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