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
  37 #undef SCRAMBLE_DELAYED_REFS
  38
  39
  40 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  41                                struct btrfs_delayed_ref_node *node, u64 parent,
  42                                u64 root_objectid, u64 owner_objectid,
  43                                u64 owner_offset, int refs_to_drop,
  44                                struct btrfs_delayed_extent_op *extra_op);
  45 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  46                                     struct extent_buffer *leaf,
  47                                     struct btrfs_extent_item *ei);
  48 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  49                                       u64 parent, u64 root_objectid,
  50                                       u64 flags, u64 owner, u64 offset,
  51                                       struct btrfs_key *ins, int ref_mod);
  52 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  53                                      struct btrfs_delayed_ref_node *node,
  54                                      struct btrfs_delayed_extent_op *extent_op);
  55 static int find_next_key(struct btrfs_path *path, int level,
  56                          struct btrfs_key *key);
  57
  58 static int block_group_bits(struct btrfs_block_group *cache, u64 bits)
  59 {
  60         return (cache->flags & bits) == bits;
  61 }
  62
  63 int btrfs_add_excluded_extent(struct btrfs_fs_info *fs_info,
  64                               u64 start, u64 num_bytes)
  65 {
  66         u64 end = start + num_bytes - 1;
  67         set_extent_bits(&fs_info->excluded_extents, start, end,
  68                         EXTENT_UPTODATE);
  69         return 0;
  70 }
  71
  72 void btrfs_free_excluded_extents(struct btrfs_block_group *cache)
  73 {
  74         struct btrfs_fs_info *fs_info = cache->fs_info;
  75         u64 start, end;
  76
  77         start = cache->start;
  78         end = start + cache->length - 1;
  79
  80         clear_extent_bits(&fs_info->excluded_extents, start, end,
  81                           EXTENT_UPTODATE);
  82 }
  83
  84 static u64 generic_ref_to_space_flags(struct btrfs_ref *ref)
  85 {
  86         if (ref->type == BTRFS_REF_METADATA) {
  87                 if (ref->tree_ref.root == BTRFS_CHUNK_TREE_OBJECTID)
  88                         return BTRFS_BLOCK_GROUP_SYSTEM;
  89                 else
  90                         return BTRFS_BLOCK_GROUP_METADATA;
  91         }
  92         return BTRFS_BLOCK_GROUP_DATA;
  93 }
  94
  95 static void add_pinned_bytes(struct btrfs_fs_info *fs_info,
  96                              struct btrfs_ref *ref)
  97 {
  98         struct btrfs_space_info *space_info;
  99         u64 flags = generic_ref_to_space_flags(ref);
 100
 101         space_info = btrfs_find_space_info(fs_info, flags);
 102         ASSERT(space_info);
 103         percpu_counter_add_batch(&space_info->total_bytes_pinned, ref->len,
 104                     BTRFS_TOTAL_BYTES_PINNED_BATCH);
 105 }
 106
 107 static void sub_pinned_bytes(struct btrfs_fs_info *fs_info,
 108                              struct btrfs_ref *ref)
 109 {
 110         struct btrfs_space_info *space_info;
 111         u64 flags = generic_ref_to_space_flags(ref);
 112
 113         space_info = btrfs_find_space_info(fs_info, flags);
 114         ASSERT(space_info);
 115         percpu_counter_add_batch(&space_info->total_bytes_pinned, -ref->len,
 116                     BTRFS_TOTAL_BYTES_PINNED_BATCH);
 117 }
 118
 119 /* simple helper to search for an existing data extent at a given offset */
 120 int btrfs_lookup_data_extent(struct btrfs_fs_info *fs_info, u64 start, u64 len)
 121 {
 122         int ret;
 123         struct btrfs_key key;
 124         struct btrfs_path *path;
 125
 126         path = btrfs_alloc_path();
 127         if (!path)
 128                 return -ENOMEM;
 129
 130         key.objectid = start;
 131         key.offset = len;
 132         key.type = BTRFS_EXTENT_ITEM_KEY;
 133         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
 134         btrfs_free_path(path);
 135         return ret;
 136 }
 137
 138 /*
 139  * helper function to lookup reference count and flags of a tree block.
 140  *
 141  * the head node for delayed ref is used to store the sum of all the
 142  * reference count modifications queued up in the rbtree. the head
 143  * node may also store the extent flags to set. This way you can check
 144  * to see what the reference count and extent flags would be if all of
 145  * the delayed refs are not processed.
 146  */
 147 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 148                              struct btrfs_fs_info *fs_info, u64 bytenr,
 149                              u64 offset, int metadata, u64 *refs, u64 *flags)
 150 {
 151         struct btrfs_delayed_ref_head *head;
 152         struct btrfs_delayed_ref_root *delayed_refs;
 153         struct btrfs_path *path;
 154         struct btrfs_extent_item *ei;
 155         struct extent_buffer *leaf;
 156         struct btrfs_key key;
 157         u32 item_size;
 158         u64 num_refs;
 159         u64 extent_flags;
 160         int ret;
 161
 162         /*
 163          * If we don't have skinny metadata, don't bother doing anything
 164          * different
 165          */
 166         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA)) {
 167                 offset = fs_info->nodesize;
 168                 metadata = 0;
 169         }
 170
 171         path = btrfs_alloc_path();
 172         if (!path)
 173                 return -ENOMEM;
 174
 175         if (!trans) {
 176                 path->skip_locking = 1;
 177                 path->search_commit_root = 1;
 178         }
 179
 180 search_again:
 181         key.objectid = bytenr;
 182         key.offset = offset;
 183         if (metadata)
 184                 key.type = BTRFS_METADATA_ITEM_KEY;
 185         else
 186                 key.type = BTRFS_EXTENT_ITEM_KEY;
 187
 188         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 0);
 189         if (ret < 0)
 190                 goto out_free;
 191
 192         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 193                 if (path->slots[0]) {
 194                         path->slots[0]--;
 195                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 196                                               path->slots[0]);
 197                         if (key.objectid == bytenr &&
 198                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 199                             key.offset == fs_info->nodesize)
 200                                 ret = 0;
 201                 }
 202         }
 203
 204         if (ret == 0) {
 205                 leaf = path->nodes[0];
 206                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 207                 if (item_size >= sizeof(*ei)) {
 208                         ei = btrfs_item_ptr(leaf, path->slots[0],
 209                                             struct btrfs_extent_item);
 210                         num_refs = btrfs_extent_refs(leaf, ei);
 211                         extent_flags = btrfs_extent_flags(leaf, ei);
 212                 } else {
 213                         ret = -EINVAL;
 214                         btrfs_print_v0_err(fs_info);
 215                         if (trans)
 216                                 btrfs_abort_transaction(trans, ret);
 217                         else
 218                                 btrfs_handle_fs_error(fs_info, ret, NULL);
 219
 220                         goto out_free;
 221                 }
 222
 223                 BUG_ON(num_refs == 0);
 224         } else {
 225                 num_refs = 0;
 226                 extent_flags = 0;
 227                 ret = 0;
 228         }
 229
 230         if (!trans)
 231                 goto out;
 232
 233         delayed_refs = &trans->transaction->delayed_refs;
 234         spin_lock(&delayed_refs->lock);
 235         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
 236         if (head) {
 237                 if (!mutex_trylock(&head->mutex)) {
 238                         refcount_inc(&head->refs);
 239                         spin_unlock(&delayed_refs->lock);
 240
 241                         btrfs_release_path(path);
 242
 243                         /*
 244                          * Mutex was contended, block until it's released and try
 245                          * again
 246                          */
 247                         mutex_lock(&head->mutex);
 248                         mutex_unlock(&head->mutex);
 249                         btrfs_put_delayed_ref_head(head);
 250                         goto search_again;
 251                 }
 252                 spin_lock(&head->lock);
 253                 if (head->extent_op && head->extent_op->update_flags)
 254                         extent_flags |= head->extent_op->flags_to_set;
 255                 else
 256                         BUG_ON(num_refs == 0);
 257
 258                 num_refs += head->ref_mod;
 259                 spin_unlock(&head->lock);
 260                 mutex_unlock(&head->mutex);
 261         }
 262         spin_unlock(&delayed_refs->lock);
 263 out:
 264         WARN_ON(num_refs == 0);
 265         if (refs)
 266                 *refs = num_refs;
 267         if (flags)
 268                 *flags = extent_flags;
 269 out_free:
 270         btrfs_free_path(path);
 271         return ret;
 272 }
 273
 274 /*
 275  * Back reference rules.  Back refs have three main goals:
 276  *
 277  * 1) differentiate between all holders of references to an extent so that
 278  *    when a reference is dropped we can make sure it was a valid reference
 279  *    before freeing the extent.
 280  *
 281  * 2) Provide enough information to quickly find the holders of an extent
 282  *    if we notice a given block is corrupted or bad.
 283  *
 284  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 285  *    maintenance.  This is actually the same as #2, but with a slightly
 286  *    different use case.
 287  *
 288  * There are two kinds of back refs. The implicit back refs is optimized
 289  * for pointers in non-shared tree blocks. For a given pointer in a block,
 290  * back refs of this kind provide information about the block's owner tree
 291  * and the pointer's key. These information allow us to find the block by
 292  * b-tree searching. The full back refs is for pointers in tree blocks not
 293  * referenced by their owner trees. The location of tree block is recorded
 294  * in the back refs. Actually the full back refs is generic, and can be
 295  * used in all cases the implicit back refs is used. The major shortcoming
 296  * of the full back refs is its overhead. Every time a tree block gets
 297  * COWed, we have to update back refs entry for all pointers in it.
 298  *
 299  * For a newly allocated tree block, we use implicit back refs for
 300  * pointers in it. This means most tree related operations only involve
 301  * implicit back refs. For a tree block created in old transaction, the
 302  * only way to drop a reference to it is COW it. So we can detect the
 303  * event that tree block loses its owner tree's reference and do the
 304  * back refs conversion.
 305  *
 306  * When a tree block is COWed through a tree, there are four cases:
 307  *
 308  * The reference count of the block is one and the tree is the block's
 309  * owner tree. Nothing to do in this case.
 310  *
 311  * The reference count of the block is one and the tree is not the
 312  * block's owner tree. In this case, full back refs is used for pointers
 313  * in the block. Remove these full back refs, add implicit back refs for
 314  * every pointers in the new block.
 315  *
 316  * The reference count of the block is greater than one and the tree is
 317  * the block's owner tree. In this case, implicit back refs is used for
 318  * pointers in the block. Add full back refs for every pointers in the
 319  * block, increase lower level extents' reference counts. The original
 320  * implicit back refs are entailed to the new block.
 321  *
 322  * The reference count of the block is greater than one and the tree is
 323  * not the block's owner tree. Add implicit back refs for every pointer in
 324  * the new block, increase lower level extents' reference count.
 325  *
 326  * Back Reference Key composing:
 327  *
 328  * The key objectid corresponds to the first byte in the extent,
 329  * The key type is used to differentiate between types of back refs.
 330  * There are different meanings of the key offset for different types
 331  * of back refs.
 332  *
 333  * File extents can be referenced by:
 334  *
 335  * - multiple snapshots, subvolumes, or different generations in one subvol
 336  * - different files inside a single subvolume
 337  * - different offsets inside a file (bookend extents in file.c)
 338  *
 339  * The extent ref structure for the implicit back refs has fields for:
 340  *
 341  * - Objectid of the subvolume root
 342  * - objectid of the file holding the reference
 343  * - original offset in the file
 344  * - how many bookend extents
 345  *
 346  * The key offset for the implicit back refs is hash of the first
 347  * three fields.
 348  *
 349  * The extent ref structure for the full back refs has field for:
 350  *
 351  * - number of pointers in the tree leaf
 352  *
 353  * The key offset for the implicit back refs is the first byte of
 354  * the tree leaf
 355  *
 356  * When a file extent is allocated, The implicit back refs is used.
 357  * the fields are filled in:
 358  *
 359  *     (root_key.objectid, inode objectid, offset in file, 1)
 360  *
 361  * When a file extent is removed file truncation, we find the
 362  * corresponding implicit back refs and check the following fields:
 363  *
 364  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
 365  *
 366  * Btree extents can be referenced by:
 367  *
 368  * - Different subvolumes
 369  *
 370  * Both the implicit back refs and the full back refs for tree blocks
 371  * only consist of key. The key offset for the implicit back refs is
 372  * objectid of block's owner tree. The key offset for the full back refs
 373  * is the first byte of parent block.
 374  *
 375  * When implicit back refs is used, information about the lowest key and
 376  * level of the tree block are required. These information are stored in
 377  * tree block info structure.
 378  */
 379
 380 /*
 381  * is_data == BTRFS_REF_TYPE_BLOCK, tree block type is required,
 382  * is_data == BTRFS_REF_TYPE_DATA, data type is requiried,
 383  * is_data == BTRFS_REF_TYPE_ANY, either type is OK.
 384  */
 385 int btrfs_get_extent_inline_ref_type(const struct extent_buffer *eb,
 386                                      struct btrfs_extent_inline_ref *iref,
 387                                      enum btrfs_inline_ref_type is_data)
 388 {
 389         int type = btrfs_extent_inline_ref_type(eb, iref);
 390         u64 offset = btrfs_extent_inline_ref_offset(eb, iref);
 391
 392         if (type == BTRFS_TREE_BLOCK_REF_KEY ||
 393             type == BTRFS_SHARED_BLOCK_REF_KEY ||
 394             type == BTRFS_SHARED_DATA_REF_KEY ||
 395             type == BTRFS_EXTENT_DATA_REF_KEY) {
 396                 if (is_data == BTRFS_REF_TYPE_BLOCK) {
 397                         if (type == BTRFS_TREE_BLOCK_REF_KEY)
 398                                 return type;
 399                         if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
 400                                 ASSERT(eb->fs_info);
 401                                 /*
 402                                  * Every shared one has parent tree
 403                                  * block, which must be aligned to
 404                                  * nodesize.
 405                                  */
 406                                 if (offset &&
 407                                     IS_ALIGNED(offset, eb->fs_info->nodesize))
 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
 417                                  * block, which must be aligned to
 418                                  * nodesize.
 419                                  */
 420                                 if (offset &&
 421                                     IS_ALIGNED(offset, eb->fs_info->nodesize))
 422                                         return type;
 423                         }
 424                 } else {
 425                         ASSERT(is_data == BTRFS_REF_TYPE_ANY);
 426                         return type;
 427                 }
 428         }
 429
 430         btrfs_print_leaf((struct extent_buffer *)eb);
 431         btrfs_err(eb->fs_info, "eb %llu invalid extent inline ref type %d",
 432                   eb->start, 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 insert_extent_backref(struct btrfs_trans_handle *trans,
1193                                  struct btrfs_path *path,
1194                                  u64 bytenr, u64 parent, u64 root_objectid,
1195                                  u64 owner, u64 offset, int refs_to_add)
1196 {
1197         int ret;
1198         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1199                 BUG_ON(refs_to_add != 1);
1200                 ret = insert_tree_block_ref(trans, path, bytenr, parent,
1201                                             root_objectid);
1202         } else {
1203                 ret = insert_extent_data_ref(trans, path, bytenr, parent,
1204                                              root_objectid, owner, offset,
1205                                              refs_to_add);
1206         }
1207         return ret;
1208 }
1209
1210 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1211                                  struct btrfs_path *path,
1212                                  struct btrfs_extent_inline_ref *iref,
1213                                  int refs_to_drop, int is_data, int *last_ref)
1214 {
1215         int ret = 0;
1216
1217         BUG_ON(!is_data && refs_to_drop != 1);
1218         if (iref) {
1219                 update_inline_extent_backref(path, iref, -refs_to_drop, NULL,
1220                                              last_ref);
1221         } else if (is_data) {
1222                 ret = remove_extent_data_ref(trans, path, refs_to_drop,
1223                                              last_ref);
1224         } else {
1225                 *last_ref = 1;
1226                 ret = btrfs_del_item(trans, trans->fs_info->extent_root, path);
1227         }
1228         return ret;
1229 }
1230
1231 static int btrfs_issue_discard(struct block_device *bdev, u64 start, u64 len,
1232                                u64 *discarded_bytes)
1233 {
1234         int j, ret = 0;
1235         u64 bytes_left, end;
1236         u64 aligned_start = ALIGN(start, 1 << 9);
1237
1238         if (WARN_ON(start != aligned_start)) {
1239                 len -= aligned_start - start;
1240                 len = round_down(len, 1 << 9);
1241                 start = aligned_start;
1242         }
1243
1244         *discarded_bytes = 0;
1245
1246         if (!len)
1247                 return 0;
1248
1249         end = start + len;
1250         bytes_left = len;
1251
1252         /* Skip any superblocks on this device. */
1253         for (j = 0; j < BTRFS_SUPER_MIRROR_MAX; j++) {
1254                 u64 sb_start = btrfs_sb_offset(j);
1255                 u64 sb_end = sb_start + BTRFS_SUPER_INFO_SIZE;
1256                 u64 size = sb_start - start;
1257
1258                 if (!in_range(sb_start, start, bytes_left) &&
1259                     !in_range(sb_end, start, bytes_left) &&
1260                     !in_range(start, sb_start, BTRFS_SUPER_INFO_SIZE))
1261                         continue;
1262
1263                 /*
1264                  * Superblock spans beginning of range.  Adjust start and
1265                  * try again.
1266                  */
1267                 if (sb_start <= start) {
1268                         start += sb_end - start;
1269                         if (start > end) {
1270                                 bytes_left = 0;
1271                                 break;
1272                         }
1273                         bytes_left = end - start;
1274                         continue;
1275                 }
1276
1277                 if (size) {
1278                         ret = blkdev_issue_discard(bdev, start >> 9, size >> 9,
1279                                                    GFP_NOFS, 0);
1280                         if (!ret)
1281                                 *discarded_bytes += size;
1282                         else if (ret != -EOPNOTSUPP)
1283                                 return ret;
1284                 }
1285
1286                 start = sb_end;
1287                 if (start > end) {
1288                         bytes_left = 0;
1289                         break;
1290                 }
1291                 bytes_left = end - start;
1292         }
1293
1294         if (bytes_left) {
1295                 ret = blkdev_issue_discard(bdev, start >> 9, bytes_left >> 9,
1296                                            GFP_NOFS, 0);
1297                 if (!ret)
1298                         *discarded_bytes += bytes_left;
1299         }
1300         return ret;
1301 }
1302
1303 int btrfs_discard_extent(struct btrfs_fs_info *fs_info, u64 bytenr,
1304                          u64 num_bytes, u64 *actual_bytes)
1305 {
1306         int ret = 0;
1307         u64 discarded_bytes = 0;
1308         u64 end = bytenr + num_bytes;
1309         u64 cur = bytenr;
1310         struct btrfs_bio *bbio = NULL;
1311
1312
1313         /*
1314          * Avoid races with device replace and make sure our bbio has devices
1315          * associated to its stripes that don't go away while we are discarding.
1316          */
1317         btrfs_bio_counter_inc_blocked(fs_info);
1318         while (cur < end) {
1319                 struct btrfs_bio_stripe *stripe;
1320                 int i;
1321
1322                 num_bytes = end - cur;
1323                 /* Tell the block device(s) that the sectors can be discarded */
1324                 ret = btrfs_map_block(fs_info, BTRFS_MAP_DISCARD, cur,
1325                                       &num_bytes, &bbio, 0);
1326                 /*
1327                  * Error can be -ENOMEM, -ENOENT (no such chunk mapping) or
1328                  * -EOPNOTSUPP. For any such error, @num_bytes is not updated,
1329                  * thus we can't continue anyway.
1330                  */
1331                 if (ret < 0)
1332                         goto out;
1333
1334                 stripe = bbio->stripes;
1335                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
1336                         u64 bytes;
1337                         struct request_queue *req_q;
1338
1339                         if (!stripe->dev->bdev) {
1340                                 ASSERT(btrfs_test_opt(fs_info, DEGRADED));
1341                                 continue;
1342                         }
1343                         req_q = bdev_get_queue(stripe->dev->bdev);
1344                         if (!blk_queue_discard(req_q))
1345                                 continue;
1346
1347                         ret = btrfs_issue_discard(stripe->dev->bdev,
1348                                                   stripe->physical,
1349                                                   stripe->length,
1350                                                   &bytes);
1351                         if (!ret) {
1352                                 discarded_bytes += bytes;
1353                         } else if (ret != -EOPNOTSUPP) {
1354                                 /*
1355                                  * Logic errors or -ENOMEM, or -EIO, but
1356                                  * unlikely to happen.
1357                                  *
1358                                  * And since there are two loops, explicitly
1359                                  * go to out to avoid confusion.
1360                                  */
1361                                 btrfs_put_bbio(bbio);
1362                                 goto out;
1363                         }
1364
1365                         /*
1366                          * Just in case we get back EOPNOTSUPP for some reason,
1367                          * just ignore the return value so we don't screw up
1368                          * people calling discard_extent.
1369                          */
1370                         ret = 0;
1371                 }
1372                 btrfs_put_bbio(bbio);
1373                 cur += num_bytes;
1374         }
1375 out:
1376         btrfs_bio_counter_dec(fs_info);
1377
1378         if (actual_bytes)
1379                 *actual_bytes = discarded_bytes;
1380
1381
1382         if (ret == -EOPNOTSUPP)
1383                 ret = 0;
1384         return ret;
1385 }
1386
1387 /* Can return -ENOMEM */
1388 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1389                          struct btrfs_ref *generic_ref)
1390 {
1391         struct btrfs_fs_info *fs_info = trans->fs_info;
1392         int old_ref_mod, new_ref_mod;
1393         int ret;
1394
1395         ASSERT(generic_ref->type != BTRFS_REF_NOT_SET &&
1396                generic_ref->action);
1397         BUG_ON(generic_ref->type == BTRFS_REF_METADATA &&
1398                generic_ref->tree_ref.root == BTRFS_TREE_LOG_OBJECTID);
1399
1400         if (generic_ref->type == BTRFS_REF_METADATA)
1401                 ret = btrfs_add_delayed_tree_ref(trans, generic_ref,
1402                                 NULL, &old_ref_mod, &new_ref_mod);
1403         else
1404                 ret = btrfs_add_delayed_data_ref(trans, generic_ref, 0,
1405                                                  &old_ref_mod, &new_ref_mod);
1406
1407         btrfs_ref_tree_mod(fs_info, generic_ref);
1408
1409         if (ret == 0 && old_ref_mod < 0 && new_ref_mod >= 0)
1410                 sub_pinned_bytes(fs_info, generic_ref);
1411
1412         return ret;
1413 }
1414
1415 /*
1416  * __btrfs_inc_extent_ref - insert backreference for a given extent
1417  *
1418  * @trans:          Handle of transaction
1419  *
1420  * @node:           The delayed ref node used to get the bytenr/length for
1421  *                  extent whose references are incremented.
1422  *
1423  * @parent:         If this is a shared extent (BTRFS_SHARED_DATA_REF_KEY/
1424  *                  BTRFS_SHARED_BLOCK_REF_KEY) then it holds the logical
1425  *                  bytenr of the parent block. Since new extents are always
1426  *                  created with indirect references, this will only be the case
1427  *                  when relocating a shared extent. In that case, root_objectid
1428  *                  will be BTRFS_TREE_RELOC_OBJECTID. Otheriwse, parent must
1429  *                  be 0
1430  *
1431  * @root_objectid:  The id of the root where this modification has originated,
1432  *                  this can be either one of the well-known metadata trees or
1433  *                  the subvolume id which references this extent.
1434  *
1435  * @owner:          For data extents it is the inode number of the owning file.
1436  *                  For metadata extents this parameter holds the level in the
1437  *                  tree of the extent.
1438  *
1439  * @offset:         For metadata extents the offset is ignored and is currently
1440  *                  always passed as 0. For data extents it is the fileoffset
1441  *                  this extent belongs to.
1442  *
1443  * @refs_to_add     Number of references to add
1444  *
1445  * @extent_op       Pointer to a structure, holding information necessary when
1446  *                  updating a tree block's flags
1447  *
1448  */
1449 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1450                                   struct btrfs_delayed_ref_node *node,
1451                                   u64 parent, u64 root_objectid,
1452                                   u64 owner, u64 offset, int refs_to_add,
1453                                   struct btrfs_delayed_extent_op *extent_op)
1454 {
1455         struct btrfs_path *path;
1456         struct extent_buffer *leaf;
1457         struct btrfs_extent_item *item;
1458         struct btrfs_key key;
1459         u64 bytenr = node->bytenr;
1460         u64 num_bytes = node->num_bytes;
1461         u64 refs;
1462         int ret;
1463
1464         path = btrfs_alloc_path();
1465         if (!path)
1466                 return -ENOMEM;
1467
1468         path->reada = READA_FORWARD;
1469         path->leave_spinning = 1;
1470         /* this will setup the path even if it fails to insert the back ref */
1471         ret = insert_inline_extent_backref(trans, path, bytenr, num_bytes,
1472                                            parent, root_objectid, owner,
1473                                            offset, refs_to_add, extent_op);
1474         if ((ret < 0 && ret != -EAGAIN) || !ret)
1475                 goto out;
1476
1477         /*
1478          * Ok we had -EAGAIN which means we didn't have space to insert and
1479          * inline extent ref, so just update the reference count and add a
1480          * normal backref.
1481          */
1482         leaf = path->nodes[0];
1483         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1484         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1485         refs = btrfs_extent_refs(leaf, item);
1486         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
1487         if (extent_op)
1488                 __run_delayed_extent_op(extent_op, leaf, item);
1489
1490         btrfs_mark_buffer_dirty(leaf);
1491         btrfs_release_path(path);
1492
1493         path->reada = READA_FORWARD;
1494         path->leave_spinning = 1;
1495         /* now insert the actual backref */
1496         ret = insert_extent_backref(trans, path, bytenr, parent, root_objectid,
1497                                     owner, offset, refs_to_add);
1498         if (ret)
1499                 btrfs_abort_transaction(trans, ret);
1500 out:
1501         btrfs_free_path(path);
1502         return ret;
1503 }
1504
1505 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
1506                                 struct btrfs_delayed_ref_node *node,
1507                                 struct btrfs_delayed_extent_op *extent_op,
1508                                 int insert_reserved)
1509 {
1510         int ret = 0;
1511         struct btrfs_delayed_data_ref *ref;
1512         struct btrfs_key ins;
1513         u64 parent = 0;
1514         u64 ref_root = 0;
1515         u64 flags = 0;
1516
1517         ins.objectid = node->bytenr;
1518         ins.offset = node->num_bytes;
1519         ins.type = BTRFS_EXTENT_ITEM_KEY;
1520
1521         ref = btrfs_delayed_node_to_data_ref(node);
1522         trace_run_delayed_data_ref(trans->fs_info, node, ref, node->action);
1523
1524         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
1525                 parent = ref->parent;
1526         ref_root = ref->root;
1527
1528         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1529                 if (extent_op)
1530                         flags |= extent_op->flags_to_set;
1531                 ret = alloc_reserved_file_extent(trans, parent, ref_root,
1532                                                  flags, ref->objectid,
1533                                                  ref->offset, &ins,
1534                                                  node->ref_mod);
1535         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1536                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
1537                                              ref->objectid, ref->offset,
1538                                              node->ref_mod, extent_op);
1539         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1540                 ret = __btrfs_free_extent(trans, node, parent,
1541                                           ref_root, ref->objectid,
1542                                           ref->offset, node->ref_mod,
1543                                           extent_op);
1544         } else {
1545                 BUG();
1546         }
1547         return ret;
1548 }
1549
1550 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
1551                                     struct extent_buffer *leaf,
1552                                     struct btrfs_extent_item *ei)
1553 {
1554         u64 flags = btrfs_extent_flags(leaf, ei);
1555         if (extent_op->update_flags) {
1556                 flags |= extent_op->flags_to_set;
1557                 btrfs_set_extent_flags(leaf, ei, flags);
1558         }
1559
1560         if (extent_op->update_key) {
1561                 struct btrfs_tree_block_info *bi;
1562                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
1563                 bi = (struct btrfs_tree_block_info *)(ei + 1);
1564                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
1565         }
1566 }
1567
1568 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
1569                                  struct btrfs_delayed_ref_head *head,
1570                                  struct btrfs_delayed_extent_op *extent_op)
1571 {
1572         struct btrfs_fs_info *fs_info = trans->fs_info;
1573         struct btrfs_key key;
1574         struct btrfs_path *path;
1575         struct btrfs_extent_item *ei;
1576         struct extent_buffer *leaf;
1577         u32 item_size;
1578         int ret;
1579         int err = 0;
1580         int metadata = !extent_op->is_data;
1581
1582         if (TRANS_ABORTED(trans))
1583                 return 0;
1584
1585         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1586                 metadata = 0;
1587
1588         path = btrfs_alloc_path();
1589         if (!path)
1590                 return -ENOMEM;
1591
1592         key.objectid = head->bytenr;
1593
1594         if (metadata) {
1595                 key.type = BTRFS_METADATA_ITEM_KEY;
1596                 key.offset = extent_op->level;
1597         } else {
1598                 key.type = BTRFS_EXTENT_ITEM_KEY;
1599                 key.offset = head->num_bytes;
1600         }
1601
1602 again:
1603         path->reada = READA_FORWARD;
1604         path->leave_spinning = 1;
1605         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 1);
1606         if (ret < 0) {
1607                 err = ret;
1608                 goto out;
1609         }
1610         if (ret > 0) {
1611                 if (metadata) {
1612                         if (path->slots[0] > 0) {
1613                                 path->slots[0]--;
1614                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
1615                                                       path->slots[0]);
1616                                 if (key.objectid == head->bytenr &&
1617                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
1618                                     key.offset == head->num_bytes)
1619                                         ret = 0;
1620                         }
1621                         if (ret > 0) {
1622                                 btrfs_release_path(path);
1623                                 metadata = 0;
1624
1625                                 key.objectid = head->bytenr;
1626                                 key.offset = head->num_bytes;
1627                                 key.type = BTRFS_EXTENT_ITEM_KEY;
1628                                 goto again;
1629                         }
1630                 } else {
1631                         err = -EIO;
1632                         goto out;
1633                 }
1634         }
1635
1636         leaf = path->nodes[0];
1637         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1638
1639         if (unlikely(item_size < sizeof(*ei))) {
1640                 err = -EINVAL;
1641                 btrfs_print_v0_err(fs_info);
1642                 btrfs_abort_transaction(trans, err);
1643                 goto out;
1644         }
1645
1646         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1647         __run_delayed_extent_op(extent_op, leaf, ei);
1648
1649         btrfs_mark_buffer_dirty(leaf);
1650 out:
1651         btrfs_free_path(path);
1652         return err;
1653 }
1654
1655 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
1656                                 struct btrfs_delayed_ref_node *node,
1657                                 struct btrfs_delayed_extent_op *extent_op,
1658                                 int insert_reserved)
1659 {
1660         int ret = 0;
1661         struct btrfs_delayed_tree_ref *ref;
1662         u64 parent = 0;
1663         u64 ref_root = 0;
1664
1665         ref = btrfs_delayed_node_to_tree_ref(node);
1666         trace_run_delayed_tree_ref(trans->fs_info, node, ref, node->action);
1667
1668         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
1669                 parent = ref->parent;
1670         ref_root = ref->root;
1671
1672         if (node->ref_mod != 1) {
1673                 btrfs_err(trans->fs_info,
1674         "btree block(%llu) has %d references rather than 1: action %d ref_root %llu parent %llu",
1675                           node->bytenr, node->ref_mod, node->action, ref_root,
1676                           parent);
1677                 return -EIO;
1678         }
1679         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1680                 BUG_ON(!extent_op || !extent_op->update_flags);
1681                 ret = alloc_reserved_tree_block(trans, node, extent_op);
1682         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1683                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
1684                                              ref->level, 0, 1, extent_op);
1685         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1686                 ret = __btrfs_free_extent(trans, node, parent, ref_root,
1687                                           ref->level, 0, 1, extent_op);
1688         } else {
1689                 BUG();
1690         }
1691         return ret;
1692 }
1693
1694 /* helper function to actually process a single delayed ref entry */
1695 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
1696                                struct btrfs_delayed_ref_node *node,
1697                                struct btrfs_delayed_extent_op *extent_op,
1698                                int insert_reserved)
1699 {
1700         int ret = 0;
1701
1702         if (TRANS_ABORTED(trans)) {
1703                 if (insert_reserved)
1704                         btrfs_pin_extent(trans, node->bytenr, node->num_bytes, 1);
1705                 return 0;
1706         }
1707
1708         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
1709             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
1710                 ret = run_delayed_tree_ref(trans, node, extent_op,
1711                                            insert_reserved);
1712         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
1713                  node->type == BTRFS_SHARED_DATA_REF_KEY)
1714                 ret = run_delayed_data_ref(trans, node, extent_op,
1715                                            insert_reserved);
1716         else
1717                 BUG();
1718         if (ret && insert_reserved)
1719                 btrfs_pin_extent(trans, node->bytenr, node->num_bytes, 1);
1720         return ret;
1721 }
1722
1723 static inline struct btrfs_delayed_ref_node *
1724 select_delayed_ref(struct btrfs_delayed_ref_head *head)
1725 {
1726         struct btrfs_delayed_ref_node *ref;
1727
1728         if (RB_EMPTY_ROOT(&head->ref_tree.rb_root))
1729                 return NULL;
1730
1731         /*
1732          * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
1733          * This is to prevent a ref count from going down to zero, which deletes
1734          * the extent item from the extent tree, when there still are references
1735          * to add, which would fail because they would not find the extent item.
1736          */
1737         if (!list_empty(&head->ref_add_list))
1738                 return list_first_entry(&head->ref_add_list,
1739                                 struct btrfs_delayed_ref_node, add_list);
1740
1741         ref = rb_entry(rb_first_cached(&head->ref_tree),
1742                        struct btrfs_delayed_ref_node, ref_node);
1743         ASSERT(list_empty(&ref->add_list));
1744         return ref;
1745 }
1746
1747 static void unselect_delayed_ref_head(struct btrfs_delayed_ref_root *delayed_refs,
1748                                       struct btrfs_delayed_ref_head *head)
1749 {
1750         spin_lock(&delayed_refs->lock);
1751         head->processing = 0;
1752         delayed_refs->num_heads_ready++;
1753         spin_unlock(&delayed_refs->lock);
1754         btrfs_delayed_ref_unlock(head);
1755 }
1756
1757 static struct btrfs_delayed_extent_op *cleanup_extent_op(
1758                                 struct btrfs_delayed_ref_head *head)
1759 {
1760         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
1761
1762         if (!extent_op)
1763                 return NULL;
1764
1765         if (head->must_insert_reserved) {
1766                 head->extent_op = NULL;
1767                 btrfs_free_delayed_extent_op(extent_op);
1768                 return NULL;
1769         }
1770         return extent_op;
1771 }
1772
1773 static int run_and_cleanup_extent_op(struct btrfs_trans_handle *trans,
1774                                      struct btrfs_delayed_ref_head *head)
1775 {
1776         struct btrfs_delayed_extent_op *extent_op;
1777         int ret;
1778
1779         extent_op = cleanup_extent_op(head);
1780         if (!extent_op)
1781                 return 0;
1782         head->extent_op = NULL;
1783         spin_unlock(&head->lock);
1784         ret = run_delayed_extent_op(trans, head, extent_op);
1785         btrfs_free_delayed_extent_op(extent_op);
1786         return ret ? ret : 1;
1787 }
1788
1789 void btrfs_cleanup_ref_head_accounting(struct btrfs_fs_info *fs_info,
1790                                   struct btrfs_delayed_ref_root *delayed_refs,
1791                                   struct btrfs_delayed_ref_head *head)
1792 {
1793         int nr_items = 1;       /* Dropping this ref head update. */
1794
1795         if (head->total_ref_mod < 0) {
1796                 struct btrfs_space_info *space_info;
1797                 u64 flags;
1798
1799                 if (head->is_data)
1800                         flags = BTRFS_BLOCK_GROUP_DATA;
1801                 else if (head->is_system)
1802                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
1803                 else
1804                         flags = BTRFS_BLOCK_GROUP_METADATA;
1805                 space_info = btrfs_find_space_info(fs_info, flags);
1806                 ASSERT(space_info);
1807                 percpu_counter_add_batch(&space_info->total_bytes_pinned,
1808                                    -head->num_bytes,
1809                                    BTRFS_TOTAL_BYTES_PINNED_BATCH);
1810
1811                 /*
1812                  * We had csum deletions accounted for in our delayed refs rsv,
1813                  * we need to drop the csum leaves for this update from our
1814                  * delayed_refs_rsv.
1815                  */
1816                 if (head->is_data) {
1817                         spin_lock(&delayed_refs->lock);
1818                         delayed_refs->pending_csums -= head->num_bytes;
1819                         spin_unlock(&delayed_refs->lock);
1820                         nr_items += btrfs_csum_bytes_to_leaves(fs_info,
1821                                 head->num_bytes);
1822                 }
1823         }
1824
1825         btrfs_delayed_refs_rsv_release(fs_info, nr_items);
1826 }
1827
1828 static int cleanup_ref_head(struct btrfs_trans_handle *trans,
1829                             struct btrfs_delayed_ref_head *head)
1830 {
1831
1832         struct btrfs_fs_info *fs_info = trans->fs_info;
1833         struct btrfs_delayed_ref_root *delayed_refs;
1834         int ret;
1835
1836         delayed_refs = &trans->transaction->delayed_refs;
1837
1838         ret = run_and_cleanup_extent_op(trans, head);
1839         if (ret < 0) {
1840                 unselect_delayed_ref_head(delayed_refs, head);
1841                 btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
1842                 return ret;
1843         } else if (ret) {
1844                 return ret;
1845         }
1846
1847         /*
1848          * Need to drop our head ref lock and re-acquire the delayed ref lock
1849          * and then re-check to make sure nobody got added.
1850          */
1851         spin_unlock(&head->lock);
1852         spin_lock(&delayed_refs->lock);
1853         spin_lock(&head->lock);
1854         if (!RB_EMPTY_ROOT(&head->ref_tree.rb_root) || head->extent_op) {
1855                 spin_unlock(&head->lock);
1856                 spin_unlock(&delayed_refs->lock);
1857                 return 1;
1858         }
1859         btrfs_delete_ref_head(delayed_refs, head);
1860         spin_unlock(&head->lock);
1861         spin_unlock(&delayed_refs->lock);
1862
1863         if (head->must_insert_reserved) {
1864                 btrfs_pin_extent(trans, head->bytenr, head->num_bytes, 1);
1865                 if (head->is_data) {
1866                         ret = btrfs_del_csums(trans, fs_info->csum_root,
1867                                               head->bytenr, head->num_bytes);
1868                 }
1869         }
1870
1871         btrfs_cleanup_ref_head_accounting(fs_info, delayed_refs, head);
1872
1873         trace_run_delayed_ref_head(fs_info, head, 0);
1874         btrfs_delayed_ref_unlock(head);
1875         btrfs_put_delayed_ref_head(head);
1876         return 0;
1877 }
1878
1879 static struct btrfs_delayed_ref_head *btrfs_obtain_ref_head(
1880                                         struct btrfs_trans_handle *trans)
1881 {
1882         struct btrfs_delayed_ref_root *delayed_refs =
1883                 &trans->transaction->delayed_refs;
1884         struct btrfs_delayed_ref_head *head = NULL;
1885         int ret;
1886
1887         spin_lock(&delayed_refs->lock);
1888         head = btrfs_select_ref_head(delayed_refs);
1889         if (!head) {
1890                 spin_unlock(&delayed_refs->lock);
1891                 return head;
1892         }
1893
1894         /*
1895          * Grab the lock that says we are going to process all the refs for
1896          * this head
1897          */
1898         ret = btrfs_delayed_ref_lock(delayed_refs, head);
1899         spin_unlock(&delayed_refs->lock);
1900
1901         /*
1902          * We may have dropped the spin lock to get the head mutex lock, and
1903          * that might have given someone else time to free the head.  If that's
1904          * true, it has been removed from our list and we can move on.
1905          */
1906         if (ret == -EAGAIN)
1907                 head = ERR_PTR(-EAGAIN);
1908
1909         return head;
1910 }
1911
1912 static int btrfs_run_delayed_refs_for_head(struct btrfs_trans_handle *trans,
1913                                     struct btrfs_delayed_ref_head *locked_ref,
1914                                     unsigned long *run_refs)
1915 {
1916         struct btrfs_fs_info *fs_info = trans->fs_info;
1917         struct btrfs_delayed_ref_root *delayed_refs;
1918         struct btrfs_delayed_extent_op *extent_op;
1919         struct btrfs_delayed_ref_node *ref;
1920         int must_insert_reserved = 0;
1921         int ret;
1922
1923         delayed_refs = &trans->transaction->delayed_refs;
1924
1925         lockdep_assert_held(&locked_ref->mutex);
1926         lockdep_assert_held(&locked_ref->lock);
1927
1928         while ((ref = select_delayed_ref(locked_ref))) {
1929                 if (ref->seq &&
1930                     btrfs_check_delayed_seq(fs_info, ref->seq)) {
1931                         spin_unlock(&locked_ref->lock);
1932                         unselect_delayed_ref_head(delayed_refs, locked_ref);
1933                         return -EAGAIN;
1934                 }
1935
1936                 (*run_refs)++;
1937                 ref->in_tree = 0;
1938                 rb_erase_cached(&ref->ref_node, &locked_ref->ref_tree);
1939                 RB_CLEAR_NODE(&ref->ref_node);
1940                 if (!list_empty(&ref->add_list))
1941                         list_del(&ref->add_list);
1942                 /*
1943                  * When we play the delayed ref, also correct the ref_mod on
1944                  * head
1945                  */
1946                 switch (ref->action) {
1947                 case BTRFS_ADD_DELAYED_REF:
1948                 case BTRFS_ADD_DELAYED_EXTENT:
1949                         locked_ref->ref_mod -= ref->ref_mod;
1950                         break;
1951                 case BTRFS_DROP_DELAYED_REF:
1952                         locked_ref->ref_mod += ref->ref_mod;
1953                         break;
1954                 default:
1955                         WARN_ON(1);
1956                 }
1957                 atomic_dec(&delayed_refs->num_entries);
1958
1959                 /*
1960                  * Record the must_insert_reserved flag before we drop the
1961                  * spin lock.
1962                  */
1963                 must_insert_reserved = locked_ref->must_insert_reserved;
1964                 locked_ref->must_insert_reserved = 0;
1965
1966                 extent_op = locked_ref->extent_op;
1967                 locked_ref->extent_op = NULL;
1968                 spin_unlock(&locked_ref->lock);
1969
1970                 ret = run_one_delayed_ref(trans, ref, extent_op,
1971                                           must_insert_reserved);
1972
1973                 btrfs_free_delayed_extent_op(extent_op);
1974                 if (ret) {
1975                         unselect_delayed_ref_head(delayed_refs, locked_ref);
1976                         btrfs_put_delayed_ref(ref);
1977                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d",
1978                                     ret);
1979                         return ret;
1980                 }
1981
1982                 btrfs_put_delayed_ref(ref);
1983                 cond_resched();
1984
1985                 spin_lock(&locked_ref->lock);
1986                 btrfs_merge_delayed_refs(trans, delayed_refs, locked_ref);
1987         }
1988
1989         return 0;
1990 }
1991
1992 /*
1993  * Returns 0 on success or if called with an already aborted transaction.
1994  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
1995  */
1996 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
1997                                              unsigned long nr)
1998 {
1999         struct btrfs_fs_info *fs_info = trans->fs_info;
2000         struct btrfs_delayed_ref_root *delayed_refs;
2001         struct btrfs_delayed_ref_head *locked_ref = NULL;
2002         ktime_t start = ktime_get();
2003         int ret;
2004         unsigned long count = 0;
2005         unsigned long actual_count = 0;
2006
2007         delayed_refs = &trans->transaction->delayed_refs;
2008         do {
2009                 if (!locked_ref) {
2010                         locked_ref = btrfs_obtain_ref_head(trans);
2011                         if (IS_ERR_OR_NULL(locked_ref)) {
2012                                 if (PTR_ERR(locked_ref) == -EAGAIN) {
2013                                         continue;
2014                                 } else {
2015                                         break;
2016                                 }
2017                         }
2018                         count++;
2019                 }
2020                 /*
2021                  * We need to try and merge add/drops of the same ref since we
2022                  * can run into issues with relocate dropping the implicit ref
2023                  * and then it being added back again before the drop can
2024                  * finish.  If we merged anything we need to re-loop so we can
2025                  * get a good ref.
2026                  * Or we can get node references of the same type that weren't
2027                  * merged when created due to bumps in the tree mod seq, and
2028                  * we need to merge them to prevent adding an inline extent
2029                  * backref before dropping it (triggering a BUG_ON at
2030                  * insert_inline_extent_backref()).
2031                  */
2032                 spin_lock(&locked_ref->lock);
2033                 btrfs_merge_delayed_refs(trans, delayed_refs, locked_ref);
2034
2035                 ret = btrfs_run_delayed_refs_for_head(trans, locked_ref,
2036                                                       &actual_count);
2037                 if (ret < 0 && ret != -EAGAIN) {
2038                         /*
2039                          * Error, btrfs_run_delayed_refs_for_head already
2040                          * unlocked everything so just bail out
2041                          */
2042                         return ret;
2043                 } else if (!ret) {
2044                         /*
2045                          * Success, perform the usual cleanup of a processed
2046                          * head
2047                          */
2048                         ret = cleanup_ref_head(trans, locked_ref);
2049                         if (ret > 0 ) {
2050                                 /* We dropped our lock, we need to loop. */
2051                                 ret = 0;
2052                                 continue;
2053                         } else if (ret) {
2054                                 return ret;
2055                         }
2056                 }
2057
2058                 /*
2059                  * Either success case or btrfs_run_delayed_refs_for_head
2060                  * returned -EAGAIN, meaning we need to select another head
2061                  */
2062
2063                 locked_ref = NULL;
2064                 cond_resched();
2065         } while ((nr != -1 && count < nr) || locked_ref);
2066
2067         /*
2068          * We don't want to include ref heads since we can have empty ref heads
2069          * and those will drastically skew our runtime down since we just do
2070          * accounting, no actual extent tree updates.
2071          */
2072         if (actual_count > 0) {
2073                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2074                 u64 avg;
2075
2076                 /*
2077                  * We weigh the current average higher than our current runtime
2078                  * to avoid large swings in the average.
2079                  */
2080                 spin_lock(&delayed_refs->lock);
2081                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2082                 fs_info->avg_delayed_ref_runtime = avg >> 2;    /* div by 4 */
2083                 spin_unlock(&delayed_refs->lock);
2084         }
2085         return 0;
2086 }
2087
2088 #ifdef SCRAMBLE_DELAYED_REFS
2089 /*
2090  * Normally delayed refs get processed in ascending bytenr order. This
2091  * correlates in most cases to the order added. To expose dependencies on this
2092  * order, we start to process the tree in the middle instead of the beginning
2093  */
2094 static u64 find_middle(struct rb_root *root)
2095 {
2096         struct rb_node *n = root->rb_node;
2097         struct btrfs_delayed_ref_node *entry;
2098         int alt = 1;
2099         u64 middle;
2100         u64 first = 0, last = 0;
2101
2102         n = rb_first(root);
2103         if (n) {
2104                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2105                 first = entry->bytenr;
2106         }
2107         n = rb_last(root);
2108         if (n) {
2109                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2110                 last = entry->bytenr;
2111         }
2112         n = root->rb_node;
2113
2114         while (n) {
2115                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2116                 WARN_ON(!entry->in_tree);
2117
2118                 middle = entry->bytenr;
2119
2120                 if (alt)
2121                         n = n->rb_left;
2122                 else
2123                         n = n->rb_right;
2124
2125                 alt = 1 - alt;
2126         }
2127         return middle;
2128 }
2129 #endif
2130
2131 static inline u64 heads_to_leaves(struct btrfs_fs_info *fs_info, u64 heads)
2132 {
2133         u64 num_bytes;
2134
2135         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2136                              sizeof(struct btrfs_extent_inline_ref));
2137         if (!btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2138                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2139
2140         /*
2141          * We don't ever fill up leaves all the way so multiply by 2 just to be
2142          * closer to what we're really going to want to use.
2143          */
2144         return div_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(fs_info));
2145 }
2146
2147 /*
2148  * Takes the number of bytes to be csumm'ed and figures out how many leaves it
2149  * would require to store the csums for that many bytes.
2150  */
2151 u64 btrfs_csum_bytes_to_leaves(struct btrfs_fs_info *fs_info, u64 csum_bytes)
2152 {
2153         u64 csum_size;
2154         u64 num_csums_per_leaf;
2155         u64 num_csums;
2156
2157         csum_size = BTRFS_MAX_ITEM_SIZE(fs_info);
2158         num_csums_per_leaf = div64_u64(csum_size,
2159                         (u64)btrfs_super_csum_size(fs_info->super_copy));
2160         num_csums = div64_u64(csum_bytes, fs_info->sectorsize);
2161         num_csums += num_csums_per_leaf - 1;
2162         num_csums = div64_u64(num_csums, num_csums_per_leaf);
2163         return num_csums;
2164 }
2165
2166 /*
2167  * this starts processing the delayed reference count updates and
2168  * extent insertions we have queued up so far.  count can be
2169  * 0, which means to process everything in the tree at the start
2170  * of the run (but not newly added entries), or it can be some target
2171  * number you'd like to process.
2172  *
2173  * Returns 0 on success or if called with an aborted transaction
2174  * Returns <0 on error and aborts the transaction
2175  */
2176 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2177                            unsigned long count)
2178 {
2179         struct btrfs_fs_info *fs_info = trans->fs_info;
2180         struct rb_node *node;
2181         struct btrfs_delayed_ref_root *delayed_refs;
2182         struct btrfs_delayed_ref_head *head;
2183         int ret;
2184         int run_all = count == (unsigned long)-1;
2185
2186         /* We'll clean this up in btrfs_cleanup_transaction */
2187         if (TRANS_ABORTED(trans))
2188                 return 0;
2189
2190         if (test_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags))
2191                 return 0;
2192
2193         delayed_refs = &trans->transaction->delayed_refs;
2194         if (count == 0)
2195                 count = atomic_read(&delayed_refs->num_entries) * 2;
2196
2197 again:
2198 #ifdef SCRAMBLE_DELAYED_REFS
2199         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2200 #endif
2201         ret = __btrfs_run_delayed_refs(trans, count);
2202         if (ret < 0) {
2203                 btrfs_abort_transaction(trans, ret);
2204                 return ret;
2205         }
2206
2207         if (run_all) {
2208                 btrfs_create_pending_block_groups(trans);
2209
2210                 spin_lock(&delayed_refs->lock);
2211                 node = rb_first_cached(&delayed_refs->href_root);
2212                 if (!node) {
2213                         spin_unlock(&delayed_refs->lock);
2214                         goto out;
2215                 }
2216                 head = rb_entry(node, struct btrfs_delayed_ref_head,
2217                                 href_node);
2218                 refcount_inc(&head->refs);
2219                 spin_unlock(&delayed_refs->lock);
2220
2221                 /* Mutex was contended, block until it's released and retry. */
2222                 mutex_lock(&head->mutex);
2223                 mutex_unlock(&head->mutex);
2224
2225                 btrfs_put_delayed_ref_head(head);
2226                 cond_resched();
2227                 goto again;
2228         }
2229 out:
2230         return 0;
2231 }
2232
2233 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2234                                 struct extent_buffer *eb, u64 flags,
2235                                 int level, int is_data)
2236 {
2237         struct btrfs_delayed_extent_op *extent_op;
2238         int ret;
2239
2240         extent_op = btrfs_alloc_delayed_extent_op();
2241         if (!extent_op)
2242                 return -ENOMEM;
2243
2244         extent_op->flags_to_set = flags;
2245         extent_op->update_flags = true;
2246         extent_op->update_key = false;
2247         extent_op->is_data = is_data ? true : false;
2248         extent_op->level = level;
2249
2250         ret = btrfs_add_delayed_extent_op(trans, eb->start, eb->len, extent_op);
2251         if (ret)
2252                 btrfs_free_delayed_extent_op(extent_op);
2253         return ret;
2254 }
2255
2256 static noinline int check_delayed_ref(struct btrfs_root *root,
2257                                       struct btrfs_path *path,
2258                                       u64 objectid, u64 offset, u64 bytenr)
2259 {
2260         struct btrfs_delayed_ref_head *head;
2261         struct btrfs_delayed_ref_node *ref;
2262         struct btrfs_delayed_data_ref *data_ref;
2263         struct btrfs_delayed_ref_root *delayed_refs;
2264         struct btrfs_transaction *cur_trans;
2265         struct rb_node *node;
2266         int ret = 0;
2267
2268         spin_lock(&root->fs_info->trans_lock);
2269         cur_trans = root->fs_info->running_transaction;
2270         if (cur_trans)
2271                 refcount_inc(&cur_trans->use_count);
2272         spin_unlock(&root->fs_info->trans_lock);
2273         if (!cur_trans)
2274                 return 0;
2275
2276         delayed_refs = &cur_trans->delayed_refs;
2277         spin_lock(&delayed_refs->lock);
2278         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
2279         if (!head) {
2280                 spin_unlock(&delayed_refs->lock);
2281                 btrfs_put_transaction(cur_trans);
2282                 return 0;
2283         }
2284
2285         if (!mutex_trylock(&head->mutex)) {
2286                 refcount_inc(&head->refs);
2287                 spin_unlock(&delayed_refs->lock);
2288
2289                 btrfs_release_path(path);
2290
2291                 /*
2292                  * Mutex was contended, block until it's released and let
2293                  * caller try again
2294                  */
2295                 mutex_lock(&head->mutex);
2296                 mutex_unlock(&head->mutex);
2297                 btrfs_put_delayed_ref_head(head);
2298                 btrfs_put_transaction(cur_trans);
2299                 return -EAGAIN;
2300         }
2301         spin_unlock(&delayed_refs->lock);
2302
2303         spin_lock(&head->lock);
2304         /*
2305          * XXX: We should replace this with a proper search function in the
2306          * future.
2307          */
2308         for (node = rb_first_cached(&head->ref_tree); node;
2309              node = rb_next(node)) {
2310                 ref = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
2311                 /* If it's a shared ref we know a cross reference exists */
2312                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
2313                         ret = 1;
2314                         break;
2315                 }
2316
2317                 data_ref = btrfs_delayed_node_to_data_ref(ref);
2318
2319                 /*
2320                  * If our ref doesn't match the one we're currently looking at
2321                  * then we have a cross reference.
2322                  */
2323                 if (data_ref->root != root->root_key.objectid ||
2324                     data_ref->objectid != objectid ||
2325                     data_ref->offset != offset) {
2326                         ret = 1;
2327                         break;
2328                 }
2329         }
2330         spin_unlock(&head->lock);
2331         mutex_unlock(&head->mutex);
2332         btrfs_put_transaction(cur_trans);
2333         return ret;
2334 }
2335
2336 static noinline int check_committed_ref(struct btrfs_root *root,
2337                                         struct btrfs_path *path,
2338                                         u64 objectid, u64 offset, u64 bytenr)
2339 {
2340         struct btrfs_fs_info *fs_info = root->fs_info;
2341         struct btrfs_root *extent_root = fs_info->extent_root;
2342         struct extent_buffer *leaf;
2343         struct btrfs_extent_data_ref *ref;
2344         struct btrfs_extent_inline_ref *iref;
2345         struct btrfs_extent_item *ei;
2346         struct btrfs_key key;
2347         u32 item_size;
2348         int type;
2349         int ret;
2350
2351         key.objectid = bytenr;
2352         key.offset = (u64)-1;
2353         key.type = BTRFS_EXTENT_ITEM_KEY;
2354
2355         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2356         if (ret < 0)
2357                 goto out;
2358         BUG_ON(ret == 0); /* Corruption */
2359
2360         ret = -ENOENT;
2361         if (path->slots[0] == 0)
2362                 goto out;
2363
2364         path->slots[0]--;
2365         leaf = path->nodes[0];
2366         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2367
2368         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2369                 goto out;
2370
2371         ret = 1;
2372         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2373         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2374
2375         /* If extent item has more than 1 inline ref then it's shared */
2376         if (item_size != sizeof(*ei) +
2377             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2378                 goto out;
2379
2380         /* If extent created before last snapshot => it's definitely shared */
2381         if (btrfs_extent_generation(leaf, ei) <=
2382             btrfs_root_last_snapshot(&root->root_item))
2383                 goto out;
2384
2385         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2386
2387         /* If this extent has SHARED_DATA_REF then it's shared */
2388         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
2389         if (type != BTRFS_EXTENT_DATA_REF_KEY)
2390                 goto out;
2391
2392         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2393         if (btrfs_extent_refs(leaf, ei) !=
2394             btrfs_extent_data_ref_count(leaf, ref) ||
2395             btrfs_extent_data_ref_root(leaf, ref) !=
2396             root->root_key.objectid ||
2397             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2398             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2399                 goto out;
2400
2401         ret = 0;
2402 out:
2403         return ret;
2404 }
2405
2406 int btrfs_cross_ref_exist(struct btrfs_root *root, u64 objectid, u64 offset,
2407                           u64 bytenr)
2408 {
2409         struct btrfs_path *path;
2410         int ret;
2411
2412         path = btrfs_alloc_path();
2413         if (!path)
2414                 return -ENOMEM;
2415
2416         do {
2417                 ret = check_committed_ref(root, path, objectid,
2418                                           offset, bytenr);
2419                 if (ret && ret != -ENOENT)
2420                         goto out;
2421
2422                 ret = check_delayed_ref(root, path, objectid, offset, bytenr);
2423         } while (ret == -EAGAIN);
2424
2425 out:
2426         btrfs_free_path(path);
2427         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2428                 WARN_ON(ret > 0);
2429         return ret;
2430 }
2431
2432 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2433                            struct btrfs_root *root,
2434                            struct extent_buffer *buf,
2435                            int full_backref, int inc)
2436 {
2437         struct btrfs_fs_info *fs_info = root->fs_info;
2438         u64 bytenr;
2439         u64 num_bytes;
2440         u64 parent;
2441         u64 ref_root;
2442         u32 nritems;
2443         struct btrfs_key key;
2444         struct btrfs_file_extent_item *fi;
2445         struct btrfs_ref generic_ref = { 0 };
2446         bool for_reloc = btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC);
2447         int i;
2448         int action;
2449         int level;
2450         int ret = 0;
2451
2452         if (btrfs_is_testing(fs_info))
2453                 return 0;
2454
2455         ref_root = btrfs_header_owner(buf);
2456         nritems = btrfs_header_nritems(buf);
2457         level = btrfs_header_level(buf);
2458
2459         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state) && level == 0)
2460                 return 0;
2461
2462         if (full_backref)
2463                 parent = buf->start;
2464         else
2465                 parent = 0;
2466         if (inc)
2467                 action = BTRFS_ADD_DELAYED_REF;
2468         else
2469                 action = BTRFS_DROP_DELAYED_REF;
2470
2471         for (i = 0; i < nritems; i++) {
2472                 if (level == 0) {
2473                         btrfs_item_key_to_cpu(buf, &key, i);
2474                         if (key.type != BTRFS_EXTENT_DATA_KEY)
2475                                 continue;
2476                         fi = btrfs_item_ptr(buf, i,
2477                                             struct btrfs_file_extent_item);
2478                         if (btrfs_file_extent_type(buf, fi) ==
2479                             BTRFS_FILE_EXTENT_INLINE)
2480                                 continue;
2481                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2482                         if (bytenr == 0)
2483                                 continue;
2484
2485                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
2486                         key.offset -= btrfs_file_extent_offset(buf, fi);
2487                         btrfs_init_generic_ref(&generic_ref, action, bytenr,
2488                                                num_bytes, parent);
2489                         generic_ref.real_root = root->root_key.objectid;
2490                         btrfs_init_data_ref(&generic_ref, ref_root, key.objectid,
2491                                             key.offset);
2492                         generic_ref.skip_qgroup = for_reloc;
2493                         if (inc)
2494                                 ret = btrfs_inc_extent_ref(trans, &generic_ref);
2495                         else
2496                                 ret = btrfs_free_extent(trans, &generic_ref);
2497                         if (ret)
2498                                 goto fail;
2499                 } else {
2500                         bytenr = btrfs_node_blockptr(buf, i);
2501                         num_bytes = fs_info->nodesize;
2502                         btrfs_init_generic_ref(&generic_ref, action, bytenr,
2503                                                num_bytes, parent);
2504                         generic_ref.real_root = root->root_key.objectid;
2505                         btrfs_init_tree_ref(&generic_ref, level - 1, ref_root);
2506                         generic_ref.skip_qgroup = for_reloc;
2507                         if (inc)
2508                                 ret = btrfs_inc_extent_ref(trans, &generic_ref);
2509                         else
2510                                 ret = btrfs_free_extent(trans, &generic_ref);
2511                         if (ret)
2512                                 goto fail;
2513                 }
2514         }
2515         return 0;
2516 fail:
2517         return ret;
2518 }
2519
2520 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2521                   struct extent_buffer *buf, int full_backref)
2522 {
2523         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
2524 }
2525
2526 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2527                   struct extent_buffer *buf, int full_backref)
2528 {
2529         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
2530 }
2531
2532 int btrfs_extent_readonly(struct btrfs_fs_info *fs_info, u64 bytenr)
2533 {
2534         struct btrfs_block_group *block_group;
2535         int readonly = 0;
2536
2537         block_group = btrfs_lookup_block_group(fs_info, bytenr);
2538         if (!block_group || block_group->ro)
2539                 readonly = 1;
2540         if (block_group)
2541                 btrfs_put_block_group(block_group);
2542         return readonly;
2543 }
2544
2545 static u64 get_alloc_profile_by_root(struct btrfs_root *root, int data)
2546 {
2547         struct btrfs_fs_info *fs_info = root->fs_info;
2548         u64 flags;
2549         u64 ret;
2550
2551         if (data)
2552                 flags = BTRFS_BLOCK_GROUP_DATA;
2553         else if (root == fs_info->chunk_root)
2554                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
2555         else
2556                 flags = BTRFS_BLOCK_GROUP_METADATA;
2557
2558         ret = btrfs_get_alloc_profile(fs_info, flags);
2559         return ret;
2560 }
2561
2562 static u64 first_logical_byte(struct btrfs_fs_info *fs_info, u64 search_start)
2563 {
2564         struct btrfs_block_group *cache;
2565         u64 bytenr;
2566
2567         spin_lock(&fs_info->block_group_cache_lock);
2568         bytenr = fs_info->first_logical_byte;
2569         spin_unlock(&fs_info->block_group_cache_lock);
2570
2571         if (bytenr < (u64)-1)
2572                 return bytenr;
2573
2574         cache = btrfs_lookup_first_block_group(fs_info, search_start);
2575         if (!cache)
2576                 return 0;
2577
2578         bytenr = cache->start;
2579         btrfs_put_block_group(cache);
2580
2581         return bytenr;
2582 }
2583
2584 static int pin_down_extent(struct btrfs_trans_handle *trans,
2585                            struct btrfs_block_group *cache,
2586                            u64 bytenr, u64 num_bytes, int reserved)
2587 {
2588         struct btrfs_fs_info *fs_info = cache->fs_info;
2589
2590         spin_lock(&cache->space_info->lock);
2591         spin_lock(&cache->lock);
2592         cache->pinned += num_bytes;
2593         btrfs_space_info_update_bytes_pinned(fs_info, cache->space_info,
2594                                              num_bytes);
2595         if (reserved) {
2596                 cache->reserved -= num_bytes;
2597                 cache->space_info->bytes_reserved -= num_bytes;
2598         }
2599         spin_unlock(&cache->lock);
2600         spin_unlock(&cache->space_info->lock);
2601
2602         percpu_counter_add_batch(&cache->space_info->total_bytes_pinned,
2603                     num_bytes, BTRFS_TOTAL_BYTES_PINNED_BATCH);
2604         set_extent_dirty(&trans->transaction->pinned_extents, bytenr,
2605                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
2606         return 0;
2607 }
2608
2609 int btrfs_pin_extent(struct btrfs_trans_handle *trans,
2610                      u64 bytenr, u64 num_bytes, int reserved)
2611 {
2612         struct btrfs_block_group *cache;
2613
2614         cache = btrfs_lookup_block_group(trans->fs_info, bytenr);
2615         BUG_ON(!cache); /* Logic error */
2616
2617         pin_down_extent(trans, cache, bytenr, num_bytes, reserved);
2618
2619         btrfs_put_block_group(cache);
2620         return 0;
2621 }
2622
2623 /*
2624  * this function must be called within transaction
2625  */
2626 int btrfs_pin_extent_for_log_replay(struct btrfs_trans_handle *trans,
2627                                     u64 bytenr, u64 num_bytes)
2628 {
2629         struct btrfs_block_group *cache;
2630         int ret;
2631
2632         btrfs_add_excluded_extent(trans->fs_info, bytenr, num_bytes);
2633
2634         cache = btrfs_lookup_block_group(trans->fs_info, bytenr);
2635         if (!cache)
2636                 return -EINVAL;
2637
2638         /*
2639          * pull in the free space cache (if any) so that our pin
2640          * removes the free space from the cache.  We have load_only set
2641          * to one because the slow code to read in the free extents does check
2642          * the pinned extents.
2643          */
2644         btrfs_cache_block_group(cache, 1);
2645
2646         pin_down_extent(trans, cache, bytenr, num_bytes, 0);
2647
2648         /* remove us from the free space cache (if we're there at all) */
2649         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
2650         btrfs_put_block_group(cache);
2651         return ret;
2652 }
2653
2654 static int __exclude_logged_extent(struct btrfs_fs_info *fs_info,
2655                                    u64 start, u64 num_bytes)
2656 {
2657         int ret;
2658         struct btrfs_block_group *block_group;
2659         struct btrfs_caching_control *caching_ctl;
2660
2661         block_group = btrfs_lookup_block_group(fs_info, start);
2662         if (!block_group)
2663                 return -EINVAL;
2664
2665         btrfs_cache_block_group(block_group, 0);
2666         caching_ctl = btrfs_get_caching_control(block_group);
2667
2668         if (!caching_ctl) {
2669                 /* Logic error */
2670                 BUG_ON(!btrfs_block_group_done(block_group));
2671                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
2672         } else {
2673                 mutex_lock(&caching_ctl->mutex);
2674
2675                 if (start >= caching_ctl->progress) {
2676                         ret = btrfs_add_excluded_extent(fs_info, start,
2677                                                         num_bytes);
2678                 } else if (start + num_bytes <= caching_ctl->progress) {
2679                         ret = btrfs_remove_free_space(block_group,
2680                                                       start, num_bytes);
2681                 } else {
2682                         num_bytes = caching_ctl->progress - start;
2683                         ret = btrfs_remove_free_space(block_group,
2684                                                       start, num_bytes);
2685                         if (ret)
2686                                 goto out_lock;
2687
2688                         num_bytes = (start + num_bytes) -
2689                                 caching_ctl->progress;
2690                         start = caching_ctl->progress;
2691                         ret = btrfs_add_excluded_extent(fs_info, start,
2692                                                         num_bytes);
2693                 }
2694 out_lock:
2695                 mutex_unlock(&caching_ctl->mutex);
2696                 btrfs_put_caching_control(caching_ctl);
2697         }
2698         btrfs_put_block_group(block_group);
2699         return ret;
2700 }
2701
2702 int btrfs_exclude_logged_extents(struct extent_buffer *eb)
2703 {
2704         struct btrfs_fs_info *fs_info = eb->fs_info;
2705         struct btrfs_file_extent_item *item;
2706         struct btrfs_key key;
2707         int found_type;
2708         int i;
2709         int ret = 0;
2710
2711         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS))
2712                 return 0;
2713
2714         for (i = 0; i < btrfs_header_nritems(eb); i++) {
2715                 btrfs_item_key_to_cpu(eb, &key, i);
2716                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2717                         continue;
2718                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
2719                 found_type = btrfs_file_extent_type(eb, item);
2720                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
2721                         continue;
2722                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
2723                         continue;
2724                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
2725                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
2726                 ret = __exclude_logged_extent(fs_info, key.objectid, key.offset);
2727                 if (ret)
2728                         break;
2729         }
2730
2731         return ret;
2732 }
2733
2734 static void
2735 btrfs_inc_block_group_reservations(struct btrfs_block_group *bg)
2736 {
2737         atomic_inc(&bg->reservations);
2738 }
2739
2740 void btrfs_prepare_extent_commit(struct btrfs_fs_info *fs_info)
2741 {
2742         struct btrfs_caching_control *next;
2743         struct btrfs_caching_control *caching_ctl;
2744         struct btrfs_block_group *cache;
2745
2746         down_write(&fs_info->commit_root_sem);
2747
2748         list_for_each_entry_safe(caching_ctl, next,
2749                                  &fs_info->caching_block_groups, list) {
2750                 cache = caching_ctl->block_group;
2751                 if (btrfs_block_group_done(cache)) {
2752                         cache->last_byte_to_unpin = (u64)-1;
2753                         list_del_init(&caching_ctl->list);
2754                         btrfs_put_caching_control(caching_ctl);
2755                 } else {
2756                         cache->last_byte_to_unpin = caching_ctl->progress;
2757                 }
2758         }
2759
2760         up_write(&fs_info->commit_root_sem);
2761
2762         btrfs_update_global_block_rsv(fs_info);
2763 }
2764
2765 /*
2766  * Returns the free cluster for the given space info and sets empty_cluster to
2767  * what it should be based on the mount options.
2768  */
2769 static struct btrfs_free_cluster *
2770 fetch_cluster_info(struct btrfs_fs_info *fs_info,
2771                    struct btrfs_space_info *space_info, u64 *empty_cluster)
2772 {
2773         struct btrfs_free_cluster *ret = NULL;
2774
2775         *empty_cluster = 0;
2776         if (btrfs_mixed_space_info(space_info))
2777                 return ret;
2778
2779         if (space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
2780                 ret = &fs_info->meta_alloc_cluster;
2781                 if (btrfs_test_opt(fs_info, SSD))
2782                         *empty_cluster = SZ_2M;
2783                 else
2784                         *empty_cluster = SZ_64K;
2785         } else if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) &&
2786                    btrfs_test_opt(fs_info, SSD_SPREAD)) {
2787                 *empty_cluster = SZ_2M;
2788                 ret = &fs_info->data_alloc_cluster;
2789         }
2790
2791         return ret;
2792 }
2793
2794 static int unpin_extent_range(struct btrfs_fs_info *fs_info,
2795                               u64 start, u64 end,
2796                               const bool return_free_space)
2797 {
2798         struct btrfs_block_group *cache = NULL;
2799         struct btrfs_space_info *space_info;
2800         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
2801         struct btrfs_free_cluster *cluster = NULL;
2802         u64 len;
2803         u64 total_unpinned = 0;
2804         u64 empty_cluster = 0;
2805         bool readonly;
2806
2807         while (start <= end) {
2808                 readonly = false;
2809                 if (!cache ||
2810                     start >= cache->start + cache->length) {
2811                         if (cache)
2812                                 btrfs_put_block_group(cache);
2813                         total_unpinned = 0;
2814                         cache = btrfs_lookup_block_group(fs_info, start);
2815                         BUG_ON(!cache); /* Logic error */
2816
2817                         cluster = fetch_cluster_info(fs_info,
2818                                                      cache->space_info,
2819                                                      &empty_cluster);
2820                         empty_cluster <<= 1;
2821                 }
2822
2823                 len = cache->start + cache->length - start;
2824                 len = min(len, end + 1 - start);
2825
2826                 if (start < cache->last_byte_to_unpin) {
2827                         len = min(len, cache->last_byte_to_unpin - start);
2828                         if (return_free_space)
2829                                 btrfs_add_free_space(cache, start, len);
2830                 }
2831
2832                 start += len;
2833                 total_unpinned += len;
2834                 space_info = cache->space_info;
2835
2836                 /*
2837                  * If this space cluster has been marked as fragmented and we've
2838                  * unpinned enough in this block group to potentially allow a
2839                  * cluster to be created inside of it go ahead and clear the
2840                  * fragmented check.
2841                  */
2842                 if (cluster && cluster->fragmented &&
2843                     total_unpinned > empty_cluster) {
2844                         spin_lock(&cluster->lock);
2845                         cluster->fragmented = 0;
2846                         spin_unlock(&cluster->lock);
2847                 }
2848
2849                 spin_lock(&space_info->lock);
2850                 spin_lock(&cache->lock);
2851                 cache->pinned -= len;
2852                 btrfs_space_info_update_bytes_pinned(fs_info, space_info, -len);
2853                 space_info->max_extent_size = 0;
2854                 percpu_counter_add_batch(&space_info->total_bytes_pinned,
2855                             -len, BTRFS_TOTAL_BYTES_PINNED_BATCH);
2856                 if (cache->ro) {
2857                         space_info->bytes_readonly += len;
2858                         readonly = true;
2859                 }
2860                 spin_unlock(&cache->lock);
2861                 if (!readonly && return_free_space &&
2862                     global_rsv->space_info == space_info) {
2863                         u64 to_add = len;
2864
2865                         spin_lock(&global_rsv->lock);
2866                         if (!global_rsv->full) {
2867                                 to_add = min(len, global_rsv->size -
2868                                              global_rsv->reserved);
2869                                 global_rsv->reserved += to_add;
2870                                 btrfs_space_info_update_bytes_may_use(fs_info,
2871                                                 space_info, to_add);
2872                                 if (global_rsv->reserved >= global_rsv->size)
2873                                         global_rsv->full = 1;
2874                                 len -= to_add;
2875                         }
2876                         spin_unlock(&global_rsv->lock);
2877                         /* Add to any tickets we may have */
2878                         if (len)
2879                                 btrfs_try_granting_tickets(fs_info,
2880                                                            space_info);
2881                 }
2882                 spin_unlock(&space_info->lock);
2883         }
2884
2885         if (cache)
2886                 btrfs_put_block_group(cache);
2887         return 0;
2888 }
2889
2890 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans)
2891 {
2892         struct btrfs_fs_info *fs_info = trans->fs_info;
2893         struct btrfs_block_group *block_group, *tmp;
2894         struct list_head *deleted_bgs;
2895         struct extent_io_tree *unpin;
2896         u64 start;
2897         u64 end;
2898         int ret;
2899
2900         unpin = &trans->transaction->pinned_extents;
2901
2902         while (!TRANS_ABORTED(trans)) {
2903                 struct extent_state *cached_state = NULL;
2904
2905                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
2906                 ret = find_first_extent_bit(unpin, 0, &start, &end,
2907                                             EXTENT_DIRTY, &cached_state);
2908                 if (ret) {
2909                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
2910                         break;
2911                 }
2912                 if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
2913                         clear_extent_bits(&fs_info->excluded_extents, start,
2914                                           end, EXTENT_UPTODATE);
2915
2916                 if (btrfs_test_opt(fs_info, DISCARD_SYNC))
2917                         ret = btrfs_discard_extent(fs_info, start,
2918                                                    end + 1 - start, NULL);
2919
2920                 clear_extent_dirty(unpin, start, end, &cached_state);
2921                 unpin_extent_range(fs_info, start, end, true);
2922                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
2923                 free_extent_state(cached_state);
2924                 cond_resched();
2925         }
2926
2927         if (btrfs_test_opt(fs_info, DISCARD_ASYNC)) {
2928                 btrfs_discard_calc_delay(&fs_info->discard_ctl);
2929                 btrfs_discard_schedule_work(&fs_info->discard_ctl, true);
2930         }
2931
2932         /*
2933          * Transaction is finished.  We don't need the lock anymore.  We
2934          * do need to clean up the block groups in case of a transaction
2935          * abort.
2936          */
2937         deleted_bgs = &trans->transaction->deleted_bgs;
2938         list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
2939                 u64 trimmed = 0;
2940
2941                 ret = -EROFS;
2942                 if (!TRANS_ABORTED(trans))
2943                         ret = btrfs_discard_extent(fs_info,
2944                                                    block_group->start,
2945                                                    block_group->length,
2946                                                    &trimmed);
2947
2948                 list_del_init(&block_group->bg_list);
2949                 btrfs_put_block_group_trimming(block_group);
2950                 btrfs_put_block_group(block_group);
2951
2952                 if (ret) {
2953                         const char *errstr = btrfs_decode_error(ret);
2954                         btrfs_warn(fs_info,
2955                            "discard failed while removing blockgroup: errno=%d %s",
2956                                    ret, errstr);
2957                 }
2958         }
2959
2960         return 0;
2961 }
2962
2963 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
2964                                struct btrfs_delayed_ref_node *node, u64 parent,
2965                                u64 root_objectid, u64 owner_objectid,
2966                                u64 owner_offset, int refs_to_drop,
2967                                struct btrfs_delayed_extent_op *extent_op)
2968 {
2969         struct btrfs_fs_info *info = trans->fs_info;
2970         struct btrfs_key key;
2971         struct btrfs_path *path;
2972         struct btrfs_root *extent_root = info->extent_root;
2973         struct extent_buffer *leaf;
2974         struct btrfs_extent_item *ei;
2975         struct btrfs_extent_inline_ref *iref;
2976         int ret;
2977         int is_data;
2978         int extent_slot = 0;
2979         int found_extent = 0;
2980         int num_to_del = 1;
2981         u32 item_size;
2982         u64 refs;
2983         u64 bytenr = node->bytenr;
2984         u64 num_bytes = node->num_bytes;
2985         int last_ref = 0;
2986         bool skinny_metadata = btrfs_fs_incompat(info, SKINNY_METADATA);
2987
2988         path = btrfs_alloc_path();
2989         if (!path)
2990                 return -ENOMEM;
2991
2992         path->reada = READA_FORWARD;
2993         path->leave_spinning = 1;
2994
2995         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
2996         BUG_ON(!is_data && refs_to_drop != 1);
2997
2998         if (is_data)
2999                 skinny_metadata = false;
3000
3001         ret = lookup_extent_backref(trans, path, &iref, bytenr, num_bytes,
3002                                     parent, root_objectid, owner_objectid,
3003                                     owner_offset);
3004         if (ret == 0) {
3005                 extent_slot = path->slots[0];
3006                 while (extent_slot >= 0) {
3007                         btrfs_item_key_to_cpu(path->nodes[0], &key,
3008                                               extent_slot);
3009                         if (key.objectid != bytenr)
3010                                 break;
3011                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
3012                             key.offset == num_bytes) {
3013                                 found_extent = 1;
3014                                 break;
3015                         }
3016                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
3017                             key.offset == owner_objectid) {
3018                                 found_extent = 1;
3019                                 break;
3020                         }
3021                         if (path->slots[0] - extent_slot > 5)
3022                                 break;
3023                         extent_slot--;
3024                 }
3025
3026                 if (!found_extent) {
3027                         BUG_ON(iref);
3028                         ret = remove_extent_backref(trans, path, NULL,
3029                                                     refs_to_drop,
3030                                                     is_data, &last_ref);
3031                         if (ret) {
3032                                 btrfs_abort_transaction(trans, ret);
3033                                 goto out;
3034                         }
3035                         btrfs_release_path(path);
3036                         path->leave_spinning = 1;
3037
3038                         key.objectid = bytenr;
3039                         key.type = BTRFS_EXTENT_ITEM_KEY;
3040                         key.offset = num_bytes;
3041
3042                         if (!is_data && skinny_metadata) {
3043                                 key.type = BTRFS_METADATA_ITEM_KEY;
3044                                 key.offset = owner_objectid;
3045                         }
3046
3047                         ret = btrfs_search_slot(trans, extent_root,
3048                                                 &key, path, -1, 1);
3049                         if (ret > 0 && skinny_metadata && path->slots[0]) {
3050                                 /*
3051                                  * Couldn't find our skinny metadata item,
3052                                  * see if we have ye olde extent item.
3053                                  */
3054                                 path->slots[0]--;
3055                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
3056                                                       path->slots[0]);
3057                                 if (key.objectid == bytenr &&
3058                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
3059                                     key.offset == num_bytes)
3060                                         ret = 0;
3061                         }
3062
3063                         if (ret > 0 && skinny_metadata) {
3064                                 skinny_metadata = false;
3065                                 key.objectid = bytenr;
3066                                 key.type = BTRFS_EXTENT_ITEM_KEY;
3067                                 key.offset = num_bytes;
3068                                 btrfs_release_path(path);
3069                                 ret = btrfs_search_slot(trans, extent_root,
3070                                                         &key, path, -1, 1);
3071                         }
3072
3073                         if (ret) {
3074                                 btrfs_err(info,
3075                                           "umm, got %d back from search, was looking for %llu",
3076                                           ret, bytenr);
3077                                 if (ret > 0)
3078                                         btrfs_print_leaf(path->nodes[0]);
3079                         }
3080                         if (ret < 0) {
3081                                 btrfs_abort_transaction(trans, ret);
3082                                 goto out;
3083                         }
3084                         extent_slot = path->slots[0];
3085                 }
3086         } else if (WARN_ON(ret == -ENOENT)) {
3087                 btrfs_print_leaf(path->nodes[0]);
3088                 btrfs_err(info,
3089                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
3090                         bytenr, parent, root_objectid, owner_objectid,
3091                         owner_offset);
3092                 btrfs_abort_transaction(trans, ret);
3093                 goto out;
3094         } else {
3095                 btrfs_abort_transaction(trans, ret);
3096                 goto out;
3097         }
3098
3099         leaf = path->nodes[0];
3100         item_size = btrfs_item_size_nr(leaf, extent_slot);
3101         if (unlikely(item_size < sizeof(*ei))) {
3102                 ret = -EINVAL;
3103                 btrfs_print_v0_err(info);
3104                 btrfs_abort_transaction(trans, ret);
3105                 goto out;
3106         }
3107         ei = btrfs_item_ptr(leaf, extent_slot,
3108                             struct btrfs_extent_item);
3109         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
3110             key.type == BTRFS_EXTENT_ITEM_KEY) {
3111                 struct btrfs_tree_block_info *bi;
3112                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
3113                 bi = (struct btrfs_tree_block_info *)(ei + 1);
3114                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
3115         }
3116
3117         refs = btrfs_extent_refs(leaf, ei);
3118         if (refs < refs_to_drop) {
3119                 btrfs_err(info,
3120                           "trying to drop %d refs but we only have %Lu for bytenr %Lu",
3121                           refs_to_drop, refs, bytenr);
3122                 ret = -EINVAL;
3123                 btrfs_abort_transaction(trans, ret);
3124                 goto out;
3125         }
3126         refs -= refs_to_drop;
3127
3128         if (refs > 0) {
3129                 if (extent_op)
3130                         __run_delayed_extent_op(extent_op, leaf, ei);
3131                 /*
3132                  * In the case of inline back ref, reference count will
3133                  * be updated by remove_extent_backref
3134                  */
3135                 if (iref) {
3136                         BUG_ON(!found_extent);
3137                 } else {
3138                         btrfs_set_extent_refs(leaf, ei, refs);
3139                         btrfs_mark_buffer_dirty(leaf);
3140                 }
3141                 if (found_extent) {
3142                         ret = remove_extent_backref(trans, path, iref,
3143                                                     refs_to_drop, is_data,
3144                                                     &last_ref);
3145                         if (ret) {
3146                                 btrfs_abort_transaction(trans, ret);
3147                                 goto out;
3148                         }
3149                 }
3150         } else {
3151                 if (found_extent) {
3152                         BUG_ON(is_data && refs_to_drop !=
3153                                extent_data_ref_count(path, iref));
3154                         if (iref) {
3155                                 BUG_ON(path->slots[0] != extent_slot);
3156                         } else {
3157                                 BUG_ON(path->slots[0] != extent_slot + 1);
3158                                 path->slots[0] = extent_slot;
3159                                 num_to_del = 2;
3160                         }
3161                 }
3162
3163                 last_ref = 1;
3164                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
3165                                       num_to_del);
3166                 if (ret) {
3167                         btrfs_abort_transaction(trans, ret);
3168                         goto out;
3169                 }
3170                 btrfs_release_path(path);
3171
3172                 if (is_data) {
3173                         ret = btrfs_del_csums(trans, info->csum_root, bytenr,
3174                                               num_bytes);
3175                         if (ret) {
3176                                 btrfs_abort_transaction(trans, ret);
3177                                 goto out;
3178                         }
3179                 }
3180
3181                 ret = add_to_free_space_tree(trans, bytenr, num_bytes);
3182                 if (ret) {
3183                         btrfs_abort_transaction(trans, ret);
3184                         goto out;
3185                 }
3186
3187                 ret = btrfs_update_block_group(trans, bytenr, num_bytes, 0);
3188                 if (ret) {
3189                         btrfs_abort_transaction(trans, ret);
3190                         goto out;
3191                 }
3192         }
3193         btrfs_release_path(path);
3194
3195 out:
3196         btrfs_free_path(path);
3197         return ret;
3198 }
3199
3200 /*
3201  * when we free an block, it is possible (and likely) that we free the last
3202  * delayed ref for that extent as well.  This searches the delayed ref tree for
3203  * a given extent, and if there are no other delayed refs to be processed, it
3204  * removes it from the tree.
3205  */
3206 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
3207                                       u64 bytenr)
3208 {
3209         struct btrfs_delayed_ref_head *head;
3210         struct btrfs_delayed_ref_root *delayed_refs;
3211         int ret = 0;
3212
3213         delayed_refs = &trans->transaction->delayed_refs;
3214         spin_lock(&delayed_refs->lock);
3215         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
3216         if (!head)
3217                 goto out_delayed_unlock;
3218
3219         spin_lock(&head->lock);
3220         if (!RB_EMPTY_ROOT(&head->ref_tree.rb_root))
3221                 goto out;
3222
3223         if (cleanup_extent_op(head) != NULL)
3224                 goto out;
3225
3226         /*
3227          * waiting for the lock here would deadlock.  If someone else has it
3228          * locked they are already in the process of dropping it anyway
3229          */
3230         if (!mutex_trylock(&head->mutex))
3231                 goto out;
3232
3233         btrfs_delete_ref_head(delayed_refs, head);
3234         head->processing = 0;
3235
3236         spin_unlock(&head->lock);
3237         spin_unlock(&delayed_refs->lock);
3238
3239         BUG_ON(head->extent_op);
3240         if (head->must_insert_reserved)
3241                 ret = 1;
3242
3243         btrfs_cleanup_ref_head_accounting(trans->fs_info, delayed_refs, head);
3244         mutex_unlock(&head->mutex);
3245         btrfs_put_delayed_ref_head(head);
3246         return ret;
3247 out:
3248         spin_unlock(&head->lock);
3249
3250 out_delayed_unlock:
3251         spin_unlock(&delayed_refs->lock);
3252         return 0;
3253 }
3254
3255 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
3256                            struct btrfs_root *root,
3257                            struct extent_buffer *buf,
3258                            u64 parent, int last_ref)
3259 {
3260         struct btrfs_fs_info *fs_info = root->fs_info;
3261         struct btrfs_ref generic_ref = { 0 };
3262         int pin = 1;
3263         int ret;
3264
3265         btrfs_init_generic_ref(&generic_ref, BTRFS_DROP_DELAYED_REF,
3266                                buf->start, buf->len, parent);
3267         btrfs_init_tree_ref(&generic_ref, btrfs_header_level(buf),
3268                             root->root_key.objectid);
3269
3270         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
3271                 int old_ref_mod, new_ref_mod;
3272
3273                 btrfs_ref_tree_mod(fs_info, &generic_ref);
3274                 ret = btrfs_add_delayed_tree_ref(trans, &generic_ref, NULL,
3275                                                  &old_ref_mod, &new_ref_mod);
3276                 BUG_ON(ret); /* -ENOMEM */
3277                 pin = old_ref_mod >= 0 && new_ref_mod < 0;
3278         }
3279
3280         if (last_ref && btrfs_header_generation(buf) == trans->transid) {
3281                 struct btrfs_block_group *cache;
3282
3283                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
3284                         ret = check_ref_cleanup(trans, buf->start);
3285                         if (!ret)
3286                                 goto out;
3287                 }
3288
3289                 pin = 0;
3290                 cache = btrfs_lookup_block_group(fs_info, buf->start);
3291
3292                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
3293                         pin_down_extent(trans, cache, buf->start, buf->len, 1);
3294                         btrfs_put_block_group(cache);
3295                         goto out;
3296                 }
3297
3298                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
3299
3300                 btrfs_add_free_space(cache, buf->start, buf->len);
3301                 btrfs_free_reserved_bytes(cache, buf->len, 0);
3302                 btrfs_put_block_group(cache);
3303                 trace_btrfs_reserved_extent_free(fs_info, buf->start, buf->len);
3304         }
3305 out:
3306         if (pin)
3307                 add_pinned_bytes(fs_info, &generic_ref);
3308
3309         if (last_ref) {
3310                 /*
3311                  * Deleting the buffer, clear the corrupt flag since it doesn't
3312                  * matter anymore.
3313                  */
3314                 clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
3315         }
3316 }
3317
3318 /* Can return -ENOMEM */
3319 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_ref *ref)
3320 {
3321         struct btrfs_fs_info *fs_info = trans->fs_info;
3322         int old_ref_mod, new_ref_mod;
3323         int ret;
3324
3325         if (btrfs_is_testing(fs_info))
3326                 return 0;
3327
3328         /*
3329          * tree log blocks never actually go into the extent allocation
3330          * tree, just update pinning info and exit early.
3331          */
3332         if ((ref->type == BTRFS_REF_METADATA &&
3333              ref->tree_ref.root == BTRFS_TREE_LOG_OBJECTID) ||
3334             (ref->type == BTRFS_REF_DATA &&
3335              ref->data_ref.ref_root == BTRFS_TREE_LOG_OBJECTID)) {
3336                 /* unlocks the pinned mutex */
3337                 btrfs_pin_extent(trans, ref->bytenr, ref->len, 1);
3338                 old_ref_mod = new_ref_mod = 0;
3339                 ret = 0;
3340         } else if (ref->type == BTRFS_REF_METADATA) {
3341                 ret = btrfs_add_delayed_tree_ref(trans, ref, NULL,
3342                                                  &old_ref_mod, &new_ref_mod);
3343         } else {
3344                 ret = btrfs_add_delayed_data_ref(trans, ref, 0,
3345                                                  &old_ref_mod, &new_ref_mod);
3346         }
3347
3348         if (!((ref->type == BTRFS_REF_METADATA &&
3349                ref->tree_ref.root == BTRFS_TREE_LOG_OBJECTID) ||
3350               (ref->type == BTRFS_REF_DATA &&
3351                ref->data_ref.ref_root == BTRFS_TREE_LOG_OBJECTID)))
3352                 btrfs_ref_tree_mod(fs_info, ref);
3353
3354         if (ret == 0 && old_ref_mod >= 0 && new_ref_mod < 0)
3355                 add_pinned_bytes(fs_info, ref);
3356
3357         return ret;
3358 }
3359
3360 enum btrfs_loop_type {
3361         LOOP_CACHING_NOWAIT,
3362         LOOP_CACHING_WAIT,
3363         LOOP_ALLOC_CHUNK,
3364         LOOP_NO_EMPTY_SIZE,
3365 };
3366
3367 static inline void
3368 btrfs_lock_block_group(struct btrfs_block_group *cache,
3369                        int delalloc)
3370 {
3371         if (delalloc)
3372                 down_read(&cache->data_rwsem);
3373 }
3374
3375 static inline void btrfs_grab_block_group(struct btrfs_block_group *cache,
3376                        int delalloc)
3377 {
3378         btrfs_get_block_group(cache);
3379         if (delalloc)
3380                 down_read(&cache->data_rwsem);
3381 }
3382
3383 static struct btrfs_block_group *btrfs_lock_cluster(
3384                    struct btrfs_block_group *block_group,
3385                    struct btrfs_free_cluster *cluster,
3386                    int delalloc)
3387 {
3388         struct btrfs_block_group *used_bg = NULL;
3389
3390         spin_lock(&cluster->refill_lock);
3391         while (1) {
3392                 used_bg = cluster->block_group;
3393                 if (!used_bg)
3394                         return NULL;
3395
3396                 if (used_bg == block_group)
3397                         return used_bg;
3398
3399                 btrfs_get_block_group(used_bg);
3400
3401                 if (!delalloc)
3402                         return used_bg;
3403
3404                 if (down_read_trylock(&used_bg->data_rwsem))
3405                         return used_bg;
3406
3407                 spin_unlock(&cluster->refill_lock);
3408
3409                 /* We should only have one-level nested. */
3410                 down_read_nested(&used_bg->data_rwsem, SINGLE_DEPTH_NESTING);
3411
3412                 spin_lock(&cluster->refill_lock);
3413                 if (used_bg == cluster->block_group)
3414                         return used_bg;
3415
3416                 up_read(&used_bg->data_rwsem);
3417                 btrfs_put_block_group(used_bg);
3418         }
3419 }
3420
3421 static inline void
3422 btrfs_release_block_group(struct btrfs_block_group *cache,
3423                          int delalloc)
3424 {
3425         if (delalloc)
3426                 up_read(&cache->data_rwsem);
3427         btrfs_put_block_group(cache);
3428 }
3429
3430 enum btrfs_extent_allocation_policy {
3431         BTRFS_EXTENT_ALLOC_CLUSTERED,
3432 };
3433
3434 /*
3435  * Structure used internally for find_free_extent() function.  Wraps needed
3436  * parameters.
3437  */
3438 struct find_free_extent_ctl {
3439         /* Basic allocation info */
3440         u64 num_bytes;
3441         u64 empty_size;
3442         u64 flags;
3443         int delalloc;
3444
3445         /* Where to start the search inside the bg */
3446         u64 search_start;
3447
3448         /* For clustered allocation */
3449         u64 empty_cluster;
3450         struct btrfs_free_cluster *last_ptr;
3451         bool use_cluster;
3452
3453         bool have_caching_bg;
3454         bool orig_have_caching_bg;
3455
3456         /* RAID index, converted from flags */
3457         int index;
3458
3459         /*
3460          * Current loop number, check find_free_extent_update_loop() for details
3461          */
3462         int loop;
3463
3464         /*
3465          * Whether we're refilling a cluster, if true we need to re-search
3466          * current block group but don't try to refill the cluster again.
3467          */
3468         bool retry_clustered;
3469
3470         /*
3471          * Whether we're updating free space cache, if true we need to re-search
3472          * current block group but don't try updating free space cache again.
3473          */
3474         bool retry_unclustered;
3475
3476         /* If current block group is cached */
3477         int cached;
3478
3479         /* Max contiguous hole found */
3480         u64 max_extent_size;
3481
3482         /* Total free space from free space cache, not always contiguous */
3483         u64 total_free_space;
3484
3485         /* Found result */
3486         u64 found_offset;
3487
3488         /* Hint where to start looking for an empty space */
3489         u64 hint_byte;
3490
3491         /* Allocation policy */
3492         enum btrfs_extent_allocation_policy policy;
3493 };
3494
3495
3496 /*
3497  * Helper function for find_free_extent().
3498  *
3499  * Return -ENOENT to inform caller that we need fallback to unclustered mode.
3500  * Return -EAGAIN to inform caller that we need to re-search this block group
3501  * Return >0 to inform caller that we find nothing
3502  * Return 0 means we have found a location and set ffe_ctl->found_offset.
3503  */
3504 static int find_free_extent_clustered(struct btrfs_block_group *bg,
3505                 struct btrfs_free_cluster *last_ptr,
3506                 struct find_free_extent_ctl *ffe_ctl,
3507                 struct btrfs_block_group **cluster_bg_ret)
3508 {
3509         struct btrfs_block_group *cluster_bg;
3510         u64 aligned_cluster;
3511         u64 offset;
3512         int ret;
3513
3514         cluster_bg = btrfs_lock_cluster(bg, last_ptr, ffe_ctl->delalloc);
3515         if (!cluster_bg)
3516                 goto refill_cluster;
3517         if (cluster_bg != bg && (cluster_bg->ro ||
3518             !block_group_bits(cluster_bg, ffe_ctl->flags)))
3519                 goto release_cluster;
3520
3521         offset = btrfs_alloc_from_cluster(cluster_bg, last_ptr,
3522                         ffe_ctl->num_bytes, cluster_bg->start,
3523                         &ffe_ctl->max_extent_size);
3524         if (offset) {
3525                 /* We have a block, we're done */
3526                 spin_unlock(&last_ptr->refill_lock);
3527                 trace_btrfs_reserve_extent_cluster(cluster_bg,
3528                                 ffe_ctl->search_start, ffe_ctl->num_bytes);
3529                 *cluster_bg_ret = cluster_bg;
3530                 ffe_ctl->found_offset = offset;
3531                 return 0;
3532         }
3533         WARN_ON(last_ptr->block_group != cluster_bg);
3534
3535 release_cluster:
3536         /*
3537          * If we are on LOOP_NO_EMPTY_SIZE, we can't set up a new clusters, so
3538          * lets just skip it and let the allocator find whatever block it can
3539          * find. If we reach this point, we will have tried the cluster
3540          * allocator plenty of times and not have found anything, so we are
3541          * likely way too fragmented for the clustering stuff to find anything.
3542          *
3543          * However, if the cluster is taken from the current block group,
3544          * release the cluster first, so that we stand a better chance of
3545          * succeeding in the unclustered allocation.
3546          */
3547         if (ffe_ctl->loop >= LOOP_NO_EMPTY_SIZE && cluster_bg != bg) {
3548                 spin_unlock(&last_ptr->refill_lock);
3549                 btrfs_release_block_group(cluster_bg, ffe_ctl->delalloc);
3550                 return -ENOENT;
3551         }
3552
3553         /* This cluster didn't work out, free it and start over */
3554         btrfs_return_cluster_to_free_space(NULL, last_ptr);
3555
3556         if (cluster_bg != bg)
3557                 btrfs_release_block_group(cluster_bg, ffe_ctl->delalloc);
3558
3559 refill_cluster:
3560         if (ffe_ctl->loop >= LOOP_NO_EMPTY_SIZE) {
3561                 spin_unlock(&last_ptr->refill_lock);
3562                 return -ENOENT;
3563         }
3564
3565         aligned_cluster = max_t(u64,
3566                         ffe_ctl->empty_cluster + ffe_ctl->empty_size,
3567                         bg->full_stripe_len);
3568         ret = btrfs_find_space_cluster(bg, last_ptr, ffe_ctl->search_start,
3569                         ffe_ctl->num_bytes, aligned_cluster);
3570         if (ret == 0) {
3571                 /* Now pull our allocation out of this cluster */
3572                 offset = btrfs_alloc_from_cluster(bg, last_ptr,
3573                                 ffe_ctl->num_bytes, ffe_ctl->search_start,
3574                                 &ffe_ctl->max_extent_size);
3575                 if (offset) {
3576                         /* We found one, proceed */
3577                         spin_unlock(&last_ptr->refill_lock);
3578                         trace_btrfs_reserve_extent_cluster(bg,
3579                                         ffe_ctl->search_start,
3580                                         ffe_ctl->num_bytes);
3581                         ffe_ctl->found_offset = offset;
3582                         return 0;
3583                 }
3584         } else if (!ffe_ctl->cached && ffe_ctl->loop > LOOP_CACHING_NOWAIT &&
3585                    !ffe_ctl->retry_clustered) {
3586                 spin_unlock(&last_ptr->refill_lock);
3587
3588                 ffe_ctl->retry_clustered = true;
3589                 btrfs_wait_block_group_cache_progress(bg, ffe_ctl->num_bytes +
3590                                 ffe_ctl->empty_cluster + ffe_ctl->empty_size);
3591                 return -EAGAIN;
3592         }
3593         /*
3594          * At this point we either didn't find a cluster or we weren't able to
3595          * allocate a block from our cluster.  Free the cluster we've been
3596          * trying to use, and go to the next block group.
3597          */
3598         btrfs_return_cluster_to_free_space(NULL, last_ptr);
3599         spin_unlock(&last_ptr->refill_lock);
3600         return 1;
3601 }
3602
3603 /*
3604  * Return >0 to inform caller that we find nothing
3605  * Return 0 when we found an free extent and set ffe_ctrl->found_offset
3606  * Return -EAGAIN to inform caller that we need to re-search this block group
3607  */
3608 static int find_free_extent_unclustered(struct btrfs_block_group *bg,
3609                 struct btrfs_free_cluster *last_ptr,
3610                 struct find_free_extent_ctl *ffe_ctl)
3611 {
3612         u64 offset;
3613
3614         /*
3615          * We are doing an unclustered allocation, set the fragmented flag so
3616          * we don't bother trying to setup a cluster again until we get more
3617          * space.
3618          */
3619         if (unlikely(last_ptr)) {
3620                 spin_lock(&last_ptr->lock);
3621                 last_ptr->fragmented = 1;
3622                 spin_unlock(&last_ptr->lock);
3623         }
3624         if (ffe_ctl->cached) {
3625                 struct btrfs_free_space_ctl *free_space_ctl;
3626
3627                 free_space_ctl = bg->free_space_ctl;
3628                 spin_lock(&free_space_ctl->tree_lock);
3629                 if (free_space_ctl->free_space <
3630                     ffe_ctl->num_bytes + ffe_ctl->empty_cluster +
3631                     ffe_ctl->empty_size) {
3632                         ffe_ctl->total_free_space = max_t(u64,
3633                                         ffe_ctl->total_free_space,
3634                                         free_space_ctl->free_space);
3635                         spin_unlock(&free_space_ctl->tree_lock);
3636                         return 1;
3637                 }
3638                 spin_unlock(&free_space_ctl->tree_lock);
3639         }
3640
3641         offset = btrfs_find_space_for_alloc(bg, ffe_ctl->search_start,
3642                         ffe_ctl->num_bytes, ffe_ctl->empty_size,
3643                         &ffe_ctl->max_extent_size);
3644
3645         /*
3646          * If we didn't find a chunk, and we haven't failed on this block group
3647          * before, and this block group is in the middle of caching and we are
3648          * ok with waiting, then go ahead and wait for progress to be made, and
3649          * set @retry_unclustered to true.
3650          *
3651          * If @retry_unclustered is true then we've already waited on this
3652          * block group once and should move on to the next block group.
3653          */
3654         if (!offset && !ffe_ctl->retry_unclustered && !ffe_ctl->cached &&
3655             ffe_ctl->loop > LOOP_CACHING_NOWAIT) {
3656                 btrfs_wait_block_group_cache_progress(bg, ffe_ctl->num_bytes +
3657                                                       ffe_ctl->empty_size);
3658                 ffe_ctl->retry_unclustered = true;
3659                 return -EAGAIN;
3660         } else if (!offset) {
3661                 return 1;
3662         }
3663         ffe_ctl->found_offset = offset;
3664         return 0;
3665 }
3666
3667 /*
3668  * Return >0 means caller needs to re-search for free extent
3669  * Return 0 means we have the needed free extent.
3670  * Return <0 means we failed to locate any free extent.
3671  */
3672 static int find_free_extent_update_loop(struct btrfs_fs_info *fs_info,
3673                                         struct btrfs_free_cluster *last_ptr,
3674                                         struct btrfs_key *ins,
3675                                         struct find_free_extent_ctl *ffe_ctl,
3676                                         bool full_search, bool use_cluster)
3677 {
3678         struct btrfs_root *root = fs_info->extent_root;
3679         int ret;
3680
3681         if ((ffe_ctl->loop == LOOP_CACHING_NOWAIT) &&
3682             ffe_ctl->have_caching_bg && !ffe_ctl->orig_have_caching_bg)
3683                 ffe_ctl->orig_have_caching_bg = true;
3684
3685         if (!ins->objectid && ffe_ctl->loop >= LOOP_CACHING_WAIT &&
3686             ffe_ctl->have_caching_bg)
3687                 return 1;
3688
3689         if (!ins->objectid && ++(ffe_ctl->index) < BTRFS_NR_RAID_TYPES)
3690                 return 1;
3691
3692         if (ins->objectid) {
3693                 if (!use_cluster && last_ptr) {
3694                         spin_lock(&last_ptr->lock);
3695                         last_ptr->window_start = ins->objectid;
3696                         spin_unlock(&last_ptr->lock);
3697                 }
3698                 return 0;
3699         }
3700
3701         /*
3702          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
3703          *                      caching kthreads as we move along
3704          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
3705          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
3706          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
3707          *                     again
3708          */
3709         if (ffe_ctl->loop < LOOP_NO_EMPTY_SIZE) {
3710                 ffe_ctl->index = 0;
3711                 if (ffe_ctl->loop == LOOP_CACHING_NOWAIT) {
3712                         /*
3713                          * We want to skip the LOOP_CACHING_WAIT step if we
3714                          * don't have any uncached bgs and we've already done a
3715                          * full search through.
3716                          */
3717                         if (ffe_ctl->orig_have_caching_bg || !full_search)
3718                                 ffe_ctl->loop = LOOP_CACHING_WAIT;
3719                         else
3720                                 ffe_ctl->loop = LOOP_ALLOC_CHUNK;
3721                 } else {
3722                         ffe_ctl->loop++;
3723                 }
3724
3725                 if (ffe_ctl->loop == LOOP_ALLOC_CHUNK) {
3726                         struct btrfs_trans_handle *trans;
3727                         int exist = 0;
3728
3729                         trans = current->journal_info;
3730                         if (trans)
3731                                 exist = 1;
3732                         else
3733                                 trans = btrfs_join_transaction(root);
3734
3735                         if (IS_ERR(trans)) {
3736                                 ret = PTR_ERR(trans);
3737                                 return ret;
3738                         }
3739
3740                         ret = btrfs_chunk_alloc(trans, ffe_ctl->flags,
3741                                                 CHUNK_ALLOC_FORCE);
3742
3743                         /*
3744                          * If we can't allocate a new chunk we've already looped
3745                          * through at least once, move on to the NO_EMPTY_SIZE
3746                          * case.
3747                          */
3748                         if (ret == -ENOSPC)
3749                                 ffe_ctl->loop = LOOP_NO_EMPTY_SIZE;
3750
3751                         /* Do not bail out on ENOSPC since we can do more. */
3752                         if (ret < 0 && ret != -ENOSPC)
3753                                 btrfs_abort_transaction(trans, ret);
3754                         else
3755                                 ret = 0;
3756                         if (!exist)
3757                                 btrfs_end_transaction(trans);
3758                         if (ret)
3759                                 return ret;
3760                 }
3761
3762                 if (ffe_ctl->loop == LOOP_NO_EMPTY_SIZE) {
3763                         /*
3764                          * Don't loop again if we already have no empty_size and
3765                          * no empty_cluster.
3766                          */
3767                         if (ffe_ctl->empty_size == 0 &&
3768                             ffe_ctl->empty_cluster == 0)
3769                                 return -ENOSPC;
3770                         ffe_ctl->empty_size = 0;
3771                         ffe_ctl->empty_cluster = 0;
3772                 }
3773                 return 1;
3774         }
3775         return -ENOSPC;
3776 }
3777
3778 /*
3779  * walks the btree of allocated extents and find a hole of a given size.
3780  * The key ins is changed to record the hole:
3781  * ins->objectid == start position
3782  * ins->flags = BTRFS_EXTENT_ITEM_KEY
3783  * ins->offset == the size of the hole.
3784  * Any available blocks before search_start are skipped.
3785  *
3786  * If there is no suitable free space, we will record the max size of
3787  * the free space extent currently.
3788  *
3789  * The overall logic and call chain:
3790  *
3791  * find_free_extent()
3792  * |- Iterate through all block groups
3793  * |  |- Get a valid block group
3794  * |  |- Try to do clustered allocation in that block group
3795  * |  |- Try to do unclustered allocation in that block group
3796  * |  |- Check if the result is valid
3797  * |  |  |- If valid, then exit
3798  * |  |- Jump to next block group
3799  * |
3800  * |- Push harder to find free extents
3801  *    |- If not found, re-iterate all block groups
3802  */
3803 static noinline int find_free_extent(struct btrfs_fs_info *fs_info,
3804                                 u64 ram_bytes, u64 num_bytes, u64 empty_size,
3805                                 u64 hint_byte_orig, struct btrfs_key *ins,
3806                                 u64 flags, int delalloc)
3807 {
3808         int ret = 0;
3809         int cache_block_group_error = 0;
3810         struct btrfs_block_group *block_group = NULL;
3811         struct find_free_extent_ctl ffe_ctl = {0};
3812         struct btrfs_space_info *space_info;
3813         bool full_search = false;
3814
3815         WARN_ON(num_bytes < fs_info->sectorsize);
3816
3817         ffe_ctl.num_bytes = num_bytes;
3818         ffe_ctl.empty_size = empty_size;
3819         ffe_ctl.flags = flags;
3820         ffe_ctl.search_start = 0;
3821         ffe_ctl.delalloc = delalloc;
3822         ffe_ctl.index = btrfs_bg_flags_to_raid_index(flags);
3823         ffe_ctl.have_caching_bg = false;
3824         ffe_ctl.orig_have_caching_bg = false;
3825         ffe_ctl.found_offset = 0;
3826         ffe_ctl.hint_byte = hint_byte_orig;
3827         ffe_ctl.policy = BTRFS_EXTENT_ALLOC_CLUSTERED;
3828
3829         /* For clustered allocation */
3830         ffe_ctl.retry_clustered = false;
3831         ffe_ctl.retry_unclustered = false;
3832         ffe_ctl.last_ptr = NULL;
3833         ffe_ctl.use_cluster = true;
3834
3835         ins->type = BTRFS_EXTENT_ITEM_KEY;
3836         ins->objectid = 0;
3837         ins->offset = 0;
3838
3839         trace_find_free_extent(fs_info, num_bytes, empty_size, flags);
3840
3841         space_info = btrfs_find_space_info(fs_info, flags);
3842         if (!space_info) {
3843                 btrfs_err(fs_info, "No space info for %llu", flags);
3844                 return -ENOSPC;
3845         }
3846
3847         /*
3848          * If our free space is heavily fragmented we may not be able to make
3849          * big contiguous allocations, so instead of doing the expensive search
3850          * for free space, simply return ENOSPC with our max_extent_size so we
3851          * can go ahead and search for a more manageable chunk.
3852          *
3853          * If our max_extent_size is large enough for our allocation simply
3854          * disable clustering since we will likely not be able to find enough
3855          * space to create a cluster and induce latency trying.
3856          */
3857         if (unlikely(space_info->max_extent_size)) {
3858                 spin_lock(&space_info->lock);
3859                 if (space_info->max_extent_size &&
3860                     num_bytes > space_info->max_extent_size) {
3861                         ins->offset = space_info->max_extent_size;
3862                         spin_unlock(&space_info->lock);
3863                         return -ENOSPC;
3864                 } else if (space_info->max_extent_size) {
3865                         ffe_ctl.use_cluster = false;
3866                 }
3867                 spin_unlock(&space_info->lock);
3868         }
3869
3870         ffe_ctl.last_ptr = fetch_cluster_info(fs_info, space_info,
3871                                               &ffe_ctl.empty_cluster);
3872         if (ffe_ctl.last_ptr) {
3873                 struct btrfs_free_cluster *last_ptr = ffe_ctl.last_ptr;
3874
3875                 spin_lock(&last_ptr->lock);
3876                 if (last_ptr->block_group)
3877                         ffe_ctl.hint_byte = last_ptr->window_start;
3878                 if (last_ptr->fragmented) {
3879                         /*
3880                          * We still set window_start so we can keep track of the
3881                          * last place we found an allocation to try and save
3882                          * some time.
3883                          */
3884                         ffe_ctl.hint_byte = last_ptr->window_start;
3885                         ffe_ctl.use_cluster = false;
3886                 }
3887                 spin_unlock(&last_ptr->lock);
3888         }
3889
3890         ffe_ctl.search_start = max(ffe_ctl.search_start,
3891                                    first_logical_byte(fs_info, 0));
3892         ffe_ctl.search_start = max(ffe_ctl.search_start, ffe_ctl.hint_byte);
3893         if (ffe_ctl.search_start == ffe_ctl.hint_byte) {
3894                 block_group = btrfs_lookup_block_group(fs_info,
3895                                                        ffe_ctl.search_start);
3896                 /*
3897                  * we don't want to use the block group if it doesn't match our
3898                  * allocation bits, or if its not cached.
3899                  *
3900                  * However if we are re-searching with an ideal block group
3901                  * picked out then we don't care that the block group is cached.
3902                  */
3903                 if (block_group && block_group_bits(block_group, flags) &&
3904                     block_group->cached != BTRFS_CACHE_NO) {
3905                         down_read(&space_info->groups_sem);
3906                         if (list_empty(&block_group->list) ||
3907                             block_group->ro) {
3908                                 /*
3909                                  * someone is removing this block group,
3910                                  * we can't jump into the have_block_group
3911                                  * target because our list pointers are not
3912                                  * valid
3913                                  */
3914                                 btrfs_put_block_group(block_group);
3915                                 up_read(&space_info->groups_sem);
3916                         } else {
3917                                 ffe_ctl.index = btrfs_bg_flags_to_raid_index(
3918                                                 block_group->flags);
3919                                 btrfs_lock_block_group(block_group, delalloc);
3920                                 goto have_block_group;
3921                         }
3922                 } else if (block_group) {
3923                         btrfs_put_block_group(block_group);
3924                 }
3925         }
3926 search:
3927         ffe_ctl.have_caching_bg = false;
3928         if (ffe_ctl.index == btrfs_bg_flags_to_raid_index(flags) ||
3929             ffe_ctl.index == 0)
3930                 full_search = true;
3931         down_read(&space_info->groups_sem);
3932         list_for_each_entry(block_group,
3933                             &space_info->block_groups[ffe_ctl.index], list) {
3934                 /* If the block group is read-only, we can skip it entirely. */
3935                 if (unlikely(block_group->ro))
3936                         continue;
3937
3938                 btrfs_grab_block_group(block_group, delalloc);
3939                 ffe_ctl.search_start = block_group->start;
3940
3941                 /*
3942                  * this can happen if we end up cycling through all the
3943                  * raid types, but we want to make sure we only allocate
3944                  * for the proper type.
3945                  */
3946                 if (!block_group_bits(block_group, flags)) {
3947                         u64 extra = BTRFS_BLOCK_GROUP_DUP |
3948                                 BTRFS_BLOCK_GROUP_RAID1_MASK |
3949                                 BTRFS_BLOCK_GROUP_RAID56_MASK |
3950                                 BTRFS_BLOCK_GROUP_RAID10;
3951
3952                         /*
3953                          * if they asked for extra copies and this block group
3954                          * doesn't provide them, bail.  This does allow us to
3955                          * fill raid0 from raid1.
3956                          */
3957                         if ((flags & extra) && !(block_group->flags & extra))
3958                                 goto loop;
3959
3960                         /*
3961                          * This block group has different flags than we want.
3962                          * It's possible that we have MIXED_GROUP flag but no
3963                          * block group is mixed.  Just skip such block group.
3964                          */
3965                         btrfs_release_block_group(block_group, delalloc);
3966                         continue;
3967                 }
3968
3969 have_block_group:
3970                 ffe_ctl.cached = btrfs_block_group_done(block_group);
3971                 if (unlikely(!ffe_ctl.cached)) {
3972                         ffe_ctl.have_caching_bg = true;
3973                         ret = btrfs_cache_block_group(block_group, 0);
3974
3975                         /*
3976                          * If we get ENOMEM here or something else we want to
3977                          * try other block groups, because it may not be fatal.
3978                          * However if we can't find anything else we need to
3979                          * save our return here so that we return the actual
3980                          * error that caused problems, not ENOSPC.
3981                          */
3982                         if (ret < 0) {
3983                                 if (!cache_block_group_error)
3984                                         cache_block_group_error = ret;
3985                                 ret = 0;
3986                                 goto loop;
3987                         }
3988                         ret = 0;
3989                 }
3990
3991                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
3992                         goto loop;
3993
3994                 /*
3995                  * Ok we want to try and use the cluster allocator, so
3996                  * lets look there
3997                  */
3998                 if (ffe_ctl.last_ptr && ffe_ctl.use_cluster) {
3999                         struct btrfs_block_group *cluster_bg = NULL;
4000
4001                         ret = find_free_extent_clustered(block_group,
4002                                                          ffe_ctl.last_ptr,
4003                                                          &ffe_ctl, &cluster_bg);
4004
4005                         if (ret == 0) {
4006                                 if (cluster_bg && cluster_bg != block_group) {
4007                                         btrfs_release_block_group(block_group,
4008                                                                   delalloc);
4009                                         block_group = cluster_bg;
4010                                 }
4011                                 goto checks;
4012                         } else if (ret == -EAGAIN) {
4013                                 goto have_block_group;
4014                         } else if (ret > 0) {
4015                                 goto loop;
4016                         }
4017                         /* ret == -ENOENT case falls through */
4018                 }
4019
4020                 ret = find_free_extent_unclustered(block_group,
4021                                                    ffe_ctl.last_ptr, &ffe_ctl);
4022                 if (ret == -EAGAIN)
4023                         goto have_block_group;
4024                 else if (ret > 0)
4025                         goto loop;
4026                 /* ret == 0 case falls through */
4027 checks:
4028                 ffe_ctl.search_start = round_up(ffe_ctl.found_offset,
4029                                              fs_info->stripesize);
4030
4031                 /* move on to the next group */
4032                 if (ffe_ctl.search_start + num_bytes >
4033                     block_group->start + block_group->length) {
4034                         btrfs_add_free_space(block_group, ffe_ctl.found_offset,
4035                                              num_bytes);
4036                         goto loop;
4037                 }
4038
4039                 if (ffe_ctl.found_offset < ffe_ctl.search_start)
4040                         btrfs_add_free_space(block_group, ffe_ctl.found_offset,
4041                                 ffe_ctl.search_start - ffe_ctl.found_offset);
4042
4043                 ret = btrfs_add_reserved_bytes(block_group, ram_bytes,
4044                                 num_bytes, delalloc);
4045                 if (ret == -EAGAIN) {
4046                         btrfs_add_free_space(block_group, ffe_ctl.found_offset,
4047                                              num_bytes);
4048                         goto loop;
4049                 }
4050                 btrfs_inc_block_group_reservations(block_group);
4051
4052                 /* we are all good, lets return */
4053                 ins->objectid = ffe_ctl.search_start;
4054                 ins->offset = num_bytes;
4055
4056                 trace_btrfs_reserve_extent(block_group, ffe_ctl.search_start,
4057                                            num_bytes);
4058                 btrfs_release_block_group(block_group, delalloc);
4059                 break;
4060 loop:
4061                 ffe_ctl.retry_clustered = false;
4062                 ffe_ctl.retry_unclustered = false;
4063                 BUG_ON(btrfs_bg_flags_to_raid_index(block_group->flags) !=
4064                        ffe_ctl.index);
4065                 btrfs_release_block_group(block_group, delalloc);
4066                 cond_resched();
4067         }
4068         up_read(&space_info->groups_sem);
4069
4070         ret = find_free_extent_update_loop(fs_info, ffe_ctl.last_ptr, ins,
4071                                            &ffe_ctl, full_search,
4072                                            ffe_ctl.use_cluster);
4073         if (ret > 0)
4074                 goto search;
4075
4076         if (ret == -ENOSPC && !cache_block_group_error) {
4077                 /*
4078                  * Use ffe_ctl->total_free_space as fallback if we can't find
4079                  * any contiguous hole.
4080                  */
4081                 if (!ffe_ctl.max_extent_size)
4082                         ffe_ctl.max_extent_size = ffe_ctl.total_free_space;
4083                 spin_lock(&space_info->lock);
4084                 space_info->max_extent_size = ffe_ctl.max_extent_size;
4085                 spin_unlock(&space_info->lock);
4086                 ins->offset = ffe_ctl.max_extent_size;
4087         } else if (ret == -ENOSPC) {
4088                 ret = cache_block_group_error;
4089         }
4090         return ret;
4091 }
4092
4093 /*
4094  * btrfs_reserve_extent - entry point to the extent allocator. Tries to find a
4095  *                        hole that is at least as big as @num_bytes.
4096  *
4097  * @root           -    The root that will contain this extent
4098  *
4099  * @ram_bytes      -    The amount of space in ram that @num_bytes take. This
4100  *                      is used for accounting purposes. This value differs
4101  *                      from @num_bytes only in the case of compressed extents.
4102  *
4103  * @num_bytes      -    Number of bytes to allocate on-disk.
4104  *
4105  * @min_alloc_size -    Indicates the minimum amount of space that the
4106  *                      allocator should try to satisfy. In some cases
4107  *                      @num_bytes may be larger than what is required and if
4108  *                      the filesystem is fragmented then allocation fails.
4109  *                      However, the presence of @min_alloc_size gives a
4110  *                      chance to try and satisfy the smaller allocation.
4111  *
4112  * @empty_size     -    A hint that you plan on doing more COW. This is the
4113  *                      size in bytes the allocator should try to find free
4114  *                      next to the block it returns.  This is just a hint and
4115  *                      may be ignored by the allocator.
4116  *
4117  * @hint_byte      -    Hint to the allocator to start searching above the byte
4118  *                      address passed. It might be ignored.
4119  *
4120  * @ins            -    This key is modified to record the found hole. It will
4121  *                      have the following values:
4122  *                      ins->objectid == start position
4123  *                      ins->flags = BTRFS_EXTENT_ITEM_KEY
4124  *                      ins->offset == the size of the hole.
4125  *
4126  * @is_data        -    Boolean flag indicating whether an extent is
4127  *                      allocated for data (true) or metadata (false)
4128  *
4129  * @delalloc       -    Boolean flag indicating whether this allocation is for
4130  *                      delalloc or not. If 'true' data_rwsem of block groups
4131  *                      is going to be acquired.
4132  *
4133  *
4134  * Returns 0 when an allocation succeeded or < 0 when an error occurred. In
4135  * case -ENOSPC is returned then @ins->offset will contain the size of the
4136  * largest available hole the allocator managed to find.
4137  */
4138 int btrfs_reserve_extent(struct btrfs_root *root, u64 ram_bytes,
4139                          u64 num_bytes, u64 min_alloc_size,
4140                          u64 empty_size, u64 hint_byte,
4141                          struct btrfs_key *ins, int is_data, int delalloc)
4142 {
4143         struct btrfs_fs_info *fs_info = root->fs_info;
4144         bool final_tried = num_bytes == min_alloc_size;
4145         u64 flags;
4146         int ret;
4147
4148         flags = get_alloc_profile_by_root(root, is_data);
4149 again:
4150         WARN_ON(num_bytes < fs_info->sectorsize);
4151         ret = find_free_extent(fs_info, ram_bytes, num_bytes, empty_size,
4152                                hint_byte, ins, flags, delalloc);
4153         if (!ret && !is_data) {
4154                 btrfs_dec_block_group_reservations(fs_info, ins->objectid);
4155         } else if (ret == -ENOSPC) {
4156                 if (!final_tried && ins->offset) {
4157                         num_bytes = min(num_bytes >> 1, ins->offset);
4158                         num_bytes = round_down(num_bytes,
4159                                                fs_info->sectorsize);
4160                         num_bytes = max(num_bytes, min_alloc_size);
4161                         ram_bytes = num_bytes;
4162                         if (num_bytes == min_alloc_size)
4163                                 final_tried = true;
4164                         goto again;
4165                 } else if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
4166                         struct btrfs_space_info *sinfo;
4167
4168                         sinfo = btrfs_find_space_info(fs_info, flags);
4169                         btrfs_err(fs_info,
4170                                   "allocation failed flags %llu, wanted %llu",
4171                                   flags, num_bytes);
4172                         if (sinfo)
4173                                 btrfs_dump_space_info(fs_info, sinfo,
4174                                                       num_bytes, 1);
4175                 }
4176         }
4177
4178         return ret;
4179 }
4180
4181 int btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
4182                                u64 start, u64 len, int delalloc)
4183 {
4184         struct btrfs_block_group *cache;
4185
4186         cache = btrfs_lookup_block_group(fs_info, start);
4187         if (!cache) {
4188                 btrfs_err(fs_info, "Unable to find block group for %llu",
4189                           start);
4190                 return -ENOSPC;
4191         }
4192
4193         btrfs_add_free_space(cache, start, len);
4194         btrfs_free_reserved_bytes(cache, len, delalloc);
4195         trace_btrfs_reserved_extent_free(fs_info, start, len);
4196
4197         btrfs_put_block_group(cache);
4198         return 0;
4199 }
4200
4201 int btrfs_pin_reserved_extent(struct btrfs_trans_handle *trans, u64 start,
4202                               u64 len)
4203 {
4204         struct btrfs_block_group *cache;
4205         int ret = 0;
4206
4207         cache = btrfs_lookup_block_group(trans->fs_info, start);
4208         if (!cache) {
4209                 btrfs_err(trans->fs_info, "unable to find block group for %llu",
4210                           start);
4211                 return -ENOSPC;
4212         }
4213
4214         ret = pin_down_extent(trans, cache, start, len, 1);
4215         btrfs_put_block_group(cache);
4216         return ret;
4217 }
4218
4219 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
4220                                       u64 parent, u64 root_objectid,
4221                                       u64 flags, u64 owner, u64 offset,
4222                                       struct btrfs_key *ins, int ref_mod)
4223 {
4224         struct btrfs_fs_info *fs_info = trans->fs_info;
4225         int ret;
4226         struct btrfs_extent_item *extent_item;
4227         struct btrfs_extent_inline_ref *iref;
4228         struct btrfs_path *path;
4229         struct extent_buffer *leaf;
4230         int type;
4231         u32 size;
4232
4233         if (parent > 0)
4234                 type = BTRFS_SHARED_DATA_REF_KEY;
4235         else
4236                 type = BTRFS_EXTENT_DATA_REF_KEY;
4237
4238         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
4239
4240         path = btrfs_alloc_path();
4241         if (!path)
4242                 return -ENOMEM;
4243
4244         path->leave_spinning = 1;
4245         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
4246                                       ins, size);
4247         if (ret) {
4248                 btrfs_free_path(path);
4249                 return ret;
4250         }
4251
4252         leaf = path->nodes[0];
4253         extent_item = btrfs_item_ptr(leaf, path->slots[0],
4254                                      struct btrfs_extent_item);
4255         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
4256         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
4257         btrfs_set_extent_flags(leaf, extent_item,
4258                                flags | BTRFS_EXTENT_FLAG_DATA);
4259
4260         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
4261         btrfs_set_extent_inline_ref_type(leaf, iref, type);
4262         if (parent > 0) {
4263                 struct btrfs_shared_data_ref *ref;
4264                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
4265                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
4266                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
4267         } else {
4268                 struct btrfs_extent_data_ref *ref;
4269                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
4270                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
4271                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
4272                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
4273                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
4274         }
4275
4276         btrfs_mark_buffer_dirty(path->nodes[0]);
4277         btrfs_free_path(path);
4278
4279         ret = remove_from_free_space_tree(trans, ins->objectid, ins->offset);
4280         if (ret)
4281                 return ret;
4282
4283         ret = btrfs_update_block_group(trans, ins->objectid, ins->offset, 1);
4284         if (ret) { /* -ENOENT, logic error */
4285                 btrfs_err(fs_info, "update block group failed for %llu %llu",
4286                         ins->objectid, ins->offset);
4287                 BUG();
4288         }
4289         trace_btrfs_reserved_extent_alloc(fs_info, ins->objectid, ins->offset);
4290         return ret;
4291 }
4292
4293 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
4294                                      struct btrfs_delayed_ref_node *node,
4295                                      struct btrfs_delayed_extent_op *extent_op)
4296 {
4297         struct btrfs_fs_info *fs_info = trans->fs_info;
4298         int ret;
4299         struct btrfs_extent_item *extent_item;
4300         struct btrfs_key extent_key;
4301         struct btrfs_tree_block_info *block_info;
4302         struct btrfs_extent_inline_ref *iref;
4303         struct btrfs_path *path;
4304         struct extent_buffer *leaf;
4305         struct btrfs_delayed_tree_ref *ref;
4306         u32 size = sizeof(*extent_item) + sizeof(*iref);
4307         u64 num_bytes;
4308         u64 flags = extent_op->flags_to_set;
4309         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
4310
4311         ref = btrfs_delayed_node_to_tree_ref(node);
4312
4313         extent_key.objectid = node->bytenr;
4314         if (skinny_metadata) {
4315                 extent_key.offset = ref->level;
4316                 extent_key.type = BTRFS_METADATA_ITEM_KEY;
4317                 num_bytes = fs_info->nodesize;
4318         } else {
4319                 extent_key.offset = node->num_bytes;
4320                 extent_key.type = BTRFS_EXTENT_ITEM_KEY;
4321                 size += sizeof(*block_info);
4322                 num_bytes = node->num_bytes;
4323         }
4324
4325         path = btrfs_alloc_path();
4326         if (!path)
4327                 return -ENOMEM;
4328
4329         path->leave_spinning = 1;
4330         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
4331                                       &extent_key, size);
4332         if (ret) {
4333                 btrfs_free_path(path);
4334                 return ret;
4335         }
4336
4337         leaf = path->nodes[0];
4338         extent_item = btrfs_item_ptr(leaf, path->slots[0],
4339                                      struct btrfs_extent_item);
4340         btrfs_set_extent_refs(leaf, extent_item, 1);
4341         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
4342         btrfs_set_extent_flags(leaf, extent_item,
4343                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
4344
4345         if (skinny_metadata) {
4346                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
4347         } else {
4348                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
4349                 btrfs_set_tree_block_key(leaf, block_info, &extent_op->key);
4350                 btrfs_set_tree_block_level(leaf, block_info, ref->level);
4351                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
4352         }
4353
4354         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY) {
4355                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
4356                 btrfs_set_extent_inline_ref_type(leaf, iref,
4357                                                  BTRFS_SHARED_BLOCK_REF_KEY);
4358                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->parent);
4359         } else {
4360                 btrfs_set_extent_inline_ref_type(leaf, iref,
4361                                                  BTRFS_TREE_BLOCK_REF_KEY);
4362                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->root);
4363         }
4364
4365         btrfs_mark_buffer_dirty(leaf);
4366         btrfs_free_path(path);
4367
4368         ret = remove_from_free_space_tree(trans, extent_key.objectid,
4369                                           num_bytes);
4370         if (ret)
4371                 return ret;
4372
4373         ret = btrfs_update_block_group(trans, extent_key.objectid,
4374                                        fs_info->nodesize, 1);
4375         if (ret) { /* -ENOENT, logic error */
4376                 btrfs_err(fs_info, "update block group failed for %llu %llu",
4377                         extent_key.objectid, extent_key.offset);
4378                 BUG();
4379         }
4380
4381         trace_btrfs_reserved_extent_alloc(fs_info, extent_key.objectid,
4382                                           fs_info->nodesize);
4383         return ret;
4384 }
4385
4386 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
4387                                      struct btrfs_root *root, u64 owner,
4388                                      u64 offset, u64 ram_bytes,
4389                                      struct btrfs_key *ins)
4390 {
4391         struct btrfs_ref generic_ref = { 0 };
4392         int ret;
4393
4394         BUG_ON(root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
4395
4396         btrfs_init_generic_ref(&generic_ref, BTRFS_ADD_DELAYED_EXTENT,
4397                                ins->objectid, ins->offset, 0);
4398         btrfs_init_data_ref(&generic_ref, root->root_key.objectid, owner, offset);
4399         btrfs_ref_tree_mod(root->fs_info, &generic_ref);
4400         ret = btrfs_add_delayed_data_ref(trans, &generic_ref,
4401                                          ram_bytes, NULL, NULL);
4402         return ret;
4403 }
4404
4405 /*
4406  * this is used by the tree logging recovery code.  It records that
4407  * an extent has been allocated and makes sure to clear the free
4408  * space cache bits as well
4409  */
4410 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
4411                                    u64 root_objectid, u64 owner, u64 offset,
4412                                    struct btrfs_key *ins)
4413 {
4414         struct btrfs_fs_info *fs_info = trans->fs_info;
4415         int ret;
4416         struct btrfs_block_group *block_group;
4417         struct btrfs_space_info *space_info;
4418
4419         /*
4420          * Mixed block groups will exclude before processing the log so we only
4421          * need to do the exclude dance if this fs isn't mixed.
4422          */
4423         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS)) {
4424                 ret = __exclude_logged_extent(fs_info, ins->objectid,
4425                                               ins->offset);
4426                 if (ret)
4427                         return ret;
4428         }
4429
4430         block_group = btrfs_lookup_block_group(fs_info, ins->objectid);
4431         if (!block_group)
4432                 return -EINVAL;
4433
4434         space_info = block_group->space_info;
4435         spin_lock(&space_info->lock);
4436         spin_lock(&block_group->lock);
4437         space_info->bytes_reserved += ins->offset;
4438         block_group->reserved += ins->offset;
4439         spin_unlock(&block_group->lock);
4440         spin_unlock(&space_info->lock);
4441
4442         ret = alloc_reserved_file_extent(trans, 0, root_objectid, 0, owner,
4443                                          offset, ins, 1);
4444         if (ret)
4445                 btrfs_pin_extent(trans, ins->objectid, ins->offset, 1);
4446         btrfs_put_block_group(block_group);
4447         return ret;
4448 }
4449
4450 static struct extent_buffer *
4451 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
4452                       u64 bytenr, int level, u64 owner)
4453 {
4454         struct btrfs_fs_info *fs_info = root->fs_info;
4455         struct extent_buffer *buf;
4456
4457         buf = btrfs_find_create_tree_block(fs_info, bytenr);
4458         if (IS_ERR(buf))
4459                 return buf;
4460
4461         /*
4462          * Extra safety check in case the extent tree is corrupted and extent
4463          * allocator chooses to use a tree block which is already used and
4464          * locked.
4465          */
4466         if (buf->lock_owner == current->pid) {
4467                 btrfs_err_rl(fs_info,
4468 "tree block %llu owner %llu already locked by pid=%d, extent tree corruption detected",
4469                         buf->start, btrfs_header_owner(buf), current->pid);
4470                 free_extent_buffer(buf);
4471                 return ERR_PTR(-EUCLEAN);
4472         }
4473
4474         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
4475         btrfs_tree_lock(buf);
4476         btrfs_clean_tree_block(buf);
4477         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
4478
4479         btrfs_set_lock_blocking_write(buf);
4480         set_extent_buffer_uptodate(buf);
4481
4482         memzero_extent_buffer(buf, 0, sizeof(struct btrfs_header));
4483         btrfs_set_header_level(buf, level);
4484         btrfs_set_header_bytenr(buf, buf->start);
4485         btrfs_set_header_generation(buf, trans->transid);
4486         btrfs_set_header_backref_rev(buf, BTRFS_MIXED_BACKREF_REV);
4487         btrfs_set_header_owner(buf, owner);
4488         write_extent_buffer_fsid(buf, fs_info->fs_devices->metadata_uuid);
4489         write_extent_buffer_chunk_tree_uuid(buf, fs_info->chunk_tree_uuid);
4490         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
4491                 buf->log_index = root->log_transid % 2;
4492                 /*
4493                  * we allow two log transactions at a time, use different
4494                  * EXTENT bit to differentiate dirty pages.
4495                  */
4496                 if (buf->log_index == 0)
4497                         set_extent_dirty(&root->dirty_log_pages, buf->start,
4498                                         buf->start + buf->len - 1, GFP_NOFS);
4499                 else
4500                         set_extent_new(&root->dirty_log_pages, buf->start,
4501                                         buf->start + buf->len - 1);
4502         } else {
4503                 buf->log_index = -1;
4504                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
4505                          buf->start + buf->len - 1, GFP_NOFS);
4506         }
4507         trans->dirty = true;
4508         /* this returns a buffer locked for blocking */
4509         return buf;
4510 }
4511
4512 /*
4513  * finds a free extent and does all the dirty work required for allocation
4514  * returns the tree buffer or an ERR_PTR on error.
4515  */
4516 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
4517                                              struct btrfs_root *root,
4518                                              u64 parent, u64 root_objectid,
4519                                              const struct btrfs_disk_key *key,
4520                                              int level, u64 hint,
4521                                              u64 empty_size)
4522 {
4523         struct btrfs_fs_info *fs_info = root->fs_info;
4524         struct btrfs_key ins;
4525         struct btrfs_block_rsv *block_rsv;
4526         struct extent_buffer *buf;
4527         struct btrfs_delayed_extent_op *extent_op;
4528         struct btrfs_ref generic_ref = { 0 };
4529         u64 flags = 0;
4530         int ret;
4531         u32 blocksize = fs_info->nodesize;
4532         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
4533
4534 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4535         if (btrfs_is_testing(fs_info)) {
4536                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
4537                                             level, root_objectid);
4538                 if (!IS_ERR(buf))
4539                         root->alloc_bytenr += blocksize;
4540                 return buf;
4541         }
4542 #endif
4543
4544         block_rsv = btrfs_use_block_rsv(trans, root, blocksize);
4545         if (IS_ERR(block_rsv))
4546                 return ERR_CAST(block_rsv);
4547
4548         ret = btrfs_reserve_extent(root, blocksize, blocksize, blocksize,
4549                                    empty_size, hint, &ins, 0, 0);
4550         if (ret)
4551                 goto out_unuse;
4552
4553         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level,
4554                                     root_objectid);
4555         if (IS_ERR(buf)) {
4556                 ret = PTR_ERR(buf);
4557                 goto out_free_reserved;
4558         }
4559
4560         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
4561                 if (parent == 0)
4562                         parent = ins.objectid;
4563                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
4564         } else
4565                 BUG_ON(parent > 0);
4566
4567         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
4568                 extent_op = btrfs_alloc_delayed_extent_op();
4569                 if (!extent_op) {
4570                         ret = -ENOMEM;
4571                         goto out_free_buf;
4572                 }
4573                 if (key)
4574                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
4575                 else
4576                         memset(&extent_op->key, 0, sizeof(extent_op->key));
4577                 extent_op->flags_to_set = flags;
4578                 extent_op->update_key = skinny_metadata ? false : true;
4579                 extent_op->update_flags = true;
4580                 extent_op->is_data = false;
4581                 extent_op->level = level;
4582
4583                 btrfs_init_generic_ref(&generic_ref, BTRFS_ADD_DELAYED_EXTENT,
4584                                        ins.objectid, ins.offset, parent);
4585                 generic_ref.real_root = root->root_key.objectid;
4586                 btrfs_init_tree_ref(&generic_ref, level, root_objectid);
4587                 btrfs_ref_tree_mod(fs_info, &generic_ref);
4588                 ret = btrfs_add_delayed_tree_ref(trans, &generic_ref,
4589                                                  extent_op, NULL, NULL);
4590                 if (ret)
4591                         goto out_free_delayed;
4592         }
4593         return buf;
4594
4595 out_free_delayed:
4596         btrfs_free_delayed_extent_op(extent_op);
4597 out_free_buf:
4598         free_extent_buffer(buf);
4599 out_free_reserved:
4600         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 0);
4601 out_unuse:
4602         btrfs_unuse_block_rsv(fs_info, block_rsv, blocksize);
4603         return ERR_PTR(ret);
4604 }
4605
4606 struct walk_control {
4607         u64 refs[BTRFS_MAX_LEVEL];
4608         u64 flags[BTRFS_MAX_LEVEL];
4609         struct btrfs_key update_progress;
4610         struct btrfs_key drop_progress;
4611         int drop_level;
4612         int stage;
4613         int level;
4614         int shared_level;
4615         int update_ref;
4616         int keep_locks;
4617         int reada_slot;
4618         int reada_count;
4619         int restarted;
4620 };
4621
4622 #define DROP_REFERENCE  1
4623 #define UPDATE_BACKREF  2
4624
4625 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
4626                                      struct btrfs_root *root,
4627                                      struct walk_control *wc,
4628                                      struct btrfs_path *path)
4629 {
4630         struct btrfs_fs_info *fs_info = root->fs_info;
4631         u64 bytenr;
4632         u64 generation;
4633         u64 refs;
4634         u64 flags;
4635         u32 nritems;
4636         struct btrfs_key key;
4637         struct extent_buffer *eb;
4638         int ret;
4639         int slot;
4640         int nread = 0;
4641
4642         if (path->slots[wc->level] < wc->reada_slot) {
4643                 wc->reada_count = wc->reada_count * 2 / 3;
4644                 wc->reada_count = max(wc->reada_count, 2);
4645         } else {
4646                 wc->reada_count = wc->reada_count * 3 / 2;
4647                 wc->reada_count = min_t(int, wc->reada_count,
4648                                         BTRFS_NODEPTRS_PER_BLOCK(fs_info));
4649         }
4650
4651         eb = path->nodes[wc->level];
4652         nritems = btrfs_header_nritems(eb);
4653
4654         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
4655                 if (nread >= wc->reada_count)
4656                         break;
4657
4658                 cond_resched();
4659                 bytenr = btrfs_node_blockptr(eb, slot);
4660                 generation = btrfs_node_ptr_generation(eb, slot);
4661
4662                 if (slot == path->slots[wc->level])
4663                         goto reada;
4664
4665                 if (wc->stage == UPDATE_BACKREF &&
4666                     generation <= root->root_key.offset)
4667                         continue;
4668
4669                 /* We don't lock the tree block, it's OK to be racy here */
4670                 ret = btrfs_lookup_extent_info(trans, fs_info, bytenr,
4671                                                wc->level - 1, 1, &refs,
4672                                                &flags);
4673                 /* We don't care about errors in readahead. */
4674                 if (ret < 0)
4675                         continue;
4676                 BUG_ON(refs == 0);
4677
4678                 if (wc->stage == DROP_REFERENCE) {
4679                         if (refs == 1)
4680                                 goto reada;
4681
4682                         if (wc->level == 1 &&
4683                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
4684                                 continue;
4685                         if (!wc->update_ref ||
4686                             generation <= root->root_key.offset)
4687                                 continue;
4688                         btrfs_node_key_to_cpu(eb, &key, slot);
4689                         ret = btrfs_comp_cpu_keys(&key,
4690                                                   &wc->update_progress);
4691                         if (ret < 0)
4692                                 continue;
4693                 } else {
4694                         if (wc->level == 1 &&
4695                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
4696                                 continue;
4697                 }
4698 reada:
4699                 readahead_tree_block(fs_info, bytenr);
4700                 nread++;
4701         }
4702         wc->reada_slot = slot;
4703 }
4704
4705 /*
4706  * helper to process tree block while walking down the tree.
4707  *
4708  * when wc->stage == UPDATE_BACKREF, this function updates
4709  * back refs for pointers in the block.
4710  *
4711  * NOTE: return value 1 means we should stop walking down.
4712  */
4713 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
4714                                    struct btrfs_root *root,
4715                                    struct btrfs_path *path,
4716                                    struct walk_control *wc, int lookup_info)
4717 {
4718         struct btrfs_fs_info *fs_info = root->fs_info;
4719         int level = wc->level;
4720         struct extent_buffer *eb = path->nodes[level];
4721         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
4722         int ret;
4723
4724         if (wc->stage == UPDATE_BACKREF &&
4725             btrfs_header_owner(eb) != root->root_key.objectid)
4726                 return 1;
4727
4728         /*
4729          * when reference count of tree block is 1, it won't increase
4730          * again. once full backref flag is set, we never clear it.
4731          */
4732         if (lookup_info &&
4733             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
4734              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
4735                 BUG_ON(!path->locks[level]);
4736                 ret = btrfs_lookup_extent_info(trans, fs_info,
4737                                                eb->start, level, 1,
4738                                                &wc->refs[level],
4739                                                &wc->flags[level]);
4740                 BUG_ON(ret == -ENOMEM);
4741                 if (ret)
4742                         return ret;
4743                 BUG_ON(wc->refs[level] == 0);
4744         }
4745
4746         if (wc->stage == DROP_REFERENCE) {
4747                 if (wc->refs[level] > 1)
4748                         return 1;
4749
4750                 if (path->locks[level] && !wc->keep_locks) {
4751                         btrfs_tree_unlock_rw(eb, path->locks[level]);
4752                         path->locks[level] = 0;
4753                 }
4754                 return 0;
4755         }
4756
4757         /* wc->stage == UPDATE_BACKREF */
4758         if (!(wc->flags[level] & flag)) {
4759                 BUG_ON(!path->locks[level]);
4760                 ret = btrfs_inc_ref(trans, root, eb, 1);
4761                 BUG_ON(ret); /* -ENOMEM */
4762                 ret = btrfs_dec_ref(trans, root, eb, 0);
4763                 BUG_ON(ret); /* -ENOMEM */
4764                 ret = btrfs_set_disk_extent_flags(trans, eb, flag,
4765                                                   btrfs_header_level(eb), 0);
4766                 BUG_ON(ret); /* -ENOMEM */
4767                 wc->flags[level] |= flag;
4768         }
4769
4770         /*
4771          * the block is shared by multiple trees, so it's not good to
4772          * keep the tree lock
4773          */
4774         if (path->locks[level] && level > 0) {
4775                 btrfs_tree_unlock_rw(eb, path->locks[level]);
4776                 path->locks[level] = 0;
4777         }
4778         return 0;
4779 }
4780
4781 /*
4782  * This is used to verify a ref exists for this root to deal with a bug where we
4783  * would have a drop_progress key that hadn't been updated properly.
4784  */
4785 static int check_ref_exists(struct btrfs_trans_handle *trans,
4786                             struct btrfs_root *root, u64 bytenr, u64 parent,
4787                             int level)
4788 {
4789         struct btrfs_path *path;
4790         struct btrfs_extent_inline_ref *iref;
4791         int ret;
4792
4793         path = btrfs_alloc_path();
4794         if (!path)
4795                 return -ENOMEM;
4796
4797         ret = lookup_extent_backref(trans, path, &iref, bytenr,
4798                                     root->fs_info->nodesize, parent,
4799                                     root->root_key.objectid, level, 0);
4800         btrfs_free_path(path);
4801         if (ret == -ENOENT)
4802                 return 0;
4803         if (ret < 0)
4804                 return ret;
4805         return 1;
4806 }
4807
4808 /*
4809  * helper to process tree block pointer.
4810  *
4811  * when wc->stage == DROP_REFERENCE, this function checks
4812  * reference count of the block pointed to. if the block
4813  * is shared and we need update back refs for the subtree
4814  * rooted at the block, this function changes wc->stage to
4815  * UPDATE_BACKREF. if the block is shared and there is no
4816  * need to update back, this function drops the reference
4817  * to the block.
4818  *
4819  * NOTE: return value 1 means we should stop walking down.
4820  */
4821 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
4822                                  struct btrfs_root *root,
4823                                  struct btrfs_path *path,
4824                                  struct walk_control *wc, int *lookup_info)
4825 {
4826         struct btrfs_fs_info *fs_info = root->fs_info;
4827         u64 bytenr;
4828         u64 generation;
4829         u64 parent;
4830         struct btrfs_key key;
4831         struct btrfs_key first_key;
4832         struct btrfs_ref ref = { 0 };
4833         struct extent_buffer *next;
4834         int level = wc->level;
4835         int reada = 0;
4836         int ret = 0;
4837         bool need_account = false;
4838
4839         generation = btrfs_node_ptr_generation(path->nodes[level],
4840                                                path->slots[level]);
4841         /*
4842          * if the lower level block was created before the snapshot
4843          * was created, we know there is no need to update back refs
4844          * for the subtree
4845          */
4846         if (wc->stage == UPDATE_BACKREF &&
4847             generation <= root->root_key.offset) {
4848                 *lookup_info = 1;
4849                 return 1;
4850         }
4851
4852         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
4853         btrfs_node_key_to_cpu(path->nodes[level], &first_key,
4854                               path->slots[level]);
4855
4856         next = find_extent_buffer(fs_info, bytenr);
4857         if (!next) {
4858                 next = btrfs_find_create_tree_block(fs_info, bytenr);
4859                 if (IS_ERR(next))
4860                         return PTR_ERR(next);
4861
4862                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
4863                                                level - 1);
4864                 reada = 1;
4865         }
4866         btrfs_tree_lock(next);
4867         btrfs_set_lock_blocking_write(next);
4868
4869         ret = btrfs_lookup_extent_info(trans, fs_info, bytenr, level - 1, 1,
4870                                        &wc->refs[level - 1],
4871                                        &wc->flags[level - 1]);
4872         if (ret < 0)
4873                 goto out_unlock;
4874
4875         if (unlikely(wc->refs[level - 1] == 0)) {
4876                 btrfs_err(fs_info, "Missing references.");
4877                 ret = -EIO;
4878                 goto out_unlock;
4879         }
4880         *lookup_info = 0;
4881
4882         if (wc->stage == DROP_REFERENCE) {
4883                 if (wc->refs[level - 1] > 1) {
4884                         need_account = true;
4885                         if (level == 1 &&
4886                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
4887                                 goto skip;
4888
4889                         if (!wc->update_ref ||
4890                             generation <= root->root_key.offset)
4891                                 goto skip;
4892
4893                         btrfs_node_key_to_cpu(path->nodes[level], &key,
4894                                               path->slots[level]);
4895                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
4896                         if (ret < 0)
4897                                 goto skip;
4898
4899                         wc->stage = UPDATE_BACKREF;
4900                         wc->shared_level = level - 1;
4901                 }
4902         } else {
4903                 if (level == 1 &&
4904                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
4905                         goto skip;
4906         }
4907
4908         if (!btrfs_buffer_uptodate(next, generation, 0)) {
4909                 btrfs_tree_unlock(next);
4910                 free_extent_buffer(next);
4911                 next = NULL;
4912                 *lookup_info = 1;
4913         }
4914
4915         if (!next) {
4916                 if (reada && level == 1)
4917                         reada_walk_down(trans, root, wc, path);
4918                 next = read_tree_block(fs_info, bytenr, generation, level - 1,
4919                                        &first_key);
4920                 if (IS_ERR(next)) {
4921                         return PTR_ERR(next);
4922                 } else if (!extent_buffer_uptodate(next)) {
4923                         free_extent_buffer(next);
4924                         return -EIO;
4925                 }
4926                 btrfs_tree_lock(next);
4927                 btrfs_set_lock_blocking_write(next);
4928         }
4929
4930         level--;
4931         ASSERT(level == btrfs_header_level(next));
4932         if (level != btrfs_header_level(next)) {
4933                 btrfs_err(root->fs_info, "mismatched level");
4934                 ret = -EIO;
4935                 goto out_unlock;
4936         }
4937         path->nodes[level] = next;
4938         path->slots[level] = 0;
4939         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
4940         wc->level = level;
4941         if (wc->level == 1)
4942                 wc->reada_slot = 0;
4943         return 0;
4944 skip:
4945         wc->refs[level - 1] = 0;
4946         wc->flags[level - 1] = 0;
4947         if (wc->stage == DROP_REFERENCE) {
4948                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
4949                         parent = path->nodes[level]->start;
4950                 } else {
4951                         ASSERT(root->root_key.objectid ==
4952                                btrfs_header_owner(path->nodes[level]));
4953                         if (root->root_key.objectid !=
4954                             btrfs_header_owner(path->nodes[level])) {
4955                                 btrfs_err(root->fs_info,
4956                                                 "mismatched block owner");
4957                                 ret = -EIO;
4958                                 goto out_unlock;
4959                         }
4960                         parent = 0;
4961                 }
4962
4963                 /*
4964                  * If we had a drop_progress we need to verify the refs are set
4965                  * as expected.  If we find our ref then we know that from here
4966                  * on out everything should be correct, and we can clear the
4967                  * ->restarted flag.
4968                  */
4969                 if (wc->restarted) {
4970                         ret = check_ref_exists(trans, root, bytenr, parent,
4971                                                level - 1);
4972                         if (ret < 0)
4973                                 goto out_unlock;
4974                         if (ret == 0)
4975                                 goto no_delete;
4976                         ret = 0;
4977                         wc->restarted = 0;
4978                 }
4979
4980                 /*
4981                  * Reloc tree doesn't contribute to qgroup numbers, and we have
4982                  * already accounted them at merge time (replace_path),
4983                  * thus we could skip expensive subtree trace here.
4984                  */
4985                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
4986                     need_account) {
4987                         ret = btrfs_qgroup_trace_subtree(trans, next,
4988                                                          generation, level - 1);
4989                         if (ret) {
4990                                 btrfs_err_rl(fs_info,
4991                                              "Error %d accounting shared subtree. Quota is out of sync, rescan required.",
4992                                              ret);
4993                         }
4994                 }
4995
4996                 /*
4997                  * We need to update the next key in our walk control so we can
4998                  * update the drop_progress key accordingly.  We don't care if
4999                  * find_next_key doesn't find a key because that means we're at
5000                  * the end and are going to clean up now.
5001                  */
5002                 wc->drop_level = level;
5003                 find_next_key(path, level, &wc->drop_progress);
5004
5005                 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
5006                                        fs_info->nodesize, parent);
5007                 btrfs_init_tree_ref(&ref, level - 1, root->root_key.objectid);
5008                 ret = btrfs_free_extent(trans, &ref);
5009                 if (ret)
5010                         goto out_unlock;
5011         }
5012 no_delete:
5013         *lookup_info = 1;
5014         ret = 1;
5015
5016 out_unlock:
5017         btrfs_tree_unlock(next);
5018         free_extent_buffer(next);
5019
5020         return ret;
5021 }
5022
5023 /*
5024  * helper to process tree block while walking up the tree.
5025  *
5026  * when wc->stage == DROP_REFERENCE, this function drops
5027  * reference count on the block.
5028  *
5029  * when wc->stage == UPDATE_BACKREF, this function changes
5030  * wc->stage back to DROP_REFERENCE if we changed wc->stage
5031  * to UPDATE_BACKREF previously while processing the block.
5032  *
5033  * NOTE: return value 1 means we should stop walking up.
5034  */
5035 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
5036                                  struct btrfs_root *root,
5037                                  struct btrfs_path *path,
5038                                  struct walk_control *wc)
5039 {
5040         struct btrfs_fs_info *fs_info = root->fs_info;
5041         int ret;
5042         int level = wc->level;
5043         struct extent_buffer *eb = path->nodes[level];
5044         u64 parent = 0;
5045
5046         if (wc->stage == UPDATE_BACKREF) {
5047                 BUG_ON(wc->shared_level < level);
5048                 if (level < wc->shared_level)
5049                         goto out;
5050
5051                 ret = find_next_key(path, level + 1, &wc->update_progress);
5052                 if (ret > 0)
5053                         wc->update_ref = 0;
5054
5055                 wc->stage = DROP_REFERENCE;
5056                 wc->shared_level = -1;
5057                 path->slots[level] = 0;
5058
5059                 /*
5060                  * check reference count again if the block isn't locked.
5061                  * we should start walking down the tree again if reference
5062                  * count is one.
5063                  */
5064                 if (!path->locks[level]) {
5065                         BUG_ON(level == 0);
5066                         btrfs_tree_lock(eb);
5067                         btrfs_set_lock_blocking_write(eb);
5068                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
5069
5070                         ret = btrfs_lookup_extent_info(trans, fs_info,
5071                                                        eb->start, level, 1,
5072                                                        &wc->refs[level],
5073                                                        &wc->flags[level]);
5074                         if (ret < 0) {
5075                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
5076                                 path->locks[level] = 0;
5077                                 return ret;
5078                         }
5079                         BUG_ON(wc->refs[level] == 0);
5080                         if (wc->refs[level] == 1) {
5081                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
5082                                 path->locks[level] = 0;
5083                                 return 1;
5084                         }
5085                 }
5086         }
5087
5088         /* wc->stage == DROP_REFERENCE */
5089         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
5090
5091         if (wc->refs[level] == 1) {
5092                 if (level == 0) {
5093                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
5094                                 ret = btrfs_dec_ref(trans, root, eb, 1);
5095                         else
5096                                 ret = btrfs_dec_ref(trans, root, eb, 0);
5097                         BUG_ON(ret); /* -ENOMEM */
5098                         if (is_fstree(root->root_key.objectid)) {
5099                                 ret = btrfs_qgroup_trace_leaf_items(trans, eb);
5100                                 if (ret) {
5101                                         btrfs_err_rl(fs_info,
5102         "error %d accounting leaf items, quota is out of sync, rescan required",
5103                                              ret);
5104                                 }
5105                         }
5106                 }
5107                 /* make block locked assertion in btrfs_clean_tree_block happy */
5108                 if (!path->locks[level] &&
5109                     btrfs_header_generation(eb) == trans->transid) {
5110                         btrfs_tree_lock(eb);
5111                         btrfs_set_lock_blocking_write(eb);
5112                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
5113                 }
5114                 btrfs_clean_tree_block(eb);
5115         }
5116
5117         if (eb == root->node) {
5118                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
5119                         parent = eb->start;
5120                 else if (root->root_key.objectid != btrfs_header_owner(eb))
5121                         goto owner_mismatch;
5122         } else {
5123                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
5124                         parent = path->nodes[level + 1]->start;
5125                 else if (root->root_key.objectid !=
5126                          btrfs_header_owner(path->nodes[level + 1]))
5127                         goto owner_mismatch;
5128         }
5129
5130         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
5131 out:
5132         wc->refs[level] = 0;
5133         wc->flags[level] = 0;
5134         return 0;
5135
5136 owner_mismatch:
5137         btrfs_err_rl(fs_info, "unexpected tree owner, have %llu expect %llu",
5138                      btrfs_header_owner(eb), root->root_key.objectid);
5139         return -EUCLEAN;
5140 }
5141
5142 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
5143                                    struct btrfs_root *root,
5144                                    struct btrfs_path *path,
5145                                    struct walk_control *wc)
5146 {
5147         int level = wc->level;
5148         int lookup_info = 1;
5149         int ret;
5150
5151         while (level >= 0) {
5152                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
5153                 if (ret > 0)
5154                         break;
5155
5156                 if (level == 0)
5157                         break;
5158
5159                 if (path->slots[level] >=
5160                     btrfs_header_nritems(path->nodes[level]))
5161                         break;
5162
5163                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
5164                 if (ret > 0) {
5165                         path->slots[level]++;
5166                         continue;
5167                 } else if (ret < 0)
5168                         return ret;
5169                 level = wc->level;
5170         }
5171         return 0;
5172 }
5173
5174 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
5175                                  struct btrfs_root *root,
5176                                  struct btrfs_path *path,
5177                                  struct walk_control *wc, int max_level)
5178 {
5179         int level = wc->level;
5180         int ret;
5181
5182         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
5183         while (level < max_level && path->nodes[level]) {
5184                 wc->level = level;
5185                 if (path->slots[level] + 1 <
5186                     btrfs_header_nritems(path->nodes[level])) {
5187                         path->slots[level]++;
5188                         return 0;
5189                 } else {
5190                         ret = walk_up_proc(trans, root, path, wc);
5191                         if (ret > 0)
5192                                 return 0;
5193                         if (ret < 0)
5194                                 return ret;
5195
5196                         if (path->locks[level]) {
5197                                 btrfs_tree_unlock_rw(path->nodes[level],
5198                                                      path->locks[level]);
5199                                 path->locks[level] = 0;
5200                         }
5201                         free_extent_buffer(path->nodes[level]);
5202                         path->nodes[level] = NULL;
5203                         level++;
5204                 }
5205         }
5206         return 1;
5207 }
5208
5209 /*
5210  * drop a subvolume tree.
5211  *
5212  * this function traverses the tree freeing any blocks that only
5213  * referenced by the tree.
5214  *
5215  * when a shared tree block is found. this function decreases its
5216  * reference count by one. if update_ref is true, this function
5217  * also make sure backrefs for the shared block and all lower level
5218  * blocks are properly updated.
5219  *
5220  * If called with for_reloc == 0, may exit early with -EAGAIN
5221  */
5222 int btrfs_drop_snapshot(struct btrfs_root *root,
5223                          struct btrfs_block_rsv *block_rsv, int update_ref,
5224                          int for_reloc)
5225 {
5226         struct btrfs_fs_info *fs_info = root->fs_info;
5227         struct btrfs_path *path;
5228         struct btrfs_trans_handle *trans;
5229         struct btrfs_root *tree_root = fs_info->tree_root;
5230         struct btrfs_root_item *root_item = &root->root_item;
5231         struct walk_control *wc;
5232         struct btrfs_key key;
5233         int err = 0;
5234         int ret;
5235         int level;
5236         bool root_dropped = false;
5237
5238         btrfs_debug(fs_info, "Drop subvolume %llu", root->root_key.objectid);
5239
5240         path = btrfs_alloc_path();
5241         if (!path) {
5242                 err = -ENOMEM;
5243                 goto out;
5244         }
5245
5246         wc = kzalloc(sizeof(*wc), GFP_NOFS);
5247         if (!wc) {
5248                 btrfs_free_path(path);
5249                 err = -ENOMEM;
5250                 goto out;
5251         }
5252
5253         trans = btrfs_start_transaction(tree_root, 0);
5254         if (IS_ERR(trans)) {
5255                 err = PTR_ERR(trans);
5256                 goto out_free;
5257         }
5258
5259         err = btrfs_run_delayed_items(trans);
5260         if (err)
5261                 goto out_end_trans;
5262
5263         if (block_rsv)
5264                 trans->block_rsv = block_rsv;
5265
5266         /*
5267          * This will help us catch people modifying the fs tree while we're
5268          * dropping it.  It is unsafe to mess with the fs tree while it's being
5269          * dropped as we unlock the root node and parent nodes as we walk down
5270          * the tree, assuming nothing will change.  If something does change
5271          * then we'll have stale information and drop references to blocks we've
5272          * already dropped.
5273          */
5274         set_bit(BTRFS_ROOT_DELETING, &root->state);
5275         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
5276                 level = btrfs_header_level(root->node);
5277                 path->nodes[level] = btrfs_lock_root_node(root);
5278                 btrfs_set_lock_blocking_write(path->nodes[level]);
5279                 path->slots[level] = 0;
5280                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
5281                 memset(&wc->update_progress, 0,
5282                        sizeof(wc->update_progress));
5283         } else {
5284                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
5285                 memcpy(&wc->update_progress, &key,
5286                        sizeof(wc->update_progress));
5287
5288                 level = root_item->drop_level;
5289                 BUG_ON(level == 0);
5290                 path->lowest_level = level;
5291                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5292                 path->lowest_level = 0;
5293                 if (ret < 0) {
5294                         err = ret;
5295                         goto out_end_trans;
5296                 }
5297                 WARN_ON(ret > 0);
5298
5299                 /*
5300                  * unlock our path, this is safe because only this
5301                  * function is allowed to delete this snapshot
5302                  */
5303                 btrfs_unlock_up_safe(path, 0);
5304
5305                 level = btrfs_header_level(root->node);
5306                 while (1) {
5307                         btrfs_tree_lock(path->nodes[level]);
5308                         btrfs_set_lock_blocking_write(path->nodes[level]);
5309                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
5310
5311                         ret = btrfs_lookup_extent_info(trans, fs_info,
5312                                                 path->nodes[level]->start,
5313                                                 level, 1, &wc->refs[level],
5314                                                 &wc->flags[level]);
5315                         if (ret < 0) {
5316                                 err = ret;
5317                                 goto out_end_trans;
5318                         }
5319                         BUG_ON(wc->refs[level] == 0);
5320
5321                         if (level == root_item->drop_level)
5322                                 break;
5323
5324                         btrfs_tree_unlock(path->nodes[level]);
5325                         path->locks[level] = 0;
5326                         WARN_ON(wc->refs[level] != 1);
5327                         level--;
5328                 }
5329         }
5330
5331         wc->restarted = test_bit(BTRFS_ROOT_DEAD_TREE, &root->state);
5332         wc->level = level;
5333         wc->shared_level = -1;
5334         wc->stage = DROP_REFERENCE;
5335         wc->update_ref = update_ref;
5336         wc->keep_locks = 0;
5337         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
5338
5339         while (1) {
5340
5341                 ret = walk_down_tree(trans, root, path, wc);
5342                 if (ret < 0) {
5343                         err = ret;
5344                         break;
5345                 }
5346
5347                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
5348                 if (ret < 0) {
5349                         err = ret;
5350                         break;
5351                 }
5352
5353                 if (ret > 0) {
5354                         BUG_ON(wc->stage != DROP_REFERENCE);
5355                         break;
5356                 }
5357
5358                 if (wc->stage == DROP_REFERENCE) {
5359                         wc->drop_level = wc->level;
5360                         btrfs_node_key_to_cpu(path->nodes[wc->drop_level],
5361                                               &wc->drop_progress,
5362                                               path->slots[wc->drop_level]);
5363                 }
5364                 btrfs_cpu_key_to_disk(&root_item->drop_progress,
5365                                       &wc->drop_progress);
5366                 root_item->drop_level = wc->drop_level;
5367
5368                 BUG_ON(wc->level == 0);
5369                 if (btrfs_should_end_transaction(trans) ||
5370                     (!for_reloc && btrfs_need_cleaner_sleep(fs_info))) {
5371                         ret = btrfs_update_root(trans, tree_root,
5372                                                 &root->root_key,
5373                                                 root_item);
5374                         if (ret) {
5375                                 btrfs_abort_transaction(trans, ret);
5376                                 err = ret;
5377                                 goto out_end_trans;
5378                         }
5379
5380                         btrfs_end_transaction_throttle(trans);
5381                         if (!for_reloc && btrfs_need_cleaner_sleep(fs_info)) {
5382                                 btrfs_debug(fs_info,
5383                                             "drop snapshot early exit");
5384                                 err = -EAGAIN;
5385                                 goto out_free;
5386                         }
5387
5388                         trans = btrfs_start_transaction(tree_root, 0);
5389                         if (IS_ERR(trans)) {
5390                                 err = PTR_ERR(trans);
5391                                 goto out_free;
5392                         }
5393                         if (block_rsv)
5394                                 trans->block_rsv = block_rsv;
5395                 }
5396         }
5397         btrfs_release_path(path);
5398         if (err)
5399                 goto out_end_trans;
5400
5401         ret = btrfs_del_root(trans, &root->root_key);
5402         if (ret) {
5403                 btrfs_abort_transaction(trans, ret);
5404                 err = ret;
5405                 goto out_end_trans;
5406         }
5407
5408         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
5409                 ret = btrfs_find_root(tree_root, &root->root_key, path,
5410                                       NULL, NULL);
5411                 if (ret < 0) {
5412                         btrfs_abort_transaction(trans, ret);
5413                         err = ret;
5414                         goto out_end_trans;
5415                 } else if (ret > 0) {
5416                         /* if we fail to delete the orphan item this time
5417                          * around, it'll get picked up the next time.
5418                          *
5419                          * The most common failure here is just -ENOENT.
5420                          */
5421                         btrfs_del_orphan_item(trans, tree_root,
5422                                               root->root_key.objectid);
5423                 }
5424         }
5425
5426         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
5427                 btrfs_add_dropped_root(trans, root);
5428         } else {
5429                 free_extent_buffer(root->node);
5430                 free_extent_buffer(root->commit_root);
5431                 btrfs_put_root(root);
5432         }
5433         root_dropped = true;
5434 out_end_trans:
5435         btrfs_end_transaction_throttle(trans);
5436 out_free:
5437         kfree(wc);
5438         btrfs_free_path(path);
5439 out:
5440         /*
5441          * So if we need to stop dropping the snapshot for whatever reason we
5442          * need to make sure to add it back to the dead root list so that we
5443          * keep trying to do the work later.  This also cleans up roots if we
5444          * don't have it in the radix (like when we recover after a power fail
5445          * or unmount) so we don't leak memory.
5446          */
5447         if (!for_reloc && !root_dropped)
5448                 btrfs_add_dead_root(root);
5449         if (err && err != -EAGAIN)
5450                 btrfs_handle_fs_error(fs_info, err, NULL);
5451         return err;
5452 }
5453
5454 /*
5455  * drop subtree rooted at tree block 'node'.
5456  *
5457  * NOTE: this function will unlock and release tree block 'node'
5458  * only used by relocation code
5459  */
5460 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
5461                         struct btrfs_root *root,
5462                         struct extent_buffer *node,
5463                         struct extent_buffer *parent)
5464 {
5465         struct btrfs_fs_info *fs_info = root->fs_info;
5466         struct btrfs_path *path;
5467         struct walk_control *wc;
5468         int level;
5469         int parent_level;
5470         int ret = 0;
5471         int wret;
5472
5473         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
5474
5475         path = btrfs_alloc_path();
5476         if (!path)
5477                 return -ENOMEM;
5478
5479         wc = kzalloc(sizeof(*wc), GFP_NOFS);
5480         if (!wc) {
5481                 btrfs_free_path(path);
5482                 return -ENOMEM;
5483         }
5484
5485         btrfs_assert_tree_locked(parent);
5486         parent_level = btrfs_header_level(parent);
5487         atomic_inc(&parent->refs);
5488         path->nodes[parent_level] = parent;
5489         path->slots[parent_level] = btrfs_header_nritems(parent);
5490
5491         btrfs_assert_tree_locked(node);
5492         level = btrfs_header_level(node);
5493         path->nodes[level] = node;
5494         path->slots[level] = 0;
5495         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
5496
5497         wc->refs[parent_level] = 1;
5498         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
5499         wc->level = level;
5500         wc->shared_level = -1;
5501         wc->stage = DROP_REFERENCE;
5502         wc->update_ref = 0;
5503         wc->keep_locks = 1;
5504         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
5505
5506         while (1) {
5507                 wret = walk_down_tree(trans, root, path, wc);
5508                 if (wret < 0) {
5509                         ret = wret;
5510                         break;
5511                 }
5512
5513                 wret = walk_up_tree(trans, root, path, wc, parent_level);
5514                 if (wret < 0)
5515                         ret = wret;
5516                 if (wret != 0)
5517                         break;
5518         }
5519
5520         kfree(wc);
5521         btrfs_free_path(path);
5522         return ret;
5523 }
5524
5525 /*
5526  * helper to account the unused space of all the readonly block group in the
5527  * space_info. takes mirrors into account.
5528  */
5529 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
5530 {
5531         struct btrfs_block_group *block_group;
5532         u64 free_bytes = 0;
5533         int factor;
5534
5535         /* It's df, we don't care if it's racy */
5536         if (list_empty(&sinfo->ro_bgs))
5537                 return 0;
5538
5539         spin_lock(&sinfo->lock);
5540         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
5541                 spin_lock(&block_group->lock);
5542
5543                 if (!block_group->ro) {
5544                         spin_unlock(&block_group->lock);
5545                         continue;
5546                 }
5547
5548                 factor = btrfs_bg_type_to_factor(block_group->flags);
5549                 free_bytes += (block_group->length -
5550                                block_group->used) * factor;
5551
5552                 spin_unlock(&block_group->lock);
5553         }
5554         spin_unlock(&sinfo->lock);
5555
5556         return free_bytes;
5557 }
5558
5559 int btrfs_error_unpin_extent_range(struct btrfs_fs_info *fs_info,
5560                                    u64 start, u64 end)
5561 {
5562         return unpin_extent_range(fs_info, start, end, false);
5563 }
5564
5565 /*
5566  * It used to be that old block groups would be left around forever.
5567  * Iterating over them would be enough to trim unused space.  Since we
5568  * now automatically remove them, we also need to iterate over unallocated
5569  * space.
5570  *
5571  * We don't want a transaction for this since the discard may take a
5572  * substantial amount of time.  We don't require that a transaction be
5573  * running, but we do need to take a running transaction into account
5574  * to ensure that we're not discarding chunks that were released or
5575  * allocated in the current transaction.
5576  *
5577  * Holding the chunks lock will prevent other threads from allocating
5578  * or releasing chunks, but it won't prevent a running transaction
5579  * from committing and releasing the memory that the pending chunks
5580  * list head uses.  For that, we need to take a reference to the
5581  * transaction and hold the commit root sem.  We only need to hold
5582  * it while performing the free space search since we have already
5583  * held back allocations.
5584  */
5585 static int btrfs_trim_free_extents(struct btrfs_device *device, u64 *trimmed)
5586 {
5587         u64 start = SZ_1M, len = 0, end = 0;
5588         int ret;
5589
5590         *trimmed = 0;
5591
5592         /* Discard not supported = nothing to do. */
5593         if (!blk_queue_discard(bdev_get_queue(device->bdev)))
5594                 return 0;
5595
5596         /* Not writable = nothing to do. */
5597         if (!test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state))
5598                 return 0;
5599
5600         /* No free space = nothing to do. */
5601         if (device->total_bytes <= device->bytes_used)
5602                 return 0;
5603
5604         ret = 0;
5605
5606         while (1) {
5607                 struct btrfs_fs_info *fs_info = device->fs_info;
5608                 u64 bytes;
5609
5610                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
5611                 if (ret)
5612                         break;
5613
5614                 find_first_clear_extent_bit(&device->alloc_state, start,
5615                                             &start, &end,
5616                                             CHUNK_TRIMMED | CHUNK_ALLOCATED);
5617
5618                 /* Ensure we skip the reserved area in the first 1M */
5619                 start = max_t(u64, start, SZ_1M);
5620
5621                 /*
5622                  * If find_first_clear_extent_bit find a range that spans the
5623                  * end of the device it will set end to -1, in this case it's up
5624                  * to the caller to trim the value to the size of the device.
5625                  */
5626                 end = min(end, device->total_bytes - 1);
5627
5628                 len = end - start + 1;
5629
5630                 /* We didn't find any extents */
5631                 if (!len) {
5632                         mutex_unlock(&fs_info->chunk_mutex);
5633                         ret = 0;
5634                         break;
5635                 }
5636
5637                 ret = btrfs_issue_discard(device->bdev, start, len,
5638                                           &bytes);
5639                 if (!ret)
5640                         set_extent_bits(&device->alloc_state, start,
5641                                         start + bytes - 1,
5642                                         CHUNK_TRIMMED);
5643                 mutex_unlock(&fs_info->chunk_mutex);
5644
5645                 if (ret)
5646                         break;
5647
5648                 start += len;
5649                 *trimmed += bytes;
5650
5651                 if (fatal_signal_pending(current)) {
5652                         ret = -ERESTARTSYS;
5653                         break;
5654                 }
5655
5656                 cond_resched();
5657         }
5658
5659         return ret;
5660 }
5661
5662 /*
5663  * Trim the whole filesystem by:
5664  * 1) trimming the free space in each block group
5665  * 2) trimming the unallocated space on each device
5666  *
5667  * This will also continue trimming even if a block group or device encounters
5668  * an error.  The return value will be the last error, or 0 if nothing bad
5669  * happens.
5670  */
5671 int btrfs_trim_fs(struct btrfs_fs_info *fs_info, struct fstrim_range *range)
5672 {
5673         struct btrfs_block_group *cache = NULL;
5674         struct btrfs_device *device;
5675         struct list_head *devices;
5676         u64 group_trimmed;
5677         u64 range_end = U64_MAX;
5678         u64 start;
5679         u64 end;
5680         u64 trimmed = 0;
5681         u64 bg_failed = 0;
5682         u64 dev_failed = 0;
5683         int bg_ret = 0;
5684         int dev_ret = 0;
5685         int ret = 0;
5686
5687         /*
5688          * Check range overflow if range->len is set.
5689          * The default range->len is U64_MAX.
5690          */
5691         if (range->len != U64_MAX &&
5692             check_add_overflow(range->start, range->len, &range_end))
5693                 return -EINVAL;
5694
5695         cache = btrfs_lookup_first_block_group(fs_info, range->start);
5696         for (; cache; cache = btrfs_next_block_group(cache)) {
5697                 if (cache->start >= range_end) {
5698                         btrfs_put_block_group(cache);
5699                         break;
5700                 }
5701
5702                 start = max(range->start, cache->start);
5703                 end = min(range_end, cache->start + cache->length);
5704
5705                 if (end - start >= range->minlen) {
5706                         if (!btrfs_block_group_done(cache)) {
5707                                 ret = btrfs_cache_block_group(cache, 0);
5708                                 if (ret) {
5709                                         bg_failed++;
5710                                         bg_ret = ret;
5711                                         continue;
5712                                 }
5713                                 ret = btrfs_wait_block_group_cache_done(cache);
5714                                 if (ret) {
5715                                         bg_failed++;
5716                                         bg_ret = ret;
5717                                         continue;
5718                                 }
5719                         }
5720                         ret = btrfs_trim_block_group(cache,
5721                                                      &group_trimmed,
5722                                                      start,
5723                                                      end,
5724                                                      range->minlen);
5725
5726                         trimmed += group_trimmed;
5727                         if (ret) {
5728                                 bg_failed++;
5729                                 bg_ret = ret;
5730                                 continue;
5731                         }
5732                 }
5733         }
5734
5735         if (bg_failed)
5736                 btrfs_warn(fs_info,
5737                         "failed to trim %llu block group(s), last error %d",
5738                         bg_failed, bg_ret);
5739         mutex_lock(&fs_info->fs_devices->device_list_mutex);
5740         devices = &fs_info->fs_devices->devices;
5741         list_for_each_entry(device, devices, dev_list) {
5742                 ret = btrfs_trim_free_extents(device, &group_trimmed);
5743                 if (ret) {
5744                         dev_failed++;
5745                         dev_ret = ret;
5746                         break;
5747                 }
5748
5749                 trimmed += group_trimmed;
5750         }
5751         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
5752
5753         if (dev_failed)
5754                 btrfs_warn(fs_info,
5755                         "failed to trim %llu device(s), last error %d",
5756                         dev_failed, dev_ret);
5757         range->len = trimmed;
5758         if (bg_ret)
5759                 return bg_ret;
5760         return dev_ret;
5761 }