptr->lock);
3877         }
3878
3879         return 0;
3880 }
3881
3882 static int prepare_allocation(struct btrfs_fs_info *fs_info,
3883                               struct find_free_extent_ctl *ffe_ctl,
3884                               struct btrfs_space_info *space_info,
3885                               struct btrfs_key *ins)
3886 {
3887         switch (ffe_ctl->policy) {
3888         case BTRFS_EXTENT_ALLOC_CLUSTERED:
3889                 return prepare_allocation_clustered(fs_info, ffe_ctl,
3890                                                     space_info, ins);
3891         default:
3892                 BUG();
3893         }
3894 }
3895
3896 /*
3897  * walks the btree of allocated extents and find a hole of a given size.
3898  * The key ins is changed to record the hole:
3899  * ins->objectid == start position
3900  * ins->flags = BTRFS_EXTENT_ITEM_KEY
3901  * ins->offset == the size of the hole.
3902  * Any available blocks before search_start are skipped.
3903  *
3904  * If there is no suitable free space, we will record the max size of
3905  * the free space extent currently.
3906  *
3907  * The overall logic and call chain:
3908  *
3909  * find_free_extent()
3910  * |- Iterate through all block groups
3911  * |  |- Get a valid block group
3912  * |  |- Try to do clustered allocation in that block group
3913  * |  |- Try to do unclustered allocation in that block group
3914  * |  |- Check if the result is valid
3915  * |  |  |- If valid, then exit
3916  * |  |- Jump to next block group
3917  * |
3918  * |- Push harder to find free extents
3919  *    |- If not found, re-iterate all block groups
3920  */
3921 static noinline int find_free_extent(struct btrfs_fs_info *fs_info,
3922                                 u64 ram_bytes, u64 num_bytes, u64 empty_size,
3923                                 u64 hint_byte_orig, struct btrfs_key *ins,
3924                                 u64 flags, int delalloc)
3925 {
3926         int ret = 0;
3927         int cache_block_group_error = 0;
3928         struct btrfs_block_group *block_group = NULL;
3929         struct find_free_extent_ctl ffe_ctl = {0};
3930         struct btrfs_space_info *space_info;
3931         bool full_search = false;
3932
3933         WARN_ON(num_bytes < fs_info->sectorsize);
3934
3935         ffe_ctl.num_bytes = num_bytes;
3936         ffe_ctl.empty_size = empty_size;
3937         ffe_ctl.flags = flags;
3938         ffe_ctl.search_start = 0;
3939         ffe_ctl.delalloc = delalloc;
3940         ffe_ctl.index = btrfs_bg_flags_to_raid_index(flags);
3941         ffe_ctl.have_caching_bg = false;
3942         ffe_ctl.orig_have_caching_bg = false;
3943         ffe_ctl.found_offset = 0;
3944         ffe_ctl.hint_byte = hint_byte_orig;
3945         ffe_ctl.policy = BTRFS_EXTENT_ALLOC_CLUSTERED;
3946
3947         /* For clustered allocation */
3948         ffe_ctl.retry_clustered = false;
3949         ffe_ctl.retry_unclustered = false;
3950         ffe_ctl.last_ptr = NULL;
3951         ffe_ctl.use_cluster = true;
3952
3953         ins->type = BTRFS_EXTENT_ITEM_KEY;
3954         ins->objectid = 0;
3955         ins->offset = 0;
3956
3957         trace_find_free_extent(fs_info, num_bytes, empty_size, flags);
3958
3959         space_info = btrfs_find_space_info(fs_info, flags);
3960         if (!space_info) {
3961                 btrfs_err(fs_info, "No space info for %llu", flags);
3962                 return -ENOSPC;
3963         }
3964
3965         ret = prepare_allocation(fs_info, &ffe_ctl, space_info, ins);
3966         if (ret < 0)
3967                 return ret;
3968
3969         ffe_ctl.search_start = max(ffe_ctl.search_start,
3970                                    first_logical_byte(fs_info, 0));
3971         ffe_ctl.search_start = max(ffe_ctl.search_start, ffe_ctl.hint_byte);
3972         if (ffe_ctl.search_start == ffe_ctl.hint_byte) {
3973                 block_group = btrfs_lookup_block_group(fs_info,
3974                                                        ffe_ctl.search_start);
3975                 /*
3976                  * we don't want to use the block group if it doesn't match our
3977                  * allocation bits, or if its not cached.
3978                  *
3979                  * However if we are re-searching with an ideal block group
3980                  * picked out then we don't care that the block group is cached.
3981                  */
3982                 if (block_group && block_group_bits(block_group, flags) &&
3983                     block_group->cached != BTRFS_CACHE_NO) {
3984                         down_read(&space_info->groups_sem);
3985                         if (list_empty(&block_group->list) ||
3986                             block_group->ro) {
3987                                 /*
3988                                  * someone is removing this block group,
3989                                  * we can't jump into the have_block_group
3990                                  * target because our list pointers are not
3991                                  * valid
3992                                  */
3993                                 btrfs_put_block_group(block_group);
3994                                 up_read(&space_info->groups_sem);
3995                         } else {
3996                                 ffe_ctl.index = btrfs_bg_flags_to_raid_index(
3997                                                 block_group->flags);
3998                                 btrfs_lock_block_group(block_group, delalloc);
3999                                 goto have_block_group;
4000                         }
4001                 } else if (block_group) {
4002                         btrfs_put_block_group(block_group);
4003                 }
4004         }
4005 search:
4006         ffe_ctl.have_caching_bg = false;
4007         if (ffe_ctl.index == btrfs_bg_flags_to_raid_index(flags) ||
4008             ffe_ctl.index == 0)
4009                 full_search = true;
4010         down_read(&space_info->groups_sem);
4011         list_for_each_entry(block_group,
4012                             &space_info->block_groups[ffe_ctl.index], list) {
4013                 struct btrfs_block_group *bg_ret;
4014
4015                 /* If the block group is read-only, we can skip it entirely. */
4016                 if (unlikely(block_group->ro))
4017                         continue;
4018
4019                 btrfs_grab_block_group(block_group, delalloc);
4020                 ffe_ctl.search_start = block_group->start;
4021
4022                 /*
4023                  * this can happen if we end up cycling through all the
4024                  * raid types, but we want to make sure we only allocate
4025                  * for the proper type.
4026                  */
4027                 if (!block_group_bits(block_group, flags)) {
4028                         u64 extra = BTRFS_BLOCK_GROUP_DUP |
4029                                 BTRFS_BLOCK_GROUP_RAID1_MASK |
4030                                 BTRFS_BLOCK_GROUP_RAID56_MASK |
4031                                 BTRFS_BLOCK_GROUP_RAID10;
4032
4033                         /*
4034                          * if they asked for extra copies and this block group
4035                          * doesn't provide them, bail.  This does allow us to
4036                          * fill raid0 from raid1.
4037                          */
4038                         if ((flags & extra) && !(block_group->flags & extra))
4039                                 goto loop;
4040
4041                         /*
4042                          * This block group has different flags than we want.
4043                          * It's possible that we have MIXED_GROUP flag but no
4044                          * block group is mixed.  Just skip such block group.
4045                          */
4046                         btrfs_release_block_group(block_group, delalloc);
4047                         continue;
4048                 }
4049
4050 have_block_group:
4051                 ffe_ctl.cached = btrfs_block_group_done(block_group);
4052                 if (unlikely(!ffe_ctl.cached)) {
4053                         ffe_ctl.have_caching_bg = true;
4054                         ret = btrfs_cache_block_group(block_group, 0);
4055
4056                         /*
4057                          * If we get ENOMEM here or something else we want to
4058                          * try other block groups, because it may not be fatal.
4059                          * However if we can't find anything else we need to
4060                          * save our return here so that we return the actual
4061                          * error that caused problems, not ENOSPC.
4062                          */
4063                         if (ret < 0) {
4064                                 if (!cache_block_group_error)
4065                                         cache_block_group_error = ret;
4066                                 ret = 0;
4067                                 goto loop;
4068                         }
4069                         ret = 0;
4070                 }
4071
4072                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
4073                         goto loop;
4074
4075                 bg_ret = NULL;
4076                 ret = do_allocation(block_group, &ffe_ctl, &bg_ret);
4077                 if (ret == 0) {
4078                         if (bg_ret && bg_ret != block_group) {
4079                                 btrfs_release_block_group(block_group, delalloc);
4080                                 block_group = bg_ret;
4081                         }
4082                 } else if (ret == -EAGAIN) {
4083                         goto have_block_group;
4084                 } else if (ret > 0) {
4085                         goto loop;
4086                 }
4087
4088                 /* Checks */
4089                 ffe_ctl.search_start = round_up(ffe_ctl.found_offset,
4090                                              fs_info->stripesize);
4091
4092                 /* move on to the next group */
4093                 if (ffe_ctl.search_start + num_bytes >
4094                     block_group->start + block_group->length) {
4095                         btrfs_add_free_space(block_group, ffe_ctl.found_offset,
4096                                              num_bytes);
4097                         goto loop;
4098                 }
4099
4100                 if (ffe_ctl.found_offset < ffe_ctl.search_start)
4101                         btrfs_add_free_space(block_group, ffe_ctl.found_offset,
4102                                 ffe_ctl.search_start - ffe_ctl.found_offset);
4103
4104                 ret = btrfs_add_reserved_bytes(block_group, ram_bytes,
4105                                 num_bytes, delalloc);
4106                 if (ret == -EAGAIN) {
4107                         btrfs_add_free_space(block_group, ffe_ctl.found_offset,
4108                                              num_bytes);
4109                         goto loop;
4110                 }
4111                 btrfs_inc_block_group_reservations(block_group);
4112
4113                 /* we are all good, lets return */
4114                 ins->objectid = ffe_ctl.search_start;
4115                 ins->offset = num_bytes;
4116
4117                 trace_btrfs_reserve_extent(block_group, ffe_ctl.search_start,
4118                                            num_bytes);
4119                 btrfs_release_block_group(block_group, delalloc);
4120                 break;
4121 loop:
4122                 release_block_group(block_group, &ffe_ctl, delalloc);
4123                 cond_resched();
4124         }
4125         up_read(&space_info->groups_sem);
4126
4127         ret = find_free_extent_update_loop(fs_info, ins, &ffe_ctl, full_search);
4128         if (ret > 0)
4129                 goto search;
4130
4131         if (ret == -ENOSPC && !cache_block_group_error) {
4132                 /*
4133                  * Use ffe_ctl->total_free_space as fallback if we can't find
4134                  * any contiguous hole.
4135                  */
4136                 if (!ffe_ctl.max_extent_size)
4137                         ffe_ctl.max_extent_size = ffe_ctl.total_free_space;
4138                 spin_lock(&space_info->lock);
4139                 space_info->max_extent_size = ffe_ctl.max_extent_size;
4140                 spin_unlock(&space_info->lock);
4141                 ins->offset = ffe_ctl.max_extent_size;
4142         } else if (ret == -ENOSPC) {
4143                 ret = cache_block_group_error;
4144         }
4145         return ret;
4146 }
4147
4148 /*
4149  * btrfs_reserve_extent - entry point to the extent allocator. Tries to find a
4150  *                        hole that is at least as big as @num_bytes.
4151  *
4152  * @root           -    The root that will contain this extent
4153  *
4154  * @ram_bytes      -    The amount of space in ram that @num_bytes take. This
4155  *                      is used for accounting purposes. This value differs
4156  *                      from @num_bytes only in the case of compressed extents.
4157  *
4158  * @num_bytes      -    Number of bytes to allocate on-disk.
4159  *
4160  * @min_alloc_size -    Indicates the minimum amount of space that the
4161  *                      allocator should try to satisfy. In some cases
4162  *                      @num_bytes may be larger than what is required and if
4163  *                      the filesystem is fragmented then allocation fails.
4164  *                      However, the presence of @min_alloc_size gives a
4165  *                      chance to try and satisfy the smaller allocation.
4166  *
4167  * @empty_size     -    A hint that you plan on doing more COW. This is the
4168  *                      size in bytes the allocator should try to find free
4169  *                      next to the block it returns.  This is just a hint and
4170  *                      may be ignored by the allocator.
4171  *
4172  * @hint_byte      -    Hint to the allocator to start searching above the byte
4173  *                      address passed. It might be ignored.
4174  *
4175  * @ins            -    This key is modified to record the found hole. It will
4176  *                      have the following values:
4177  *                      ins->objectid == start position
4178  *                      ins->flags = BTRFS_EXTENT_ITEM_KEY
4179  *                      ins->offset == the size of the hole.
4180  *
4181  * @is_data        -    Boolean flag indicating whether an extent is
4182  *                      allocated for data (true) or metadata (false)
4183  *
4184  * @delalloc       -    Boolean flag indicating whether this allocation is for
4185  *                      delalloc or not. If 'true' data_rwsem of block groups
4186  *                      is going to be acquired.
4187  *
4188  *
4189  * Returns 0 when an allocation succeeded or < 0 when an error occurred. In
4190  * case -ENOSPC is returned then @ins->offset will contain the size of the
4191  * largest available hole the allocator managed to find.
4192  */
4193 int btrfs_reserve_extent(struct btrfs_root *root, u64 ram_bytes,
4194                          u64 num_bytes, u64 min_alloc_size,
4195                          u64 empty_size, u64 hint_byte,
4196                          struct btrfs_key *ins, int is_data, int delalloc)
4197 {
4198         struct btrfs_fs_info *fs_info = root->fs_info;
4199         bool final_tried = num_bytes == min_alloc_size;
4200         u64 flags;
4201         int ret;
4202
4203         flags = get_alloc_profile_by_root(root, is_data);
4204 again:
4205         WARN_ON(num_bytes < fs_info->sectorsize);
4206         ret = find_free_extent(fs_info, ram_bytes, num_bytes, empty_size,
4207                                hint_byte, ins, flags, delalloc);
4208         if (!ret && !is_data) {
4209                 btrfs_dec_block_group_reservations(fs_info, ins->objectid);
4210         } else if (ret == -ENOSPC) {
4211                 if (!final_tried && ins->offset) {
4212                         num_bytes = min(num_bytes >> 1, ins->offset);
4213                         num_bytes = round_down(num_bytes,
4214                                                fs_info->sectorsize);
4215                         num_bytes = max(num_bytes, min_alloc_size);
4216                         ram_bytes = num_bytes;
4217                         if (num_bytes == min_alloc_size)
4218                                 final_tried = true;
4219                         goto again;
4220                 } else if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
4221                         struct btrfs_space_info *sinfo;
4222
4223                         sinfo = btrfs_find_space_info(fs_info, flags);
4224                         btrfs_err(fs_info,
4225                                   "allocation failed flags %llu, wanted %llu",
4226                                   flags, num_bytes);
4227                         if (sinfo)
4228                                 btrfs_dump_space_info(fs_info, sinfo,
4229                                                       num_bytes, 1);
4230                 }
4231         }
4232
4233         return ret;
4234 }
4235
4236 int btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
4237                                u64 start, u64 len, int delalloc)
4238 {
4239         struct btrfs_block_group *cache;
4240
4241         cache = btrfs_lookup_block_group(fs_info, start);
4242         if (!cache) {
4243                 btrfs_err(fs_info, "Unable to find block group for %llu",
4244                           start);
4245                 return -ENOSPC;
4246         }
4247
4248         btrfs_add_free_space(cache, start, len);
4249         btrfs_free_reserved_bytes(cache, len, delalloc);
4250         trace_btrfs_reserved_extent_free(fs_info, start, len);
4251
4252         btrfs_put_block_group(cache);
4253         return 0;
4254 }
4255
4256 int btrfs_pin_reserved_extent(struct btrfs_trans_handle *trans, u64 start,
4257                               u64 len)
4258 {
4259         struct btrfs_block_group *cache;
4260         int ret = 0;
4261
4262         cache = btrfs_lookup_block_group(trans->fs_info, start);
4263         if (!cache) {
4264                 btrfs_err(trans->fs_info, "unable to find block group for %llu",
4265                           start);
4266                 return -ENOSPC;
4267         }
4268
4269         ret = pin_down_extent(trans, cache, start, len, 1);
4270         btrfs_put_block_group(cache);
4271         return ret;
4272 }
4273
4274 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
4275                                       u64 parent, u64 root_objectid,
4276                                       u64 flags, u64 owner, u64 offset,
4277                                       struct btrfs_key *ins, int ref_mod)
4278 {
4279         struct btrfs_fs_info *fs_info = trans->fs_info;
4280         int ret;
4281         struct btrfs_extent_item *extent_item;
4282         struct btrfs_extent_inline_ref *iref;
4283         struct btrfs_path *path;
4284         struct extent_buffer *leaf;
4285         int type;
4286         u32 size;
4287
4288         if (parent > 0)
4289                 type = BTRFS_SHARED_DATA_REF_KEY;
4290         else
4291                 type = BTRFS_EXTENT_DATA_REF_KEY;
4292
4293         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
4294
4295         path = btrfs_alloc_path();
4296         if (!path)
4297                 return -ENOMEM;
4298
4299         path->leave_spinning = 1;
4300         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
4301                                       ins, size);
4302         if (ret) {
4303                 btrfs_free_path(path);
4304                 return ret;
4305         }
4306
4307         leaf = path->nodes[0];
4308         extent_item = btrfs_item_ptr(leaf, path->slots[0],
4309                                      struct btrfs_extent_item);
4310         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
4311         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
4312         btrfs_set_extent_flags(leaf, extent_item,
4313                                flags | BTRFS_EXTENT_FLAG_DATA);
4314
4315         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
4316         btrfs_set_extent_inline_ref_type(leaf, iref, type);
4317         if (parent > 0) {
4318                 struct btrfs_shared_data_ref *ref;
4319                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
4320                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
4321                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
4322         } else {
4323                 struct btrfs_extent_data_ref *ref;
4324                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
4325                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
4326                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
4327                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
4328                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
4329         }
4330
4331         btrfs_mark_buffer_dirty(path->nodes[0]);
4332         btrfs_free_path(path);
4333
4334         ret = remove_from_free_space_tree(trans, ins->objectid, ins->offset);
4335         if (ret)
4336                 return ret;
4337
4338         ret = btrfs_update_block_group(trans, ins->objectid, ins->offset, 1);
4339         if (ret) { /* -ENOENT, logic error */
4340                 btrfs_err(fs_info, "update block group failed for %llu %llu",
4341                         ins->objectid, ins->offset);
4342                 BUG();
4343         }
4344         trace_btrfs_reserved_extent_alloc(fs_info, ins->objectid, ins->offset);
4345         return ret;
4346 }
4347
4348 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
4349                                      struct btrfs_delayed_ref_node *node,
4350                                      struct btrfs_delayed_extent_op *extent_op)
4351 {
4352         struct btrfs_fs_info *fs_info = trans->fs_info;
4353         int ret;
4354         struct btrfs_extent_item *extent_item;
4355         struct btrfs_key extent_key;
4356         struct btrfs_tree_block_info *block_info;
4357         struct btrfs_extent_inline_ref *iref;
4358         struct btrfs_path *path;
4359         struct extent_buffer *leaf;
4360         struct btrfs_delayed_tree_ref *ref;
4361         u32 size = sizeof(*extent_item) + sizeof(*iref);
4362         u64 num_bytes;
4363         u64 flags = extent_op->flags_to_set;
4364         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
4365
4366         ref = btrfs_delayed_node_to_tree_ref(node);
4367
4368         extent_key.objectid = node->bytenr;
4369         if (skinny_metadata) {
4370                 extent_key.offset = ref->level;
4371                 extent_key.type = BTRFS_METADATA_ITEM_KEY;
4372                 num_bytes = fs_info->nodesize;
4373         } else {
4374                 extent_key.offset = node->num_bytes;
4375                 extent_key.type = BTRFS_EXTENT_ITEM_KEY;
4376                 size += sizeof(*block_info);
4377                 num_bytes = node->num_bytes;
4378         }
4379
4380         path = btrfs_alloc_path();
4381         if (!path)
4382                 return -ENOMEM;
4383
4384         path->leave_spinning = 1;
4385         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
4386                                       &extent_key, size);
4387         if (ret) {
4388                 btrfs_free_path(path);
4389                 return ret;
4390         }
4391
4392         leaf = path->nodes[0];
4393         extent_item = btrfs_item_ptr(leaf, path->slots[0],
4394                                      struct btrfs_extent_item);
4395         btrfs_set_extent_refs(leaf, extent_item, 1);
4396         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
4397         btrfs_set_extent_flags(leaf, extent_item,
4398                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
4399
4400         if (skinny_metadata) {
4401                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
4402         } else {
4403                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
4404                 btrfs_set_tree_block_key(leaf, block_info, &extent_op->key);
4405                 btrfs_set_tree_block_level(leaf, block_info, ref->level);
4406                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
4407         }
4408
4409         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY) {
4410                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
4411                 btrfs_set_extent_inline_ref_type(leaf, iref,
4412                                                  BTRFS_SHARED_BLOCK_REF_KEY);
4413                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->parent);
4414         } else {
4415                 btrfs_set_extent_inline_ref_type(leaf, iref,
4416                                                  BTRFS_TREE_BLOCK_REF_KEY);
4417                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->root);
4418         }
4419
4420         btrfs_mark_buffer_dirty(leaf);
4421         btrfs_free_path(path);
4422
4423         ret = remove_from_free_space_tree(trans, extent_key.objectid,
4424                                           num_bytes);
4425         if (ret)
4426                 return ret;
4427
4428         ret = btrfs_update_block_group(trans, extent_key.objectid,
4429                                        fs_info->nodesize, 1);
4430         if (ret) { /* -ENOENT, logic error */
4431                 btrfs_err(fs_info, "update block group failed for %llu %llu",
4432                         extent_key.objectid, extent_key.offset);
4433                 BUG();
4434         }
4435
4436         trace_btrfs_reserved_extent_alloc(fs_info, extent_key.objectid,
4437                                           fs_info->nodesize);
4438         return ret;
4439 }
4440
4441 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
4442                                      struct btrfs_root *root, u64 owner,
4443                                      u64 offset, u64 ram_bytes,
4444                                      struct btrfs_key *ins)
4445 {
4446         struct btrfs_ref generic_ref = { 0 };
4447         int ret;
4448
4449         BUG_ON(root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
4450
4451         btrfs_init_generic_ref(&generic_ref, BTRFS_ADD_DELAYED_EXTENT,
4452                                ins->objectid, ins->offset, 0);
4453         btrfs_init_data_ref(&generic_ref, root->root_key.objectid, owner, offset);
4454         btrfs_ref_tree_mod(root->fs_info, &generic_ref);
4455         ret = btrfs_add_delayed_data_ref(trans, &generic_ref,
4456                                          ram_bytes, NULL, NULL);
4457         return ret;
4458 }
4459
4460 /*
4461  * this is used by the tree logging recovery code.  It records that
4462  * an extent has been allocated and makes sure to clear the free
4463  * space cache bits as well
4464  */
4465 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
4466                                    u64 root_objectid, u64 owner, u64 offset,
4467                                    struct btrfs_key *ins)
4468 {
4469         struct btrfs_fs_info *fs_info = trans->fs_info;
4470         int ret;
4471         struct btrfs_block_group *block_group;
4472         struct btrfs_space_info *space_info;
4473
4474         /*
4475          * Mixed block groups will exclude before processing the log so we only
4476          * need to do the exclude dance if this fs isn't mixed.
4477          */
4478         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS)) {
4479                 ret = __exclude_logged_extent(fs_info, ins->objectid,
4480                                               ins->offset);
4481                 if (ret)
4482                         return ret;
4483         }
4484
4485         block_group = btrfs_lookup_block_group(fs_info, ins->objectid);
4486         if (!block_group)
4487                 return -EINVAL;
4488
4489         space_info = block_group->space_info;
4490         spin_lock(&space_info->lock);
4491         spin_lock(&block_group->lock);
4492         space_info->bytes_reserved += ins->offset;
4493         block_group->reserved += ins->offset;
4494         spin_unlock(&block_group->lock);
4495         spin_unlock(&space_info->lock);
4496
4497         ret = alloc_reserved_file_extent(trans, 0, root_objectid, 0, owner,
4498                                          offset, ins, 1);
4499         if (ret)
4500                 btrfs_pin_extent(trans, ins->objectid, ins->offset, 1);
4501         btrfs_put_block_group(block_group);
4502         return ret;
4503 }
4504
4505 static struct extent_buffer *
4506 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
4507                       u64 bytenr, int level, u64 owner)
4508 {
4509         struct btrfs_fs_info *fs_info = root->fs_info;
4510         struct extent_buffer *buf;
4511
4512         buf = btrfs_find_create_tree_block(fs_info, bytenr);
4513         if (IS_ERR(buf))
4514                 return buf;
4515
4516         /*
4517          * Extra safety check in case the extent tree is corrupted and extent
4518          * allocator chooses to use a tree block which is already used and
4519          * locked.
4520          */
4521         if (buf->lock_owner == current->pid) {
4522                 btrfs_err_rl(fs_info,
4523 "tree block %llu owner %llu already locked by pid=%d, extent tree corruption detected",
4524                         buf->start, btrfs_header_owner(buf), current->pid);
4525                 free_extent_buffer(buf);
4526                 return ERR_PTR(-EUCLEAN);
4527         }
4528
4529         btrfs_set_buffer_lockdep_class(owner, buf, level);
4530         btrfs_tree_lock(buf);
4531         btrfs_clean_tree_block(buf);
4532         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
4533
4534         btrfs_set_lock_blocking_write(buf);
4535         set_extent_buffer_uptodate(buf);
4536
4537         memzero_extent_buffer(buf, 0, sizeof(struct btrfs_header));
4538         btrfs_set_header_level(buf, level);
4539         btrfs_set_header_bytenr(buf, buf->start);
4540         btrfs_set_header_generation(buf, trans->transid);
4541         btrfs_set_header_backref_rev(buf, BTRFS_MIXED_BACKREF_REV);
4542         btrfs_set_header_owner(buf, owner);
4543         write_extent_buffer_fsid(buf, fs_info->fs_devices->metadata_uuid);
4544         write_extent_buffer_chunk_tree_uuid(buf, fs_info->chunk_tree_uuid);
4545         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
4546                 buf->log_index = root->log_transid % 2;
4547                 /*
4548                  * we allow two log transactions at a time, use different
4549                  * EXTENT bit to differentiate dirty pages.
4550                  */
4551                 if (buf->log_index == 0)
4552                         set_extent_dirty(&root->dirty_log_pages, buf->start,
4553                                         buf->start + buf->len - 1, GFP_NOFS);
4554                 else
4555                         set_extent_new(&root->dirty_log_pages, buf->start,
4556                                         buf->start + buf->len - 1);
4557         } else {
4558                 buf->log_index = -1;
4559                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
4560                          buf->start + buf->len - 1, GFP_NOFS);
4561         }
4562         trans->dirty = true;
4563         /* this returns a buffer locked for blocking */
4564         return buf;
4565 }
4566
4567 /*
4568  * finds a free extent and does all the dirty work required for allocation
4569  * returns the tree buffer or an ERR_PTR on error.
4570  */
4571 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
4572                                              struct btrfs_root *root,
4573                                              u64 parent, u64 root_objectid,
4574                                              const struct btrfs_disk_key *key,
4575                                              int level, u64 hint,
4576                                              u64 empty_size)
4577 {
4578         struct btrfs_fs_info *fs_info = root->fs_info;
4579         struct btrfs_key ins;
4580         struct btrfs_block_rsv *block_rsv;
4581         struct extent_buffer *buf;
4582         struct btrfs_delayed_extent_op *extent_op;
4583         struct btrfs_ref generic_ref = { 0 };
4584         u64 flags = 0;
4585         int ret;
4586         u32 blocksize = fs_info->nodesize;
4587         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
4588
4589 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4590         if (btrfs_is_testing(fs_info)) {
4591                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
4592                                             level, root_objectid);
4593                 if (!IS_ERR(buf))
4594                         root->alloc_bytenr += blocksize;
4595                 return buf;
4596         }
4597 #endif
4598
4599         block_rsv = btrfs_use_block_rsv(trans, root, blocksize);
4600         if (IS_ERR(block_rsv))
4601                 return ERR_CAST(block_rsv);
4602
4603         ret = btrfs_reserve_extent(root, blocksize, blocksize, blocksize,
4604                                    empty_size, hint, &ins, 0, 0);
4605         if (ret)
4606                 goto out_unuse;
4607
4608         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level,
4609                                     root_objectid);
4610         if (IS_ERR(buf)) {
4611                 ret = PTR_ERR(buf);
4612                 goto out_free_reserved;
4613         }
4614
4615         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
4616                 if (parent == 0)
4617                         parent = ins.objectid;
4618                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
4619         } else
4620                 BUG_ON(parent > 0);
4621
4622         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
4623                 extent_op = btrfs_alloc_delayed_extent_op();
4624                 if (!extent_op) {
4625                         ret = -ENOMEM;
4626                         goto out_free_buf;
4627                 }
4628                 if (key)
4629                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
4630                 else
4631                         memset(&extent_op->key, 0, sizeof(extent_op->key));
4632                 extent_op->flags_to_set = flags;
4633                 extent_op->update_key = skinny_metadata ? false : true;
4634                 extent_op->update_flags = true;
4635                 extent_op->is_data = false;
4636                 extent_op->level = level;
4637
4638                 btrfs_init_generic_ref(&generic_ref, BTRFS_ADD_DELAYED_EXTENT,
4639                                        ins.objectid, ins.offset, parent);
4640                 generic_ref.real_root = root->root_key.objectid;
4641                 btrfs_init_tree_ref(&generic_ref, level, root_objectid);
4642                 btrfs_ref_tree_mod(fs_info, &generic_ref);
4643                 ret = btrfs_add_delayed_tree_ref(trans, &generic_ref,
4644                                                  extent_op, NULL, NULL);
4645                 if (ret)
4646                         goto out_free_delayed;
4647         }
4648         return buf;
4649
4650 out_free_delayed:
4651         btrfs_free_delayed_extent_op(extent_op);
4652 out_free_buf:
4653         free_extent_buffer(buf);
4654 out_free_reserved:
4655         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 0);
4656 out_unuse:
4657         btrfs_unuse_block_rsv(fs_info, block_rsv, blocksize);
4658         return ERR_PTR(ret);
4659 }
4660
4661 struct walk_control {
4662         u64 refs[BTRFS_MAX_LEVEL];
4663         u64 flags[BTRFS_MAX_LEVEL];
4664         struct btrfs_key update_progress;
4665         struct btrfs_key drop_progress;
4666         int drop_level;
4667         int stage;
4668         int level;
4669         int shared_level;
4670         int update_ref;
4671         int keep_locks;
4672         int reada_slot;
4673         int reada_count;
4674         int restarted;
4675 };
4676
4677 #define DROP_REFERENCE  1
4678 #define UPDATE_BACKREF  2
4679
4680 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
4681                                      struct btrfs_root *root,
4682                                      struct walk_control *wc,
4683                                      struct btrfs_path *path)
4684 {
4685         struct btrfs_fs_info *fs_info = root->fs_info;
4686         u64 bytenr;
4687         u64 generation;
4688         u64 refs;
4689         u64 flags;
4690         u32 nritems;
4691         struct btrfs_key key;
4692         struct extent_buffer *eb;
4693         int ret;
4694         int slot;
4695         int nread = 0;
4696
4697         if (path->slots[wc->level] < wc->reada_slot) {
4698                 wc->reada_count = wc->reada_count * 2 / 3;
4699                 wc->reada_count = max(wc->reada_count, 2);
4700         } else {
4701                 wc->reada_count = wc->reada_count * 3 / 2;
4702                 wc->reada_count = min_t(int, wc->reada_count,
4703                                         BTRFS_NODEPTRS_PER_BLOCK(fs_info));
4704         }
4705
4706         eb = path->nodes[wc->level];
4707         nritems = btrfs_header_nritems(eb);
4708
4709         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
4710                 if (nread >= wc->reada_count)
4711                         break;
4712
4713                 cond_resched();
4714                 bytenr = btrfs_node_blockptr(eb, slot);
4715                 generation = btrfs_node_ptr_generation(eb, slot);
4716
4717                 if (slot == path->slots[wc->level])
4718                         goto reada;
4719
4720                 if (wc->stage == UPDATE_BACKREF &&
4721                     generation <= root->root_key.offset)
4722                         continue;
4723
4724                 /* We don't lock the tree block, it's OK to be racy here */
4725                 ret = btrfs_lookup_extent_info(trans, fs_info, bytenr,
4726                                                wc->level - 1, 1, &refs,
4727                                                &flags);
4728                 /* We don't care about errors in readahead. */
4729                 if (ret < 0)
4730                         continue;
4731                 BUG_ON(refs == 0);
4732
4733                 if (wc->stage == DROP_REFERENCE) {
4734                         if (refs == 1)
4735                                 goto reada;
4736
4737                         if (wc->level == 1 &&
4738                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
4739                                 continue;
4740                         if (!wc->update_ref ||
4741                             generation <= root->root_key.offset)
4742                                 continue;
4743                         btrfs_node_key_to_cpu(eb, &key, slot);
4744                         ret = btrfs_comp_cpu_keys(&key,
4745                                                   &wc->update_progress);
4746                         if (ret < 0)
4747                                 continue;
4748                 } else {
4749                         if (wc->level == 1 &&
4750                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
4751                                 continue;
4752                 }
4753 reada:
4754                 readahead_tree_block(fs_info, bytenr);
4755                 nread++;
4756         }
4757         wc->reada_slot = slot;
4758 }
4759
4760 /*
4761  * helper to process tree block while walking down the tree.
4762  *
4763  * when wc->stage == UPDATE_BACKREF, this function updates
4764  * back refs for pointers in the block.
4765  *
4766  * NOTE: return value 1 means we should stop walking down.
4767  */
4768 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
4769                                    struct btrfs_root *root,
4770                                    struct btrfs_path *path,
4771                                    struct walk_control *wc, int lookup_info)
4772 {
4773         struct btrfs_fs_info *fs_info = root->fs_info;
4774         int level = wc->level;
4775         struct extent_buffer *eb = path->nodes[level];
4776         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
4777         int ret;
4778
4779         if (wc->stage == UPDATE_BACKREF &&
4780             btrfs_header_owner(eb) != root->root_key.objectid)
4781                 return 1;
4782
4783         /*
4784          * when reference count of tree block is 1, it won't increase
4785          * again. once full backref flag is set, we never clear it.
4786          */
4787         if (lookup_info &&
4788             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
4789              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
4790                 BUG_ON(!path->locks[level]);
4791                 ret = btrfs_lookup_extent_info(trans, fs_info,
4792                                                eb->start, level, 1,
4793                                                &wc->refs[level],
4794                                                &wc->flags[level]);
4795                 BUG_ON(ret == -ENOMEM);
4796                 if (ret)
4797                         return ret;
4798                 BUG_ON(wc->refs[level] == 0);
4799         }
4800
4801         if (wc->stage == DROP_REFERENCE) {
4802                 if (wc->refs[level] > 1)
4803                         return 1;
4804
4805                 if (path->locks[level] && !wc->keep_locks) {
4806                         btrfs_tree_unlock_rw(eb, path->locks[level]);
4807                         path->locks[level] = 0;
4808                 }
4809                 return 0;
4810         }
4811
4812         /* wc->stage == UPDATE_BACKREF */
4813         if (!(wc->flags[level] & flag)) {
4814                 BUG_ON(!path->locks[level]);
4815                 ret = btrfs_inc_ref(trans, root, eb, 1);
4816                 BUG_ON(ret); /* -ENOMEM */
4817                 ret = btrfs_dec_ref(trans, root, eb, 0);
4818                 BUG_ON(ret); /* -ENOMEM */
4819                 ret = btrfs_set_disk_extent_flags(trans, eb, flag,
4820                                                   btrfs_header_level(eb), 0);
4821                 BUG_ON(ret); /* -ENOMEM */
4822                 wc->flags[level] |= flag;
4823         }
4824
4825         /*
4826          * the block is shared by multiple trees, so it's not good to
4827          * keep the tree lock
4828          */
4829         if (path->locks[level] && level > 0) {
4830                 btrfs_tree_unlock_rw(eb, path->locks[level]);
4831                 path->locks[level] = 0;
4832         }
4833         return 0;
4834 }
4835
4836 /*
4837  * This is used to verify a ref exists for this root to deal with a bug where we
4838  * would have a drop_progress key that hadn't been updated properly.
4839  */
4840 static int check_ref_exists(struct btrfs_trans_handle *trans,
4841                             struct btrfs_root *root, u64 bytenr, u64 parent,
4842                             int level)
4843 {
4844         struct btrfs_path *path;
4845         struct btrfs_extent_inline_ref *iref;
4846         int ret;
4847
4848         path = btrfs_alloc_path();
4849         if (!path)
4850                 return -ENOMEM;
4851
4852         ret = lookup_extent_backref(trans, path, &iref, bytenr,
4853                                     root->fs_info->nodesize, parent,
4854                                     root->root_key.objectid, level, 0);
4855         btrfs_free_path(path);
4856         if (ret == -ENOENT)
4857                 return 0;
4858         if (ret < 0)
4859                 return ret;
4860         return 1;
4861 }
4862
4863 /*
4864  * helper to process tree block pointer.
4865  *
4866  * when wc->stage == DROP_REFERENCE, this function checks
4867  * reference count of the block pointed to. if the block
4868  * is shared and we need update back refs for the subtree
4869  * rooted at the block, this function changes wc->stage to
4870  * UPDATE_BACKREF. if the block is shared and there is no
4871  * need to update back, this function drops the reference
4872  * to the block.
4873  *
4874  * NOTE: return value 1 means we should stop walking down.
4875  */
4876 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
4877                                  struct btrfs_root *root,
4878                                  struct btrfs_path *path,
4879                                  struct walk_control *wc, int *lookup_info)
4880 {
4881         struct btrfs_fs_info *fs_info = root->fs_info;
4882         u64 bytenr;
4883         u64 generation;
4884         u64 parent;
4885         struct btrfs_key key;
4886         struct btrfs_key first_key;
4887         struct btrfs_ref ref = { 0 };
4888         struct extent_buffer *next;
4889         int level = wc->level;
4890         int reada = 0;
4891         int ret = 0;
4892         bool need_account = false;
4893
4894         generation = btrfs_node_ptr_generation(path->nodes[level],
4895                                                path->slots[level]);
4896         /*
4897          * if the lower level block was created before the snapshot
4898          * was created, we know there is no need to update back refs
4899          * for the subtree
4900          */
4901         if (wc->stage == UPDATE_BACKREF &&
4902             generation <= root->root_key.offset) {
4903                 *lookup_info = 1;
4904                 return 1;
4905         }
4906
4907         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
4908         btrfs_node_key_to_cpu(path->nodes[level], &first_key,
4909                               path->slots[level]);
4910
4911         next = find_extent_buffer(fs_info, bytenr);
4912         if (!next) {
4913                 next = btrfs_find_create_tree_block(fs_info, bytenr);
4914                 if (IS_ERR(next))
4915                         return PTR_ERR(next);
4916
4917                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
4918                                                level - 1);
4919                 reada = 1;
4920         }
4921         btrfs_tree_lock(next);
4922         btrfs_set_lock_blocking_write(next);
4923
4924         ret = btrfs_lookup_extent_info(trans, fs_info, bytenr, level - 1, 1,
4925                                        &wc->refs[level - 1],
4926                                        &wc->flags[level - 1]);
4927         if (ret < 0)
4928                 goto out_unlock;
4929
4930         if (unlikely(wc->refs[level - 1] == 0)) {
4931                 btrfs_err(fs_info, "Missing references.");
4932                 ret = -EIO;
4933                 goto out_unlock;
4934         }
4935         *lookup_info = 0;
4936
4937         if (wc->stage == DROP_REFERENCE) {
4938                 if (wc->refs[level - 1] > 1) {
4939                         need_account = true;
4940                         if (level == 1 &&
4941                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
4942                                 goto skip;
4943
4944                         if (!wc->update_ref ||
4945                             generation <= root->root_key.offset)
4946                                 goto skip;
4947
4948                         btrfs_node_key_to_cpu(path->nodes[level], &key,
4949                                               path->slots[level]);
4950                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
4951                         if (ret < 0)
4952                                 goto skip;
4953
4954                         wc->stage = UPDATE_BACKREF;
4955                         wc->shared_level = level - 1;
4956                 }
4957         } else {
4958                 if (level == 1 &&
4959                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
4960                         goto skip;
4961         }
4962
4963         if (!btrfs_buffer_uptodate(next, generation, 0)) {
4964                 btrfs_tree_unlock(next);
4965                 free_extent_buffer(next);
4966                 next = NULL;
4967                 *lookup_info = 1;
4968         }
4969
4970         if (!next) {
4971                 if (reada && level == 1)
4972                         reada_walk_down(trans, root, wc, path);
4973                 next = read_tree_block(fs_info, bytenr, generation, level - 1,
4974                                        &first_key);
4975                 if (IS_ERR(next)) {
4976                         return PTR_ERR(next);
4977                 } else if (!extent_buffer_uptodate(next)) {
4978                         free_extent_buffer(next);
4979                         return -EIO;
4980                 }
4981                 btrfs_tree_lock(next);
4982                 btrfs_set_lock_blocking_write(next);
4983         }
4984
4985         level--;
4986         ASSERT(level == btrfs_header_level(next));
4987         if (level != btrfs_header_level(next)) {
4988                 btrfs_err(root->fs_info, "mismatched level");
4989                 ret = -EIO;
4990                 goto out_unlock;
4991         }
4992         path->nodes[level] = next;
4993         path->slots[level] = 0;
4994         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
4995         wc->level = level;
4996         if (wc->level == 1)
4997                 wc->reada_slot = 0;
4998         return 0;
4999 skip:
5000         wc->refs[level - 1] = 0;
5001         wc->flags[level - 1] = 0;
5002         if (wc->stage == DROP_REFERENCE) {
5003                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
5004                         parent = path->nodes[level]->start;
5005                 } else {
5006                         ASSERT(root->root_key.objectid ==
5007                                btrfs_header_owner(path->nodes[level]));
5008                         if (root->root_key.objectid !=
5009                             btrfs_header_owner(path->nodes[level])) {
5010                                 btrfs_err(root->fs_info,
5011                                                 "mismatched block owner");
5012                                 ret = -EIO;
5013                                 goto out_unlock;
5014                         }
5015                         parent = 0;
5016                 }
5017
5018                 /*
5019                  * If we had a drop_progress we need to verify the refs are set
5020                  * as expected.  If we find our ref then we know that from here
5021                  * on out everything should be correct, and we can clear the
5022                  * ->restarted flag.
5023                  */
5024                 if (wc->restarted) {
5025                         ret = check_ref_exists(trans, root, bytenr, parent,
5026                                                level - 1);
5027                         if (ret < 0)
5028                                 goto out_unlock;
5029                         if (ret == 0)
5030                                 goto no_delete;
5031                         ret = 0;
5032                         wc->restarted = 0;
5033                 }
5034
5035                 /*
5036                  * Reloc tree doesn't contribute to qgroup numbers, and we have
5037                  * already accounted them at merge time (replace_path),
5038                  * thus we could skip expensive subtree trace here.
5039                  */
5040                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
5041                     need_account) {
5042                         ret = btrfs_qgroup_trace_subtree(trans, next,
5043                                                          generation, level - 1);
5044                         if (ret) {
5045                                 btrfs_err_rl(fs_info,
5046                                              "Error %d accounting shared subtree. Quota is out of sync, rescan required.",
5047                                              ret);
5048                         }
5049                 }
5050
5051                 /*
5052                  * We need to update the next key in our walk control so we can
5053                  * update the drop_progress key accordingly.  We don't care if
5054                  * find_next_key doesn't find a key because that means we're at
5055                  * the end and are going to clean up now.
5056                  */
5057                 wc->drop_level = level;
5058                 find_next_key(path, level, &wc->drop_progress);
5059
5060                 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
5061                                        fs_info->nodesize, parent);
5062                 btrfs_init_tree_ref(&ref, level - 1, root->root_key.objectid);
5063                 ret = btrfs_free_extent(trans, &ref);
5064                 if (ret)
5065                         goto out_unlock;
5066         }
5067 no_delete:
5068         *lookup_info = 1;
5069         ret = 1;
5070
5071 out_unlock:
5072         btrfs_tree_unlock(next);
5073         free_extent_buffer(next);
5074
5075         return ret;
5076 }
5077
5078 /*
5079  * helper to process tree block while walking up the tree.
5080  *
5081  * when wc->stage == DROP_REFERENCE, this function drops
5082  * reference count on the block.
5083  *
5084  * when wc->stage == UPDATE_BACKREF, this function changes
5085  * wc->stage back to DROP_REFERENCE if we changed wc->stage
5086  * to UPDATE_BACKREF previously while processing the block.
5087  *
5088  * NOTE: return value 1 means we should stop walking up.
5089  */
5090 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
5091                                  struct btrfs_root *root,
5092                                  struct btrfs_path *path,
5093                                  struct walk_control *wc)
5094 {
5095         struct btrfs_fs_info *fs_info = root->fs_info;
5096         int ret;
5097         int level = wc->level;
5098         struct extent_buffer *eb = path->nodes[level];
5099         u64 parent = 0;
5100
5101         if (wc->stage == UPDATE_BACKREF) {
5102                 BUG_ON(wc->shared_level < level);
5103                 if (level < wc->shared_level)
5104                         goto out;
5105
5106                 ret = find_next_key(path, level + 1, &wc->update_progress);
5107                 if (ret > 0)
5108                         wc->update_ref = 0;
5109
5110                 wc->stage = DROP_REFERENCE;
5111                 wc->shared_level = -1;
5112                 path->slots[level] = 0;
5113
5114                 /*
5115                  * check reference count again if the block isn't locked.
5116                  * we should start walking down the tree again if reference
5117                  * count is one.
5118                  */
5119                 if (!path->locks[level]) {
5120                         BUG_ON(level == 0);
5121                         btrfs_tree_lock(eb);
5122                         btrfs_set_lock_blocking_write(eb);
5123                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
5124
5125                         ret = btrfs_lookup_extent_info(trans, fs_info,
5126                                                        eb->start, level, 1,
5127                                                        &wc->refs[level],
5128                                                        &wc->flags[level]);
5129                         if (ret < 0) {
5130                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
5131                                 path->locks[level] = 0;
5132                                 return ret;
5133                         }
5134                         BUG_ON(wc->refs[level] == 0);
5135                         if (wc->refs[level] == 1) {
5136                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
5137                                 path->locks[level] = 0;
5138                                 return 1;
5139                         }
5140                 }
5141         }
5142
5143         /* wc->stage == DROP_REFERENCE */
5144         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
5145
5146         if (wc->refs[level] == 1) {
5147                 if (level == 0) {
5148                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
5149                                 ret = btrfs_dec_ref(trans, root, eb, 1);
5150                         else
5151                                 ret = btrfs_dec_ref(trans, root, eb, 0);
5152                         BUG_ON(ret); /* -ENOMEM */
5153                         if (is_fstree(root->root_key.objectid)) {
5154                                 ret = btrfs_qgroup_trace_leaf_items(trans, eb);
5155                                 if (ret) {
5156                                         btrfs_err_rl(fs_info,
5157         "error %d accounting leaf items, quota is out of sync, rescan required",
5158                                              ret);
5159                                 }
5160                         }
5161                 }
5162                 /* make block locked assertion in btrfs_clean_tree_block happy */
5163                 if (!path->locks[level] &&
5164                     btrfs_header_generation(eb) == trans->transid) {
5165                         btrfs_tree_lock(eb);
5166                         btrfs_set_lock_blocking_write(eb);
5167                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
5168                 }
5169                 btrfs_clean_tree_block(eb);
5170         }
5171
5172         if (eb == root->node) {
5173                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
5174                         parent = eb->start;
5175                 else if (root->root_key.objectid != btrfs_header_owner(eb))
5176                         goto owner_mismatch;
5177         } else {
5178                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
5179                         parent = path->nodes[level + 1]->start;
5180                 else if (root->root_key.objectid !=
5181                          btrfs_header_owner(path->nodes[level + 1]))
5182                         goto owner_mismatch;
5183         }
5184
5185         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
5186 out:
5187         wc->refs[level] = 0;
5188         wc->flags[level] = 0;
5189         return 0;
5190
5191 owner_mismatch:
5192         btrfs_err_rl(fs_info, "unexpected tree owner, have %llu expect %llu",
5193                      btrfs_header_owner(eb), root->root_key.objectid);
5194         return -EUCLEAN;
5195 }
5196
5197 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
5198                                    struct btrfs_root *root,
5199                                    struct btrfs_path *path,
5200                                    struct walk_control *wc)
5201 {
5202         int level = wc->level;
5203         int lookup_info = 1;
5204         int ret;
5205
5206         while (level >= 0) {
5207                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
5208                 if (ret > 0)
5209                         break;
5210
5211                 if (level == 0)
5212                         break;
5213
5214                 if (path->slots[level] >=
5215                     btrfs_header_nritems(path->nodes[level]))
5216                         break;
5217
5218                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
5219                 if (ret > 0) {
5220                         path->slots[level]++;
5221                         continue;
5222                 } else if (ret < 0)
5223                         return ret;
5224                 level = wc->level;
5225         }
5226         return 0;
5227 }
5228
5229 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
5230                                  struct btrfs_root *root,
5231                                  struct btrfs_path *path,
5232                                  struct walk_control *wc, int max_level)
5233 {
5234         int level = wc->level;
5235         int ret;
5236
5237         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
5238         while (level < max_level && path->nodes[level]) {
5239                 wc->level = level;
5240                 if (path->slots[level] + 1 <
5241                     btrfs_header_nritems(path->nodes[level])) {
5242                         path->slots[level]++;
5243                         return 0;
5244                 } else {
5245                         ret = walk_up_proc(trans, root, path, wc);
5246                         if (ret > 0)
5247                                 return 0;
5248                         if (ret < 0)
5249                                 return ret;
5250
5251                         if (path->locks[level]) {
5252                                 btrfs_tree_unlock_rw(path->nodes[level],
5253                                                      path->locks[level]);
5254                                 path->locks[level] = 0;
5255                         }
5256                         free_extent_buffer(path->nodes[level]);
5257                         path->nodes[level] = NULL;
5258                         level++;
5259                 }
5260         }
5261         return 1;
5262 }
5263
5264 /*
5265  * drop a subvolume tree.
5266  *
5267  * this function traverses the tree freeing any blocks that only
5268  * referenced by the tree.
5269  *
5270  * when a shared tree block is found. this function decreases its
5271  * reference count by one. if update_ref is true, this function
5272  * also make sure backrefs for the shared block and all lower level
5273  * blocks are properly updated.
5274  *
5275  * If called with for_reloc == 0, may exit early with -EAGAIN
5276  */
5277 int btrfs_drop_snapshot(struct btrfs_root *root, int update_ref, int for_reloc)
5278 {
5279         struct btrfs_fs_info *fs_info = root->fs_info;
5280         struct btrfs_path *path;
5281         struct btrfs_trans_handle *trans;
5282         struct btrfs_root *tree_root = fs_info->tree_root;
5283         struct btrfs_root_item *root_item = &root->root_item;
5284         struct walk_control *wc;
5285         struct btrfs_key key;
5286         int err = 0;
5287         int ret;
5288         int level;
5289         bool root_dropped = false;
5290
5291         btrfs_debug(fs_info, "Drop subvolume %llu", root->root_key.objectid);
5292
5293         path = btrfs_alloc_path();
5294         if (!path) {
5295                 err = -ENOMEM;
5296                 goto out;
5297         }
5298
5299         wc = kzalloc(sizeof(*wc), GFP_NOFS);
5300         if (!wc) {
5301                 btrfs_free_path(path);
5302                 err = -ENOMEM;
5303                 goto out;
5304         }
5305
5306         /*
5307          * Use join to avoid potential EINTR from transaction start. See
5308          * wait_reserve_ticket and the whole reservation callchain.
5309          */
5310         if (for_reloc)
5311                 trans = btrfs_join_transaction(tree_root);
5312         else
5313                 trans = btrfs_start_transaction(tree_root, 0);
5314         if (IS_ERR(trans)) {
5315                 err = PTR_ERR(trans);
5316                 goto out_free;
5317         }
5318
5319         err = btrfs_run_delayed_items(trans);
5320         if (err)
5321                 goto out_end_trans;
5322
5323         /*
5324          * This will help us catch people modifying the fs tree while we're
5325          * dropping it.  It is unsafe to mess with the fs tree while it's being
5326          * dropped as we unlock the root node and parent nodes as we walk down
5327          * the tree, assuming nothing will change.  If something does change
5328          * then we'll have stale information and drop references to blocks we've
5329          * already dropped.
5330          */
5331         set_bit(BTRFS_ROOT_DELETING, &root->state);
5332         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
5333                 level = btrfs_header_level(root->node);
5334                 path->nodes[level] = btrfs_lock_root_node(root);
5335                 btrfs_set_lock_blocking_write(path->nodes[level]);
5336                 path->slots[level] = 0;
5337                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
5338                 memset(&wc->update_progress, 0,
5339                        sizeof(wc->update_progress));
5340         } else {
5341                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
5342                 memcpy(&wc->update_progress, &key,
5343                        sizeof(wc->update_progress));
5344
5345                 level = root_item->drop_level;
5346                 BUG_ON(level == 0);
5347                 path->lowest_level = level;
5348                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5349                 path->lowest_level = 0;
5350                 if (ret < 0) {
5351                         err = ret;
5352                         goto out_end_trans;
5353                 }
5354                 WARN_ON(ret > 0);
5355
5356                 /*
5357                  * unlock our path, this is safe because only this
5358                  * function is allowed to delete this snapshot
5359                  */
5360                 btrfs_unlock_up_safe(path, 0);
5361
5362                 level = btrfs_header_level(root->node);
5363                 while (1) {
5364                         btrfs_tree_lock(path->nodes[level]);
5365                         btrfs_set_lock_blocking_write(path->nodes[level]);
5366                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
5367
5368                         ret = btrfs_lookup_extent_info(trans, fs_info,
5369                                                 path->nodes[level]->start,
5370                                                 level, 1, &wc->refs[level],
5371                                                 &wc->flags[level]);
5372                         if (ret < 0) {
5373                                 err = ret;
5374                                 goto out_end_trans;
5375                         }
5376                         BUG_ON(wc->refs[level] == 0);
5377
5378                         if (level == root_item->drop_level)
5379                                 break;
5380
5381                         btrfs_tree_unlock(path->nodes[level]);
5382                         path->locks[level] = 0;
5383                         WARN_ON(wc->refs[level] != 1);
5384                         level--;
5385                 }
5386         }
5387
5388         wc->restarted = test_bit(BTRFS_ROOT_DEAD_TREE, &root->state);
5389         wc->level = level;
5390         wc->shared_level = -1;
5391         wc->stage = DROP_REFERENCE;
5392         wc->update_ref = update_ref;
5393         wc->keep_locks = 0;
5394         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
5395
5396         while (1) {
5397
5398                 ret = walk_down_tree(trans, root, path, wc);
5399                 if (ret < 0) {
5400                         err = ret;
5401                         break;
5402                 }
5403
5404                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
5405                 if (ret < 0) {
5406                         err = ret;
5407                         break;
5408                 }
5409
5410                 if (ret > 0) {
5411                         BUG_ON(wc->stage != DROP_REFERENCE);
5412                         break;
5413                 }
5414
5415                 if (wc->stage == DROP_REFERENCE) {
5416                         wc->drop_level = wc->level;
5417                         btrfs_node_key_to_cpu(path->nodes[wc->drop_level],
5418                                               &wc->drop_progress,
5419                                               path->slots[wc->drop_level]);
5420                 }
5421                 btrfs_cpu_key_to_disk(&root_item->drop_progress,
5422                                       &wc->drop_progress);
5423                 root_item->drop_level = wc->drop_level;
5424
5425                 BUG_ON(wc->level == 0);
5426                 if (btrfs_should_end_transaction(trans) ||
5427                     (!for_reloc && btrfs_need_cleaner_sleep(fs_info))) {
5428                         ret = btrfs_update_root(trans, tree_root,
5429                                                 &root->root_key,
5430                                                 root_item);
5431                         if (ret) {
5432                                 btrfs_abort_transaction(trans, ret);
5433                                 err = ret;
5434                                 goto out_end_trans;
5435                         }
5436
5437                         btrfs_end_transaction_throttle(trans);
5438                         if (!for_reloc && btrfs_need_cleaner_sleep(fs_info)) {
5439                                 btrfs_debug(fs_info,
5440                                             "drop snapshot early exit");
5441                                 err = -EAGAIN;
5442                                 goto out_free;
5443                         }
5444
5445                         trans = btrfs_start_transaction(tree_root, 0);
5446                         if (IS_ERR(trans)) {
5447                                 err = PTR_ERR(trans);
5448                                 goto out_free;
5449                         }
5450                 }
5451         }
5452         btrfs_release_path(path);
5453         if (err)
5454                 goto out_end_trans;
5455
5456         ret = btrfs_del_root(trans, &root->root_key);
5457         if (ret) {
5458                 btrfs_abort_transaction(trans, ret);
5459                 err = ret;
5460                 goto out_end_trans;
5461         }
5462
5463         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
5464                 ret = btrfs_find_root(tree_root, &root->root_key, path,
5465                                       NULL, NULL);
5466                 if (ret < 0) {
5467                         btrfs_abort_transaction(trans, ret);
5468                         err = ret;
5469                         goto out_end_trans;
5470                 } else if (ret > 0) {
5471                         /* if we fail to delete the orphan item this time
5472                          * around, it'll get picked up the next time.
5473                          *
5474                          * The most common failure here is just -ENOENT.
5475                          */
5476                         btrfs_del_orphan_item(trans, tree_root,
5477                                               root->root_key.objectid);
5478                 }
5479         }
5480
5481         /*
5482          * This subvolume is going to be completely dropped, and won't be
5483          * recorded as dirty roots, thus pertrans meta rsv will not be freed at
5484          * commit transaction time.  So free it here manually.
5485          */
5486         btrfs_qgroup_convert_reserved_meta(root, INT_MAX);
5487         btrfs_qgroup_free_meta_all_pertrans(root);
5488
5489         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state))
5490                 btrfs_add_dropped_root(trans, root);
5491         else
5492                 btrfs_put_root(root);
5493         root_dropped = true;
5494 out_end_trans:
5495         btrfs_end_transaction_throttle(trans);
5496 out_free:
5497         kfree(wc);
5498         btrfs_free_path(path);
5499 out:
5500         /*
5501          * So if we need to stop dropping the snapshot for whatever reason we
5502          * need to make sure to add it back to the dead root list so that we
5503          * keep trying to do the work later.  This also cleans up roots if we
5504          * don't have it in the radix (like when we recover after a power fail
5505          * or unmount) so we don't leak memory.
5506          */
5507         if (!for_reloc && !root_dropped)
5508                 btrfs_add_dead_root(root);
5509         return err;
5510 }
5511
5512 /*
5513  * drop subtree rooted at tree block 'node'.
5514  *
5515  * NOTE: this function will unlock and release tree block 'node'
5516  * only used by relocation code
5517  */
5518 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
5519                         struct btrfs_root *root,
5520                         struct extent_buffer *node,
5521                         struct extent_buffer *parent)
5522 {
5523         struct btrfs_fs_info *fs_info = root->fs_info;
5524         struct btrfs_path *path;
5525         struct walk_control *wc;
5526         int level;
5527         int parent_level;
5528         int ret = 0;
5529         int wret;
5530
5531         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
5532
5533         path = btrfs_alloc_path();
5534         if (!path)
5535                 return -ENOMEM;
5536
5537         wc = kzalloc(sizeof(*wc), GFP_NOFS);
5538         if (!wc) {
5539                 btrfs_free_path(path);
5540                 return -ENOMEM;
5541         }
5542
5543         btrfs_assert_tree_locked(parent);
5544         parent_level = btrfs_header_level(parent);
5545         atomic_inc(&parent->refs);
5546         path->nodes[parent_level] = parent;
5547         path->slots[parent_level] = btrfs_header_nritems(parent);
5548
5549         btrfs_assert_tree_locked(node);
5550         level = btrfs_header_level(node);
5551         path->nodes[level] = node;
5552         path->slots[level] = 0;
5553         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
5554
5555         wc->refs[parent_level] = 1;
5556         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
5557         wc->level = level;
5558         wc->shared_level = -1;
5559         wc->stage = DROP_REFERENCE;
5560         wc->update_ref = 0;
5561         wc->keep_locks = 1;
5562         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
5563
5564         while (1) {
5565                 wret = walk_down_tree(trans, root, path, wc);
5566                 if (wret < 0) {
5567                         ret = wret;
5568                         break;
5569                 }
5570
5571                 wret = walk_up_tree(trans, root, path, wc, parent_level);
5572                 if (wret < 0)
5573                         ret = wret;
5574                 if (wret != 0)
5575                         break;
5576         }
5577
5578         kfree(wc);
5579         btrfs_free_path(path);
5580         return ret;
5581 }
5582
5583 /*
5584  * helper to account the unused space of all the readonly block group in the
5585  * space_info. takes mirrors into account.
5586  */
5587 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
5588 {
5589         struct btrfs_block_group *block_group;
5590         u64 free_bytes = 0;
5591         int factor;
5592
5593         /* It's df, we don't care if it's racy */
5594         if (list_empty(&sinfo->ro_bgs))
5595                 return 0;
5596
5597         spin_lock(&sinfo->lock);
5598         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
5599                 spin_lock(&block_group->lock);
5600
5601                 if (!block_group->ro) {
5602                         spin_unlock(&block_group->lock);
5603                         continue;
5604                 }
5605
5606                 factor = btrfs_bg_type_to_factor(block_group->flags);
5607                 free_bytes += (block_group->length -
5608                                block_group->used) * factor;
5609
5610                 spin_unlock(&block_group->lock);
5611         }
5612         spin_unlock(&sinfo->lock);
5613
5614         return free_bytes;
5615 }
5616
5617 int btrfs_error_unpin_extent_range(struct btrfs_fs_info *fs_info,
5618                                    u64 start, u64 end)
5619 {
5620         return unpin_extent_range(fs_info, start, end, false);
5621 }
5622
5623 /*
5624  * It used to be that old block groups would be left around forever.
5625  * Iterating over them would be enough to trim unused space.  Since we
5626  * now automatically remove them, we also need to iterate over unallocated
5627  * space.
5628  *
5629  * We don't want a transaction for this since the discard may take a
5630  * substantial amount of time.  We don't require that a transaction be
5631  * running, but we do need to take a running transaction into account
5632  * to ensure that we're not discarding chunks that were released or
5633  * allocated in the current transaction.
5634  *
5635  * Holding the chunks lock will prevent other threads from allocating
5636  * or releasing chunks, but it won't prevent a running transaction
5637  * from committing and releasing the memory that the pending chunks
5638  * list head uses.  For that, we need to take a reference to the
5639  * transaction and hold the commit root sem.  We only need to hold
5640  * it while performing the free space search since we have already
5641  * held back allocations.
5642  */
5643 static int btrfs_trim_free_extents(struct btrfs_device *device, u64 *trimmed)
5644 {
5645         u64 start = SZ_1M, len = 0, end = 0;
5646         int ret;
5647
5648         *trimmed = 0;
5649
5650         /* Discard not supported = nothing to do. */
5651         if (!blk_queue_discard(bdev_get_queue(device->bdev)))
5652                 return 0;
5653
5654         /* Not writable = nothing to do. */
5655         if (!test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state))
5656                 return 0;
5657
5658         /* No free space = nothing to do. */
5659         if (device->total_bytes <= device->bytes_used)
5660                 return 0;
5661
5662         ret = 0;
5663
5664         while (1) {
5665                 struct btrfs_fs_info *fs_info = device->fs_info;
5666                 u64 bytes;
5667
5668                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
5669                 if (ret)
5670                         break;
5671
5672                 find_first_clear_extent_bit(&device->alloc_state, start,
5673                                             &start, &end,
5674                                             CHUNK_TRIMMED | CHUNK_ALLOCATED);
5675
5676                 /* Check if there are any CHUNK_* bits left */
5677                 if (start > device->total_bytes) {
5678                         WARN_ON(IS_ENABLED(CONFIG_BTRFS_DEBUG));
5679                         btrfs_warn_in_rcu(fs_info,
5680 "ignoring attempt to trim beyond device size: offset %llu length %llu device %s device size %llu",
5681                                           start, end - start + 1,
5682                                           rcu_str_deref(device->name),
5683                                           device->total_bytes);
5684                         mutex_unlock(&fs_info->chunk_mutex);
5685                         ret = 0;
5686                         break;
5687                 }
5688
5689                 /* Ensure we skip the reserved area in the first 1M */
5690                 start = max_t(u64, start, SZ_1M);
5691
5692                 /*
5693                  * If find_first_clear_extent_bit find a range that spans the
5694                  * end of the device it will set end to -1, in this case it's up
5695                  * to the caller to trim the value to the size of the device.
5696                  */
5697                 end = min(end, device->total_bytes - 1);
5698
5699                 len = end - start + 1;
5700
5701                 /* We didn't find any extents */
5702                 if (!len) {
5703                         mutex_unlock(&fs_info->chunk_mutex);
5704                         ret = 0;
5705                         break;
5706                 }
5707
5708                 ret = btrfs_issue_discard(device->bdev, start, len,
5709                                           &bytes);
5710                 if (!ret)
5711                         set_extent_bits(&device->alloc_state, start,
5712                                         start + bytes - 1,
5713                                         CHUNK_TRIMMED);
5714                 mutex_unlock(&fs_info->chunk_mutex);
5715
5716                 if (ret)
5717                         break;
5718
5719                 start += len;
5720                 *trimmed += bytes;
5721
5722                 if (fatal_signal_pending(current)) {
5723                         ret = -ERESTARTSYS;
5724                         break;
5725                 }
5726
5727                 cond_resched();
5728         }
5729
5730         return ret;
5731 }
5732
5733 /*
5734  * Trim the whole filesystem by:
5735  * 1) trimming the free space in each block group
5736  * 2) trimming the unallocated space on each device
5737  *
5738  * This will also continue trimming even if a block group or device encounters
5739  * an error.  The return value will be the last error, or 0 if nothing bad
5740  * happens.
5741  */
5742 int btrfs_trim_fs(struct btrfs_fs_info *fs_info, struct fstrim_range *range)
5743 {
5744         struct btrfs_block_group *cache = NULL;
5745         struct btrfs_device *device;
5746         struct list_head *devices;
5747         u64 group_trimmed;
5748         u64 range_end = U64_MAX;
5749         u64 start;
5750         u64 end;
5751         u64 trimmed = 0;
5752         u64 bg_failed = 0;
5753         u64 dev_failed = 0;
5754         int bg_ret = 0;
5755         int dev_ret = 0;
5756         int ret = 0;
5757
5758         /*
5759          * Check range overflow if range->len is set.
5760          * The default range->len is U64_MAX.
5761          */
5762         if (range->len != U64_MAX &&
5763             check_add_overflow(range->start, range->len, &range_end))
5764                 return -EINVAL;
5765
5766         cache = btrfs_lookup_first_block_group(fs_info, range->start);
5767         for (; cache; cache = btrfs_next_block_group(cache)) {
5768                 if (cache->start >= range_end) {
5769                         btrfs_put_block_group(cache);
5770                         break;
5771                 }
5772
5773                 start = max(range->start, cache->start);
5774                 end = min(range_end, cache->start + cache->length);
5775
5776                 if (end - start >= range->minlen) {
5777                         if (!btrfs_block_group_done(cache)) {
5778                                 ret = btrfs_cache_block_group(cache, 0);
5779                                 if (ret) {
5780                                         bg_failed++;
5781                                         bg_ret = ret;
5782                                         continue;
5783                                 }
5784                                 ret = btrfs_wait_block_group_cache_done(cache);
5785                                 if (ret) {
5786                                         bg_failed++;
5787                                         bg_ret = ret;
5788                                         continue;
5789                                 }
5790                         }
5791                         ret = btrfs_trim_block_group(cache,
5792                                                      &group_trimmed,
5793                                                      start,
5794                                                      end,
5795                                                      range->minlen);
5796
5797                         trimmed += group_trimmed;
5798                         if (ret) {
5799                                 bg_failed++;
5800                                 bg_ret = ret;
5801                                 continue;
5802                         }
5803                 }
5804         }
5805
5806         if (bg_failed)
5807                 btrfs_warn(fs_info,
5808                         "failed to trim %llu block group(s), last error %d",
5809                         bg_failed, bg_ret);
5810         mutex_lock(&fs_info->fs_devices->device_list_mutex);
5811         devices = &fs_info->fs_devices->devices;
5812         list_for_each_entry(device, devices, dev_list) {
5813                 ret = btrfs_trim_free_extents(device, &group_trimmed);
5814                 if (ret) {
5815                         dev_failed++;
5816                         dev_ret = ret;
5817                         break;
5818                 }
5819
5820                 trimmed += group_trimmed;
5821         }
5822         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
5823
5824         if (dev_failed)
5825                 btrfs_warn(fs_info,
5826                         "failed to trim %llu device(s), last error %d",
5827                         dev_failed, dev_ret);
5828         range->len = trimmed;
5829         if (bg_ret)
5830                 return bg_ret;
5831         return dev_ret;
5832 }