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 "ctree.h"
  20 #include "extent-tree.h"
  21 #include "tree-log.h"
  22 #include "disk-io.h"
  23 #include "print-tree.h"
  24 #include "volumes.h"
  25 #include "raid56.h"
  26 #include "locking.h"
  27 #include "free-space-cache.h"
  28 #include "free-space-tree.h"
  29 #include "sysfs.h"
  30 #include "qgroup.h"
  31 #include "ref-verify.h"
  32 #include "space-info.h"
  33 #include "block-rsv.h"
  34 #include "delalloc-space.h"
  35 #include "discard.h"
  36 #include "rcu-string.h"
  37 #include "zoned.h"
  38 #include "dev-replace.h"
  39 #include "fs.h"
  40 #include "accessors.h"
  41 #include "root-tree.h"
  42 #include "file-item.h"
  43 #include "orphan.h"
  44 #include "tree-checker.h"
  45
  46 #undef SCRAMBLE_DELAYED_REFS
  47
  48
  49 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  50                                struct btrfs_delayed_ref_node *node, u64 parent,
  51                                u64 root_objectid, u64 owner_objectid,
  52                                u64 owner_offset,
  53                                struct btrfs_delayed_extent_op *extra_op);
  54 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  55                                     struct extent_buffer *leaf,
  56                                     struct btrfs_extent_item *ei);
  57 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  58                                       u64 parent, u64 root_objectid,
  59                                       u64 flags, u64 owner, u64 offset,
  60                                       struct btrfs_key *ins, int ref_mod);
  61 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  62                                      struct btrfs_delayed_ref_node *node,
  63                                      struct btrfs_delayed_extent_op *extent_op);
  64 static int find_next_key(struct btrfs_path *path, int level,
  65                          struct btrfs_key *key);
  66
  67 static int block_group_bits(struct btrfs_block_group *cache, u64 bits)
  68 {
  69         return (cache->flags & bits) == bits;
  70 }
  71
  72 /* simple helper to search for an existing data extent at a given offset */
  73 int btrfs_lookup_data_extent(struct btrfs_fs_info *fs_info, u64 start, u64 len)
  74 {
  75         struct btrfs_root *root = btrfs_extent_root(fs_info, start);
  76         int ret;
  77         struct btrfs_key key;
  78         struct btrfs_path *path;
  79
  80         path = btrfs_alloc_path();
  81         if (!path)
  82                 return -ENOMEM;
  83
  84         key.objectid = start;
  85         key.offset = len;
  86         key.type = BTRFS_EXTENT_ITEM_KEY;
  87         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
  88         btrfs_free_path(path);
  89         return ret;
  90 }
  91
  92 /*
  93  * helper function to lookup reference count and flags of a tree block.
  94  *
  95  * the head node for delayed ref is used to store the sum of all the
  96  * reference count modifications queued up in the rbtree. the head
  97  * node may also store the extent flags to set. This way you can check
  98  * to see what the reference count and extent flags would be if all of
  99  * the delayed refs are not processed.
 100  */
 101 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 102                              struct btrfs_fs_info *fs_info, u64 bytenr,
 103                              u64 offset, int metadata, u64 *refs, u64 *flags)
 104 {
 105         struct btrfs_root *extent_root;
 106         struct btrfs_delayed_ref_head *head;
 107         struct btrfs_delayed_ref_root *delayed_refs;
 108         struct btrfs_path *path;
 109         struct btrfs_extent_item *ei;
 110         struct extent_buffer *leaf;
 111         struct btrfs_key key;
 112         u32 item_size;
 113         u64 num_refs;
 114         u64 extent_flags;
 115         int ret;
 116
 117         /*
 118          * If we don't have skinny metadata, don't bother doing anything
 119          * different
 120          */
 121         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA)) {
 122                 offset = fs_info->nodesize;
 123                 metadata = 0;
 124         }
 125
 126         path = btrfs_alloc_path();
 127         if (!path)
 128                 return -ENOMEM;
 129
 130         if (!trans) {
 131                 path->skip_locking = 1;
 132                 path->search_commit_root = 1;
 133         }
 134
 135 search_again:
 136         key.objectid = bytenr;
 137         key.offset = offset;
 138         if (metadata)
 139                 key.type = BTRFS_METADATA_ITEM_KEY;
 140         else
 141                 key.type = BTRFS_EXTENT_ITEM_KEY;
 142
 143         extent_root = btrfs_extent_root(fs_info, bytenr);
 144         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 145         if (ret < 0)
 146                 goto out_free;
 147
 148         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 149                 if (path->slots[0]) {
 150                         path->slots[0]--;
 151                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 152                                               path->slots[0]);
 153                         if (key.objectid == bytenr &&
 154                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 155                             key.offset == fs_info->nodesize)
 156                                 ret = 0;
 157                 }
 158         }
 159
 160         if (ret == 0) {
 161                 leaf = path->nodes[0];
 162                 item_size = btrfs_item_size(leaf, path->slots[0]);
 163                 if (item_size >= sizeof(*ei)) {
 164                         ei = btrfs_item_ptr(leaf, path->slots[0],
 165                                             struct btrfs_extent_item);
 166                         num_refs = btrfs_extent_refs(leaf, ei);
 167                         extent_flags = btrfs_extent_flags(leaf, ei);
 168                 } else {
 169                         ret = -EUCLEAN;
 170                         btrfs_err(fs_info,
 171                         "unexpected extent item size, has %u expect >= %zu",
 172                                   item_size, sizeof(*ei));
 173                         if (trans)
 174                                 btrfs_abort_transaction(trans, ret);
 175                         else
 176                                 btrfs_handle_fs_error(fs_info, ret, NULL);
 177
 178                         goto out_free;
 179                 }
 180
 181                 BUG_ON(num_refs == 0);
 182         } else {
 183                 num_refs = 0;
 184                 extent_flags = 0;
 185                 ret = 0;
 186         }
 187
 188         if (!trans)
 189                 goto out;
 190
 191         delayed_refs = &trans->transaction->delayed_refs;
 192         spin_lock(&delayed_refs->lock);
 193         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
 194         if (head) {
 195                 if (!mutex_trylock(&head->mutex)) {
 196                         refcount_inc(&head->refs);
 197                         spin_unlock(&delayed_refs->lock);
 198
 199                         btrfs_release_path(path);
 200
 201                         /*
 202                          * Mutex was contended, block until it's released and try
 203                          * again
 204                          */
 205                         mutex_lock(&head->mutex);
 206                         mutex_unlock(&head->mutex);
 207                         btrfs_put_delayed_ref_head(head);
 208                         goto search_again;
 209                 }
 210                 spin_lock(&head->lock);
 211                 if (head->extent_op && head->extent_op->update_flags)
 212                         extent_flags |= head->extent_op->flags_to_set;
 213                 else
 214                         BUG_ON(num_refs == 0);
 215
 216                 num_refs += head->ref_mod;
 217                 spin_unlock(&head->lock);
 218                 mutex_unlock(&head->mutex);
 219         }
 220         spin_unlock(&delayed_refs->lock);
 221 out:
 222         WARN_ON(num_refs == 0);
 223         if (refs)
 224                 *refs = num_refs;
 225         if (flags)
 226                 *flags = extent_flags;
 227 out_free:
 228         btrfs_free_path(path);
 229         return ret;
 230 }
 231
 232 /*
 233  * Back reference rules.  Back refs have three main goals:
 234  *
 235  * 1) differentiate between all holders of references to an extent so that
 236  *    when a reference is dropped we can make sure it was a valid reference
 237  *    before freeing the extent.
 238  *
 239  * 2) Provide enough information to quickly find the holders of an extent
 240  *    if we notice a given block is corrupted or bad.
 241  *
 242  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 243  *    maintenance.  This is actually the same as #2, but with a slightly
 244  *    different use case.
 245  *
 246  * There are two kinds of back refs. The implicit back refs is optimized
 247  * for pointers in non-shared tree blocks. For a given pointer in a block,
 248  * back refs of this kind provide information about the block's owner tree
 249  * and the pointer's key. These information allow us to find the block by
 250  * b-tree searching. The full back refs is for pointers in tree blocks not
 251  * referenced by their owner trees. The location of tree block is recorded
 252  * in the back refs. Actually the full back refs is generic, and can be
 253  * used in all cases the implicit back refs is used. The major shortcoming
 254  * of the full back refs is its overhead. Every time a tree block gets
 255  * COWed, we have to update back refs entry for all pointers in it.
 256  *
 257  * For a newly allocated tree block, we use implicit back refs for
 258  * pointers in it. This means most tree related operations only involve
 259  * implicit back refs. For a tree block created in old transaction, the
 260  * only way to drop a reference to it is COW it. So we can detect the
 261  * event that tree block loses its owner tree's reference and do the
 262  * back refs conversion.
 263  *
 264  * When a tree block is COWed through a tree, there are four cases:
 265  *
 266  * The reference count of the block is one and the tree is the block's
 267  * owner tree. Nothing to do in this case.
 268  *
 269  * The reference count of the block is one and the tree is not the
 270  * block's owner tree. In this case, full back refs is used for pointers
 271  * in the block. Remove these full back refs, add implicit back refs for
 272  * every pointers in the new block.
 273  *
 274  * The reference count of the block is greater than one and the tree is
 275  * the block's owner tree. In this case, implicit back refs is used for
 276  * pointers in the block. Add full back refs for every pointers in the
 277  * block, increase lower level extents' reference counts. The original
 278  * implicit back refs are entailed to the new block.
 279  *
 280  * The reference count of the block is greater than one and the tree is
 281  * not the block's owner tree. Add implicit back refs for every pointer in
 282  * the new block, increase lower level extents' reference count.
 283  *
 284  * Back Reference Key composing:
 285  *
 286  * The key objectid corresponds to the first byte in the extent,
 287  * The key type is used to differentiate between types of back refs.
 288  * There are different meanings of the key offset for different types
 289  * of back refs.
 290  *
 291  * File extents can be referenced by:
 292  *
 293  * - multiple snapshots, subvolumes, or different generations in one subvol
 294  * - different files inside a single subvolume
 295  * - different offsets inside a file (bookend extents in file.c)
 296  *
 297  * The extent ref structure for the implicit back refs has fields for:
 298  *
 299  * - Objectid of the subvolume root
 300  * - objectid of the file holding the reference
 301  * - original offset in the file
 302  * - how many bookend extents
 303  *
 304  * The key offset for the implicit back refs is hash of the first
 305  * three fields.
 306  *
 307  * The extent ref structure for the full back refs has field for:
 308  *
 309  * - number of pointers in the tree leaf
 310  *
 311  * The key offset for the implicit back refs is the first byte of
 312  * the tree leaf
 313  *
 314  * When a file extent is allocated, The implicit back refs is used.
 315  * the fields are filled in:
 316  *
 317  *     (root_key.objectid, inode objectid, offset in file, 1)
 318  *
 319  * When a file extent is removed file truncation, we find the
 320  * corresponding implicit back refs and check the following fields:
 321  *
 322  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
 323  *
 324  * Btree extents can be referenced by:
 325  *
 326  * - Different subvolumes
 327  *
 328  * Both the implicit back refs and the full back refs for tree blocks
 329  * only consist of key. The key offset for the implicit back refs is
 330  * objectid of block's owner tree. The key offset for the full back refs
 331  * is the first byte of parent block.
 332  *
 333  * When implicit back refs is used, information about the lowest key and
 334  * level of the tree block are required. These information are stored in
 335  * tree block info structure.
 336  */
 337
 338 /*
 339  * is_data == BTRFS_REF_TYPE_BLOCK, tree block type is required,
 340  * is_data == BTRFS_REF_TYPE_DATA, data type is requiried,
 341  * is_data == BTRFS_REF_TYPE_ANY, either type is OK.
 342  */
 343 int btrfs_get_extent_inline_ref_type(const struct extent_buffer *eb,
 344                                      struct btrfs_extent_inline_ref *iref,
 345                                      enum btrfs_inline_ref_type is_data)
 346 {
 347         int type = btrfs_extent_inline_ref_type(eb, iref);
 348         u64 offset = btrfs_extent_inline_ref_offset(eb, iref);
 349
 350         if (type == BTRFS_TREE_BLOCK_REF_KEY ||
 351             type == BTRFS_SHARED_BLOCK_REF_KEY ||
 352             type == BTRFS_SHARED_DATA_REF_KEY ||
 353             type == BTRFS_EXTENT_DATA_REF_KEY) {
 354                 if (is_data == BTRFS_REF_TYPE_BLOCK) {
 355                         if (type == BTRFS_TREE_BLOCK_REF_KEY)
 356                                 return type;
 357                         if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
 358                                 ASSERT(eb->fs_info);
 359                                 /*
 360                                  * Every shared one has parent tree block,
 361                                  * which must be aligned to sector size.
 362                                  */
 363                                 if (offset &&
 364                                     IS_ALIGNED(offset, eb->fs_info->sectorsize))
 365                                         return type;
 366                         }
 367                 } else if (is_data == BTRFS_REF_TYPE_DATA) {
 368                         if (type == BTRFS_EXTENT_DATA_REF_KEY)
 369                                 return type;
 370                         if (type == BTRFS_SHARED_DATA_REF_KEY) {
 371                                 ASSERT(eb->fs_info);
 372                                 /*
 373                                  * Every shared one has parent tree block,
 374                                  * which must be aligned to sector size.
 375                                  */
 376                                 if (offset &&
 377                                     IS_ALIGNED(offset, eb->fs_info->sectorsize))
 378                                         return type;
 379                         }
 380                 } else {
 381                         ASSERT(is_data == BTRFS_REF_TYPE_ANY);
 382                         return type;
 383                 }
 384         }
 385
 386         WARN_ON(1);
 387         btrfs_print_leaf(eb);
 388         btrfs_err(eb->fs_info,
 389                   "eb %llu iref 0x%lx invalid extent inline ref type %d",
 390                   eb->start, (unsigned long)iref, type);
 391
 392         return BTRFS_REF_TYPE_INVALID;
 393 }
 394
 395 u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
 396 {
 397         u32 high_crc = ~(u32)0;
 398         u32 low_crc = ~(u32)0;
 399         __le64 lenum;
 400
 401         lenum = cpu_to_le64(root_objectid);
 402         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
 403         lenum = cpu_to_le64(owner);
 404         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
 405         lenum = cpu_to_le64(offset);
 406         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
 407
 408         return ((u64)high_crc << 31) ^ (u64)low_crc;
 409 }
 410
 411 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
 412                                      struct btrfs_extent_data_ref *ref)
 413 {
 414         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
 415                                     btrfs_extent_data_ref_objectid(leaf, ref),
 416                                     btrfs_extent_data_ref_offset(leaf, ref));
 417 }
 418
 419 static int match_extent_data_ref(struct extent_buffer *leaf,
 420                                  struct btrfs_extent_data_ref *ref,
 421                                  u64 root_objectid, u64 owner, u64 offset)
 422 {
 423         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
 424             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
 425             btrfs_extent_data_ref_offset(leaf, ref) != offset)
 426                 return 0;
 427         return 1;
 428 }
 429
 430 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
 431                                            struct btrfs_path *path,
 432                                            u64 bytenr, u64 parent,
 433                                            u64 root_objectid,
 434                                            u64 owner, u64 offset)
 435 {
 436         struct btrfs_root *root = btrfs_extent_root(trans->fs_info, bytenr);
 437         struct btrfs_key key;
 438         struct btrfs_extent_data_ref *ref;
 439         struct extent_buffer *leaf;
 440         u32 nritems;
 441         int ret;
 442         int recow;
 443         int err = -ENOENT;
 444
 445         key.objectid = bytenr;
 446         if (parent) {
 447                 key.type = BTRFS_SHARED_DATA_REF_KEY;
 448                 key.offset = parent;
 449         } else {
 450                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
 451                 key.offset = hash_extent_data_ref(root_objectid,
 452                                                   owner, offset);
 453         }
 454 again:
 455         recow = 0;
 456         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
 457         if (ret < 0) {
 458                 err = ret;
 459                 goto fail;
 460         }
 461
 462         if (parent) {
 463                 if (!ret)
 464                         return 0;
 465                 goto fail;
 466         }
 467
 468         leaf = path->nodes[0];
 469         nritems = btrfs_header_nritems(leaf);
 470         while (1) {
 471                 if (path->slots[0] >= nritems) {
 472                         ret = btrfs_next_leaf(root, path);
 473                         if (ret < 0)
 474                                 err = ret;
 475                         if (ret)
 476                                 goto fail;
 477
 478                         leaf = path->nodes[0];
 479                         nritems = btrfs_header_nritems(leaf);
 480                         recow = 1;
 481                 }
 482
 483                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 484                 if (key.objectid != bytenr ||
 485                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
 486                         goto fail;
 487
 488                 ref = btrfs_item_ptr(leaf, path->slots[0],
 489                                      struct btrfs_extent_data_ref);
 490
 491                 if (match_extent_data_ref(leaf, ref, root_objectid,
 492                                           owner, offset)) {
 493                         if (recow) {
 494                                 btrfs_release_path(path);
 495                                 goto again;
 496                         }
 497                         err = 0;
 498                         break;
 499                 }
 500                 path->slots[0]++;
 501         }
 502 fail:
 503         return err;
 504 }
 505
 506 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
 507                                            struct btrfs_path *path,
 508                                            u64 bytenr, u64 parent,
 509                                            u64 root_objectid, u64 owner,
 510                                            u64 offset, int refs_to_add)
 511 {
 512         struct btrfs_root *root = btrfs_extent_root(trans->fs_info, bytenr);
 513         struct btrfs_key key;
 514         struct extent_buffer *leaf;
 515         u32 size;
 516         u32 num_refs;
 517         int ret;
 518
 519         key.objectid = bytenr;
 520         if (parent) {
 521                 key.type = BTRFS_SHARED_DATA_REF_KEY;
 522                 key.offset = parent;
 523                 size = sizeof(struct btrfs_shared_data_ref);
 524         } else {
 525                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
 526                 key.offset = hash_extent_data_ref(root_objectid,
 527                                                   owner, offset);
 528                 size = sizeof(struct btrfs_extent_data_ref);
 529         }
 530
 531         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
 532         if (ret && ret != -EEXIST)
 533                 goto fail;
 534
 535         leaf = path->nodes[0];
 536         if (parent) {
 537                 struct btrfs_shared_data_ref *ref;
 538                 ref = btrfs_item_ptr(leaf, path->slots[0],
 539                                      struct btrfs_shared_data_ref);
 540                 if (ret == 0) {
 541                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
 542                 } else {
 543                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
 544                         num_refs += refs_to_add;
 545                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
 546                 }
 547         } else {
 548                 struct btrfs_extent_data_ref *ref;
 549                 while (ret == -EEXIST) {
 550                         ref = btrfs_item_ptr(leaf, path->slots[0],
 551                                              struct btrfs_extent_data_ref);
 552                         if (match_extent_data_ref(leaf, ref, root_objectid,
 553                                                   owner, offset))
 554                                 break;
 555                         btrfs_release_path(path);
 556                         key.offset++;
 557                         ret = btrfs_insert_empty_item(trans, root, path, &key,
 558                                                       size);
 559                         if (ret && ret != -EEXIST)
 560                                 goto fail;
 561
 562                         leaf = path->nodes[0];
 563                 }
 564                 ref = btrfs_item_ptr(leaf, path->slots[0],
 565                                      struct btrfs_extent_data_ref);
 566                 if (ret == 0) {
 567                         btrfs_set_extent_data_ref_root(leaf, ref,
 568                                                        root_objectid);
 569                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
 570                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
 571                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
 572                 } else {
 573                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
 574                         num_refs += refs_to_add;
 575                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
 576                 }
 577         }
 578         btrfs_mark_buffer_dirty(trans, leaf);
 579         ret = 0;
 580 fail:
 581         btrfs_release_path(path);
 582         return ret;
 583 }
 584
 585 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
 586                                            struct btrfs_root *root,
 587                                            struct btrfs_path *path,
 588                                            int refs_to_drop)
 589 {
 590         struct btrfs_key key;
 591         struct btrfs_extent_data_ref *ref1 = NULL;
 592         struct btrfs_shared_data_ref *ref2 = NULL;
 593         struct extent_buffer *leaf;
 594         u32 num_refs = 0;
 595         int ret = 0;
 596
 597         leaf = path->nodes[0];
 598         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 599
 600         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
 601                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
 602                                       struct btrfs_extent_data_ref);
 603                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
 604         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
 605                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
 606                                       struct btrfs_shared_data_ref);
 607                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
 608         } else {
 609                 btrfs_err(trans->fs_info,
 610                           "unrecognized backref key (%llu %u %llu)",
 611                           key.objectid, key.type, key.offset);
 612                 btrfs_abort_transaction(trans, -EUCLEAN);
 613                 return -EUCLEAN;
 614         }
 615
 616         BUG_ON(num_refs < refs_to_drop);
 617         num_refs -= refs_to_drop;
 618
 619         if (num_refs == 0) {
 620                 ret = btrfs_del_item(trans, root, path);
 621         } else {
 622                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
 623                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
 624                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
 625                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
 626                 btrfs_mark_buffer_dirty(trans, leaf);
 627         }
 628         return ret;
 629 }
 630
 631 static noinline u32 extent_data_ref_count(struct btrfs_path *path,
 632                                           struct btrfs_extent_inline_ref *iref)
 633 {
 634         struct btrfs_key key;
 635         struct extent_buffer *leaf;
 636         struct btrfs_extent_data_ref *ref1;
 637         struct btrfs_shared_data_ref *ref2;
 638         u32 num_refs = 0;
 639         int type;
 640
 641         leaf = path->nodes[0];
 642         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 643
 644         if (iref) {
 645                 /*
 646                  * If type is invalid, we should have bailed out earlier than
 647                  * this call.
 648                  */
 649                 type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
 650                 ASSERT(type != BTRFS_REF_TYPE_INVALID);
 651                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
 652                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
 653                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
 654                 } else {
 655                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
 656                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
 657                 }
 658         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
 659                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
 660                                       struct btrfs_extent_data_ref);
 661                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
 662         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
 663                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
 664                                       struct btrfs_shared_data_ref);
 665                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
 666         } else {
 667                 WARN_ON(1);
 668         }
 669         return num_refs;
 670 }
 671
 672 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
 673                                           struct btrfs_path *path,
 674                                           u64 bytenr, u64 parent,
 675                                           u64 root_objectid)
 676 {
 677         struct btrfs_root *root = btrfs_extent_root(trans->fs_info, bytenr);
 678         struct btrfs_key key;
 679         int ret;
 680
 681         key.objectid = bytenr;
 682         if (parent) {
 683                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
 684                 key.offset = parent;
 685         } else {
 686                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
 687                 key.offset = root_objectid;
 688         }
 689
 690         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
 691         if (ret > 0)
 692                 ret = -ENOENT;
 693         return ret;
 694 }
 695
 696 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
 697                                           struct btrfs_path *path,
 698                                           u64 bytenr, u64 parent,
 699                                           u64 root_objectid)
 700 {
 701         struct btrfs_root *root = btrfs_extent_root(trans->fs_info, bytenr);
 702         struct btrfs_key key;
 703         int ret;
 704
 705         key.objectid = bytenr;
 706         if (parent) {
 707                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
 708                 key.offset = parent;
 709         } else {
 710                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
 711                 key.offset = root_objectid;
 712         }
 713
 714         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
 715         btrfs_release_path(path);
 716         return ret;
 717 }
 718
 719 static inline int extent_ref_type(u64 parent, u64 owner)
 720 {
 721         int type;
 722         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
 723                 if (parent > 0)
 724                         type = BTRFS_SHARED_BLOCK_REF_KEY;
 725                 else
 726                         type = BTRFS_TREE_BLOCK_REF_KEY;
 727         } else {
 728                 if (parent > 0)
 729                         type = BTRFS_SHARED_DATA_REF_KEY;
 730                 else
 731                         type = BTRFS_EXTENT_DATA_REF_KEY;
 732         }
 733         return type;
 734 }
 735
 736 static int find_next_key(struct btrfs_path *path, int level,
 737                          struct btrfs_key *key)
 738
 739 {
 740         for (; level < BTRFS_MAX_LEVEL; level++) {
 741                 if (!path->nodes[level])
 742                         break;
 743                 if (path->slots[level] + 1 >=
 744                     btrfs_header_nritems(path->nodes[level]))
 745                         continue;
 746                 if (level == 0)
 747                         btrfs_item_key_to_cpu(path->nodes[level], key,
 748                                               path->slots[level] + 1);
 749                 else
 750                         btrfs_node_key_to_cpu(path->nodes[level], key,
 751                                               path->slots[level] + 1);
 752                 return 0;
 753         }
 754         return 1;
 755 }
 756
 757 /*
 758  * look for inline back ref. if back ref is found, *ref_ret is set
 759  * to the address of inline back ref, and 0 is returned.
 760  *
 761  * if back ref isn't found, *ref_ret is set to the address where it
 762  * should be inserted, and -ENOENT is returned.
 763  *
 764  * if insert is true and there are too many inline back refs, the path
 765  * points to the extent item, and -EAGAIN is returned.
 766  *
 767  * NOTE: inline back refs are ordered in the same way that back ref
 768  *       items in the tree are ordered.
 769  */
 770 static noinline_for_stack
 771 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
 772                                  struct btrfs_path *path,
 773                                  struct btrfs_extent_inline_ref **ref_ret,
 774                                  u64 bytenr, u64 num_bytes,
 775                                  u64 parent, u64 root_objectid,
 776                                  u64 owner, u64 offset, int insert)
 777 {
 778         struct btrfs_fs_info *fs_info = trans->fs_info;
 779         struct btrfs_root *root = btrfs_extent_root(fs_info, bytenr);
 780         struct btrfs_key key;
 781         struct extent_buffer *leaf;
 782         struct btrfs_extent_item *ei;
 783         struct btrfs_extent_inline_ref *iref;
 784         u64 flags;
 785         u64 item_size;
 786         unsigned long ptr;
 787         unsigned long end;
 788         int extra_size;
 789         int type;
 790         int want;
 791         int ret;
 792         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
 793         int needed;
 794
 795         key.objectid = bytenr;
 796         key.type = BTRFS_EXTENT_ITEM_KEY;
 797         key.offset = num_bytes;
 798
 799         want = extent_ref_type(parent, owner);
 800         if (insert) {
 801                 extra_size = btrfs_extent_inline_ref_size(want);
 802                 path->search_for_extension = 1;
 803                 path->keep_locks = 1;
 804         } else
 805                 extra_size = -1;
 806
 807         /*
 808          * Owner is our level, so we can just add one to get the level for the
 809          * block we are interested in.
 810          */
 811         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
 812                 key.type = BTRFS_METADATA_ITEM_KEY;
 813                 key.offset = owner;
 814         }
 815
 816 again:
 817         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
 818         if (ret < 0)
 819                 goto out;
 820
 821         /*
 822          * We may be a newly converted file system which still has the old fat
 823          * extent entries for metadata, so try and see if we have one of those.
 824          */
 825         if (ret > 0 && skinny_metadata) {
 826                 skinny_metadata = false;
 827                 if (path->slots[0]) {
 828                         path->slots[0]--;
 829                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 830                                               path->slots[0]);
 831                         if (key.objectid == bytenr &&
 832                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 833                             key.offset == num_bytes)
 834                                 ret = 0;
 835                 }
 836                 if (ret) {
 837                         key.objectid = bytenr;
 838                         key.type = BTRFS_EXTENT_ITEM_KEY;
 839                         key.offset = num_bytes;
 840                         btrfs_release_path(path);
 841                         goto again;
 842                 }
 843         }
 844
 845         if (ret && !insert) {
 846                 ret = -ENOENT;
 847                 goto out;
 848         } else if (WARN_ON(ret)) {
 849                 btrfs_print_leaf(path->nodes[0]);
 850                 btrfs_err(fs_info,
 851 "extent item not found for insert, bytenr %llu num_bytes %llu parent %llu root_objectid %llu owner %llu offset %llu",
 852                           bytenr, num_bytes, parent, root_objectid, owner,
 853                           offset);
 854                 ret = -EUCLEAN;
 855                 goto out;
 856         }
 857
 858         leaf = path->nodes[0];
 859         item_size = btrfs_item_size(leaf, path->slots[0]);
 860         if (unlikely(item_size < sizeof(*ei))) {
 861                 ret = -EUCLEAN;
 862                 btrfs_err(fs_info,
 863                           "unexpected extent item size, has %llu expect >= %zu",
 864                           item_size, sizeof(*ei));
 865                 btrfs_abort_transaction(trans, ret);
 866                 goto out;
 867         }
 868
 869         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
 870         flags = btrfs_extent_flags(leaf, ei);
 871
 872         ptr = (unsigned long)(ei + 1);
 873         end = (unsigned long)ei + item_size;
 874
 875         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
 876                 ptr += sizeof(struct btrfs_tree_block_info);
 877                 BUG_ON(ptr > end);
 878         }
 879
 880         if (owner >= BTRFS_FIRST_FREE_OBJECTID)
 881                 needed = BTRFS_REF_TYPE_DATA;
 882         else
 883                 needed = BTRFS_REF_TYPE_BLOCK;
 884
 885         ret = -ENOENT;
 886         while (ptr < end) {
 887                 iref = (struct btrfs_extent_inline_ref *)ptr;
 888                 type = btrfs_get_extent_inline_ref_type(leaf, iref, needed);
 889                 if (type == BTRFS_REF_TYPE_INVALID) {
 890                         ret = -EUCLEAN;
 891                         goto out;
 892                 }
 893
 894                 if (want < type)
 895                         break;
 896                 if (want > type) {
 897                         ptr += btrfs_extent_inline_ref_size(type);
 898                         continue;
 899                 }
 900
 901                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
 902                         struct btrfs_extent_data_ref *dref;
 903                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
 904                         if (match_extent_data_ref(leaf, dref, root_objectid,
 905                                                   owner, offset)) {
 906                                 ret = 0;
 907                                 break;
 908                         }
 909                         if (hash_extent_data_ref_item(leaf, dref) <
 910                             hash_extent_data_ref(root_objectid, owner, offset))
 911                                 break;
 912                 } else {
 913                         u64 ref_offset;
 914                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
 915                         if (parent > 0) {
 916                                 if (parent == ref_offset) {
 917                                         ret = 0;
 918                                         break;
 919                                 }
 920                                 if (ref_offset < parent)
 921                                         break;
 922                         } else {
 923                                 if (root_objectid == ref_offset) {
 924                                         ret = 0;
 925                                         break;
 926                                 }
 927                                 if (ref_offset < root_objectid)
 928                                         break;
 929                         }
 930                 }
 931                 ptr += btrfs_extent_inline_ref_size(type);
 932         }
 933
 934         if (unlikely(ptr > end)) {
 935                 ret = -EUCLEAN;
 936                 btrfs_print_leaf(path->nodes[0]);
 937                 btrfs_crit(fs_info,
 938 "overrun extent record at slot %d while looking for inline extent for root %llu owner %llu offset %llu parent %llu",
 939                            path->slots[0], root_objectid, owner, offset, parent);
 940                 goto out;
 941         }
 942
 943         if (ret == -ENOENT && insert) {
 944                 if (item_size + extra_size >=
 945                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
 946                         ret = -EAGAIN;
 947                         goto out;
 948                 }
 949                 /*
 950                  * To add new inline back ref, we have to make sure
 951                  * there is no corresponding back ref item.
 952                  * For simplicity, we just do not add new inline back
 953                  * ref if there is any kind of item for this block
 954                  */
 955                 if (find_next_key(path, 0, &key) == 0 &&
 956                     key.objectid == bytenr &&
 957                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
 958                         ret = -EAGAIN;
 959                         goto out;
 960                 }
 961         }
 962         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
 963 out:
 964         if (insert) {
 965                 path->keep_locks = 0;
 966                 path->search_for_extension = 0;
 967                 btrfs_unlock_up_safe(path, 1);
 968         }
 969         return ret;
 970 }
 971
 972 /*
 973  * helper to add new inline back ref
 974  */
 975 static noinline_for_stack
 976 void setup_inline_extent_backref(struct btrfs_trans_handle *trans,
 977                                  struct btrfs_path *path,
 978                                  struct btrfs_extent_inline_ref *iref,
 979                                  u64 parent, u64 root_objectid,
 980                                  u64 owner, u64 offset, int refs_to_add,
 981                                  struct btrfs_delayed_extent_op *extent_op)
 982 {
 983         struct extent_buffer *leaf;
 984         struct btrfs_extent_item *ei;
 985         unsigned long ptr;
 986         unsigned long end;
 987         unsigned long item_offset;
 988         u64 refs;
 989         int size;
 990         int type;
 991
 992         leaf = path->nodes[0];
 993         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
 994         item_offset = (unsigned long)iref - (unsigned long)ei;
 995
 996         type = extent_ref_type(parent, owner);
 997         size = btrfs_extent_inline_ref_size(type);
 998
 999         btrfs_extend_item(trans, path, size);
1000
1001         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1002         refs = btrfs_extent_refs(leaf, ei);
1003         refs += refs_to_add;
1004         btrfs_set_extent_refs(leaf, ei, refs);
1005         if (extent_op)
1006                 __run_delayed_extent_op(extent_op, leaf, ei);
1007
1008         ptr = (unsigned long)ei + item_offset;
1009         end = (unsigned long)ei + btrfs_item_size(leaf, path->slots[0]);
1010         if (ptr < end - size)
1011                 memmove_extent_buffer(leaf, ptr + size, ptr,
1012                                       end - size - ptr);
1013
1014         iref = (struct btrfs_extent_inline_ref *)ptr;
1015         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1016         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1017                 struct btrfs_extent_data_ref *dref;
1018                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1019                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1020                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1021                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1022                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1023         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1024                 struct btrfs_shared_data_ref *sref;
1025                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1026                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1027                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1028         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1029                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1030         } else {
1031                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1032         }
1033         btrfs_mark_buffer_dirty(trans, leaf);
1034 }
1035
1036 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1037                                  struct btrfs_path *path,
1038                                  struct btrfs_extent_inline_ref **ref_ret,
1039                                  u64 bytenr, u64 num_bytes, u64 parent,
1040                                  u64 root_objectid, u64 owner, u64 offset)
1041 {
1042         int ret;
1043
1044         ret = lookup_inline_extent_backref(trans, path, ref_ret, bytenr,
1045                                            num_bytes, parent, root_objectid,
1046                                            owner, offset, 0);
1047         if (ret != -ENOENT)
1048                 return ret;
1049
1050         btrfs_release_path(path);
1051         *ref_ret = NULL;
1052
1053         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1054                 ret = lookup_tree_block_ref(trans, path, bytenr, parent,
1055                                             root_objectid);
1056         } else {
1057                 ret = lookup_extent_data_ref(trans, path, bytenr, parent,
1058                                              root_objectid, owner, offset);
1059         }
1060         return ret;
1061 }
1062
1063 /*
1064  * helper to update/remove inline back ref
1065  */
1066 static noinline_for_stack int update_inline_extent_backref(
1067                                   struct btrfs_trans_handle *trans,
1068                                   struct btrfs_path *path,
1069                                   struct btrfs_extent_inline_ref *iref,
1070                                   int refs_to_mod,
1071                                   struct btrfs_delayed_extent_op *extent_op)
1072 {
1073         struct extent_buffer *leaf = path->nodes[0];
1074         struct btrfs_fs_info *fs_info = leaf->fs_info;
1075         struct btrfs_extent_item *ei;
1076         struct btrfs_extent_data_ref *dref = NULL;
1077         struct btrfs_shared_data_ref *sref = NULL;
1078         unsigned long ptr;
1079         unsigned long end;
1080         u32 item_size;
1081         int size;
1082         int type;
1083         u64 refs;
1084
1085         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1086         refs = btrfs_extent_refs(leaf, ei);
1087         if (unlikely(refs_to_mod < 0 && refs + refs_to_mod <= 0)) {
1088                 struct btrfs_key key;
1089                 u32 extent_size;
1090
1091                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1092                 if (key.type == BTRFS_METADATA_ITEM_KEY)
1093                         extent_size = fs_info->nodesize;
1094                 else
1095                         extent_size = key.offset;
1096                 btrfs_print_leaf(leaf);
1097                 btrfs_err(fs_info,
1098         "invalid refs_to_mod for extent %llu num_bytes %u, has %d expect >= -%llu",
1099                           key.objectid, extent_size, refs_to_mod, refs);
1100                 return -EUCLEAN;
1101         }
1102         refs += refs_to_mod;
1103         btrfs_set_extent_refs(leaf, ei, refs);
1104         if (extent_op)
1105                 __run_delayed_extent_op(extent_op, leaf, ei);
1106
1107         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_ANY);
1108         /*
1109          * Function btrfs_get_extent_inline_ref_type() has already printed
1110          * error messages.
1111          */
1112         if (unlikely(type == BTRFS_REF_TYPE_INVALID))
1113                 return -EUCLEAN;
1114
1115         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1116                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1117                 refs = btrfs_extent_data_ref_count(leaf, dref);
1118         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1119                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1120                 refs = btrfs_shared_data_ref_count(leaf, sref);
1121         } else {
1122                 refs = 1;
1123                 /*
1124                  * For tree blocks we can only drop one ref for it, and tree
1125                  * blocks should not have refs > 1.
1126                  *
1127                  * Furthermore if we're inserting a new inline backref, we
1128                  * won't reach this path either. That would be
1129                  * setup_inline_extent_backref().
1130                  */
1131                 if (unlikely(refs_to_mod != -1)) {
1132                         struct btrfs_key key;
1133
1134                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1135
1136                         btrfs_print_leaf(leaf);
1137                         btrfs_err(fs_info,
1138                         "invalid refs_to_mod for tree block %llu, has %d expect -1",
1139                                   key.objectid, refs_to_mod);
1140                         return -EUCLEAN;
1141                 }
1142         }
1143
1144         if (unlikely(refs_to_mod < 0 && refs < -refs_to_mod)) {
1145                 struct btrfs_key key;
1146                 u32 extent_size;
1147
1148                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1149                 if (key.type == BTRFS_METADATA_ITEM_KEY)
1150                         extent_size = fs_info->nodesize;
1151                 else
1152                         extent_size = key.offset;
1153                 btrfs_print_leaf(leaf);
1154                 btrfs_err(fs_info,
1155 "invalid refs_to_mod for backref entry, iref %lu extent %llu num_bytes %u, has %d expect >= -%llu",
1156                           (unsigned long)iref, key.objectid, extent_size,
1157                           refs_to_mod, refs);
1158                 return -EUCLEAN;
1159         }
1160         refs += refs_to_mod;
1161
1162         if (refs > 0) {
1163                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1164                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1165                 else
1166                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1167         } else {
1168                 size =  btrfs_extent_inline_ref_size(type);
1169                 item_size = btrfs_item_size(leaf, path->slots[0]);
1170                 ptr = (unsigned long)iref;
1171                 end = (unsigned long)ei + item_size;
1172                 if (ptr + size < end)
1173                         memmove_extent_buffer(leaf, ptr, ptr + size,
1174                                               end - ptr - size);
1175                 item_size -= size;
1176                 btrfs_truncate_item(trans, path, item_size, 1);
1177         }
1178         btrfs_mark_buffer_dirty(trans, leaf);
1179         return 0;
1180 }
1181
1182 static noinline_for_stack
1183 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1184                                  struct btrfs_path *path,
1185                                  u64 bytenr, u64 num_bytes, u64 parent,
1186                                  u64 root_objectid, u64 owner,
1187                                  u64 offset, int refs_to_add,
1188                                  struct btrfs_delayed_extent_op *extent_op)
1189 {
1190         struct btrfs_extent_inline_ref *iref;
1191         int ret;
1192
1193         ret = lookup_inline_extent_backref(trans, path, &iref, bytenr,
1194                                            num_bytes, parent, root_objectid,
1195                                            owner, offset, 1);
1196         if (ret == 0) {
1197                 /*
1198                  * We're adding refs to a tree block we already own, this
1199                  * should not happen at all.
1200                  */
1201                 if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1202                         btrfs_print_leaf(path->nodes[0]);
1203                         btrfs_crit(trans->fs_info,
1204 "adding refs to an existing tree ref, bytenr %llu num_bytes %llu root_objectid %llu slot %u",
1205                                    bytenr, num_bytes, root_objectid, path->slots[0]);
1206                         return -EUCLEAN;
1207                 }
1208                 ret = update_inline_extent_backref(trans, path, iref,
1209                                                    refs_to_add, extent_op);
1210         } else if (ret == -ENOENT) {
1211                 setup_inline_extent_backref(trans, path, iref, parent,
1212                                             root_objectid, owner, offset,
1213                                             refs_to_add, extent_op);
1214                 ret = 0;
1215         }
1216         return ret;
1217 }
1218
1219 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1220                                  struct btrfs_root *root,
1221                                  struct btrfs_path *path,
1222                                  struct btrfs_extent_inline_ref *iref,
1223                                  int refs_to_drop, int is_data)
1224 {
1225         int ret = 0;
1226
1227         BUG_ON(!is_data && refs_to_drop != 1);
1228         if (iref)
1229                 ret = update_inline_extent_backref(trans, path, iref,
1230                                                    -refs_to_drop, NULL);
1231         else if (is_data)
1232                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop);
1233         else
1234                 ret = btrfs_del_item(trans, root, path);
1235         return ret;
1236 }
1237
1238 static int btrfs_issue_discard(struct block_device *bdev, u64 start, u64 len,
1239                                u64 *discarded_bytes)
1240 {
1241         int j, ret = 0;
1242         u64 bytes_left, end;
1243         u64 aligned_start = ALIGN(start, 1 << SECTOR_SHIFT);
1244
1245         if (WARN_ON(start != aligned_start)) {
1246                 len -= aligned_start - start;
1247                 len = round_down(len, 1 << SECTOR_SHIFT);
1248                 start = aligned_start;
1249         }
1250
1251         *discarded_bytes = 0;
1252
1253         if (!len)
1254                 return 0;
1255
1256         end = start + len;
1257         bytes_left = len;
1258
1259         /* Skip any superblocks on this device. */
1260         for (j = 0; j < BTRFS_SUPER_MIRROR_MAX; j++) {
1261                 u64 sb_start = btrfs_sb_offset(j);
1262                 u64 sb_end = sb_start + BTRFS_SUPER_INFO_SIZE;
1263                 u64 size = sb_start - start;
1264
1265                 if (!in_range(sb_start, start, bytes_left) &&
1266                     !in_range(sb_end, start, bytes_left) &&
1267                     !in_range(start, sb_start, BTRFS_SUPER_INFO_SIZE))
1268                         continue;
1269
1270                 /*
1271                  * Superblock spans beginning of range.  Adjust start and
1272                  * try again.
1273                  */
1274                 if (sb_start <= start) {
1275                         start += sb_end - start;
1276                         if (start > end) {
1277                                 bytes_left = 0;
1278                                 break;
1279                         }
1280                         bytes_left = end - start;
1281                         continue;
1282                 }
1283
1284                 if (size) {
1285                         ret = blkdev_issue_discard(bdev, start >> SECTOR_SHIFT,
1286                                                    size >> SECTOR_SHIFT,
1287                                                    GFP_NOFS);
1288                         if (!ret)
1289                                 *discarded_bytes += size;
1290                         else if (ret != -EOPNOTSUPP)
1291                                 return ret;
1292                 }
1293
1294                 start = sb_end;
1295                 if (start > end) {
1296                         bytes_left = 0;
1297                         break;
1298                 }
1299                 bytes_left = end - start;
1300         }
1301
1302         if (bytes_left) {
1303                 ret = blkdev_issue_discard(bdev, start >> SECTOR_SHIFT,
1304                                            bytes_left >> SECTOR_SHIFT,
1305                                            GFP_NOFS);
1306                 if (!ret)
1307                         *discarded_bytes += bytes_left;
1308         }
1309         return ret;
1310 }
1311
1312 static int do_discard_extent(struct btrfs_discard_stripe *stripe, u64 *bytes)
1313 {
1314         struct btrfs_device *dev = stripe->dev;
1315         struct btrfs_fs_info *fs_info = dev->fs_info;
1316         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1317         u64 phys = stripe->physical;
1318         u64 len = stripe->length;
1319         u64 discarded = 0;
1320         int ret = 0;
1321
1322         /* Zone reset on a zoned filesystem */
1323         if (btrfs_can_zone_reset(dev, phys, len)) {
1324                 u64 src_disc;
1325
1326                 ret = btrfs_reset_device_zone(dev, phys, len, &discarded);
1327                 if (ret)
1328                         goto out;
1329
1330                 if (!btrfs_dev_replace_is_ongoing(dev_replace) ||
1331                     dev != dev_replace->srcdev)
1332                         goto out;
1333
1334                 src_disc = discarded;
1335
1336                 /* Send to replace target as well */
1337                 ret = btrfs_reset_device_zone(dev_replace->tgtdev, phys, len,
1338                                               &discarded);
1339                 discarded += src_disc;
1340         } else if (bdev_max_discard_sectors(stripe->dev->bdev)) {
1341                 ret = btrfs_issue_discard(dev->bdev, phys, len, &discarded);
1342         } else {
1343                 ret = 0;
1344                 *bytes = 0;
1345         }
1346
1347 out:
1348         *bytes = discarded;
1349         return ret;
1350 }
1351
1352 int btrfs_discard_extent(struct btrfs_fs_info *fs_info, u64 bytenr,
1353                          u64 num_bytes, u64 *actual_bytes)
1354 {
1355         int ret = 0;
1356         u64 discarded_bytes = 0;
1357         u64 end = bytenr + num_bytes;
1358         u64 cur = bytenr;
1359
1360         /*
1361          * Avoid races with device replace and make sure the devices in the
1362          * stripes don't go away while we are discarding.
1363          */
1364         btrfs_bio_counter_inc_blocked(fs_info);
1365         while (cur < end) {
1366                 struct btrfs_discard_stripe *stripes;
1367                 unsigned int num_stripes;
1368                 int i;
1369
1370                 num_bytes = end - cur;
1371                 stripes = btrfs_map_discard(fs_info, cur, &num_bytes, &num_stripes);
1372                 if (IS_ERR(stripes)) {
1373                         ret = PTR_ERR(stripes);
1374                         if (ret == -EOPNOTSUPP)
1375                                 ret = 0;
1376                         break;
1377                 }
1378
1379                 for (i = 0; i < num_stripes; i++) {
1380                         struct btrfs_discard_stripe *stripe = stripes + i;
1381                         u64 bytes;
1382
1383                         if (!stripe->dev->bdev) {
1384                                 ASSERT(btrfs_test_opt(fs_info, DEGRADED));
1385                                 continue;
1386                         }
1387
1388                         if (!test_bit(BTRFS_DEV_STATE_WRITEABLE,
1389                                         &stripe->dev->dev_state))
1390                                 continue;
1391
1392                         ret = do_discard_extent(stripe, &bytes);
1393                         if (ret) {
1394                                 /*
1395                                  * Keep going if discard is not supported by the
1396                                  * device.
1397                                  */
1398                                 if (ret != -EOPNOTSUPP)
1399                                         break;
1400                                 ret = 0;
1401                         } else {
1402                                 discarded_bytes += bytes;
1403                         }
1404                 }
1405                 kfree(stripes);
1406                 if (ret)
1407                         break;
1408                 cur += num_bytes;
1409         }
1410         btrfs_bio_counter_dec(fs_info);
1411         if (actual_bytes)
1412                 *actual_bytes = discarded_bytes;
1413         return ret;
1414 }
1415
1416 /* Can return -ENOMEM */
1417 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1418                          struct btrfs_ref *generic_ref)
1419 {
1420         struct btrfs_fs_info *fs_info = trans->fs_info;
1421         int ret;
1422
1423         ASSERT(generic_ref->type != BTRFS_REF_NOT_SET &&
1424                generic_ref->action);
1425         BUG_ON(generic_ref->type == BTRFS_REF_METADATA &&
1426                generic_ref->tree_ref.owning_root == BTRFS_TREE_LOG_OBJECTID);
1427
1428         if (generic_ref->type == BTRFS_REF_METADATA)
1429                 ret = btrfs_add_delayed_tree_ref(trans, generic_ref, NULL);
1430         else
1431                 ret = btrfs_add_delayed_data_ref(trans, generic_ref, 0);
1432
1433         btrfs_ref_tree_mod(fs_info, generic_ref);
1434
1435         return ret;
1436 }
1437
1438 /*
1439  * Insert backreference for a given extent.
1440  *
1441  * The counterpart is in __btrfs_free_extent(), with examples and more details
1442  * how it works.
1443  *
1444  * @trans:          Handle of transaction
1445  *
1446  * @node:           The delayed ref node used to get the bytenr/length for
1447  *                  extent whose references are incremented.
1448  *
1449  * @parent:         If this is a shared extent (BTRFS_SHARED_DATA_REF_KEY/
1450  *                  BTRFS_SHARED_BLOCK_REF_KEY) then it holds the logical
1451  *                  bytenr of the parent block. Since new extents are always
1452  *                  created with indirect references, this will only be the case
1453  *                  when relocating a shared extent. In that case, root_objectid
1454  *                  will be BTRFS_TREE_RELOC_OBJECTID. Otherwise, parent must
1455  *                  be 0
1456  *
1457  * @root_objectid:  The id of the root where this modification has originated,
1458  *                  this can be either one of the well-known metadata trees or
1459  *                  the subvolume id which references this extent.
1460  *
1461  * @owner:          For data extents it is the inode number of the owning file.
1462  *                  For metadata extents this parameter holds the level in the
1463  *                  tree of the extent.
1464  *
1465  * @offset:         For metadata extents the offset is ignored and is currently
1466  *                  always passed as 0. For data extents it is the fileoffset
1467  *                  this extent belongs to.
1468  *
1469  * @extent_op       Pointer to a structure, holding information necessary when
1470  *                  updating a tree block's flags
1471  *
1472  */
1473 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1474                                   struct btrfs_delayed_ref_node *node,
1475                                   u64 parent, u64 root_objectid,
1476                                   u64 owner, u64 offset,
1477                                   struct btrfs_delayed_extent_op *extent_op)
1478 {
1479         struct btrfs_path *path;
1480         struct extent_buffer *leaf;
1481         struct btrfs_extent_item *item;
1482         struct btrfs_key key;
1483         u64 bytenr = node->bytenr;
1484         u64 num_bytes = node->num_bytes;
1485         u64 refs;
1486         int refs_to_add = node->ref_mod;
1487         int ret;
1488
1489         path = btrfs_alloc_path();
1490         if (!path)
1491                 return -ENOMEM;
1492
1493         /* this will setup the path even if it fails to insert the back ref */
1494         ret = insert_inline_extent_backref(trans, path, bytenr, num_bytes,
1495                                            parent, root_objectid, owner,
1496                                            offset, refs_to_add, extent_op);
1497         if ((ret < 0 && ret != -EAGAIN) || !ret)
1498                 goto out;
1499
1500         /*
1501          * Ok we had -EAGAIN which means we didn't have space to insert and
1502          * inline extent ref, so just update the reference count and add a
1503          * normal backref.
1504          */
1505         leaf = path->nodes[0];
1506         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1507         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1508         refs = btrfs_extent_refs(leaf, item);
1509         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
1510         if (extent_op)
1511                 __run_delayed_extent_op(extent_op, leaf, item);
1512
1513         btrfs_mark_buffer_dirty(trans, leaf);
1514         btrfs_release_path(path);
1515
1516         /* now insert the actual backref */
1517         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1518                 ret = insert_tree_block_ref(trans, path, bytenr, parent,
1519                                             root_objectid);
1520         else
1521                 ret = insert_extent_data_ref(trans, path, bytenr, parent,
1522                                              root_objectid, owner, offset,
1523                                              refs_to_add);
1524
1525         if (ret)
1526                 btrfs_abort_transaction(trans, ret);
1527 out:
1528         btrfs_free_path(path);
1529         return ret;
1530 }
1531
1532 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
1533                                 struct btrfs_delayed_ref_node *node,
1534                                 struct btrfs_delayed_extent_op *extent_op,
1535                                 bool insert_reserved)
1536 {
1537         int ret = 0;
1538         struct btrfs_delayed_data_ref *ref;
1539         u64 parent = 0;
1540         u64 flags = 0;
1541
1542         ref = btrfs_delayed_node_to_data_ref(node);
1543         trace_run_delayed_data_ref(trans->fs_info, node, ref, node->action);
1544
1545         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
1546                 parent = ref->parent;
1547
1548         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1549                 struct btrfs_key key;
1550
1551                 if (extent_op)
1552                         flags |= extent_op->flags_to_set;
1553
1554                 key.objectid = node->bytenr;
1555                 key.type = BTRFS_EXTENT_ITEM_KEY;
1556                 key.offset = node->num_bytes;
1557
1558                 ret = alloc_reserved_file_extent(trans, parent, ref->root,
1559                                                  flags, ref->objectid,
1560                                                  ref->offset, &key,
1561                                                  node->ref_mod);
1562         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1563                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref->root,
1564                                              ref->objectid, ref->offset,
1565                                              extent_op);
1566         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1567                 ret = __btrfs_free_extent(trans, node, parent,
1568                                           ref->root, ref->objectid,
1569                                           ref->offset, extent_op);
1570         } else {
1571                 BUG();
1572         }
1573         return ret;
1574 }
1575
1576 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
1577                                     struct extent_buffer *leaf,
1578                                     struct btrfs_extent_item *ei)
1579 {
1580         u64 flags = btrfs_extent_flags(leaf, ei);
1581         if (extent_op->update_flags) {
1582                 flags |= extent_op->flags_to_set;
1583                 btrfs_set_extent_flags(leaf, ei, flags);
1584         }
1585
1586         if (extent_op->update_key) {
1587                 struct btrfs_tree_block_info *bi;
1588                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
1589                 bi = (struct btrfs_tree_block_info *)(ei + 1);
1590                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
1591         }
1592 }
1593
1594 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
1595                                  struct btrfs_delayed_ref_head *head,
1596                                  struct btrfs_delayed_extent_op *extent_op)
1597 {
1598         struct btrfs_fs_info *fs_info = trans->fs_info;
1599         struct btrfs_root *root;
1600         struct btrfs_key key;
1601         struct btrfs_path *path;
1602         struct btrfs_extent_item *ei;
1603         struct extent_buffer *leaf;
1604         u32 item_size;
1605         int ret;
1606         int metadata = 1;
1607
1608         if (TRANS_ABORTED(trans))
1609                 return 0;
1610
1611         if (!btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1612                 metadata = 0;
1613
1614         path = btrfs_alloc_path();
1615         if (!path)
1616                 return -ENOMEM;
1617
1618         key.objectid = head->bytenr;
1619
1620         if (metadata) {
1621                 key.type = BTRFS_METADATA_ITEM_KEY;
1622                 key.offset = extent_op->level;
1623         } else {
1624                 key.type = BTRFS_EXTENT_ITEM_KEY;
1625                 key.offset = head->num_bytes;
1626         }
1627
1628         root = btrfs_extent_root(fs_info, key.objectid);
1629 again:
1630         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
1631         if (ret < 0) {
1632                 goto out;
1633         } else if (ret > 0) {
1634                 if (metadata) {
1635                         if (path->slots[0] > 0) {
1636                                 path->slots[0]--;
1637                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
1638                                                       path->slots[0]);
1639                                 if (key.objectid == head->bytenr &&
1640                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
1641                                     key.offset == head->num_bytes)
1642                                         ret = 0;
1643                         }
1644                         if (ret > 0) {
1645                                 btrfs_release_path(path);
1646                                 metadata = 0;
1647
1648                                 key.objectid = head->bytenr;
1649                                 key.offset = head->num_bytes;
1650                                 key.type = BTRFS_EXTENT_ITEM_KEY;
1651                                 goto again;
1652                         }
1653                 } else {
1654                         ret = -EUCLEAN;
1655                         btrfs_err(fs_info,
1656                   "missing extent item for extent %llu num_bytes %llu level %d",
1657                                   head->bytenr, head->num_bytes, extent_op->level);
1658                         goto out;
1659                 }
1660         }
1661
1662         leaf = path->nodes[0];
1663         item_size = btrfs_item_size(leaf, path->slots[0]);
1664
1665         if (unlikely(item_size < sizeof(*ei))) {
1666                 ret = -EUCLEAN;
1667                 btrfs_err(fs_info,
1668                           "unexpected extent item size, has %u expect >= %zu",
1669                           item_size, sizeof(*ei));
1670                 btrfs_abort_transaction(trans, ret);
1671                 goto out;
1672         }
1673
1674         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1675         __run_delayed_extent_op(extent_op, leaf, ei);
1676
1677         btrfs_mark_buffer_dirty(trans, leaf);
1678 out:
1679         btrfs_free_path(path);
1680         return ret;
1681 }
1682
1683 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
1684                                 struct btrfs_delayed_ref_node *node,
1685                                 struct btrfs_delayed_extent_op *extent_op,
1686                                 bool insert_reserved)
1687 {
1688         int ret = 0;
1689         struct btrfs_delayed_tree_ref *ref;
1690         u64 parent = 0;
1691         u64 ref_root = 0;
1692
1693         ref = btrfs_delayed_node_to_tree_ref(node);
1694         trace_run_delayed_tree_ref(trans->fs_info, node, ref, node->action);
1695
1696         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
1697                 parent = ref->parent;
1698         ref_root = ref->root;
1699
1700         if (unlikely(node->ref_mod != 1)) {
1701                 btrfs_err(trans->fs_info,
1702         "btree block %llu has %d references rather than 1: action %d ref_root %llu parent %llu",
1703                           node->bytenr, node->ref_mod, node->action, ref_root,
1704                           parent);
1705                 return -EUCLEAN;
1706         }
1707         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1708                 BUG_ON(!extent_op || !extent_op->update_flags);
1709                 ret = alloc_reserved_tree_block(trans, node, extent_op);
1710         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1711                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
1712                                              ref->level, 0, extent_op);
1713         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1714                 ret = __btrfs_free_extent(trans, node, parent, ref_root,
1715                                           ref->level, 0, extent_op);
1716         } else {
1717                 BUG();
1718         }
1719         return ret;
1720 }
1721
1722 /* helper function to actually process a single delayed ref entry */
1723 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
1724                                struct btrfs_delayed_ref_node *node,
1725                                struct btrfs_delayed_extent_op *extent_op,
1726                                bool insert_reserved)
1727 {
1728         int ret = 0;
1729
1730         if (TRANS_ABORTED(trans)) {
1731                 if (insert_reserved)
1732                         btrfs_pin_extent(trans, node->bytenr, node->num_bytes, 1);
1733                 return 0;
1734         }
1735
1736         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
1737             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
1738                 ret = run_delayed_tree_ref(trans, node, extent_op,
1739                                            insert_reserved);
1740         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
1741                  node->type == BTRFS_SHARED_DATA_REF_KEY)
1742                 ret = run_delayed_data_ref(trans, node, extent_op,
1743                                            insert_reserved);
1744         else
1745                 BUG();
1746         if (ret && insert_reserved)
1747                 btrfs_pin_extent(trans, node->bytenr, node->num_bytes, 1);
1748         if (ret < 0)
1749                 btrfs_err(trans->fs_info,
1750 "failed to run delayed ref for logical %llu num_bytes %llu type %u action %u ref_mod %d: %d",
1751                           node->bytenr, node->num_bytes, node->type,
1752                           node->action, node->ref_mod, ret);
1753         return ret;
1754 }
1755
1756 static inline struct btrfs_delayed_ref_node *
1757 select_delayed_ref(struct btrfs_delayed_ref_head *head)
1758 {
1759         struct btrfs_delayed_ref_node *ref;
1760
1761         if (RB_EMPTY_ROOT(&head->ref_tree.rb_root))
1762                 return NULL;
1763
1764         /*
1765          * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
1766          * This is to prevent a ref count from going down to zero, which deletes
1767          * the extent item from the extent tree, when there still are references
1768          * to add, which would fail because they would not find the extent item.
1769          */
1770         if (!list_empty(&head->ref_add_list))
1771                 return list_first_entry(&head->ref_add_list,
1772                                 struct btrfs_delayed_ref_node, add_list);
1773
1774         ref = rb_entry(rb_first_cached(&head->ref_tree),
1775                        struct btrfs_delayed_ref_node, ref_node);
1776         ASSERT(list_empty(&ref->add_list));
1777         return ref;
1778 }
1779
1780 static void unselect_delayed_ref_head(struct btrfs_delayed_ref_root *delayed_refs,
1781                                       struct btrfs_delayed_ref_head *head)
1782 {
1783         spin_lock(&delayed_refs->lock);
1784         head->processing = false;
1785         delayed_refs->num_heads_ready++;
1786         spin_unlock(&delayed_refs->lock);
1787         btrfs_delayed_ref_unlock(head);
1788 }
1789
1790 static struct btrfs_delayed_extent_op *cleanup_extent_op(
1791                                 struct btrfs_delayed_ref_head *head)
1792 {
1793         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
1794
1795         if (!extent_op)
1796                 return NULL;
1797
1798         if (head->must_insert_reserved) {
1799                 head->extent_op = NULL;
1800                 btrfs_free_delayed_extent_op(extent_op);
1801                 return NULL;
1802         }
1803         return extent_op;
1804 }
1805
1806 static int run_and_cleanup_extent_op(struct btrfs_trans_handle *trans,
1807                                      struct btrfs_delayed_ref_head *head)
1808 {
1809         struct btrfs_delayed_extent_op *extent_op;
1810         int ret;
1811
1812         extent_op = cleanup_extent_op(head);
1813         if (!extent_op)
1814                 return 0;
1815         head->extent_op = NULL;
1816         spin_unlock(&head->lock);
1817         ret = run_delayed_extent_op(trans, head, extent_op);
1818         btrfs_free_delayed_extent_op(extent_op);
1819         return ret ? ret : 1;
1820 }
1821
1822 u64 btrfs_cleanup_ref_head_accounting(struct btrfs_fs_info *fs_info,
1823                                   struct btrfs_delayed_ref_root *delayed_refs,
1824                                   struct btrfs_delayed_ref_head *head)
1825 {
1826         /*
1827          * We had csum deletions accounted for in our delayed refs rsv, we need
1828          * to drop the csum leaves for this update from our delayed_refs_rsv.
1829          */
1830         if (head->total_ref_mod < 0 && head->is_data) {
1831                 int nr_csums;
1832
1833                 spin_lock(&delayed_refs->lock);
1834                 delayed_refs->pending_csums -= head->num_bytes;
1835                 spin_unlock(&delayed_refs->lock);
1836                 nr_csums = btrfs_csum_bytes_to_leaves(fs_info, head->num_bytes);
1837
1838                 btrfs_delayed_refs_rsv_release(fs_info, 0, nr_csums);
1839
1840                 return btrfs_calc_delayed_ref_csum_bytes(fs_info, nr_csums);
1841         }
1842
1843         return 0;
1844 }
1845
1846 static int cleanup_ref_head(struct btrfs_trans_handle *trans,
1847                             struct btrfs_delayed_ref_head *head,
1848                             u64 *bytes_released)
1849 {
1850
1851         struct btrfs_fs_info *fs_info = trans->fs_info;
1852         struct btrfs_delayed_ref_root *delayed_refs;
1853         int ret;
1854
1855         delayed_refs = &trans->transaction->delayed_refs;
1856
1857         ret = run_and_cleanup_extent_op(trans, head);
1858         if (ret < 0) {
1859                 unselect_delayed_ref_head(delayed_refs, head);
1860                 btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
1861                 return ret;
1862         } else if (ret) {
1863                 return ret;
1864         }
1865
1866         /*
1867          * Need to drop our head ref lock and re-acquire the delayed ref lock
1868          * and then re-check to make sure nobody got added.
1869          */
1870         spin_unlock(&head->lock);
1871         spin_lock(&delayed_refs->lock);
1872         spin_lock(&head->lock);
1873         if (!RB_EMPTY_ROOT(&head->ref_tree.rb_root) || head->extent_op) {
1874                 spin_unlock(&head->lock);
1875                 spin_unlock(&delayed_refs->lock);
1876                 return 1;
1877         }
1878         btrfs_delete_ref_head(delayed_refs, head);
1879         spin_unlock(&head->lock);
1880         spin_unlock(&delayed_refs->lock);
1881
1882         if (head->must_insert_reserved) {
1883                 btrfs_pin_extent(trans, head->bytenr, head->num_bytes, 1);
1884                 if (head->is_data) {
1885                         struct btrfs_root *csum_root;
1886
1887                         csum_root = btrfs_csum_root(fs_info, head->bytenr);
1888                         ret = btrfs_del_csums(trans, csum_root, head->bytenr,
1889                                               head->num_bytes);
1890                 }
1891         }
1892
1893         *bytes_released += btrfs_cleanup_ref_head_accounting(fs_info, delayed_refs, head);
1894
1895         trace_run_delayed_ref_head(fs_info, head, 0);
1896         btrfs_delayed_ref_unlock(head);
1897         btrfs_put_delayed_ref_head(head);
1898         return ret;
1899 }
1900
1901 static struct btrfs_delayed_ref_head *btrfs_obtain_ref_head(
1902                                         struct btrfs_trans_handle *trans)
1903 {
1904         struct btrfs_delayed_ref_root *delayed_refs =
1905                 &trans->transaction->delayed_refs;
1906         struct btrfs_delayed_ref_head *head = NULL;
1907         int ret;
1908
1909         spin_lock(&delayed_refs->lock);
1910         head = btrfs_select_ref_head(delayed_refs);
1911         if (!head) {
1912                 spin_unlock(&delayed_refs->lock);
1913                 return head;
1914         }
1915
1916         /*
1917          * Grab the lock that says we are going to process all the refs for
1918          * this head
1919          */
1920         ret = btrfs_delayed_ref_lock(delayed_refs, head);
1921         spin_unlock(&delayed_refs->lock);
1922
1923         /*
1924          * We may have dropped the spin lock to get the head mutex lock, and
1925          * that might have given someone else time to free the head.  If that's
1926          * true, it has been removed from our list and we can move on.
1927          */
1928         if (ret == -EAGAIN)
1929                 head = ERR_PTR(-EAGAIN);
1930
1931         return head;
1932 }
1933
1934 static int btrfs_run_delayed_refs_for_head(struct btrfs_trans_handle *trans,
1935                                            struct btrfs_delayed_ref_head *locked_ref,
1936                                            u64 *bytes_released)
1937 {
1938         struct btrfs_fs_info *fs_info = trans->fs_info;
1939         struct btrfs_delayed_ref_root *delayed_refs;
1940         struct btrfs_delayed_extent_op *extent_op;
1941         struct btrfs_delayed_ref_node *ref;
1942         bool must_insert_reserved;
1943         int ret;
1944
1945         delayed_refs = &trans->transaction->delayed_refs;
1946
1947         lockdep_assert_held(&locked_ref->mutex);
1948         lockdep_assert_held(&locked_ref->lock);
1949
1950         while ((ref = select_delayed_ref(locked_ref))) {
1951                 if (ref->seq &&
1952                     btrfs_check_delayed_seq(fs_info, ref->seq)) {
1953                         spin_unlock(&locked_ref->lock);
1954                         unselect_delayed_ref_head(delayed_refs, locked_ref);
1955                         return -EAGAIN;
1956                 }
1957
1958                 rb_erase_cached(&ref->ref_node, &locked_ref->ref_tree);
1959                 RB_CLEAR_NODE(&ref->ref_node);
1960                 if (!list_empty(&ref->add_list))
1961                         list_del(&ref->add_list);
1962                 /*
1963                  * When we play the delayed ref, also correct the ref_mod on
1964                  * head
1965                  */
1966                 switch (ref->action) {
1967                 case BTRFS_ADD_DELAYED_REF:
1968                 case BTRFS_ADD_DELAYED_EXTENT:
1969                         locked_ref->ref_mod -= ref->ref_mod;
1970                         break;
1971                 case BTRFS_DROP_DELAYED_REF:
1972                         locked_ref->ref_mod += ref->ref_mod;
1973                         break;
1974                 default:
1975                         WARN_ON(1);
1976                 }
1977                 atomic_dec(&delayed_refs->num_entries);
1978
1979                 /*
1980                  * Record the must_insert_reserved flag before we drop the
1981                  * spin lock.
1982                  */
1983                 must_insert_reserved = locked_ref->must_insert_reserved;
1984                 locked_ref->must_insert_reserved = false;
1985
1986                 extent_op = locked_ref->extent_op;
1987                 locked_ref->extent_op = NULL;
1988                 spin_unlock(&locked_ref->lock);
1989
1990                 ret = run_one_delayed_ref(trans, ref, extent_op,
1991                                           must_insert_reserved);
1992                 btrfs_delayed_refs_rsv_release(fs_info, 1, 0);
1993                 *bytes_released += btrfs_calc_delayed_ref_bytes(fs_info, 1);
1994                 btrfs_free_delayed_extent_op(extent_op);
1995                 if (ret) {
1996                         unselect_delayed_ref_head(delayed_refs, locked_ref);
1997                         btrfs_put_delayed_ref(ref);
1998                         return ret;
1999                 }
2000
2001                 btrfs_put_delayed_ref(ref);
2002                 cond_resched();
2003
2004                 spin_lock(&locked_ref->lock);
2005                 btrfs_merge_delayed_refs(fs_info, delayed_refs, locked_ref);
2006         }
2007
2008         return 0;
2009 }
2010
2011 /*
2012  * Returns 0 on success or if called with an already aborted transaction.
2013  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2014  */
2015 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2016                                              u64 min_bytes)
2017 {
2018         struct btrfs_fs_info *fs_info = trans->fs_info;
2019         struct btrfs_delayed_ref_root *delayed_refs;
2020         struct btrfs_delayed_ref_head *locked_ref = NULL;
2021         int ret;
2022         unsigned long count = 0;
2023         unsigned long max_count = 0;
2024         u64 bytes_processed = 0;
2025
2026         delayed_refs = &trans->transaction->delayed_refs;
2027         if (min_bytes == 0) {
2028                 max_count = delayed_refs->num_heads_ready;
2029                 min_bytes = U64_MAX;
2030         }
2031
2032         do {
2033                 if (!locked_ref) {
2034                         locked_ref = btrfs_obtain_ref_head(trans);
2035                         if (IS_ERR_OR_NULL(locked_ref)) {
2036                                 if (PTR_ERR(locked_ref) == -EAGAIN) {
2037                                         continue;
2038                                 } else {
2039                                         break;
2040                                 }
2041                         }
2042                         count++;
2043                 }
2044                 /*
2045                  * We need to try and merge add/drops of the same ref since we
2046                  * can run into issues with relocate dropping the implicit ref
2047                  * and then it being added back again before the drop can
2048                  * finish.  If we merged anything we need to re-loop so we can
2049                  * get a good ref.
2050                  * Or we can get node references of the same type that weren't
2051                  * merged when created due to bumps in the tree mod seq, and
2052                  * we need to merge them to prevent adding an inline extent
2053                  * backref before dropping it (triggering a BUG_ON at
2054                  * insert_inline_extent_backref()).
2055                  */
2056                 spin_lock(&locked_ref->lock);
2057                 btrfs_merge_delayed_refs(fs_info, delayed_refs, locked_ref);
2058
2059                 ret = btrfs_run_delayed_refs_for_head(trans, locked_ref, &bytes_processed);
2060                 if (ret < 0 && ret != -EAGAIN) {
2061                         /*
2062                          * Error, btrfs_run_delayed_refs_for_head already
2063                          * unlocked everything so just bail out
2064                          */
2065                         return ret;
2066                 } else if (!ret) {
2067                         /*
2068                          * Success, perform the usual cleanup of a processed
2069                          * head
2070                          */
2071                         ret = cleanup_ref_head(trans, locked_ref, &bytes_processed);
2072                         if (ret > 0 ) {
2073                                 /* We dropped our lock, we need to loop. */
2074                                 ret = 0;
2075                                 continue;
2076                         } else if (ret) {
2077                                 return ret;
2078                         }
2079                 }
2080
2081                 /*
2082                  * Either success case or btrfs_run_delayed_refs_for_head
2083                  * returned -EAGAIN, meaning we need to select another head
2084                  */
2085
2086                 locked_ref = NULL;
2087                 cond_resched();
2088         } while ((min_bytes != U64_MAX && bytes_processed < min_bytes) ||
2089                  (max_count > 0 && count < max_count) ||
2090                  locked_ref);
2091
2092         return 0;
2093 }
2094
2095 #ifdef SCRAMBLE_DELAYED_REFS
2096 /*
2097  * Normally delayed refs get processed in ascending bytenr order. This
2098  * correlates in most cases to the order added. To expose dependencies on this
2099  * order, we start to process the tree in the middle instead of the beginning
2100  */
2101 static u64 find_middle(struct rb_root *root)
2102 {
2103         struct rb_node *n = root->rb_node;
2104         struct btrfs_delayed_ref_node *entry;
2105         int alt = 1;
2106         u64 middle;
2107         u64 first = 0, last = 0;
2108
2109         n = rb_first(root);
2110         if (n) {
2111                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2112                 first = entry->bytenr;
2113         }
2114         n = rb_last(root);
2115         if (n) {
2116                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2117                 last = entry->bytenr;
2118         }
2119         n = root->rb_node;
2120
2121         while (n) {
2122                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2123                 WARN_ON(!entry->in_tree);
2124
2125                 middle = entry->bytenr;
2126
2127                 if (alt)
2128                         n = n->rb_left;
2129                 else
2130                         n = n->rb_right;
2131
2132                 alt = 1 - alt;
2133         }
2134         return middle;
2135 }
2136 #endif
2137
2138 /*
2139  * Start processing the delayed reference count updates and extent insertions
2140  * we have queued up so far.
2141  *
2142  * @trans:      Transaction handle.
2143  * @min_bytes:  How many bytes of delayed references to process. After this
2144  *              many bytes we stop processing delayed references if there are
2145  *              any more. If 0 it means to run all existing delayed references,
2146  *              but not new ones added after running all existing ones.
2147  *              Use (u64)-1 (U64_MAX) to run all existing delayed references
2148  *              plus any new ones that are added.
2149  *
2150  * Returns 0 on success or if called with an aborted transaction
2151  * Returns <0 on error and aborts the transaction
2152  */
2153 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans, u64 min_bytes)
2154 {
2155         struct btrfs_fs_info *fs_info = trans->fs_info;
2156         struct btrfs_delayed_ref_root *delayed_refs;
2157         int ret;
2158
2159         /* We'll clean this up in btrfs_cleanup_transaction */
2160         if (TRANS_ABORTED(trans))
2161                 return 0;
2162
2163         if (test_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags))
2164                 return 0;
2165
2166         delayed_refs = &trans->transaction->delayed_refs;
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, min_bytes);
2172         if (ret < 0) {
2173                 btrfs_abort_transaction(trans, ret);
2174                 return ret;
2175         }
2176
2177         if (min_bytes == U64_MAX) {
2178                 btrfs_create_pending_block_groups(trans);
2179
2180                 spin_lock(&delayed_refs->lock);
2181                 if (RB_EMPTY_ROOT(&delayed_refs->href_root.rb_root)) {
2182                         spin_unlock(&delayed_refs->lock);
2183                         return 0;
2184                 }
2185                 spin_unlock(&delayed_refs->lock);
2186
2187                 cond_resched();
2188                 goto again;
2189         }
2190
2191         return 0;
2192 }
2193
2194 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2195                                 struct extent_buffer *eb, u64 flags)
2196 {
2197         struct btrfs_delayed_extent_op *extent_op;
2198         int level = btrfs_header_level(eb);
2199         int ret;
2200
2201         extent_op = btrfs_alloc_delayed_extent_op();
2202         if (!extent_op)
2203                 return -ENOMEM;
2204
2205         extent_op->flags_to_set = flags;
2206         extent_op->update_flags = true;
2207         extent_op->update_key = false;
2208         extent_op->level = level;
2209
2210         ret = btrfs_add_delayed_extent_op(trans, eb->start, eb->len, extent_op);
2211         if (ret)
2212                 btrfs_free_delayed_extent_op(extent_op);
2213         return ret;
2214 }
2215
2216 static noinline int check_delayed_ref(struct btrfs_root *root,
2217                                       struct btrfs_path *path,
2218                                       u64 objectid, u64 offset, u64 bytenr)
2219 {
2220         struct btrfs_delayed_ref_head *head;
2221         struct btrfs_delayed_ref_node *ref;
2222         struct btrfs_delayed_data_ref *data_ref;
2223         struct btrfs_delayed_ref_root *delayed_refs;
2224         struct btrfs_transaction *cur_trans;
2225         struct rb_node *node;
2226         int ret = 0;
2227
2228         spin_lock(&root->fs_info->trans_lock);
2229         cur_trans = root->fs_info->running_transaction;
2230         if (cur_trans)
2231                 refcount_inc(&cur_trans->use_count);
2232         spin_unlock(&root->fs_info->trans_lock);
2233         if (!cur_trans)
2234                 return 0;
2235
2236         delayed_refs = &cur_trans->delayed_refs;
2237         spin_lock(&delayed_refs->lock);
2238         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
2239         if (!head) {
2240                 spin_unlock(&delayed_refs->lock);
2241                 btrfs_put_transaction(cur_trans);
2242                 return 0;
2243         }
2244
2245         if (!mutex_trylock(&head->mutex)) {
2246                 if (path->nowait) {
2247                         spin_unlock(&delayed_refs->lock);
2248                         btrfs_put_transaction(cur_trans);
2249                         return -EAGAIN;
2250                 }
2251
2252                 refcount_inc(&head->refs);
2253                 spin_unlock(&delayed_refs->lock);
2254
2255                 btrfs_release_path(path);
2256
2257                 /*
2258                  * Mutex was contended, block until it's released and let
2259                  * caller try again
2260                  */
2261                 mutex_lock(&head->mutex);
2262                 mutex_unlock(&head->mutex);
2263                 btrfs_put_delayed_ref_head(head);
2264                 btrfs_put_transaction(cur_trans);
2265                 return -EAGAIN;
2266         }
2267         spin_unlock(&delayed_refs->lock);
2268
2269         spin_lock(&head->lock);
2270         /*
2271          * XXX: We should replace this with a proper search function in the
2272          * future.
2273          */
2274         for (node = rb_first_cached(&head->ref_tree); node;
2275              node = rb_next(node)) {
2276                 ref = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
2277                 /* If it's a shared ref we know a cross reference exists */
2278                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
2279                         ret = 1;
2280                         break;
2281                 }
2282
2283                 data_ref = btrfs_delayed_node_to_data_ref(ref);
2284
2285                 /*
2286                  * If our ref doesn't match the one we're currently looking at
2287                  * then we have a cross reference.
2288                  */
2289                 if (data_ref->root != root->root_key.objectid ||
2290                     data_ref->objectid != objectid ||
2291                     data_ref->offset != offset) {
2292                         ret = 1;
2293                         break;
2294                 }
2295         }
2296         spin_unlock(&head->lock);
2297         mutex_unlock(&head->mutex);
2298         btrfs_put_transaction(cur_trans);
2299         return ret;
2300 }
2301
2302 static noinline int check_committed_ref(struct btrfs_root *root,
2303                                         struct btrfs_path *path,
2304                                         u64 objectid, u64 offset, u64 bytenr,
2305                                         bool strict)
2306 {
2307         struct btrfs_fs_info *fs_info = root->fs_info;
2308         struct btrfs_root *extent_root = btrfs_extent_root(fs_info, bytenr);
2309         struct extent_buffer *leaf;
2310         struct btrfs_extent_data_ref *ref;
2311         struct btrfs_extent_inline_ref *iref;
2312         struct btrfs_extent_item *ei;
2313         struct btrfs_key key;
2314         u32 item_size;
2315         int type;
2316         int ret;
2317
2318         key.objectid = bytenr;
2319         key.offset = (u64)-1;
2320         key.type = BTRFS_EXTENT_ITEM_KEY;
2321
2322         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2323         if (ret < 0)
2324                 goto out;
2325         BUG_ON(ret == 0); /* Corruption */
2326
2327         ret = -ENOENT;
2328         if (path->slots[0] == 0)
2329                 goto out;
2330
2331         path->slots[0]--;
2332         leaf = path->nodes[0];
2333         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2334
2335         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2336                 goto out;
2337
2338         ret = 1;
2339         item_size = btrfs_item_size(leaf, path->slots[0]);
2340         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2341
2342         /* If extent item has more than 1 inline ref then it's shared */
2343         if (item_size != sizeof(*ei) +
2344             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2345                 goto out;
2346
2347         /*
2348          * If extent created before last snapshot => it's shared unless the
2349          * snapshot has been deleted. Use the heuristic if strict is false.
2350          */
2351         if (!strict &&
2352             (btrfs_extent_generation(leaf, ei) <=
2353              btrfs_root_last_snapshot(&root->root_item)))
2354                 goto out;
2355
2356         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2357
2358         /* If this extent has SHARED_DATA_REF then it's shared */
2359         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
2360         if (type != BTRFS_EXTENT_DATA_REF_KEY)
2361                 goto out;
2362
2363         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2364         if (btrfs_extent_refs(leaf, ei) !=
2365             btrfs_extent_data_ref_count(leaf, ref) ||
2366             btrfs_extent_data_ref_root(leaf, ref) !=
2367             root->root_key.objectid ||
2368             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2369             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2370                 goto out;
2371
2372         ret = 0;
2373 out:
2374         return ret;
2375 }
2376
2377 int btrfs_cross_ref_exist(struct btrfs_root *root, u64 objectid, u64 offset,
2378                           u64 bytenr, bool strict, struct btrfs_path *path)
2379 {
2380         int ret;
2381
2382         do {
2383                 ret = check_committed_ref(root, path, objectid,
2384                                           offset, bytenr, strict);
2385                 if (ret && ret != -ENOENT)
2386                         goto out;
2387
2388                 ret = check_delayed_ref(root, path, objectid, offset, bytenr);
2389         } while (ret == -EAGAIN);
2390
2391 out:
2392         btrfs_release_path(path);
2393         if (btrfs_is_data_reloc_root(root))
2394                 WARN_ON(ret > 0);
2395         return ret;
2396 }
2397
2398 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2399                            struct btrfs_root *root,
2400                            struct extent_buffer *buf,
2401                            int full_backref, int inc)
2402 {
2403         struct btrfs_fs_info *fs_info = root->fs_info;
2404         u64 bytenr;
2405         u64 num_bytes;
2406         u64 parent;
2407         u64 ref_root;
2408         u32 nritems;
2409         struct btrfs_key key;
2410         struct btrfs_file_extent_item *fi;
2411         struct btrfs_ref generic_ref = { 0 };
2412         bool for_reloc = btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC);
2413         int i;
2414         int action;
2415         int level;
2416         int ret = 0;
2417
2418         if (btrfs_is_testing(fs_info))
2419                 return 0;
2420
2421         ref_root = btrfs_header_owner(buf);
2422         nritems = btrfs_header_nritems(buf);
2423         level = btrfs_header_level(buf);
2424
2425         if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state) && level == 0)
2426                 return 0;
2427
2428         if (full_backref)
2429                 parent = buf->start;
2430         else
2431                 parent = 0;
2432         if (inc)
2433                 action = BTRFS_ADD_DELAYED_REF;
2434         else
2435                 action = BTRFS_DROP_DELAYED_REF;
2436
2437         for (i = 0; i < nritems; i++) {
2438                 if (level == 0) {
2439                         btrfs_item_key_to_cpu(buf, &key, i);
2440                         if (key.type != BTRFS_EXTENT_DATA_KEY)
2441                                 continue;
2442                         fi = btrfs_item_ptr(buf, i,
2443                                             struct btrfs_file_extent_item);
2444                         if (btrfs_file_extent_type(buf, fi) ==
2445                             BTRFS_FILE_EXTENT_INLINE)
2446                                 continue;
2447                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2448                         if (bytenr == 0)
2449                                 continue;
2450
2451                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
2452                         key.offset -= btrfs_file_extent_offset(buf, fi);
2453                         btrfs_init_generic_ref(&generic_ref, action, bytenr,
2454                                                num_bytes, parent);
2455                         btrfs_init_data_ref(&generic_ref, ref_root, key.objectid,
2456                                             key.offset, root->root_key.objectid,
2457                                             for_reloc);
2458                         if (inc)
2459                                 ret = btrfs_inc_extent_ref(trans, &generic_ref);
2460                         else
2461                                 ret = btrfs_free_extent(trans, &generic_ref);
2462                         if (ret)
2463                                 goto fail;
2464                 } else {
2465                         bytenr = btrfs_node_blockptr(buf, i);
2466                         num_bytes = fs_info->nodesize;
2467                         btrfs_init_generic_ref(&generic_ref, action, bytenr,
2468                                                num_bytes, parent);
2469                         btrfs_init_tree_ref(&generic_ref, level - 1, ref_root,
2470                                             root->root_key.objectid, for_reloc);
2471                         if (inc)
2472                                 ret = btrfs_inc_extent_ref(trans, &generic_ref);
2473                         else
2474                                 ret = btrfs_free_extent(trans, &generic_ref);
2475                         if (ret)
2476                                 goto fail;
2477                 }
2478         }
2479         return 0;
2480 fail:
2481         return ret;
2482 }
2483
2484 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2485                   struct extent_buffer *buf, int full_backref)
2486 {
2487         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
2488 }
2489
2490 int btrfs_dec_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, 0);
2494 }
2495
2496 static u64 get_alloc_profile_by_root(struct btrfs_root *root, int data)
2497 {
2498         struct btrfs_fs_info *fs_info = root->fs_info;
2499         u64 flags;
2500         u64 ret;
2501
2502         if (data)
2503                 flags = BTRFS_BLOCK_GROUP_DATA;
2504         else if (root == fs_info->chunk_root)
2505                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
2506         else
2507                 flags = BTRFS_BLOCK_GROUP_METADATA;
2508
2509         ret = btrfs_get_alloc_profile(fs_info, flags);
2510         return ret;
2511 }
2512
2513 static u64 first_logical_byte(struct btrfs_fs_info *fs_info)
2514 {
2515         struct rb_node *leftmost;
2516         u64 bytenr = 0;
2517
2518         read_lock(&fs_info->block_group_cache_lock);
2519         /* Get the block group with the lowest logical start address. */
2520         leftmost = rb_first_cached(&fs_info->block_group_cache_tree);
2521         if (leftmost) {
2522                 struct btrfs_block_group *bg;
2523
2524                 bg = rb_entry(leftmost, struct btrfs_block_group, cache_node);
2525                 bytenr = bg->start;
2526         }
2527         read_unlock(&fs_info->block_group_cache_lock);
2528
2529         return bytenr;
2530 }
2531
2532 static int pin_down_extent(struct btrfs_trans_handle *trans,
2533                            struct btrfs_block_group *cache,
2534                            u64 bytenr, u64 num_bytes, int reserved)
2535 {
2536         struct btrfs_fs_info *fs_info = cache->fs_info;
2537
2538         spin_lock(&cache->space_info->lock);
2539         spin_lock(&cache->lock);
2540         cache->pinned += num_bytes;
2541         btrfs_space_info_update_bytes_pinned(fs_info, cache->space_info,
2542                                              num_bytes);
2543         if (reserved) {
2544                 cache->reserved -= num_bytes;
2545                 cache->space_info->bytes_reserved -= num_bytes;
2546         }
2547         spin_unlock(&cache->lock);
2548         spin_unlock(&cache->space_info->lock);
2549
2550         set_extent_bit(&trans->transaction->pinned_extents, bytenr,
2551                        bytenr + num_bytes - 1, EXTENT_DIRTY, NULL);
2552         return 0;
2553 }
2554
2555 int btrfs_pin_extent(struct btrfs_trans_handle *trans,
2556                      u64 bytenr, u64 num_bytes, int reserved)
2557 {
2558         struct btrfs_block_group *cache;
2559
2560         cache = btrfs_lookup_block_group(trans->fs_info, bytenr);
2561         BUG_ON(!cache); /* Logic error */
2562
2563         pin_down_extent(trans, cache, bytenr, num_bytes, reserved);
2564
2565         btrfs_put_block_group(cache);
2566         return 0;
2567 }
2568
2569 int btrfs_pin_extent_for_log_replay(struct btrfs_trans_handle *trans,
2570                                     const struct extent_buffer *eb)
2571 {
2572         struct btrfs_block_group *cache;
2573         int ret;
2574
2575         cache = btrfs_lookup_block_group(trans->fs_info, eb->start);
2576         if (!cache)
2577                 return -EINVAL;
2578
2579         /*
2580          * Fully cache the free space first so that our pin removes the free space
2581          * from the cache.
2582          */
2583         ret = btrfs_cache_block_group(cache, true);
2584         if (ret)
2585                 goto out;
2586
2587         pin_down_extent(trans, cache, eb->start, eb->len, 0);
2588
2589         /* remove us from the free space cache (if we're there at all) */
2590         ret = btrfs_remove_free_space(cache, eb->start, eb->len);
2591 out:
2592         btrfs_put_block_group(cache);
2593         return ret;
2594 }
2595
2596 static int __exclude_logged_extent(struct btrfs_fs_info *fs_info,
2597                                    u64 start, u64 num_bytes)
2598 {
2599         int ret;
2600         struct btrfs_block_group *block_group;
2601
2602         block_group = btrfs_lookup_block_group(fs_info, start);
2603         if (!block_group)
2604                 return -EINVAL;
2605
2606         ret = btrfs_cache_block_group(block_group, true);
2607         if (ret)
2608                 goto out;
2609
2610         ret = btrfs_remove_free_space(block_group, start, num_bytes);
2611 out:
2612         btrfs_put_block_group(block_group);
2613         return ret;
2614 }
2615
2616 int btrfs_exclude_logged_extents(struct extent_buffer *eb)
2617 {
2618         struct btrfs_fs_info *fs_info = eb->fs_info;
2619         struct btrfs_file_extent_item *item;
2620         struct btrfs_key key;
2621         int found_type;
2622         int i;
2623         int ret = 0;
2624
2625         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS))
2626                 return 0;
2627
2628         for (i = 0; i < btrfs_header_nritems(eb); i++) {
2629                 btrfs_item_key_to_cpu(eb, &key, i);
2630                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2631                         continue;
2632                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
2633                 found_type = btrfs_file_extent_type(eb, item);
2634                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
2635                         continue;
2636                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
2637                         continue;
2638                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
2639                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
2640                 ret = __exclude_logged_extent(fs_info, key.objectid, key.offset);
2641                 if (ret)
2642                         break;
2643         }
2644
2645         return ret;
2646 }
2647
2648 static void
2649 btrfs_inc_block_group_reservations(struct btrfs_block_group *bg)
2650 {
2651         atomic_inc(&bg->reservations);
2652 }
2653
2654 /*
2655  * Returns the free cluster for the given space info and sets empty_cluster to
2656  * what it should be based on the mount options.
2657  */
2658 static struct btrfs_free_cluster *
2659 fetch_cluster_info(struct btrfs_fs_info *fs_info,
2660                    struct btrfs_space_info *space_info, u64 *empty_cluster)
2661 {
2662         struct btrfs_free_cluster *ret = NULL;
2663
2664         *empty_cluster = 0;
2665         if (btrfs_mixed_space_info(space_info))
2666                 return ret;
2667
2668         if (space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
2669                 ret = &fs_info->meta_alloc_cluster;
2670                 if (btrfs_test_opt(fs_info, SSD))
2671                         *empty_cluster = SZ_2M;
2672                 else
2673                         *empty_cluster = SZ_64K;
2674         } else if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) &&
2675                    btrfs_test_opt(fs_info, SSD_SPREAD)) {
2676                 *empty_cluster = SZ_2M;
2677                 ret = &fs_info->data_alloc_cluster;
2678         }
2679
2680         return ret;
2681 }
2682
2683 static int unpin_extent_range(struct btrfs_fs_info *fs_info,
2684                               u64 start, u64 end,
2685                               const bool return_free_space)
2686 {
2687         struct btrfs_block_group *cache = NULL;
2688         struct btrfs_space_info *space_info;
2689         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
2690         struct btrfs_free_cluster *cluster = NULL;
2691         u64 len;
2692         u64 total_unpinned = 0;
2693         u64 empty_cluster = 0;
2694         bool readonly;
2695
2696         while (start <= end) {
2697                 readonly = false;
2698                 if (!cache ||
2699                     start >= cache->start + cache->length) {
2700                         if (cache)
2701                                 btrfs_put_block_group(cache);
2702                         total_unpinned = 0;
2703                         cache = btrfs_lookup_block_group(fs_info, start);
2704                         BUG_ON(!cache); /* Logic error */
2705
2706                         cluster = fetch_cluster_info(fs_info,
2707                                                      cache->space_info,
2708                                                      &empty_cluster);
2709                         empty_cluster <<= 1;
2710                 }
2711
2712                 len = cache->start + cache->length - start;
2713                 len = min(len, end + 1 - start);
2714
2715                 if (return_free_space)
2716                         btrfs_add_free_space(cache, start, len);
2717
2718                 start += len;
2719                 total_unpinned += len;
2720                 space_info = cache->space_info;
2721
2722                 /*
2723                  * If this space cluster has been marked as fragmented and we've
2724                  * unpinned enough in this block group to potentially allow a
2725                  * cluster to be created inside of it go ahead and clear the
2726                  * fragmented check.
2727                  */
2728                 if (cluster && cluster->fragmented &&
2729                     total_unpinned > empty_cluster) {
2730                         spin_lock(&cluster->lock);
2731                         cluster->fragmented = 0;
2732                         spin_unlock(&cluster->lock);
2733                 }
2734
2735                 spin_lock(&space_info->lock);
2736                 spin_lock(&cache->lock);
2737                 cache->pinned -= len;
2738                 btrfs_space_info_update_bytes_pinned(fs_info, space_info, -len);
2739                 space_info->max_extent_size = 0;
2740                 if (cache->ro) {
2741                         space_info->bytes_readonly += len;
2742                         readonly = true;
2743                 } else if (btrfs_is_zoned(fs_info)) {
2744                         /* Need reset before reusing in a zoned block group */
2745                         space_info->bytes_zone_unusable += len;
2746                         readonly = true;
2747                 }
2748                 spin_unlock(&cache->lock);
2749                 if (!readonly && return_free_space &&
2750                     global_rsv->space_info == space_info) {
2751                         spin_lock(&global_rsv->lock);
2752                         if (!global_rsv->full) {
2753                                 u64 to_add = min(len, global_rsv->size -
2754                                                       global_rsv->reserved);
2755
2756                                 global_rsv->reserved += to_add;
2757                                 btrfs_space_info_update_bytes_may_use(fs_info,
2758                                                 space_info, to_add);
2759                                 if (global_rsv->reserved >= global_rsv->size)
2760                                         global_rsv->full = 1;
2761                                 len -= to_add;
2762                         }
2763                         spin_unlock(&global_rsv->lock);
2764                 }
2765                 /* Add to any tickets we may have */
2766                 if (!readonly && return_free_space && len)
2767                         btrfs_try_granting_tickets(fs_info, space_info);
2768                 spin_unlock(&space_info->lock);
2769         }
2770
2771         if (cache)
2772                 btrfs_put_block_group(cache);
2773         return 0;
2774 }
2775
2776 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans)
2777 {
2778         struct btrfs_fs_info *fs_info = trans->fs_info;
2779         struct btrfs_block_group *block_group, *tmp;
2780         struct list_head *deleted_bgs;
2781         struct extent_io_tree *unpin;
2782         u64 start;
2783         u64 end;
2784         int ret;
2785
2786         unpin = &trans->transaction->pinned_extents;
2787
2788         while (!TRANS_ABORTED(trans)) {
2789                 struct extent_state *cached_state = NULL;
2790
2791                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
2792                 if (!find_first_extent_bit(unpin, 0, &start, &end,
2793                                            EXTENT_DIRTY, &cached_state)) {
2794                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
2795                         break;
2796                 }
2797
2798                 if (btrfs_test_opt(fs_info, DISCARD_SYNC))
2799                         ret = btrfs_discard_extent(fs_info, start,
2800                                                    end + 1 - start, NULL);
2801
2802                 clear_extent_dirty(unpin, start, end, &cached_state);
2803                 unpin_extent_range(fs_info, start, end, true);
2804                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
2805                 free_extent_state(cached_state);
2806                 cond_resched();
2807         }
2808
2809         if (btrfs_test_opt(fs_info, DISCARD_ASYNC)) {
2810                 btrfs_discard_calc_delay(&fs_info->discard_ctl);
2811                 btrfs_discard_schedule_work(&fs_info->discard_ctl, true);
2812         }
2813
2814         /*
2815          * Transaction is finished.  We don't need the lock anymore.  We
2816          * do need to clean up the block groups in case of a transaction
2817          * abort.
2818          */
2819         deleted_bgs = &trans->transaction->deleted_bgs;
2820         list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
2821                 u64 trimmed = 0;
2822
2823                 ret = -EROFS;
2824                 if (!TRANS_ABORTED(trans))
2825                         ret = btrfs_discard_extent(fs_info,
2826                                                    block_group->start,
2827                                                    block_group->length,
2828                                                    &trimmed);
2829
2830                 list_del_init(&block_group->bg_list);
2831                 btrfs_unfreeze_block_group(block_group);
2832                 btrfs_put_block_group(block_group);
2833
2834                 if (ret) {
2835                         const char *errstr = btrfs_decode_error(ret);
2836                         btrfs_warn(fs_info,
2837                            "discard failed while removing blockgroup: errno=%d %s",
2838                                    ret, errstr);
2839                 }
2840         }
2841
2842         return 0;
2843 }
2844
2845 static int do_free_extent_accounting(struct btrfs_trans_handle *trans,
2846                                      u64 bytenr, u64 num_bytes, bool is_data)
2847 {
2848         int ret;
2849
2850         if (is_data) {
2851                 struct btrfs_root *csum_root;
2852
2853                 csum_root = btrfs_csum_root(trans->fs_info, bytenr);
2854                 ret = btrfs_del_csums(trans, csum_root, bytenr, num_bytes);
2855                 if (ret) {
2856                         btrfs_abort_transaction(trans, ret);
2857                         return ret;
2858                 }
2859         }
2860
2861         ret = add_to_free_space_tree(trans, bytenr, num_bytes);
2862         if (ret) {
2863                 btrfs_abort_transaction(trans, ret);
2864                 return ret;
2865         }
2866
2867         ret = btrfs_update_block_group(trans, bytenr, num_bytes, false);
2868         if (ret)
2869                 btrfs_abort_transaction(trans, ret);
2870
2871         return ret;
2872 }
2873
2874 #define abort_and_dump(trans, path, fmt, args...)       \
2875 ({                                                      \
2876         btrfs_abort_transaction(trans, -EUCLEAN);       \
2877         btrfs_print_leaf(path->nodes[0]);               \
2878         btrfs_crit(trans->fs_info, fmt, ##args);        \
2879 })
2880
2881 /*
2882  * Drop one or more refs of @node.
2883  *
2884  * 1. Locate the extent refs.
2885  *    It's either inline in EXTENT/METADATA_ITEM or in keyed SHARED_* item.
2886  *    Locate it, then reduce the refs number or remove the ref line completely.
2887  *
2888  * 2. Update the refs count in EXTENT/METADATA_ITEM
2889  *
2890  * Inline backref case:
2891  *
2892  * in extent tree we have:
2893  *
2894  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 16201 itemsize 82
2895  *              refs 2 gen 6 flags DATA
2896  *              extent data backref root FS_TREE objectid 258 offset 0 count 1
2897  *              extent data backref root FS_TREE objectid 257 offset 0 count 1
2898  *
2899  * This function gets called with:
2900  *
2901  *    node->bytenr = 13631488
2902  *    node->num_bytes = 1048576
2903  *    root_objectid = FS_TREE
2904  *    owner_objectid = 257
2905  *    owner_offset = 0
2906  *    refs_to_drop = 1
2907  *
2908  * Then we should get some like:
2909  *
2910  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 16201 itemsize 82
2911  *              refs 1 gen 6 flags DATA
2912  *              extent data backref root FS_TREE objectid 258 offset 0 count 1
2913  *
2914  * Keyed backref case:
2915  *
2916  * in extent tree we have:
2917  *
2918  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 3971 itemsize 24
2919  *              refs 754 gen 6 flags DATA
2920  *      [...]
2921  *      item 2 key (13631488 EXTENT_DATA_REF <HASH>) itemoff 3915 itemsize 28
2922  *              extent data backref root FS_TREE objectid 866 offset 0 count 1
2923  *
2924  * This function get called with:
2925  *
2926  *    node->bytenr = 13631488
2927  *    node->num_bytes = 1048576
2928  *    root_objectid = FS_TREE
2929  *    owner_objectid = 866
2930  *    owner_offset = 0
2931  *    refs_to_drop = 1
2932  *
2933  * Then we should get some like:
2934  *
2935  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 3971 itemsize 24
2936  *              refs 753 gen 6 flags DATA
2937  *
2938  * And that (13631488 EXTENT_DATA_REF <HASH>) gets removed.
2939  */
2940 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
2941                                struct btrfs_delayed_ref_node *node, u64 parent,
2942                                u64 root_objectid, u64 owner_objectid,
2943                                u64 owner_offset,
2944                                struct btrfs_delayed_extent_op *extent_op)
2945 {
2946         struct btrfs_fs_info *info = trans->fs_info;
2947         struct btrfs_key key;
2948         struct btrfs_path *path;
2949         struct btrfs_root *extent_root;
2950         struct extent_buffer *leaf;
2951         struct btrfs_extent_item *ei;
2952         struct btrfs_extent_inline_ref *iref;
2953         int ret;
2954         int is_data;
2955         int extent_slot = 0;
2956         int found_extent = 0;
2957         int num_to_del = 1;
2958         int refs_to_drop = node->ref_mod;
2959         u32 item_size;
2960         u64 refs;
2961         u64 bytenr = node->bytenr;
2962         u64 num_bytes = node->num_bytes;
2963         bool skinny_metadata = btrfs_fs_incompat(info, SKINNY_METADATA);
2964
2965         extent_root = btrfs_extent_root(info, bytenr);
2966         ASSERT(extent_root);
2967
2968         path = btrfs_alloc_path();
2969         if (!path)
2970                 return -ENOMEM;
2971
2972         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
2973
2974         if (!is_data && refs_to_drop != 1) {
2975                 btrfs_crit(info,
2976 "invalid refs_to_drop, dropping more than 1 refs for tree block %llu refs_to_drop %u",
2977                            node->bytenr, refs_to_drop);
2978                 ret = -EINVAL;
2979                 btrfs_abort_transaction(trans, ret);
2980                 goto out;
2981         }
2982
2983         if (is_data)
2984                 skinny_metadata = false;
2985
2986         ret = lookup_extent_backref(trans, path, &iref, bytenr, num_bytes,
2987                                     parent, root_objectid, owner_objectid,
2988                                     owner_offset);
2989         if (ret == 0) {
2990                 /*
2991                  * Either the inline backref or the SHARED_DATA_REF/
2992                  * SHARED_BLOCK_REF is found
2993                  *
2994                  * Here is a quick path to locate EXTENT/METADATA_ITEM.
2995                  * It's possible the EXTENT/METADATA_ITEM is near current slot.
2996                  */
2997                 extent_slot = path->slots[0];
2998                 while (extent_slot >= 0) {
2999                         btrfs_item_key_to_cpu(path->nodes[0], &key,
3000                                               extent_slot);
3001                         if (key.objectid != bytenr)
3002                                 break;
3003                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
3004                             key.offset == num_bytes) {
3005                                 found_extent = 1;
3006                                 break;
3007                         }
3008                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
3009                             key.offset == owner_objectid) {
3010                                 found_extent = 1;
3011                                 break;
3012                         }
3013
3014                         /* Quick path didn't find the EXTEMT/METADATA_ITEM */
3015                         if (path->slots[0] - extent_slot > 5)
3016                                 break;
3017                         extent_slot--;
3018                 }
3019
3020                 if (!found_extent) {
3021                         if (iref) {
3022                                 abort_and_dump(trans, path,
3023 "invalid iref slot %u, no EXTENT/METADATA_ITEM found but has inline extent ref",
3024                                            path->slots[0]);
3025                                 ret = -EUCLEAN;
3026                                 goto out;
3027                         }
3028                         /* Must be SHARED_* item, remove the backref first */
3029                         ret = remove_extent_backref(trans, extent_root, path,
3030                                                     NULL, refs_to_drop, is_data);
3031                         if (ret) {
3032                                 btrfs_abort_transaction(trans, ret);
3033                                 goto out;
3034                         }
3035                         btrfs_release_path(path);
3036
3037                         /* Slow path to locate EXTENT/METADATA_ITEM */
3038                         key.objectid = bytenr;
3039                         key.type = BTRFS_EXTENT_ITEM_KEY;
3040                         key.offset = num_bytes;
3041
3042                         if (!is_data && skinny_metadata) {
3043                                 key.type = BTRFS_METADATA_ITEM_KEY;
3044                                 key.offset = owner_objectid;
3045                         }
3046
3047                         ret = btrfs_search_slot(trans, extent_root,
3048                                                 &key, path, -1, 1);
3049                         if (ret > 0 && skinny_metadata && path->slots[0]) {
3050                                 /*
3051                                  * Couldn't find our skinny metadata item,
3052                                  * see if we have ye olde extent item.
3053                                  */
3054                                 path->slots[0]--;
3055                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
3056                                                       path->slots[0]);
3057                                 if (key.objectid == bytenr &&
3058                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
3059                                     key.offset == num_bytes)
3060                                         ret = 0;
3061                         }
3062
3063                         if (ret > 0 && skinny_metadata) {
3064                                 skinny_metadata = false;
3065                                 key.objectid = bytenr;
3066                                 key.type = BTRFS_EXTENT_ITEM_KEY;
3067                                 key.offset = num_bytes;
3068                                 btrfs_release_path(path);
3069                                 ret = btrfs_search_slot(trans, extent_root,
3070                                                         &key, path, -1, 1);
3071                         }
3072
3073                         if (ret) {
3074                                 if (ret > 0)
3075                                         btrfs_print_leaf(path->nodes[0]);
3076                                 btrfs_err(info,
3077                         "umm, got %d back from search, was looking for %llu, slot %d",
3078                                           ret, bytenr, path->slots[0]);
3079                         }
3080                         if (ret < 0) {
3081                                 btrfs_abort_transaction(trans, ret);
3082                                 goto out;
3083                         }
3084                         extent_slot = path->slots[0];
3085                 }
3086         } else if (WARN_ON(ret == -ENOENT)) {
3087                 abort_and_dump(trans, path,
3088 "unable to find ref byte nr %llu parent %llu root %llu owner %llu offset %llu slot %d",
3089                                bytenr, parent, root_objectid, owner_objectid,
3090                                owner_offset, path->slots[0]);
3091                 goto out;
3092         } else {
3093                 btrfs_abort_transaction(trans, ret);
3094                 goto out;
3095         }
3096
3097         leaf = path->nodes[0];
3098         item_size = btrfs_item_size(leaf, extent_slot);
3099         if (unlikely(item_size < sizeof(*ei))) {
3100                 ret = -EUCLEAN;
3101                 btrfs_err(trans->fs_info,
3102                           "unexpected extent item size, has %u expect >= %zu",
3103                           item_size, sizeof(*ei));
3104                 btrfs_abort_transaction(trans, ret);
3105                 goto out;
3106         }
3107         ei = btrfs_item_ptr(leaf, extent_slot,
3108                             struct btrfs_extent_item);
3109         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
3110             key.type == BTRFS_EXTENT_ITEM_KEY) {
3111                 struct btrfs_tree_block_info *bi;
3112
3113                 if (item_size < sizeof(*ei) + sizeof(*bi)) {
3114                         abort_and_dump(trans, path,
3115 "invalid extent item size for key (%llu, %u, %llu) slot %u owner %llu, has %u expect >= %zu",
3116                                        key.objectid, key.type, key.offset,
3117                                        path->slots[0], owner_objectid, item_size,
3118                                        sizeof(*ei) + sizeof(*bi));
3119                         ret = -EUCLEAN;
3120                         goto out;
3121                 }
3122                 bi = (struct btrfs_tree_block_info *)(ei + 1);
3123                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
3124         }
3125
3126         refs = btrfs_extent_refs(leaf, ei);
3127         if (refs < refs_to_drop) {
3128                 abort_and_dump(trans, path,
3129                 "trying to drop %d refs but we only have %llu for bytenr %llu slot %u",
3130                                refs_to_drop, refs, bytenr, path->slots[0]);
3131                 ret = -EUCLEAN;
3132                 goto out;
3133         }
3134         refs -= refs_to_drop;
3135
3136         if (refs > 0) {
3137                 if (extent_op)
3138                         __run_delayed_extent_op(extent_op, leaf, ei);
3139                 /*
3140                  * In the case of inline back ref, reference count will
3141                  * be updated by remove_extent_backref
3142                  */
3143                 if (iref) {
3144                         if (!found_extent) {
3145                                 abort_and_dump(trans, path,
3146 "invalid iref, got inlined extent ref but no EXTENT/METADATA_ITEM found, slot %u",
3147                                                path->slots[0]);
3148                                 ret = -EUCLEAN;
3149                                 goto out;
3150                         }
3151                 } else {
3152                         btrfs_set_extent_refs(leaf, ei, refs);
3153                         btrfs_mark_buffer_dirty(trans, leaf);
3154                 }
3155                 if (found_extent) {
3156                         ret = remove_extent_backref(trans, extent_root, path,
3157                                                     iref, refs_to_drop, is_data);
3158                         if (ret) {
3159                                 btrfs_abort_transaction(trans, ret);
3160                                 goto out;
3161                         }
3162                 }
3163         } else {
3164                 /* In this branch refs == 1 */
3165                 if (found_extent) {
3166                         if (is_data && refs_to_drop !=
3167                             extent_data_ref_count(path, iref)) {
3168                                 abort_and_dump(trans, path,
3169                 "invalid refs_to_drop, current refs %u refs_to_drop %u slot %u",
3170                                                extent_data_ref_count(path, iref),
3171                                                refs_to_drop, path->slots[0]);
3172                                 ret = -EUCLEAN;
3173                                 goto out;
3174                         }
3175                         if (iref) {
3176                                 if (path->slots[0] != extent_slot) {
3177                                         abort_and_dump(trans, path,
3178 "invalid iref, extent item key (%llu %u %llu) slot %u doesn't have wanted iref",
3179                                                        key.objectid, key.type,
3180                                                        key.offset, path->slots[0]);
3181                                         ret = -EUCLEAN;
3182                                         goto out;
3183                                 }
3184                         } else {
3185                                 /*
3186                                  * No inline ref, we must be at SHARED_* item,
3187                                  * And it's single ref, it must be:
3188                                  * |    extent_slot       ||extent_slot + 1|
3189                                  * [ EXTENT/METADATA_ITEM ][ SHARED_* ITEM ]
3190                                  */
3191                                 if (path->slots[0] != extent_slot + 1) {
3192                                         abort_and_dump(trans, path,
3193         "invalid SHARED_* item slot %u, previous item is not EXTENT/METADATA_ITEM",
3194                                                        path->slots[0]);
3195                                         ret = -EUCLEAN;
3196                                         goto out;
3197                                 }
3198                                 path->slots[0] = extent_slot;
3199                                 num_to_del = 2;
3200                         }
3201                 }
3202
3203                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
3204                                       num_to_del);
3205                 if (ret) {
3206                         btrfs_abort_transaction(trans, ret);
3207                         goto out;
3208                 }
3209                 btrfs_release_path(path);
3210
3211                 ret = do_free_extent_accounting(trans, bytenr, num_bytes, is_data);
3212         }
3213         btrfs_release_path(path);
3214
3215 out:
3216         btrfs_free_path(path);
3217         return ret;
3218 }
3219
3220 /*
3221  * when we free an block, it is possible (and likely) that we free the last
3222  * delayed ref for that extent as well.  This searches the delayed ref tree for
3223  * a given extent, and if there are no other delayed refs to be processed, it
3224  * removes it from the tree.
3225  */
3226 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
3227                                       u64 bytenr)
3228 {
3229         struct btrfs_delayed_ref_head *head;
3230         struct btrfs_delayed_ref_root *delayed_refs;
3231         int ret = 0;
3232
3233         delayed_refs = &trans->transaction->delayed_refs;
3234         spin_lock(&delayed_refs->lock);
3235         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
3236         if (!head)
3237                 goto out_delayed_unlock;
3238
3239         spin_lock(&head->lock);
3240         if (!RB_EMPTY_ROOT(&head->ref_tree.rb_root))
3241                 goto out;
3242
3243         if (cleanup_extent_op(head) != NULL)
3244                 goto out;
3245
3246         /*
3247          * waiting for the lock here would deadlock.  If someone else has it
3248          * locked they are already in the process of dropping it anyway
3249          */
3250         if (!mutex_trylock(&head->mutex))
3251                 goto out;
3252
3253         btrfs_delete_ref_head(delayed_refs, head);
3254         head->processing = false;
3255
3256         spin_unlock(&head->lock);
3257         spin_unlock(&delayed_refs->lock);
3258
3259         BUG_ON(head->extent_op);
3260         if (head->must_insert_reserved)
3261                 ret = 1;
3262
3263         btrfs_cleanup_ref_head_accounting(trans->fs_info, delayed_refs, head);
3264         mutex_unlock(&head->mutex);
3265         btrfs_put_delayed_ref_head(head);
3266         return ret;
3267 out:
3268         spin_unlock(&head->lock);
3269
3270 out_delayed_unlock:
3271         spin_unlock(&delayed_refs->lock);
3272         return 0;
3273 }
3274
3275 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
3276                            u64 root_id,
3277                            struct extent_buffer *buf,
3278                            u64 parent, int last_ref)
3279 {
3280         struct btrfs_fs_info *fs_info = trans->fs_info;
3281         struct btrfs_ref generic_ref = { 0 };
3282         int ret;
3283
3284         btrfs_init_generic_ref(&generic_ref, BTRFS_DROP_DELAYED_REF,
3285                                buf->start, buf->len, parent);
3286         btrfs_init_tree_ref(&generic_ref, btrfs_header_level(buf),
3287                             root_id, 0, false);
3288
3289         if (root_id != BTRFS_TREE_LOG_OBJECTID) {
3290                 btrfs_ref_tree_mod(fs_info, &generic_ref);
3291                 ret = btrfs_add_delayed_tree_ref(trans, &generic_ref, NULL);
3292                 BUG_ON(ret); /* -ENOMEM */
3293         }
3294
3295         if (last_ref && btrfs_header_generation(buf) == trans->transid) {
3296                 struct btrfs_block_group *cache;
3297                 bool must_pin = false;
3298
3299                 if (root_id != BTRFS_TREE_LOG_OBJECTID) {
3300                         ret = check_ref_cleanup(trans, buf->start);
3301                         if (!ret) {
3302                                 btrfs_redirty_list_add(trans->transaction, buf);
3303                                 goto out;
3304                         }
3305                 }
3306
3307                 cache = btrfs_lookup_block_group(fs_info, buf->start);
3308
3309                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
3310                         pin_down_extent(trans, cache, buf->start, buf->len, 1);
3311                         btrfs_put_block_group(cache);
3312                         goto out;
3313                 }
3314
3315                 /*
3316                  * If there are tree mod log users we may have recorded mod log
3317                  * operations for this node.  If we re-allocate this node we
3318                  * could replay operations on this node that happened when it
3319                  * existed in a completely different root.  For example if it
3320                  * was part of root A, then was reallocated to root B, and we
3321                  * are doing a btrfs_old_search_slot(root b), we could replay
3322                  * operations that happened when the block was part of root A,
3323                  * giving us an inconsistent view of the btree.
3324                  *
3325                  * We are safe from races here because at this point no other
3326                  * node or root points to this extent buffer, so if after this
3327                  * check a new tree mod log user joins we will not have an
3328                  * existing log of operations on this node that we have to
3329                  * contend with.
3330                  */
3331                 if (test_bit(BTRFS_FS_TREE_MOD_LOG_USERS, &fs_info->flags))
3332                         must_pin = true;
3333
3334                 if (must_pin || btrfs_is_zoned(fs_info)) {
3335                         btrfs_redirty_list_add(trans->transaction, buf);
3336                         pin_down_extent(trans, cache, buf->start, buf->len, 1);
3337                         btrfs_put_block_group(cache);
3338                         goto out;
3339                 }
3340
3341                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
3342
3343                 btrfs_add_free_space(cache, buf->start, buf->len);
3344                 btrfs_free_reserved_bytes(cache, buf->len, 0);
3345                 btrfs_put_block_group(cache);
3346                 trace_btrfs_reserved_extent_free(fs_info, buf->start, buf->len);
3347         }
3348 out:
3349         if (last_ref) {
3350                 /*
3351                  * Deleting the buffer, clear the corrupt flag since it doesn't
3352                  * matter anymore.
3353                  */
3354                 clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
3355         }
3356 }
3357
3358 /* Can return -ENOMEM */
3359 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_ref *ref)
3360 {
3361         struct btrfs_fs_info *fs_info = trans->fs_info;
3362         int ret;
3363
3364         if (btrfs_is_testing(fs_info))
3365                 return 0;
3366
3367         /*
3368          * tree log blocks never actually go into the extent allocation
3369          * tree, just update pinning info and exit early.
3370          */
3371         if ((ref->type == BTRFS_REF_METADATA &&
3372              ref->tree_ref.owning_root == BTRFS_TREE_LOG_OBJECTID) ||
3373             (ref->type == BTRFS_REF_DATA &&
3374              ref->data_ref.owning_root == BTRFS_TREE_LOG_OBJECTID)) {
3375                 btrfs_pin_extent(trans, ref->bytenr, ref->len, 1);
3376                 ret = 0;
3377         } else if (ref->type == BTRFS_REF_METADATA) {
3378                 ret = btrfs_add_delayed_tree_ref(trans, ref, NULL);
3379         } else {
3380                 ret = btrfs_add_delayed_data_ref(trans, ref, 0);
3381         }
3382
3383         if (!((ref->type == BTRFS_REF_METADATA &&
3384                ref->tree_ref.owning_root == BTRFS_TREE_LOG_OBJECTID) ||
3385               (ref->type == BTRFS_REF_DATA &&
3386                ref->data_ref.owning_root == BTRFS_TREE_LOG_OBJECTID)))
3387                 btrfs_ref_tree_mod(fs_info, ref);
3388
3389         return ret;
3390 }
3391
3392 enum btrfs_loop_type {
3393         /*
3394          * Start caching block groups but do not wait for progress or for them
3395          * to be done.
3396          */
3397         LOOP_CACHING_NOWAIT,
3398
3399         /*
3400          * Wait for the block group free_space >= the space we're waiting for if
3401          * the block group isn't cached.
3402          */
3403         LOOP_CACHING_WAIT,
3404
3405         /*
3406          * Allow allocations to happen from block groups that do not yet have a
3407          * size classification.
3408          */
3409         LOOP_UNSET_SIZE_CLASS,
3410
3411         /*
3412          * Allocate a chunk and then retry the allocation.
3413          */
3414         LOOP_ALLOC_CHUNK,
3415
3416         /*
3417          * Ignore the size class restrictions for this allocation.
3418          */
3419         LOOP_WRONG_SIZE_CLASS,
3420
3421         /*
3422          * Ignore the empty size, only try to allocate the number of bytes
3423          * needed for this allocation.
3424          */
3425         LOOP_NO_EMPTY_SIZE,
3426 };
3427
3428 static inline void
3429 btrfs_lock_block_group(struct btrfs_block_group *cache,
3430                        int delalloc)
3431 {
3432         if (delalloc)
3433                 down_read(&cache->data_rwsem);
3434 }
3435
3436 static inline void btrfs_grab_block_group(struct btrfs_block_group *cache,
3437                        int delalloc)
3438 {
3439         btrfs_get_block_group(cache);
3440         if (delalloc)
3441                 down_read(&cache->data_rwsem);
3442 }
3443
3444 static struct btrfs_block_group *btrfs_lock_cluster(
3445                    struct btrfs_block_group *block_group,
3446                    struct btrfs_free_cluster *cluster,
3447                    int delalloc)
3448         __acquires(&cluster->refill_lock)
3449 {
3450         struct btrfs_block_group *used_bg = NULL;
3451
3452         spin_lock(&cluster->refill_lock);
3453         while (1) {
3454                 used_bg = cluster->block_group;
3455                 if (!used_bg)
3456                         return NULL;
3457
3458                 if (used_bg == block_group)
3459                         return used_bg;
3460
3461                 btrfs_get_block_group(used_bg);
3462
3463                 if (!delalloc)
3464                         return used_bg;
3465
3466                 if (down_read_trylock(&used_bg->data_rwsem))
3467                         return used_bg;
3468
3469                 spin_unlock(&cluster->refill_lock);
3470
3471                 /* We should only have one-level nested. */
3472                 down_read_nested(&used_bg->data_rwsem, SINGLE_DEPTH_NESTING);
3473
3474                 spin_lock(&cluster->refill_lock);
3475                 if (used_bg == cluster->block_group)
3476                         return used_bg;
3477
3478                 up_read(&used_bg->data_rwsem);
3479                 btrfs_put_block_group(used_bg);
3480         }
3481 }
3482
3483 static inline void
3484 btrfs_release_block_group(struct btrfs_block_group *cache,
3485                          int delalloc)
3486 {
3487         if (delalloc)
3488                 up_read(&cache->data_rwsem);
3489         btrfs_put_block_group(cache);
3490 }
3491
3492 /*
3493  * Helper function for find_free_extent().
3494  *
3495  * Return -ENOENT to inform caller that we need fallback to unclustered mode.
3496  * Return >0 to inform caller that we find nothing
3497  * Return 0 means we have found a location and set ffe_ctl->found_offset.
3498  */
3499 static int find_free_extent_clustered(struct btrfs_block_group *bg,
3500                                       struct find_free_extent_ctl *ffe_ctl,
3501                                       struct btrfs_block_group **cluster_bg_ret)
3502 {
3503         struct btrfs_block_group *cluster_bg;
3504         struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
3505         u64 aligned_cluster;
3506         u64 offset;
3507         int ret;
3508
3509         cluster_bg = btrfs_lock_cluster(bg, last_ptr, ffe_ctl->delalloc);
3510         if (!cluster_bg)
3511                 goto refill_cluster;
3512         if (cluster_bg != bg && (cluster_bg->ro ||
3513             !block_group_bits(cluster_bg, ffe_ctl->flags)))
3514                 goto release_cluster;
3515
3516         offset = btrfs_alloc_from_cluster(cluster_bg, last_ptr,
3517                         ffe_ctl->num_bytes, cluster_bg->start,
3518                         &ffe_ctl->max_extent_size);
3519         if (offset) {
3520                 /* We have a block, we're done */
3521                 spin_unlock(&last_ptr->refill_lock);
3522                 trace_btrfs_reserve_extent_cluster(cluster_bg, ffe_ctl);
3523                 *cluster_bg_ret = cluster_bg;
3524                 ffe_ctl->found_offset = offset;
3525                 return 0;
3526         }
3527         WARN_ON(last_ptr->block_group != cluster_bg);
3528
3529 release_cluster:
3530         /*
3531          * If we are on LOOP_NO_EMPTY_SIZE, we can't set up a new clusters, so
3532          * lets just skip it and let the allocator find whatever block it can
3533          * find. If we reach this point, we will have tried the cluster
3534          * allocator plenty of times and not have found anything, so we are
3535          * likely way too fragmented for the clustering stuff to find anything.
3536          *
3537          * However, if the cluster is taken from the current block group,
3538          * release the cluster first, so that we stand a better chance of
3539          * succeeding in the unclustered allocation.
3540          */
3541         if (ffe_ctl->loop >= LOOP_NO_EMPTY_SIZE && cluster_bg != bg) {
3542                 spin_unlock(&last_ptr->refill_lock);
3543                 btrfs_release_block_group(cluster_bg, ffe_ctl->delalloc);
3544                 return -ENOENT;
3545         }
3546
3547         /* This cluster didn't work out, free it and start over */
3548         btrfs_return_cluster_to_free_space(NULL, last_ptr);
3549
3550         if (cluster_bg != bg)
3551                 btrfs_release_block_group(cluster_bg, ffe_ctl->delalloc);
3552
3553 refill_cluster:
3554         if (ffe_ctl->loop >= LOOP_NO_EMPTY_SIZE) {
3555                 spin_unlock(&last_ptr->refill_lock);
3556                 return -ENOENT;
3557         }
3558
3559         aligned_cluster = max_t(u64,
3560                         ffe_ctl->empty_cluster + ffe_ctl->empty_size,
3561                         bg->full_stripe_len);
3562         ret = btrfs_find_space_cluster(bg, last_ptr, ffe_ctl->search_start,
3563                         ffe_ctl->num_bytes, aligned_cluster);
3564         if (ret == 0) {
3565                 /* Now pull our allocation out of this cluster */
3566                 offset = btrfs_alloc_from_cluster(bg, last_ptr,
3567                                 ffe_ctl->num_bytes, ffe_ctl->search_start,
3568                                 &ffe_ctl->max_extent_size);
3569                 if (offset) {
3570                         /* We found one, proceed */
3571                         spin_unlock(&last_ptr->refill_lock);
3572                         ffe_ctl->found_offset = offset;
3573                         trace_btrfs_reserve_extent_cluster(bg, ffe_ctl);
3574                         return 0;
3575                 }
3576         }
3577         /*
3578          * At this point we either didn't find a cluster or we weren't able to
3579          * allocate a block from our cluster.  Free the cluster we've been
3580          * trying to use, and go to the next block group.
3581          */
3582         btrfs_return_cluster_to_free_space(NULL, last_ptr);
3583         spin_unlock(&last_ptr->refill_lock);
3584         return 1;
3585 }
3586
3587 /*
3588  * Return >0 to inform caller that we find nothing
3589  * Return 0 when we found an free extent and set ffe_ctrl->found_offset
3590  */
3591 static int find_free_extent_unclustered(struct btrfs_block_group *bg,
3592                                         struct find_free_extent_ctl *ffe_ctl)
3593 {
3594         struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
3595         u64 offset;
3596
3597         /*
3598          * We are doing an unclustered allocation, set the fragmented flag so
3599          * we don't bother trying to setup a cluster again until we get more
3600          * space.
3601          */
3602         if (unlikely(last_ptr)) {
3603                 spin_lock(&last_ptr->lock);
3604                 last_ptr->fragmented = 1;
3605                 spin_unlock(&last_ptr->lock);
3606         }
3607         if (ffe_ctl->cached) {
3608                 struct btrfs_free_space_ctl *free_space_ctl;
3609
3610                 free_space_ctl = bg->free_space_ctl;
3611                 spin_lock(&free_space_ctl->tree_lock);
3612                 if (free_space_ctl->free_space <
3613                     ffe_ctl->num_bytes + ffe_ctl->empty_cluster +
3614                     ffe_ctl->empty_size) {
3615                         ffe_ctl->total_free_space = max_t(u64,
3616                                         ffe_ctl->total_free_space,
3617                                         free_space_ctl->free_space);
3618                         spin_unlock(&free_space_ctl->tree_lock);
3619                         return 1;
3620                 }
3621                 spin_unlock(&free_space_ctl->tree_lock);
3622         }
3623
3624         offset = btrfs_find_space_for_alloc(bg, ffe_ctl->search_start,
3625                         ffe_ctl->num_bytes, ffe_ctl->empty_size,
3626                         &ffe_ctl->max_extent_size);
3627         if (!offset)
3628                 return 1;
3629         ffe_ctl->found_offset = offset;
3630         return 0;
3631 }
3632
3633 static int do_allocation_clustered(struct btrfs_block_group *block_group,
3634                                    struct find_free_extent_ctl *ffe_ctl,
3635                                    struct btrfs_block_group **bg_ret)
3636 {
3637         int ret;
3638
3639         /* We want to try and use the cluster allocator, so lets look there */
3640         if (ffe_ctl->last_ptr && ffe_ctl->use_cluster) {
3641                 ret = find_free_extent_clustered(block_group, ffe_ctl, bg_ret);
3642                 if (ret >= 0)
3643                         return ret;
3644                 /* ret == -ENOENT case falls through */
3645         }
3646
3647         return find_free_extent_unclustered(block_group, ffe_ctl);
3648 }
3649
3650 /*
3651  * Tree-log block group locking
3652  * ============================
3653  *
3654  * fs_info::treelog_bg_lock protects the fs_info::treelog_bg which
3655  * indicates the starting address of a block group, which is reserved only
3656  * for tree-log metadata.
3657  *
3658  * Lock nesting
3659  * ============
3660  *
3661  * space_info::lock
3662  *   block_group::lock
3663  *     fs_info::treelog_bg_lock
3664  */
3665
3666 /*
3667  * Simple allocator for sequential-only block group. It only allows sequential
3668  * allocation. No need to play with trees. This function also reserves the
3669  * bytes as in btrfs_add_reserved_bytes.
3670  */
3671 static int do_allocation_zoned(struct btrfs_block_group *block_group,
3672                                struct find_free_extent_ctl *ffe_ctl,
3673                                struct btrfs_block_group **bg_ret)
3674 {
3675         struct btrfs_fs_info *fs_info = block_group->fs_info;
3676         struct btrfs_space_info *space_info = block_group->space_info;
3677         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3678         u64 start = block_group->start;
3679         u64 num_bytes = ffe_ctl->num_bytes;
3680         u64 avail;
3681         u64 bytenr = block_group->start;
3682         u64 log_bytenr;
3683         u64 data_reloc_bytenr;
3684         int ret = 0;
3685         bool skip = false;
3686
3687         ASSERT(btrfs_is_zoned(block_group->fs_info));
3688
3689         /*
3690          * Do not allow non-tree-log blocks in the dedicated tree-log block
3691          * group, and vice versa.
3692          */
3693         spin_lock(&fs_info->treelog_bg_lock);
3694         log_bytenr = fs_info->treelog_bg;
3695         if (log_bytenr && ((ffe_ctl->for_treelog && bytenr != log_bytenr) ||
3696                            (!ffe_ctl->for_treelog && bytenr == log_bytenr)))
3697                 skip = true;
3698         spin_unlock(&fs_info->treelog_bg_lock);
3699         if (skip)
3700                 return 1;
3701
3702         /*
3703          * Do not allow non-relocation blocks in the dedicated relocation block
3704          * group, and vice versa.
3705          */
3706         spin_lock(&fs_info->relocation_bg_lock);
3707         data_reloc_bytenr = fs_info->data_reloc_bg;
3708         if (data_reloc_bytenr &&
3709             ((ffe_ctl->for_data_reloc && bytenr != data_reloc_bytenr) ||
3710              (!ffe_ctl->for_data_reloc && bytenr == data_reloc_bytenr)))
3711                 skip = true;
3712         spin_unlock(&fs_info->relocation_bg_lock);
3713         if (skip)
3714                 return 1;
3715
3716         /* Check RO and no space case before trying to activate it */
3717         spin_lock(&block_group->lock);
3718         if (block_group->ro || btrfs_zoned_bg_is_full(block_group)) {
3719                 ret = 1;
3720                 /*
3721                  * May need to clear fs_info->{treelog,data_reloc}_bg.
3722                  * Return the error after taking the locks.
3723                  */
3724         }
3725         spin_unlock(&block_group->lock);
3726
3727         /* Metadata block group is activated at write time. */
3728         if (!ret && (block_group->flags & BTRFS_BLOCK_GROUP_DATA) &&
3729             !btrfs_zone_activate(block_group)) {
3730                 ret = 1;
3731                 /*
3732                  * May need to clear fs_info->{treelog,data_reloc}_bg.
3733                  * Return the error after taking the locks.
3734                  */
3735         }
3736
3737         spin_lock(&space_info->lock);
3738         spin_lock(&block_group->lock);
3739         spin_lock(&fs_info->treelog_bg_lock);
3740         spin_lock(&fs_info->relocation_bg_lock);
3741
3742         if (ret)
3743                 goto out;
3744
3745         ASSERT(!ffe_ctl->for_treelog ||
3746                block_group->start == fs_info->treelog_bg ||
3747                fs_info->treelog_bg == 0);
3748         ASSERT(!ffe_ctl->for_data_reloc ||
3749                block_group->start == fs_info->data_reloc_bg ||
3750                fs_info->data_reloc_bg == 0);
3751
3752         if (block_group->ro ||
3753             (!ffe_ctl->for_data_reloc &&
3754              test_bit(BLOCK_GROUP_FLAG_ZONED_DATA_RELOC, &block_group->runtime_flags))) {
3755                 ret = 1;
3756                 goto out;
3757         }
3758
3759         /*
3760          * Do not allow currently using block group to be tree-log dedicated
3761          * block group.
3762          */
3763         if (ffe_ctl->for_treelog && !fs_info->treelog_bg &&
3764             (block_group->used || block_group->reserved)) {
3765                 ret = 1;
3766                 goto out;
3767         }
3768
3769         /*
3770          * Do not allow currently used block group to be the data relocation
3771          * dedicated block group.
3772          */
3773         if (ffe_ctl->for_data_reloc && !fs_info->data_reloc_bg &&
3774             (block_group->used || block_group->reserved)) {
3775                 ret = 1;
3776                 goto out;
3777         }
3778
3779         WARN_ON_ONCE(block_group->alloc_offset > block_group->zone_capacity);
3780         avail = block_group->zone_capacity - block_group->alloc_offset;
3781         if (avail < num_bytes) {
3782                 if (ffe_ctl->max_extent_size < avail) {
3783                         /*
3784                          * With sequential allocator, free space is always
3785                          * contiguous
3786                          */
3787                         ffe_ctl->max_extent_size = avail;
3788                         ffe_ctl->total_free_space = avail;
3789                 }
3790                 ret = 1;
3791                 goto out;
3792         }
3793
3794         if (ffe_ctl->for_treelog && !fs_info->treelog_bg)
3795                 fs_info->treelog_bg = block_group->start;
3796
3797         if (ffe_ctl->for_data_reloc) {
3798                 if (!fs_info->data_reloc_bg)
3799                         fs_info->data_reloc_bg = block_group->start;
3800                 /*
3801                  * Do not allow allocations from this block group, unless it is
3802                  * for data relocation. Compared to increasing the ->ro, setting
3803                  * the ->zoned_data_reloc_ongoing flag still allows nocow
3804                  * writers to come in. See btrfs_inc_nocow_writers().
3805                  *
3806                  * We need to disable an allocation to avoid an allocation of
3807                  * regular (non-relocation data) extent. With mix of relocation
3808                  * extents and regular extents, we can dispatch WRITE commands
3809                  * (for relocation extents) and ZONE APPEND commands (for
3810                  * regular extents) at the same time to the same zone, which
3811                  * easily break the write pointer.
3812                  *
3813                  * Also, this flag avoids this block group to be zone finished.
3814                  */
3815                 set_bit(BLOCK_GROUP_FLAG_ZONED_DATA_RELOC, &block_group->runtime_flags);
3816         }
3817
3818         ffe_ctl->found_offset = start + block_group->alloc_offset;
3819         block_group->alloc_offset += num_bytes;
3820         spin_lock(&ctl->tree_lock);
3821         ctl->free_space -= num_bytes;
3822         spin_unlock(&ctl->tree_lock);
3823
3824         /*
3825          * We do not check if found_offset is aligned to stripesize. The
3826          * address is anyway rewritten when using zone append writing.
3827          */
3828
3829         ffe_ctl->search_start = ffe_ctl->found_offset;
3830
3831 out:
3832         if (ret && ffe_ctl->for_treelog)
3833                 fs_info->treelog_bg = 0;
3834         if (ret && ffe_ctl->for_data_reloc)
3835                 fs_info->data_reloc_bg = 0;
3836         spin_unlock(&fs_info->relocation_bg_lock);
3837         spin_unlock(&fs_info->treelog_bg_lock);
3838         spin_unlock(&block_group->lock);
3839         spin_unlock(&space_info->lock);
3840         return ret;
3841 }
3842
3843 static int do_allocation(struct btrfs_block_group *block_group,
3844                          struct find_free_extent_ctl *ffe_ctl,
3845                          struct btrfs_block_group **bg_ret)
3846 {
3847         switch (ffe_ctl->policy) {
3848         case BTRFS_EXTENT_ALLOC_CLUSTERED:
3849                 return do_allocation_clustered(block_group, ffe_ctl, bg_ret);
3850         case BTRFS_EXTENT_ALLOC_ZONED:
3851                 return do_allocation_zoned(block_group, ffe_ctl, bg_ret);
3852         default:
3853                 BUG();
3854         }
3855 }
3856
3857 static void release_block_group(struct btrfs_block_group *block_group,
3858                                 struct find_free_extent_ctl *ffe_ctl,
3859                                 int delalloc)
3860 {
3861         switch (ffe_ctl->policy) {
3862         case BTRFS_EXTENT_ALLOC_CLUSTERED:
3863                 ffe_ctl->retry_uncached = false;
3864                 break;
3865         case BTRFS_EXTENT_ALLOC_ZONED:
3866                 /* Nothing to do */
3867                 break;
3868         default:
3869                 BUG();
3870         }
3871
3872         BUG_ON(btrfs_bg_flags_to_raid_index(block_group->flags) !=
3873                ffe_ctl->index);
3874         btrfs_release_block_group(block_group, delalloc);
3875 }
3876
3877 static void found_extent_clustered(struct find_free_extent_ctl *ffe_ctl,
3878                                    struct btrfs_key *ins)
3879 {
3880         struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
3881
3882         if (!ffe_ctl->use_cluster && last_ptr) {
3883                 spin_lock(&last_ptr->lock);
3884                 last_ptr->window_start = ins->objectid;
3885                 spin_unlock(&last_ptr->lock);
3886         }
3887 }
3888
3889 static void found_extent(struct find_free_extent_ctl *ffe_ctl,
3890                          struct btrfs_key *ins)
3891 {
3892         switch (ffe_ctl->policy) {
3893         case BTRFS_EXTENT_ALLOC_CLUSTERED:
3894                 found_extent_clustered(ffe_ctl, ins);
3895                 break;
3896         case BTRFS_EXTENT_ALLOC_ZONED:
3897                 /* Nothing to do */
3898                 break;
3899         default:
3900                 BUG();
3901         }
3902 }
3903
3904 static int can_allocate_chunk_zoned(struct btrfs_fs_info *fs_info,
3905                                     struct find_free_extent_ctl *ffe_ctl)
3906 {
3907         /* Block group's activeness is not a requirement for METADATA block groups. */
3908         if (!(ffe_ctl->flags & BTRFS_BLOCK_GROUP_DATA))
3909                 return 0;
3910
3911         /* If we can activate new zone, just allocate a chunk and use it */
3912         if (btrfs_can_activate_zone(fs_info->fs_devices, ffe_ctl->flags))
3913                 return 0;
3914
3915         /*
3916          * We already reached the max active zones. Try to finish one block
3917          * group to make a room for a new block group. This is only possible
3918          * for a data block group because btrfs_zone_finish() may need to wait
3919          * for a running transaction which can cause a deadlock for metadata
3920          * allocation.
3921          */
3922         if (ffe_ctl->flags & BTRFS_BLOCK_GROUP_DATA) {
3923                 int ret = btrfs_zone_finish_one_bg(fs_info);
3924
3925                 if (ret == 1)
3926                         return 0;
3927                 else if (ret < 0)
3928                         return ret;
3929         }
3930
3931         /*
3932          * If we have enough free space left in an already active block group
3933          * and we can't activate any other zone now, do not allow allocating a
3934          * new chunk and let find_free_extent() retry with a smaller size.
3935          */
3936         if (ffe_ctl->max_extent_size >= ffe_ctl->min_alloc_size)
3937                 return -ENOSPC;
3938
3939         /*
3940          * Even min_alloc_size is not left in any block groups. Since we cannot
3941          * activate a new block group, allocating it may not help. Let's tell a
3942          * caller to try again and hope it progress something by writing some
3943          * parts of the region. That is only possible for data block groups,
3944          * where a part of the region can be written.
3945          */
3946         if (ffe_ctl->flags & BTRFS_BLOCK_GROUP_DATA)
3947                 return -EAGAIN;
3948
3949         /*
3950          * We cannot activate a new block group and no enough space left in any
3951          * block groups. So, allocating a new block group may not help. But,
3952          * there is nothing to do anyway, so let's go with it.
3953          */
3954         return 0;
3955 }
3956
3957 static int can_allocate_chunk(struct btrfs_fs_info *fs_info,
3958                               struct find_free_extent_ctl *ffe_ctl)
3959 {
3960         switch (ffe_ctl->policy) {
3961         case BTRFS_EXTENT_ALLOC_CLUSTERED:
3962                 return 0;
3963         case BTRFS_EXTENT_ALLOC_ZONED:
3964                 return can_allocate_chunk_zoned(fs_info, ffe_ctl);
3965         default:
3966                 BUG();
3967         }
3968 }
3969
3970 /*
3971  * Return >0 means caller needs to re-search for free extent
3972  * Return 0 means we have the needed free extent.
3973  * Return <0 means we failed to locate any free extent.
3974  */
3975 static int find_free_extent_update_loop(struct btrfs_fs_info *fs_info,
3976                                         struct btrfs_key *ins,
3977                                         struct find_free_extent_ctl *ffe_ctl,
3978                                         bool full_search)
3979 {
3980         struct btrfs_root *root = fs_info->chunk_root;
3981         int ret;
3982
3983         if ((ffe_ctl->loop == LOOP_CACHING_NOWAIT) &&
3984             ffe_ctl->have_caching_bg && !ffe_ctl->orig_have_caching_bg)
3985                 ffe_ctl->orig_have_caching_bg = true;
3986
3987         if (ins->objectid) {
3988                 found_extent(ffe_ctl, ins);
3989                 return 0;
3990         }
3991
3992         if (ffe_ctl->loop >= LOOP_CACHING_WAIT && ffe_ctl->have_caching_bg)
3993                 return 1;
3994
3995         ffe_ctl->index++;
3996         if (ffe_ctl->index < BTRFS_NR_RAID_TYPES)
3997                 return 1;
3998
3999         /* See the comments for btrfs_loop_type for an explanation of the phases. */
4000         if (ffe_ctl->loop < LOOP_NO_EMPTY_SIZE) {
4001                 ffe_ctl->index = 0;
4002                 /*
4003                  * We want to skip the LOOP_CACHING_WAIT step if we don't have
4004                  * any uncached bgs and we've already done a full search
4005                  * through.
4006                  */
4007                 if (ffe_ctl->loop == LOOP_CACHING_NOWAIT &&
4008                     (!ffe_ctl->orig_have_caching_bg && full_search))
4009                         ffe_ctl->loop++;
4010                 ffe_ctl->loop++;
4011
4012                 if (ffe_ctl->loop == LOOP_ALLOC_CHUNK) {
4013                         struct btrfs_trans_handle *trans;
4014                         int exist = 0;
4015
4016                         /* Check if allocation policy allows to create a new chunk */
4017                         ret = can_allocate_chunk(fs_info, ffe_ctl);
4018                         if (ret)
4019                                 return ret;
4020
4021                         trans = current->journal_info;
4022                         if (trans)
4023                                 exist = 1;
4024                         else
4025                                 trans = btrfs_join_transaction(root);
4026
4027                         if (IS_ERR(trans)) {
4028                                 ret = PTR_ERR(trans);
4029                                 return ret;
4030                         }
4031
4032                         ret = btrfs_chunk_alloc(trans, ffe_ctl->flags,
4033                                                 CHUNK_ALLOC_FORCE_FOR_EXTENT);
4034
4035                         /* Do not bail out on ENOSPC since we can do more. */
4036                         if (ret == -ENOSPC) {
4037                                 ret = 0;
4038                                 ffe_ctl->loop++;
4039                         }
4040                         else if (ret < 0)
4041                                 btrfs_abort_transaction(trans, ret);
4042                         else
4043                                 ret = 0;
4044                         if (!exist)
4045                                 btrfs_end_transaction(trans);
4046                         if (ret)
4047                                 return ret;
4048                 }
4049
4050                 if (ffe_ctl->loop == LOOP_NO_EMPTY_SIZE) {
4051                         if (ffe_ctl->policy != BTRFS_EXTENT_ALLOC_CLUSTERED)
4052                                 return -ENOSPC;
4053
4054                         /*
4055                          * Don't loop again if we already have no empty_size and
4056                          * no empty_cluster.
4057                          */
4058                         if (ffe_ctl->empty_size == 0 &&
4059                             ffe_ctl->empty_cluster == 0)
4060                                 return -ENOSPC;
4061                         ffe_ctl->empty_size = 0;
4062                         ffe_ctl->empty_cluster = 0;
4063                 }
4064                 return 1;
4065         }
4066         return -ENOSPC;
4067 }
4068
4069 static bool find_free_extent_check_size_class(struct find_free_extent_ctl *ffe_ctl,
4070                                               struct btrfs_block_group *bg)
4071 {
4072         if (ffe_ctl->policy == BTRFS_EXTENT_ALLOC_ZONED)
4073                 return true;
4074         if (!btrfs_block_group_should_use_size_class(bg))
4075                 return true;
4076         if (ffe_ctl->loop >= LOOP_WRONG_SIZE_CLASS)
4077                 return true;
4078         if (ffe_ctl->loop >= LOOP_UNSET_SIZE_CLASS &&
4079             bg->size_class == BTRFS_BG_SZ_NONE)
4080                 return true;
4081         return ffe_ctl->size_class == bg->size_class;
4082 }
4083
4084 static int prepare_allocation_clustered(struct btrfs_fs_info *fs_info,
4085                                         struct find_free_extent_ctl *ffe_ctl,
4086                                         struct btrfs_space_info *space_info,
4087                                         struct btrfs_key *ins)
4088 {
4089         /*
4090          * If our free space is heavily fragmented we may not be able to make
4091          * big contiguous allocations, so instead of doing the expensive search
4092          * for free space, simply return ENOSPC with our max_extent_size so we
4093          * can go ahead and search for a more manageable chunk.
4094          *
4095          * If our max_extent_size is large enough for our allocation simply
4096          * disable clustering since we will likely not be able to find enough
4097          * space to create a cluster and induce latency trying.
4098          */
4099         if (space_info->max_extent_size) {
4100                 spin_lock(&space_info->lock);
4101                 if (space_info->max_extent_size &&
4102                     ffe_ctl->num_bytes > space_info->max_extent_size) {
4103                         ins->offset = space_info->max_extent_size;
4104                         spin_unlock(&space_info->lock);
4105                         return -ENOSPC;
4106                 } else if (space_info->max_extent_size) {
4107                         ffe_ctl->use_cluster = false;
4108                 }
4109                 spin_unlock(&space_info->lock);
4110         }
4111
4112         ffe_ctl->last_ptr = fetch_cluster_info(fs_info, space_info,
4113                                                &ffe_ctl->empty_cluster);
4114         if (ffe_ctl->last_ptr) {
4115                 struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
4116
4117                 spin_lock(&last_ptr->lock);
4118                 if (last_ptr->block_group)
4119                         ffe_ctl->hint_byte = last_ptr->window_start;
4120                 if (last_ptr->fragmented) {
4121                         /*
4122                          * We still set window_start so we can keep track of the
4123                          * last place we found an allocation to try and save
4124                          * some time.
4125                          */
4126                         ffe_ctl->hint_byte = last_ptr->window_start;
4127                         ffe_ctl->use_cluster = false;
4128                 }
4129                 spin_unlock(&last_ptr->lock);
4130         }
4131
4132         return 0;
4133 }
4134
4135 static int prepare_allocation(struct btrfs_fs_info *fs_info,
4136                               struct find_free_extent_ctl *ffe_ctl,
4137                               struct btrfs_space_info *space_info,
4138                               struct btrfs_key *ins)
4139 {
4140         switch (ffe_ctl->policy) {
4141         case BTRFS_EXTENT_ALLOC_CLUSTERED:
4142                 return prepare_allocation_clustered(fs_info, ffe_ctl,
4143                                                     space_info, ins);
4144         case BTRFS_EXTENT_ALLOC_ZONED:
4145                 if (ffe_ctl->for_treelog) {
4146                         spin_lock(&fs_info->treelog_bg_lock);
4147                         if (fs_info->treelog_bg)
4148                                 ffe_ctl->hint_byte = fs_info->treelog_bg;
4149                         spin_unlock(&fs_info->treelog_bg_lock);
4150                 }
4151                 if (ffe_ctl->for_data_reloc) {
4152                         spin_lock(&fs_info->relocation_bg_lock);
4153                         if (fs_info->data_reloc_bg)
4154                                 ffe_ctl->hint_byte = fs_info->data_reloc_bg;
4155                         spin_unlock(&fs_info->relocation_bg_lock);
4156                 }
4157                 return 0;
4158         default:
4159                 BUG();
4160         }
4161 }
4162
4163 /*
4164  * walks the btree of allocated extents and find a hole of a given size.
4165  * The key ins is changed to record the hole:
4166  * ins->objectid == start position
4167  * ins->flags = BTRFS_EXTENT_ITEM_KEY
4168  * ins->offset == the size of the hole.
4169  * Any available blocks before search_start are skipped.
4170  *
4171  * If there is no suitable free space, we will record the max size of
4172  * the free space extent currently.
4173  *
4174  * The overall logic and call chain:
4175  *
4176  * find_free_extent()
4177  * |- Iterate through all block groups
4178  * |  |- Get a valid block group
4179  * |  |- Try to do clustered allocation in that block group
4180  * |  |- Try to do unclustered allocation in that block group
4181  * |  |- Check if the result is valid
4182  * |  |  |- If valid, then exit
4183  * |  |- Jump to next block group
4184  * |
4185  * |- Push harder to find free extents
4186  *    |- If not found, re-iterate all block groups
4187  */
4188 static noinline int find_free_extent(struct btrfs_root *root,
4189                                      struct btrfs_key *ins,
4190                                      struct find_free_extent_ctl *ffe_ctl)
4191 {
4192         struct btrfs_fs_info *fs_info = root->fs_info;
4193         int ret = 0;
4194         int cache_block_group_error = 0;
4195         struct btrfs_block_group *block_group = NULL;
4196         struct btrfs_space_info *space_info;
4197         bool full_search = false;
4198
4199         WARN_ON(ffe_ctl->num_bytes < fs_info->sectorsize);
4200
4201         ffe_ctl->search_start = 0;
4202         /* For clustered allocation */
4203         ffe_ctl->empty_cluster = 0;
4204         ffe_ctl->last_ptr = NULL;
4205         ffe_ctl->use_cluster = true;
4206         ffe_ctl->have_caching_bg = false;
4207         ffe_ctl->orig_have_caching_bg = false;
4208         ffe_ctl->index = btrfs_bg_flags_to_raid_index(ffe_ctl->flags);
4209         ffe_ctl->loop = 0;
4210         ffe_ctl->retry_uncached = false;
4211         ffe_ctl->cached = 0;
4212         ffe_ctl->max_extent_size = 0;
4213         ffe_ctl->total_free_space = 0;
4214         ffe_ctl->found_offset = 0;
4215         ffe_ctl->policy = BTRFS_EXTENT_ALLOC_CLUSTERED;
4216         ffe_ctl->size_class = btrfs_calc_block_group_size_class(ffe_ctl->num_bytes);
4217
4218         if (btrfs_is_zoned(fs_info))
4219                 ffe_ctl->policy = BTRFS_EXTENT_ALLOC_ZONED;
4220
4221         ins->type = BTRFS_EXTENT_ITEM_KEY;
4222         ins->objectid = 0;
4223         ins->offset = 0;
4224
4225         trace_find_free_extent(root, ffe_ctl);
4226
4227         space_info = btrfs_find_space_info(fs_info, ffe_ctl->flags);
4228         if (!space_info) {
4229                 btrfs_err(fs_info, "No space info for %llu", ffe_ctl->flags);
4230                 return -ENOSPC;
4231         }
4232
4233         ret = prepare_allocation(fs_info, ffe_ctl, space_info, ins);
4234         if (ret < 0)
4235                 return ret;
4236
4237         ffe_ctl->search_start = max(ffe_ctl->search_start,
4238                                     first_logical_byte(fs_info));
4239         ffe_ctl->search_start = max(ffe_ctl->search_start, ffe_ctl->hint_byte);
4240         if (ffe_ctl->search_start == ffe_ctl->hint_byte) {
4241                 block_group = btrfs_lookup_block_group(fs_info,
4242                                                        ffe_ctl->search_start);
4243                 /*
4244                  * we don't want to use the block group if it doesn't match our
4245                  * allocation bits, or if its not cached.
4246                  *
4247                  * However if we are re-searching with an ideal block group
4248                  * picked out then we don't care that the block group is cached.
4249                  */
4250                 if (block_group && block_group_bits(block_group, ffe_ctl->flags) &&
4251                     block_group->cached != BTRFS_CACHE_NO) {
4252                         down_read(&space_info->groups_sem);
4253                         if (list_empty(&block_group->list) ||
4254                             block_group->ro) {
4255                                 /*
4256                                  * someone is removing this block group,
4257                                  * we can't jump into the have_block_group
4258                                  * target because our list pointers are not
4259                                  * valid
4260                                  */
4261                                 btrfs_put_block_group(block_group);
4262                                 up_read(&space_info->groups_sem);
4263                         } else {
4264                                 ffe_ctl->index = btrfs_bg_flags_to_raid_index(
4265                                                         block_group->flags);
4266                                 btrfs_lock_block_group(block_group,
4267                                                        ffe_ctl->delalloc);
4268                                 ffe_ctl->hinted = true;
4269                                 goto have_block_group;
4270                         }
4271                 } else if (block_group) {
4272                         btrfs_put_block_group(block_group);
4273                 }
4274         }
4275 search:
4276         trace_find_free_extent_search_loop(root, ffe_ctl);
4277         ffe_ctl->have_caching_bg = false;
4278         if (ffe_ctl->index == btrfs_bg_flags_to_raid_index(ffe_ctl->flags) ||
4279             ffe_ctl->index == 0)
4280                 full_search = true;
4281         down_read(&space_info->groups_sem);
4282         list_for_each_entry(block_group,
4283                             &space_info->block_groups[ffe_ctl->index], list) {
4284                 struct btrfs_block_group *bg_ret;
4285
4286                 ffe_ctl->hinted = false;
4287                 /* If the block group is read-only, we can skip it entirely. */
4288                 if (unlikely(block_group->ro)) {
4289                         if (ffe_ctl->for_treelog)
4290                                 btrfs_clear_treelog_bg(block_group);
4291                         if (ffe_ctl->for_data_reloc)
4292                                 btrfs_clear_data_reloc_bg(block_group);
4293                         continue;
4294                 }
4295
4296                 btrfs_grab_block_group(block_group, ffe_ctl->delalloc);
4297                 ffe_ctl->search_start = block_group->start;
4298
4299                 /*
4300                  * this can happen if we end up cycling through all the
4301                  * raid types, but we want to make sure we only allocate
4302                  * for the proper type.
4303                  */
4304                 if (!block_group_bits(block_group, ffe_ctl->flags)) {
4305                         u64 extra = BTRFS_BLOCK_GROUP_DUP |
4306                                 BTRFS_BLOCK_GROUP_RAID1_MASK |
4307                                 BTRFS_BLOCK_GROUP_RAID56_MASK |
4308                                 BTRFS_BLOCK_GROUP_RAID10;
4309
4310                         /*
4311                          * if they asked for extra copies and this block group
4312                          * doesn't provide them, bail.  This does allow us to
4313                          * fill raid0 from raid1.
4314                          */
4315                         if ((ffe_ctl->flags & extra) && !(block_group->flags & extra))
4316                                 goto loop;
4317
4318                         /*
4319                          * This block group has different flags than we want.
4320                          * It's possible that we have MIXED_GROUP flag but no
4321                          * block group is mixed.  Just skip such block group.
4322                          */
4323                         btrfs_release_block_group(block_group, ffe_ctl->delalloc);
4324                         continue;
4325                 }
4326
4327 have_block_group:
4328                 trace_find_free_extent_have_block_group(root, ffe_ctl, block_group);
4329                 ffe_ctl->cached = btrfs_block_group_done(block_group);
4330                 if (unlikely(!ffe_ctl->cached)) {
4331                         ffe_ctl->have_caching_bg = true;
4332                         ret = btrfs_cache_block_group(block_group, false);
4333
4334                         /*
4335                          * If we get ENOMEM here or something else we want to
4336                          * try other block groups, because it may not be fatal.
4337                          * However if we can't find anything else we need to
4338                          * save our return here so that we return the actual
4339                          * error that caused problems, not ENOSPC.
4340                          */
4341                         if (ret < 0) {
4342                                 if (!cache_block_group_error)
4343                                         cache_block_group_error = ret;
4344                                 ret = 0;
4345                                 goto loop;
4346                         }
4347                         ret = 0;
4348                 }
4349
4350                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR)) {
4351                         if (!cache_block_group_error)
4352                                 cache_block_group_error = -EIO;
4353                         goto loop;
4354                 }
4355
4356                 if (!find_free_extent_check_size_class(ffe_ctl, block_group))
4357                         goto loop;
4358
4359                 bg_ret = NULL;
4360                 ret = do_allocation(block_group, ffe_ctl, &bg_ret);
4361                 if (ret > 0)
4362                         goto loop;
4363
4364                 if (bg_ret && bg_ret != block_group) {
4365                         btrfs_release_block_group(block_group, ffe_ctl->delalloc);
4366                         block_group = bg_ret;
4367                 }
4368
4369                 /* Checks */
4370                 ffe_ctl->search_start = round_up(ffe_ctl->found_offset,
4371                                                  fs_info->stripesize);
4372
4373                 /* move on to the next group */
4374                 if (ffe_ctl->search_start + ffe_ctl->num_bytes >
4375                     block_group->start + block_group->length) {
4376                         btrfs_add_free_space_unused(block_group,
4377                                             ffe_ctl->found_offset,
4378                                             ffe_ctl->num_bytes);
4379                         goto loop;
4380                 }
4381
4382                 if (ffe_ctl->found_offset < ffe_ctl->search_start)
4383                         btrfs_add_free_space_unused(block_group,
4384                                         ffe_ctl->found_offset,
4385                                         ffe_ctl->search_start - ffe_ctl->found_offset);
4386
4387                 ret = btrfs_add_reserved_bytes(block_group, ffe_ctl->ram_bytes,
4388                                                ffe_ctl->num_bytes,
4389                                                ffe_ctl->delalloc,
4390                                                ffe_ctl->loop >= LOOP_WRONG_SIZE_CLASS);
4391                 if (ret == -EAGAIN) {
4392                         btrfs_add_free_space_unused(block_group,
4393                                         ffe_ctl->found_offset,
4394                                         ffe_ctl->num_bytes);
4395                         goto loop;
4396                 }
4397                 btrfs_inc_block_group_reservations(block_group);
4398
4399                 /* we are all good, lets return */
4400                 ins->objectid = ffe_ctl->search_start;
4401                 ins->offset = ffe_ctl->num_bytes;
4402
4403                 trace_btrfs_reserve_extent(block_group, ffe_ctl);
4404                 btrfs_release_block_group(block_group, ffe_ctl->delalloc);
4405                 break;
4406 loop:
4407                 if (!ffe_ctl->cached && ffe_ctl->loop > LOOP_CACHING_NOWAIT &&
4408                     !ffe_ctl->retry_uncached) {
4409                         ffe_ctl->retry_uncached = true;
4410                         btrfs_wait_block_group_cache_progress(block_group,
4411                                                 ffe_ctl->num_bytes +
4412                                                 ffe_ctl->empty_cluster +
4413                                                 ffe_ctl->empty_size);
4414                         goto have_block_group;
4415                 }
4416                 release_block_group(block_group, ffe_ctl, ffe_ctl->delalloc);
4417                 cond_resched();
4418         }
4419         up_read(&space_info->groups_sem);
4420
4421         ret = find_free_extent_update_loop(fs_info, ins, ffe_ctl, full_search);
4422         if (ret > 0)
4423                 goto search;
4424
4425         if (ret == -ENOSPC && !cache_block_group_error) {
4426                 /*
4427                  * Use ffe_ctl->total_free_space as fallback if we can't find
4428                  * any contiguous hole.
4429                  */
4430                 if (!ffe_ctl->max_extent_size)
4431                         ffe_ctl->max_extent_size = ffe_ctl->total_free_space;
4432                 spin_lock(&space_info->lock);
4433                 space_info->max_extent_size = ffe_ctl->max_extent_size;
4434                 spin_unlock(&space_info->lock);
4435                 ins->offset = ffe_ctl->max_extent_size;
4436         } else if (ret == -ENOSPC) {
4437                 ret = cache_block_group_error;
4438         }
4439         return ret;
4440 }
4441
4442 /*
4443  * Entry point to the extent allocator. Tries to find a hole that is at least
4444  * as big as @num_bytes.
4445  *
4446  * @root           -    The root that will contain this extent
4447  *
4448  * @ram_bytes      -    The amount of space in ram that @num_bytes take. This
4449  *                      is used for accounting purposes. This value differs
4450  *                      from @num_bytes only in the case of compressed extents.
4451  *
4452  * @num_bytes      -    Number of bytes to allocate on-disk.
4453  *
4454  * @min_alloc_size -    Indicates the minimum amount of space that the
4455  *                      allocator should try to satisfy. In some cases
4456  *                      @num_bytes may be larger than what is required and if
4457  *                      the filesystem is fragmented then allocation fails.
4458  *                      However, the presence of @min_alloc_size gives a
4459  *                      chance to try and satisfy the smaller allocation.
4460  *
4461  * @empty_size     -    A hint that you plan on doing more COW. This is the
4462  *                      size in bytes the allocator should try to find free
4463  *                      next to the block it returns.  This is just a hint and
4464  *                      may be ignored by the allocator.
4465  *
4466  * @hint_byte      -    Hint to the allocator to start searching above the byte
4467  *                      address passed. It might be ignored.
4468  *
4469  * @ins            -    This key is modified to record the found hole. It will
4470  *                      have the following values:
4471  *                      ins->objectid == start position
4472  *                      ins->flags = BTRFS_EXTENT_ITEM_KEY
4473  *                      ins->offset == the size of the hole.
4474  *
4475  * @is_data        -    Boolean flag indicating whether an extent is
4476  *                      allocated for data (true) or metadata (false)
4477  *
4478  * @delalloc       -    Boolean flag indicating whether this allocation is for
4479  *                      delalloc or not. If 'true' data_rwsem of block groups
4480  *                      is going to be acquired.
4481  *
4482  *
4483  * Returns 0 when an allocation succeeded or < 0 when an error occurred. In
4484  * case -ENOSPC is returned then @ins->offset will contain the size of the
4485  * largest available hole the allocator managed to find.
4486  */
4487 int btrfs_reserve_extent(struct btrfs_root *root, u64 ram_bytes,
4488                          u64 num_bytes, u64 min_alloc_size,
4489                          u64 empty_size, u64 hint_byte,
4490                          struct btrfs_key *ins, int is_data, int delalloc)
4491 {
4492         struct btrfs_fs_info *fs_info = root->fs_info;
4493         struct find_free_extent_ctl ffe_ctl = {};
4494         bool final_tried = num_bytes == min_alloc_size;
4495         u64 flags;
4496         int ret;
4497         bool for_treelog = (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
4498         bool for_data_reloc = (btrfs_is_data_reloc_root(root) && is_data);
4499
4500         flags = get_alloc_profile_by_root(root, is_data);
4501 again:
4502         WARN_ON(num_bytes < fs_info->sectorsize);
4503
4504         ffe_ctl.ram_bytes = ram_bytes;
4505         ffe_ctl.num_bytes = num_bytes;
4506         ffe_ctl.min_alloc_size = min_alloc_size;
4507         ffe_ctl.empty_size = empty_size;
4508         ffe_ctl.flags = flags;
4509         ffe_ctl.delalloc = delalloc;
4510         ffe_ctl.hint_byte = hint_byte;
4511         ffe_ctl.for_treelog = for_treelog;
4512         ffe_ctl.for_data_reloc = for_data_reloc;
4513
4514         ret = find_free_extent(root, ins, &ffe_ctl);
4515         if (!ret && !is_data) {
4516                 btrfs_dec_block_group_reservations(fs_info, ins->objectid);
4517         } else if (ret == -ENOSPC) {
4518                 if (!final_tried && ins->offset) {
4519                         num_bytes = min(num_bytes >> 1, ins->offset);
4520                         num_bytes = round_down(num_bytes,
4521                                                fs_info->sectorsize);
4522                         num_bytes = max(num_bytes, min_alloc_size);
4523                         ram_bytes = num_bytes;
4524                         if (num_bytes == min_alloc_size)
4525                                 final_tried = true;
4526                         goto again;
4527                 } else if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
4528                         struct btrfs_space_info *sinfo;
4529
4530                         sinfo = btrfs_find_space_info(fs_info, flags);
4531                         btrfs_err(fs_info,
4532         "allocation failed flags %llu, wanted %llu tree-log %d, relocation: %d",
4533                                   flags, num_bytes, for_treelog, for_data_reloc);
4534                         if (sinfo)
4535                                 btrfs_dump_space_info(fs_info, sinfo,
4536                                                       num_bytes, 1);
4537                 }
4538         }
4539
4540         return ret;
4541 }
4542
4543 int btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
4544                                u64 start, u64 len, int delalloc)
4545 {
4546         struct btrfs_block_group *cache;
4547
4548         cache = btrfs_lookup_block_group(fs_info, start);
4549         if (!cache) {
4550                 btrfs_err(fs_info, "Unable to find block group for %llu",
4551                           start);
4552                 return -ENOSPC;
4553         }
4554
4555         btrfs_add_free_space(cache, start, len);
4556         btrfs_free_reserved_bytes(cache, len, delalloc);
4557         trace_btrfs_reserved_extent_free(fs_info, start, len);
4558
4559         btrfs_put_block_group(cache);
4560         return 0;
4561 }
4562
4563 int btrfs_pin_reserved_extent(struct btrfs_trans_handle *trans,
4564                               const struct extent_buffer *eb)
4565 {
4566         struct btrfs_block_group *cache;
4567         int ret = 0;
4568
4569         cache = btrfs_lookup_block_group(trans->fs_info, eb->start);
4570         if (!cache) {
4571                 btrfs_err(trans->fs_info, "unable to find block group for %llu",
4572                           eb->start);
4573                 return -ENOSPC;
4574         }
4575
4576         ret = pin_down_extent(trans, cache, eb->start, eb->len, 1);
4577         btrfs_put_block_group(cache);
4578         return ret;
4579 }
4580
4581 static int alloc_reserved_extent(struct btrfs_trans_handle *trans, u64 bytenr,
4582                                  u64 num_bytes)
4583 {
4584         struct btrfs_fs_info *fs_info = trans->fs_info;
4585         int ret;
4586
4587         ret = remove_from_free_space_tree(trans, bytenr, num_bytes);
4588         if (ret)
4589                 return ret;
4590
4591         ret = btrfs_update_block_group(trans, bytenr, num_bytes, true);
4592         if (ret) {
4593                 ASSERT(!ret);
4594                 btrfs_err(fs_info, "update block group failed for %llu %llu",
4595                           bytenr, num_bytes);
4596                 return ret;
4597         }
4598
4599         trace_btrfs_reserved_extent_alloc(fs_info, bytenr, num_bytes);
4600         return 0;
4601 }
4602
4603 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
4604                                       u64 parent, u64 root_objectid,
4605                                       u64 flags, u64 owner, u64 offset,
4606                                       struct btrfs_key *ins, int ref_mod)
4607 {
4608         struct btrfs_fs_info *fs_info = trans->fs_info;
4609         struct btrfs_root *extent_root;
4610         int ret;
4611         struct btrfs_extent_item *extent_item;
4612         struct btrfs_extent_inline_ref *iref;
4613         struct btrfs_path *path;
4614         struct extent_buffer *leaf;
4615         int type;
4616         u32 size;
4617
4618         if (parent > 0)
4619                 type = BTRFS_SHARED_DATA_REF_KEY;
4620         else
4621                 type = BTRFS_EXTENT_DATA_REF_KEY;
4622
4623         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
4624
4625         path = btrfs_alloc_path();
4626         if (!path)
4627                 return -ENOMEM;
4628
4629         extent_root = btrfs_extent_root(fs_info, ins->objectid);
4630         ret = btrfs_insert_empty_item(trans, extent_root, path, ins, size);
4631         if (ret) {
4632                 btrfs_free_path(path);
4633                 return ret;
4634         }
4635
4636         leaf = path->nodes[0];
4637         extent_item = btrfs_item_ptr(leaf, path->slots[0],
4638                                      struct btrfs_extent_item);
4639         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
4640         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
4641         btrfs_set_extent_flags(leaf, extent_item,
4642                                flags | BTRFS_EXTENT_FLAG_DATA);
4643
4644         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
4645         btrfs_set_extent_inline_ref_type(leaf, iref, type);
4646         if (parent > 0) {
4647                 struct btrfs_shared_data_ref *ref;
4648                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
4649                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
4650                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
4651         } else {
4652                 struct btrfs_extent_data_ref *ref;
4653                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
4654                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
4655                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
4656                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
4657                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
4658         }
4659
4660         btrfs_mark_buffer_dirty(trans, path->nodes[0]);
4661         btrfs_free_path(path);
4662
4663         return alloc_reserved_extent(trans, ins->objectid, ins->offset);
4664 }
4665
4666 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
4667                                      struct btrfs_delayed_ref_node *node,
4668                                      struct btrfs_delayed_extent_op *extent_op)
4669 {
4670         struct btrfs_fs_info *fs_info = trans->fs_info;
4671         struct btrfs_root *extent_root;
4672         int ret;
4673         struct btrfs_extent_item *extent_item;
4674         struct btrfs_key extent_key;
4675         struct btrfs_tree_block_info *block_info;
4676         struct btrfs_extent_inline_ref *iref;
4677         struct btrfs_path *path;
4678         struct extent_buffer *leaf;
4679         struct btrfs_delayed_tree_ref *ref;
4680         u32 size = sizeof(*extent_item) + sizeof(*iref);
4681         u64 flags = extent_op->flags_to_set;
4682         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
4683
4684         ref = btrfs_delayed_node_to_tree_ref(node);
4685
4686         extent_key.objectid = node->bytenr;
4687         if (skinny_metadata) {
4688                 extent_key.offset = ref->level;
4689                 extent_key.type = BTRFS_METADATA_ITEM_KEY;
4690         } else {
4691                 extent_key.offset = node->num_bytes;
4692                 extent_key.type = BTRFS_EXTENT_ITEM_KEY;
4693                 size += sizeof(*block_info);
4694         }
4695
4696         path = btrfs_alloc_path();
4697         if (!path)
4698                 return -ENOMEM;
4699
4700         extent_root = btrfs_extent_root(fs_info, extent_key.objectid);
4701         ret = btrfs_insert_empty_item(trans, extent_root, path, &extent_key,
4702                                       size);
4703         if (ret) {
4704                 btrfs_free_path(path);
4705                 return ret;
4706         }
4707
4708         leaf = path->nodes[0];
4709         extent_item = btrfs_item_ptr(leaf, path->slots[0],
4710                                      struct btrfs_extent_item);
4711         btrfs_set_extent_refs(leaf, extent_item, 1);
4712         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
4713         btrfs_set_extent_flags(leaf, extent_item,
4714                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
4715
4716         if (skinny_metadata) {
4717                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
4718         } else {
4719                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
4720                 btrfs_set_tree_block_key(leaf, block_info, &extent_op->key);
4721                 btrfs_set_tree_block_level(leaf, block_info, ref->level);
4722                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
4723         }
4724
4725         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY) {
4726                 btrfs_set_extent_inline_ref_type(leaf, iref,
4727                                                  BTRFS_SHARED_BLOCK_REF_KEY);
4728                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->parent);
4729         } else {
4730                 btrfs_set_extent_inline_ref_type(leaf, iref,
4731                                                  BTRFS_TREE_BLOCK_REF_KEY);
4732                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->root);
4733         }
4734
4735         btrfs_mark_buffer_dirty(trans, leaf);
4736         btrfs_free_path(path);
4737
4738         return alloc_reserved_extent(trans, node->bytenr, fs_info->nodesize);
4739 }
4740
4741 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
4742                                      struct btrfs_root *root, u64 owner,
4743                                      u64 offset, u64 ram_bytes,
4744                                      struct btrfs_key *ins)
4745 {
4746         struct btrfs_ref generic_ref = { 0 };
4747
4748         BUG_ON(root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
4749
4750         btrfs_init_generic_ref(&generic_ref, BTRFS_ADD_DELAYED_EXTENT,
4751                                ins->objectid, ins->offset, 0);
4752         btrfs_init_data_ref(&generic_ref, root->root_key.objectid, owner,
4753                             offset, 0, false);
4754         btrfs_ref_tree_mod(root->fs_info, &generic_ref);
4755
4756         return btrfs_add_delayed_data_ref(trans, &generic_ref, ram_bytes);
4757 }
4758
4759 /*
4760  * this is used by the tree logging recovery code.  It records that
4761  * an extent has been allocated and makes sure to clear the free
4762  * space cache bits as well
4763  */
4764 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
4765                                    u64 root_objectid, u64 owner, u64 offset,
4766                                    struct btrfs_key *ins)
4767 {
4768         struct btrfs_fs_info *fs_info = trans->fs_info;
4769         int ret;
4770         struct btrfs_block_group *block_group;
4771         struct btrfs_space_info *space_info;
4772
4773         /*
4774          * Mixed block groups will exclude before processing the log so we only
4775          * need to do the exclude dance if this fs isn't mixed.
4776          */
4777         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS)) {
4778                 ret = __exclude_logged_extent(fs_info, ins->objectid,
4779                                               ins->offset);
4780                 if (ret)
4781                         return ret;
4782         }
4783
4784         block_group = btrfs_lookup_block_group(fs_info, ins->objectid);
4785         if (!block_group)
4786                 return -EINVAL;
4787
4788         space_info = block_group->space_info;
4789         spin_lock(&space_info->lock);
4790         spin_lock(&block_group->lock);
4791         space_info->bytes_reserved += ins->offset;
4792         block_group->reserved += ins->offset;
4793         spin_unlock(&block_group->lock);
4794         spin_unlock(&space_info->lock);
4795
4796         ret = alloc_reserved_file_extent(trans, 0, root_objectid, 0, owner,
4797                                          offset, ins, 1);
4798         if (ret)
4799                 btrfs_pin_extent(trans, ins->objectid, ins->offset, 1);
4800         btrfs_put_block_group(block_group);
4801         return ret;
4802 }
4803
4804 #ifdef CONFIG_BTRFS_DEBUG
4805 /*
4806  * Extra safety check in case the extent tree is corrupted and extent allocator
4807  * chooses to use a tree block which is already used and locked.
4808  */
4809 static bool check_eb_lock_owner(const struct extent_buffer *eb)
4810 {
4811         if (eb->lock_owner == current->pid) {
4812                 btrfs_err_rl(eb->fs_info,
4813 "tree block %llu owner %llu already locked by pid=%d, extent tree corruption detected",
4814                              eb->start, btrfs_header_owner(eb), current->pid);
4815                 return true;
4816         }
4817         return false;
4818 }
4819 #else
4820 static bool check_eb_lock_owner(struct extent_buffer *eb)
4821 {
4822         return false;
4823 }
4824 #endif
4825
4826 static struct extent_buffer *
4827 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
4828                       u64 bytenr, int level, u64 owner,
4829                       enum btrfs_lock_nesting nest)
4830 {
4831         struct btrfs_fs_info *fs_info = root->fs_info;
4832         struct extent_buffer *buf;
4833         u64 lockdep_owner = owner;
4834
4835         buf = btrfs_find_create_tree_block(fs_info, bytenr, owner, level);
4836         if (IS_ERR(buf))
4837                 return buf;
4838
4839         if (check_eb_lock_owner(buf)) {
4840                 free_extent_buffer(buf);
4841                 return ERR_PTR(-EUCLEAN);
4842         }
4843
4844         /*
4845          * The reloc trees are just snapshots, so we need them to appear to be
4846          * just like any other fs tree WRT lockdep.
4847          *
4848          * The exception however is in replace_path() in relocation, where we
4849          * hold the lock on the original fs root and then search for the reloc
4850          * root.  At that point we need to make sure any reloc root buffers are
4851          * set to the BTRFS_TREE_RELOC_OBJECTID lockdep class in order to make
4852          * lockdep happy.
4853          */
4854         if (lockdep_owner == BTRFS_TREE_RELOC_OBJECTID &&
4855             !test_bit(BTRFS_ROOT_RESET_LOCKDEP_CLASS, &root->state))
4856                 lockdep_owner = BTRFS_FS_TREE_OBJECTID;
4857
4858         /* btrfs_clear_buffer_dirty() accesses generation field. */
4859         btrfs_set_header_generation(buf, trans->transid);
4860
4861         /*
4862          * This needs to stay, because we could allocate a freed block from an
4863          * old tree into a new tree, so we need to make sure this new block is
4864          * set to the appropriate level and owner.
4865          */
4866         btrfs_set_buffer_lockdep_class(lockdep_owner, buf, level);
4867
4868         __btrfs_tree_lock(buf, nest);
4869         btrfs_clear_buffer_dirty(trans, buf);
4870         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
4871         clear_bit(EXTENT_BUFFER_NO_CHECK, &buf->bflags);
4872
4873         set_extent_buffer_uptodate(buf);
4874
4875         memzero_extent_buffer(buf, 0, sizeof(struct btrfs_header));
4876         btrfs_set_header_level(buf, level);
4877         btrfs_set_header_bytenr(buf, buf->start);
4878         btrfs_set_header_generation(buf, trans->transid);
4879         btrfs_set_header_backref_rev(buf, BTRFS_MIXED_BACKREF_REV);
4880         btrfs_set_header_owner(buf, owner);
4881         write_extent_buffer_fsid(buf, fs_info->fs_devices->metadata_uuid);
4882         write_extent_buffer_chunk_tree_uuid(buf, fs_info->chunk_tree_uuid);
4883         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
4884                 buf->log_index = root->log_transid % 2;
4885                 /*
4886                  * we allow two log transactions at a time, use different
4887                  * EXTENT bit to differentiate dirty pages.
4888                  */
4889                 if (buf->log_index == 0)
4890                         set_extent_bit(&root->dirty_log_pages, buf->start,
4891                                        buf->start + buf->len - 1,
4892                                        EXTENT_DIRTY, NULL);
4893                 else
4894                         set_extent_bit(&root->dirty_log_pages, buf->start,
4895                                        buf->start + buf->len - 1,
4896                                        EXTENT_NEW, NULL);
4897         } else {
4898                 buf->log_index = -1;
4899                 set_extent_bit(&trans->transaction->dirty_pages, buf->start,
4900                                buf->start + buf->len - 1, EXTENT_DIRTY, NULL);
4901         }
4902         /* this returns a buffer locked for blocking */
4903         return buf;
4904 }
4905
4906 /*
4907  * finds a free extent and does all the dirty work required for allocation
4908  * returns the tree buffer or an ERR_PTR on error.
4909  */
4910 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
4911                                              struct btrfs_root *root,
4912                                              u64 parent, u64 root_objectid,
4913                                              const struct btrfs_disk_key *key,
4914                                              int level, u64 hint,
4915                                              u64 empty_size,
4916                                              enum btrfs_lock_nesting nest)
4917 {
4918         struct btrfs_fs_info *fs_info = root->fs_info;
4919         struct btrfs_key ins;
4920         struct btrfs_block_rsv *block_rsv;
4921         struct extent_buffer *buf;
4922         struct btrfs_delayed_extent_op *extent_op;
4923         struct btrfs_ref generic_ref = { 0 };
4924         u64 flags = 0;
4925         int ret;
4926         u32 blocksize = fs_info->nodesize;
4927         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
4928
4929 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4930         if (btrfs_is_testing(fs_info)) {
4931                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
4932                                             level, root_objectid, nest);
4933                 if (!IS_ERR(buf))
4934                         root->alloc_bytenr += blocksize;
4935                 return buf;
4936         }
4937 #endif
4938
4939         block_rsv = btrfs_use_block_rsv(trans, root, blocksize);
4940         if (IS_ERR(block_rsv))
4941                 return ERR_CAST(block_rsv);
4942
4943         ret = btrfs_reserve_extent(root, blocksize, blocksize, blocksize,
4944                                    empty_size, hint, &ins, 0, 0);
4945         if (ret)
4946                 goto out_unuse;
4947
4948         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level,
4949                                     root_objectid, nest);
4950         if (IS_ERR(buf)) {
4951                 ret = PTR_ERR(buf);
4952                 goto out_free_reserved;
4953         }
4954
4955         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
4956                 if (parent == 0)
4957                         parent = ins.objectid;
4958                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
4959         } else
4960                 BUG_ON(parent > 0);
4961
4962         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
4963                 extent_op = btrfs_alloc_delayed_extent_op();
4964                 if (!extent_op) {
4965                         ret = -ENOMEM;
4966                         goto out_free_buf;
4967                 }
4968                 if (key)
4969                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
4970                 else
4971                         memset(&extent_op->key, 0, sizeof(extent_op->key));
4972                 extent_op->flags_to_set = flags;
4973                 extent_op->update_key = skinny_metadata ? false : true;
4974                 extent_op->update_flags = true;
4975                 extent_op->level = level;
4976
4977                 btrfs_init_generic_ref(&generic_ref, BTRFS_ADD_DELAYED_EXTENT,
4978                                        ins.objectid, ins.offset, parent);
4979                 btrfs_init_tree_ref(&generic_ref, level, root_objectid,
4980                                     root->root_key.objectid, false);
4981                 btrfs_ref_tree_mod(fs_info, &generic_ref);
4982                 ret = btrfs_add_delayed_tree_ref(trans, &generic_ref, extent_op);
4983                 if (ret)
4984                         goto out_free_delayed;
4985         }
4986         return buf;
4987
4988 out_free_delayed:
4989         btrfs_free_delayed_extent_op(extent_op);
4990 out_free_buf:
4991         btrfs_tree_unlock(buf);
4992         free_extent_buffer(buf);
4993 out_free_reserved:
4994         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 0);
4995 out_unuse:
4996         btrfs_unuse_block_rsv(fs_info, block_rsv, blocksize);
4997         return ERR_PTR(ret);
4998 }
4999
5000 struct walk_control {
5001         u64 refs[BTRFS_MAX_LEVEL];
5002         u64 flags[BTRFS_MAX_LEVEL];
5003         struct btrfs_key update_progress;
5004         struct btrfs_key drop_progress;
5005         int drop_level;
5006         int stage;
5007         int level;
5008         int shared_level;
5009         int update_ref;
5010         int keep_locks;
5011         int reada_slot;
5012         int reada_count;
5013         int restarted;
5014 };
5015
5016 #define DROP_REFERENCE  1
5017 #define UPDATE_BACKREF  2
5018
5019 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
5020                                      struct btrfs_root *root,
5021                                      struct walk_control *wc,
5022                                      struct btrfs_path *path)
5023 {
5024         struct btrfs_fs_info *fs_info = root->fs_info;
5025         u64 bytenr;
5026         u64 generation;
5027         u64 refs;
5028         u64 flags;
5029         u32 nritems;
5030         struct btrfs_key key;
5031         struct extent_buffer *eb;
5032         int ret;
5033         int slot;
5034         int nread = 0;
5035
5036         if (path->slots[wc->level] < wc->reada_slot) {
5037                 wc->reada_count = wc->reada_count * 2 / 3;
5038                 wc->reada_count = max(wc->reada_count, 2);
5039         } else {
5040                 wc->reada_count = wc->reada_count * 3 / 2;
5041                 wc->reada_count = min_t(int, wc->reada_count,
5042                                         BTRFS_NODEPTRS_PER_BLOCK(fs_info));
5043         }
5044
5045         eb = path->nodes[wc->level];
5046         nritems = btrfs_header_nritems(eb);
5047
5048         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
5049                 if (nread >= wc->reada_count)
5050                         break;
5051
5052                 cond_resched();
5053                 bytenr = btrfs_node_blockptr(eb, slot);
5054                 generation = btrfs_node_ptr_generation(eb, slot);
5055
5056                 if (slot == path->slots[wc->level])
5057                         goto reada;
5058
5059                 if (wc->stage == UPDATE_BACKREF &&
5060                     generation <= root->root_key.offset)
5061                         continue;
5062
5063                 /* We don't lock the tree block, it's OK to be racy here */
5064                 ret = btrfs_lookup_extent_info(trans, fs_info, bytenr,
5065                                                wc->level - 1, 1, &refs,
5066                                                &flags);
5067                 /* We don't care about errors in readahead. */
5068                 if (ret < 0)
5069                         continue;
5070                 BUG_ON(refs == 0);
5071
5072                 if (wc->stage == DROP_REFERENCE) {
5073                         if (refs == 1)
5074                                 goto reada;
5075
5076                         if (wc->level == 1 &&
5077                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5078                                 continue;
5079                         if (!wc->update_ref ||
5080                             generation <= root->root_key.offset)
5081                                 continue;
5082                         btrfs_node_key_to_cpu(eb, &key, slot);
5083                         ret = btrfs_comp_cpu_keys(&key,
5084                                                   &wc->update_progress);
5085                         if (ret < 0)
5086                                 continue;
5087                 } else {
5088                         if (wc->level == 1 &&
5089                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5090                                 continue;
5091                 }
5092 reada:
5093                 btrfs_readahead_node_child(eb, slot);
5094                 nread++;
5095         }
5096         wc->reada_slot = slot;
5097 }
5098
5099 /*
5100  * helper to process tree block while walking down the tree.
5101  *
5102  * when wc->stage == UPDATE_BACKREF, this function updates
5103  * back refs for pointers in the block.
5104  *
5105  * NOTE: return value 1 means we should stop walking down.
5106  */
5107 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
5108                                    struct btrfs_root *root,
5109                                    struct btrfs_path *path,
5110                                    struct walk_control *wc, int lookup_info)
5111 {
5112         struct btrfs_fs_info *fs_info = root->fs_info;
5113         int level = wc->level;
5114         struct extent_buffer *eb = path->nodes[level];
5115         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
5116         int ret;
5117
5118         if (wc->stage == UPDATE_BACKREF &&
5119             btrfs_header_owner(eb) != root->root_key.objectid)
5120                 return 1;
5121
5122         /*
5123          * when reference count of tree block is 1, it won't increase
5124          * again. once full backref flag is set, we never clear it.
5125          */
5126         if (lookup_info &&
5127             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
5128              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
5129                 BUG_ON(!path->locks[level]);
5130                 ret = btrfs_lookup_extent_info(trans, fs_info,
5131                                                eb->start, level, 1,
5132                                                &wc->refs[level],
5133                                                &wc->flags[level]);
5134                 BUG_ON(ret == -ENOMEM);
5135                 if (ret)
5136                         return ret;
5137                 BUG_ON(wc->refs[level] == 0);
5138         }
5139
5140         if (wc->stage == DROP_REFERENCE) {
5141                 if (wc->refs[level] > 1)
5142                         return 1;
5143
5144                 if (path->locks[level] && !wc->keep_locks) {
5145                         btrfs_tree_unlock_rw(eb, path->locks[level]);
5146                         path->locks[level] = 0;
5147                 }
5148                 return 0;
5149         }
5150
5151         /* wc->stage == UPDATE_BACKREF */
5152         if (!(wc->flags[level] & flag)) {
5153                 BUG_ON(!path->locks[level]);
5154                 ret = btrfs_inc_ref(trans, root, eb, 1);
5155                 BUG_ON(ret); /* -ENOMEM */
5156                 ret = btrfs_dec_ref(trans, root, eb, 0);
5157                 BUG_ON(ret); /* -ENOMEM */
5158                 ret = btrfs_set_disk_extent_flags(trans, eb, flag);
5159                 BUG_ON(ret); /* -ENOMEM */
5160                 wc->flags[level] |= flag;
5161         }
5162
5163         /*
5164          * the block is shared by multiple trees, so it's not good to
5165          * keep the tree lock
5166          */
5167         if (path->locks[level] && level > 0) {
5168                 btrfs_tree_unlock_rw(eb, path->locks[level]);
5169                 path->locks[level] = 0;
5170         }
5171         return 0;
5172 }
5173
5174 /*
5175  * This is used to verify a ref exists for this root to deal with a bug where we
5176  * would have a drop_progress key that hadn't been updated properly.
5177  */
5178 static int check_ref_exists(struct btrfs_trans_handle *trans,
5179                             struct btrfs_root *root, u64 bytenr, u64 parent,
5180                             int level)
5181 {
5182         struct btrfs_path *path;
5183         struct btrfs_extent_inline_ref *iref;
5184         int ret;
5185
5186         path = btrfs_alloc_path();
5187         if (!path)
5188                 return -ENOMEM;
5189
5190         ret = lookup_extent_backref(trans, path, &iref, bytenr,
5191                                     root->fs_info->nodesize, parent,
5192                                     root->root_key.objectid, level, 0);
5193         btrfs_free_path(path);
5194         if (ret == -ENOENT)
5195                 return 0;
5196         if (ret < 0)
5197                 return ret;
5198         return 1;
5199 }
5200
5201 /*
5202  * helper to process tree block pointer.
5203  *
5204  * when wc->stage == DROP_REFERENCE, this function checks
5205  * reference count of the block pointed to. if the block
5206  * is shared and we need update back refs for the subtree
5207  * rooted at the block, this function changes wc->stage to
5208  * UPDATE_BACKREF. if the block is shared and there is no
5209  * need to update back, this function drops the reference
5210  * to the block.
5211  *
5212  * NOTE: return value 1 means we should stop walking down.
5213  */
5214 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
5215                                  struct btrfs_root *root,
5216                                  struct btrfs_path *path,
5217                                  struct walk_control *wc, int *lookup_info)
5218 {
5219         struct btrfs_fs_info *fs_info = root->fs_info;
5220         u64 bytenr;
5221         u64 generation;
5222         u64 parent;
5223         struct btrfs_tree_parent_check check = { 0 };
5224         struct btrfs_key key;
5225         struct btrfs_ref ref = { 0 };
5226         struct extent_buffer *next;
5227         int level = wc->level;
5228         int reada = 0;
5229         int ret = 0;
5230         bool need_account = false;
5231
5232         generation = btrfs_node_ptr_generation(path->nodes[level],
5233                                                path->slots[level]);
5234         /*
5235          * if the lower level block was created before the snapshot
5236          * was created, we know there is no need to update back refs
5237          * for the subtree
5238          */
5239         if (wc->stage == UPDATE_BACKREF &&
5240             generation <= root->root_key.offset) {
5241                 *lookup_info = 1;
5242                 return 1;
5243         }
5244
5245         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
5246
5247         check.level = level - 1;
5248         check.transid = generation;
5249         check.owner_root = root->root_key.objectid;
5250         check.has_first_key = true;
5251         btrfs_node_key_to_cpu(path->nodes[level], &check.first_key,
5252                               path->slots[level]);
5253
5254         next = find_extent_buffer(fs_info, bytenr);
5255         if (!next) {
5256                 next = btrfs_find_create_tree_block(fs_info, bytenr,
5257                                 root->root_key.objectid, level - 1);
5258                 if (IS_ERR(next))
5259                         return PTR_ERR(next);
5260                 reada = 1;
5261         }
5262         btrfs_tree_lock(next);
5263
5264         ret = btrfs_lookup_extent_info(trans, fs_info, bytenr, level - 1, 1,
5265                                        &wc->refs[level - 1],
5266                                        &wc->flags[level - 1]);
5267         if (ret < 0)
5268                 goto out_unlock;
5269
5270         if (unlikely(wc->refs[level - 1] == 0)) {
5271                 btrfs_err(fs_info, "Missing references.");
5272                 ret = -EIO;
5273                 goto out_unlock;
5274         }
5275         *lookup_info = 0;
5276
5277         if (wc->stage == DROP_REFERENCE) {
5278                 if (wc->refs[level - 1] > 1) {
5279                         need_account = true;
5280                         if (level == 1 &&
5281                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5282                                 goto skip;
5283
5284                         if (!wc->update_ref ||
5285                             generation <= root->root_key.offset)
5286                                 goto skip;
5287
5288                         btrfs_node_key_to_cpu(path->nodes[level], &key,
5289                                               path->slots[level]);
5290                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
5291                         if (ret < 0)
5292                                 goto skip;
5293
5294                         wc->stage = UPDATE_BACKREF;
5295                         wc->shared_level = level - 1;
5296                 }
5297         } else {
5298                 if (level == 1 &&
5299                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5300                         goto skip;
5301         }
5302
5303         if (!btrfs_buffer_uptodate(next, generation, 0)) {
5304                 btrfs_tree_unlock(next);
5305                 free_extent_buffer(next);
5306                 next = NULL;
5307                 *lookup_info = 1;
5308         }
5309
5310         if (!next) {
5311                 if (reada && level == 1)
5312                         reada_walk_down(trans, root, wc, path);
5313                 next = read_tree_block(fs_info, bytenr, &check);
5314                 if (IS_ERR(next)) {
5315                         return PTR_ERR(next);
5316                 } else if (!extent_buffer_uptodate(next)) {
5317                         free_extent_buffer(next);
5318                         return -EIO;
5319                 }
5320                 btrfs_tree_lock(next);
5321         }
5322
5323         level--;
5324         ASSERT(level == btrfs_header_level(next));
5325         if (level != btrfs_header_level(next)) {
5326                 btrfs_err(root->fs_info, "mismatched level");
5327                 ret = -EIO;
5328                 goto out_unlock;
5329         }
5330         path->nodes[level] = next;
5331         path->slots[level] = 0;
5332         path->locks[level] = BTRFS_WRITE_LOCK;
5333         wc->level = level;
5334         if (wc->level == 1)
5335                 wc->reada_slot = 0;
5336         return 0;
5337 skip:
5338         wc->refs[level - 1] = 0;
5339         wc->flags[level - 1] = 0;
5340         if (wc->stage == DROP_REFERENCE) {
5341                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
5342                         parent = path->nodes[level]->start;
5343                 } else {
5344                         ASSERT(root->root_key.objectid ==
5345                                btrfs_header_owner(path->nodes[level]));
5346                         if (root->root_key.objectid !=
5347                             btrfs_header_owner(path->nodes[level])) {
5348                                 btrfs_err(root->fs_info,
5349                                                 "mismatched block owner");
5350                                 ret = -EIO;
5351                                 goto out_unlock;
5352                         }
5353                         parent = 0;
5354                 }
5355
5356                 /*
5357                  * If we had a drop_progress we need to verify the refs are set
5358                  * as expected.  If we find our ref then we know that from here
5359                  * on out everything should be correct, and we can clear the
5360                  * ->restarted flag.
5361                  */
5362                 if (wc->restarted) {
5363                         ret = check_ref_exists(trans, root, bytenr, parent,
5364                                                level - 1);
5365                         if (ret < 0)
5366                                 goto out_unlock;
5367                         if (ret == 0)
5368                                 goto no_delete;
5369                         ret = 0;
5370                         wc->restarted = 0;
5371                 }
5372
5373                 /*
5374                  * Reloc tree doesn't contribute to qgroup numbers, and we have
5375                  * already accounted them at merge time (replace_path),
5376                  * thus we could skip expensive subtree trace here.
5377                  */
5378                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
5379                     need_account) {
5380                         ret = btrfs_qgroup_trace_subtree(trans, next,
5381                                                          generation, level - 1);
5382                         if (ret) {
5383                                 btrfs_err_rl(fs_info,
5384                                              "Error %d accounting shared subtree. Quota is out of sync, rescan required.",
5385                                              ret);
5386                         }
5387                 }
5388
5389                 /*
5390                  * We need to update the next key in our walk control so we can
5391                  * update the drop_progress key accordingly.  We don't care if
5392                  * find_next_key doesn't find a key because that means we're at
5393                  * the end and are going to clean up now.
5394                  */
5395                 wc->drop_level = level;
5396                 find_next_key(path, level, &wc->drop_progress);
5397
5398                 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
5399                                        fs_info->nodesize, parent);
5400                 btrfs_init_tree_ref(&ref, level - 1, root->root_key.objectid,
5401                                     0, false);
5402                 ret = btrfs_free_extent(trans, &ref);
5403                 if (ret)
5404                         goto out_unlock;
5405         }
5406 no_delete:
5407         *lookup_info = 1;
5408         ret = 1;
5409
5410 out_unlock:
5411         btrfs_tree_unlock(next);
5412         free_extent_buffer(next);
5413
5414         return ret;
5415 }
5416
5417 /*
5418  * helper to process tree block while walking up the tree.
5419  *
5420  * when wc->stage == DROP_REFERENCE, this function drops
5421  * reference count on the block.
5422  *
5423  * when wc->stage == UPDATE_BACKREF, this function changes
5424  * wc->stage back to DROP_REFERENCE if we changed wc->stage
5425  * to UPDATE_BACKREF previously while processing the block.
5426  *
5427  * NOTE: return value 1 means we should stop walking up.
5428  */
5429 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
5430                                  struct btrfs_root *root,
5431                                  struct btrfs_path *path,
5432                                  struct walk_control *wc)
5433 {
5434         struct btrfs_fs_info *fs_info = root->fs_info;
5435         int ret;
5436         int level = wc->level;
5437         struct extent_buffer *eb = path->nodes[level];
5438         u64 parent = 0;
5439
5440         if (wc->stage == UPDATE_BACKREF) {
5441                 BUG_ON(wc->shared_level < level);
5442                 if (level < wc->shared_level)
5443                         goto out;
5444
5445                 ret = find_next_key(path, level + 1, &wc->update_progress);
5446                 if (ret > 0)
5447                         wc->update_ref = 0;
5448
5449                 wc->stage = DROP_REFERENCE;
5450                 wc->shared_level = -1;
5451                 path->slots[level] = 0;
5452
5453                 /*
5454                  * check reference count again if the block isn't locked.
5455                  * we should start walking down the tree again if reference
5456                  * count is one.
5457                  */
5458                 if (!path->locks[level]) {
5459                         BUG_ON(level == 0);
5460                         btrfs_tree_lock(eb);
5461                         path->locks[level] = BTRFS_WRITE_LOCK;
5462
5463                         ret = btrfs_lookup_extent_info(trans, fs_info,
5464                                                        eb->start, level, 1,
5465                                                        &wc->refs[level],
5466                                                        &wc->flags[level]);
5467                         if (ret < 0) {
5468                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
5469                                 path->locks[level] = 0;
5470                                 return ret;
5471                         }
5472                         BUG_ON(wc->refs[level] == 0);
5473                         if (wc->refs[level] == 1) {
5474                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
5475                                 path->locks[level] = 0;
5476                                 return 1;
5477                         }
5478                 }
5479         }
5480
5481         /* wc->stage == DROP_REFERENCE */
5482         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
5483
5484         if (wc->refs[level] == 1) {
5485                 if (level == 0) {
5486                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
5487                                 ret = btrfs_dec_ref(trans, root, eb, 1);
5488                         else
5489                                 ret = btrfs_dec_ref(trans, root, eb, 0);
5490                         BUG_ON(ret); /* -ENOMEM */
5491                         if (is_fstree(root->root_key.objectid)) {
5492                                 ret = btrfs_qgroup_trace_leaf_items(trans, eb);
5493                                 if (ret) {
5494                                         btrfs_err_rl(fs_info,
5495         "error %d accounting leaf items, quota is out of sync, rescan required",
5496                                              ret);
5497                                 }
5498                         }
5499                 }
5500                 /* Make block locked assertion in btrfs_clear_buffer_dirty happy. */
5501                 if (!path->locks[level]) {
5502                         btrfs_tree_lock(eb);
5503                         path->locks[level] = BTRFS_WRITE_LOCK;
5504                 }
5505                 btrfs_clear_buffer_dirty(trans, eb);
5506         }
5507
5508         if (eb == root->node) {
5509                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
5510                         parent = eb->start;
5511                 else if (root->root_key.objectid != btrfs_header_owner(eb))
5512                         goto owner_mismatch;
5513         } else {
5514                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
5515                         parent = path->nodes[level + 1]->start;
5516                 else if (root->root_key.objectid !=
5517                          btrfs_header_owner(path->nodes[level + 1]))
5518                         goto owner_mismatch;
5519         }
5520
5521         btrfs_free_tree_block(trans, btrfs_root_id(root), eb, parent,
5522                               wc->refs[level] == 1);
5523 out:
5524         wc->refs[level] = 0;
5525         wc->flags[level] = 0;
5526         return 0;
5527
5528 owner_mismatch:
5529         btrfs_err_rl(fs_info, "unexpected tree owner, have %llu expect %llu",
5530                      btrfs_header_owner(eb), root->root_key.objectid);
5531         return -EUCLEAN;
5532 }
5533
5534 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
5535                                    struct btrfs_root *root,
5536                                    struct btrfs_path *path,
5537                                    struct walk_control *wc)
5538 {
5539         int level = wc->level;
5540         int lookup_info = 1;
5541         int ret = 0;
5542
5543         while (level >= 0) {
5544                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
5545                 if (ret)
5546                         break;
5547
5548                 if (level == 0)
5549                         break;
5550
5551                 if (path->slots[level] >=
5552                     btrfs_header_nritems(path->nodes[level]))
5553                         break;
5554
5555                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
5556                 if (ret > 0) {
5557                         path->slots[level]++;
5558                         continue;
5559                 } else if (ret < 0)
5560                         break;
5561                 level = wc->level;
5562         }
5563         return (ret == 1) ? 0 : ret;
5564 }
5565
5566 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
5567                                  struct btrfs_root *root,
5568                                  struct btrfs_path *path,
5569                                  struct walk_control *wc, int max_level)
5570 {
5571         int level = wc->level;
5572         int ret;
5573
5574         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
5575         while (level < max_level && path->nodes[level]) {
5576                 wc->level = level;
5577                 if (path->slots[level] + 1 <
5578                     btrfs_header_nritems(path->nodes[level])) {
5579                         path->slots[level]++;
5580                         return 0;
5581                 } else {
5582                         ret = walk_up_proc(trans, root, path, wc);
5583                         if (ret > 0)
5584                                 return 0;
5585                         if (ret < 0)
5586                                 return ret;
5587
5588                         if (path->locks[level]) {
5589                                 btrfs_tree_unlock_rw(path->nodes[level],
5590                                                      path->locks[level]);
5591                                 path->locks[level] = 0;
5592                         }
5593                         free_extent_buffer(path->nodes[level]);
5594                         path->nodes[level] = NULL;
5595                         level++;
5596                 }
5597         }
5598         return 1;
5599 }
5600
5601 /*
5602  * drop a subvolume tree.
5603  *
5604  * this function traverses the tree freeing any blocks that only
5605  * referenced by the tree.
5606  *
5607  * when a shared tree block is found. this function decreases its
5608  * reference count by one. if update_ref is true, this function
5609  * also make sure backrefs for the shared block and all lower level
5610  * blocks are properly updated.
5611  *
5612  * If called with for_reloc == 0, may exit early with -EAGAIN
5613  */
5614 int btrfs_drop_snapshot(struct btrfs_root *root, int update_ref, int for_reloc)
5615 {
5616         const bool is_reloc_root = (root->root_key.objectid ==
5617                                     BTRFS_TREE_RELOC_OBJECTID);
5618         struct btrfs_fs_info *fs_info = root->fs_info;
5619         struct btrfs_path *path;
5620         struct btrfs_trans_handle *trans;
5621         struct btrfs_root *tree_root = fs_info->tree_root;
5622         struct btrfs_root_item *root_item = &root->root_item;
5623         struct walk_control *wc;
5624         struct btrfs_key key;
5625         int err = 0;
5626         int ret;
5627         int level;
5628         bool root_dropped = false;
5629         bool unfinished_drop = false;
5630
5631         btrfs_debug(fs_info, "Drop subvolume %llu", root->root_key.objectid);
5632
5633         path = btrfs_alloc_path();
5634         if (!path) {
5635                 err = -ENOMEM;
5636                 goto out;
5637         }
5638
5639         wc = kzalloc(sizeof(*wc), GFP_NOFS);
5640         if (!wc) {
5641                 btrfs_free_path(path);
5642                 err = -ENOMEM;
5643                 goto out;
5644         }
5645
5646         /*
5647          * Use join to avoid potential EINTR from transaction start. See
5648          * wait_reserve_ticket and the whole reservation callchain.
5649          */
5650         if (for_reloc)
5651                 trans = btrfs_join_transaction(tree_root);
5652         else
5653                 trans = btrfs_start_transaction(tree_root, 0);
5654         if (IS_ERR(trans)) {
5655                 err = PTR_ERR(trans);
5656                 goto out_free;
5657         }
5658
5659         err = btrfs_run_delayed_items(trans);
5660         if (err)
5661                 goto out_end_trans;
5662
5663         /*
5664          * This will help us catch people modifying the fs tree while we're
5665          * dropping it.  It is unsafe to mess with the fs tree while it's being
5666          * dropped as we unlock the root node and parent nodes as we walk down
5667          * the tree, assuming nothing will change.  If something does change
5668          * then we'll have stale information and drop references to blocks we've
5669          * already dropped.
5670          */
5671         set_bit(BTRFS_ROOT_DELETING, &root->state);
5672         unfinished_drop = test_bit(BTRFS_ROOT_UNFINISHED_DROP, &root->state);
5673
5674         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
5675                 level = btrfs_header_level(root->node);
5676                 path->nodes[level] = btrfs_lock_root_node(root);
5677                 path->slots[level] = 0;
5678                 path->locks[level] = BTRFS_WRITE_LOCK;
5679                 memset(&wc->update_progress, 0,
5680                        sizeof(wc->update_progress));
5681         } else {
5682                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
5683                 memcpy(&wc->update_progress, &key,
5684                        sizeof(wc->update_progress));
5685
5686                 level = btrfs_root_drop_level(root_item);
5687                 BUG_ON(level == 0);
5688                 path->lowest_level = level;
5689                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5690                 path->lowest_level = 0;
5691                 if (ret < 0) {
5692                         err = ret;
5693                         goto out_end_trans;
5694                 }
5695                 WARN_ON(ret > 0);
5696
5697                 /*
5698                  * unlock our path, this is safe because only this
5699                  * function is allowed to delete this snapshot
5700                  */
5701                 btrfs_unlock_up_safe(path, 0);
5702
5703                 level = btrfs_header_level(root->node);
5704                 while (1) {
5705                         btrfs_tree_lock(path->nodes[level]);
5706                         path->locks[level] = BTRFS_WRITE_LOCK;
5707
5708                         ret = btrfs_lookup_extent_info(trans, fs_info,
5709                                                 path->nodes[level]->start,
5710                                                 level, 1, &wc->refs[level],
5711                                                 &wc->flags[level]);
5712                         if (ret < 0) {
5713                                 err = ret;
5714                                 goto out_end_trans;
5715                         }
5716                         BUG_ON(wc->refs[level] == 0);
5717
5718                         if (level == btrfs_root_drop_level(root_item))
5719                                 break;
5720
5721                         btrfs_tree_unlock(path->nodes[level]);
5722                         path->locks[level] = 0;
5723                         WARN_ON(wc->refs[level] != 1);
5724                         level--;
5725                 }
5726         }
5727
5728         wc->restarted = test_bit(BTRFS_ROOT_DEAD_TREE, &root->state);
5729         wc->level = level;
5730         wc->shared_level = -1;
5731         wc->stage = DROP_REFERENCE;
5732         wc->update_ref = update_ref;
5733         wc->keep_locks = 0;
5734         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
5735
5736         while (1) {
5737
5738                 ret = walk_down_tree(trans, root, path, wc);
5739                 if (ret < 0) {
5740                         btrfs_abort_transaction(trans, ret);
5741                         err = ret;
5742                         break;
5743                 }
5744
5745                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
5746                 if (ret < 0) {
5747                         btrfs_abort_transaction(trans, ret);
5748                         err = ret;
5749                         break;
5750                 }
5751
5752                 if (ret > 0) {
5753                         BUG_ON(wc->stage != DROP_REFERENCE);
5754                         break;
5755                 }
5756
5757                 if (wc->stage == DROP_REFERENCE) {
5758                         wc->drop_level = wc->level;
5759                         btrfs_node_key_to_cpu(path->nodes[wc->drop_level],
5760                                               &wc->drop_progress,
5761                                               path->slots[wc->drop_level]);
5762                 }
5763                 btrfs_cpu_key_to_disk(&root_item->drop_progress,
5764                                       &wc->drop_progress);
5765                 btrfs_set_root_drop_level(root_item, wc->drop_level);
5766
5767                 BUG_ON(wc->level == 0);
5768                 if (btrfs_should_end_transaction(trans) ||
5769                     (!for_reloc && btrfs_need_cleaner_sleep(fs_info))) {
5770                         ret = btrfs_update_root(trans, tree_root,
5771                                                 &root->root_key,
5772                                                 root_item);
5773                         if (ret) {
5774                                 btrfs_abort_transaction(trans, ret);
5775                                 err = ret;
5776                                 goto out_end_trans;
5777                         }
5778
5779                         if (!is_reloc_root)
5780                                 btrfs_set_last_root_drop_gen(fs_info, trans->transid);
5781
5782                         btrfs_end_transaction_throttle(trans);
5783                         if (!for_reloc && btrfs_need_cleaner_sleep(fs_info)) {
5784                                 btrfs_debug(fs_info,
5785                                             "drop snapshot early exit");
5786                                 err = -EAGAIN;
5787                                 goto out_free;
5788                         }
5789
5790                        /*
5791                         * Use join to avoid potential EINTR from transaction
5792                         * start. See wait_reserve_ticket and the whole
5793                         * reservation callchain.
5794                         */
5795                         if (for_reloc)
5796                                 trans = btrfs_join_transaction(tree_root);
5797                         else
5798                                 trans = btrfs_start_transaction(tree_root, 0);
5799                         if (IS_ERR(trans)) {
5800                                 err = PTR_ERR(trans);
5801                                 goto out_free;
5802                         }
5803                 }
5804         }
5805         btrfs_release_path(path);
5806         if (err)
5807                 goto out_end_trans;
5808
5809         ret = btrfs_del_root(trans, &root->root_key);
5810         if (ret) {
5811                 btrfs_abort_transaction(trans, ret);
5812                 err = ret;
5813                 goto out_end_trans;
5814         }
5815
5816         if (!is_reloc_root) {
5817                 ret = btrfs_find_root(tree_root, &root->root_key, path,
5818                                       NULL, NULL);
5819                 if (ret < 0) {
5820                         btrfs_abort_transaction(trans, ret);
5821                         err = ret;
5822                         goto out_end_trans;
5823                 } else if (ret > 0) {
5824                         /* if we fail to delete the orphan item this time
5825                          * around, it'll get picked up the next time.
5826                          *
5827                          * The most common failure here is just -ENOENT.
5828                          */
5829                         btrfs_del_orphan_item(trans, tree_root,
5830                                               root->root_key.objectid);
5831                 }
5832         }
5833
5834         /*
5835          * This subvolume is going to be completely dropped, and won't be
5836          * recorded as dirty roots, thus pertrans meta rsv will not be freed at
5837          * commit transaction time.  So free it here manually.
5838          */
5839         btrfs_qgroup_convert_reserved_meta(root, INT_MAX);
5840         btrfs_qgroup_free_meta_all_pertrans(root);
5841
5842         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state))
5843                 btrfs_add_dropped_root(trans, root);
5844         else
5845                 btrfs_put_root(root);
5846         root_dropped = true;
5847 out_end_trans:
5848         if (!is_reloc_root)
5849                 btrfs_set_last_root_drop_gen(fs_info, trans->transid);
5850
5851         btrfs_end_transaction_throttle(trans);
5852 out_free:
5853         kfree(wc);
5854         btrfs_free_path(path);
5855 out:
5856         /*
5857          * We were an unfinished drop root, check to see if there are any
5858          * pending, and if not clear and wake up any waiters.
5859          */
5860         if (!err && unfinished_drop)
5861                 btrfs_maybe_wake_unfinished_drop(fs_info);
5862
5863         /*
5864          * So if we need to stop dropping the snapshot for whatever reason we
5865          * need to make sure to add it back to the dead root list so that we
5866          * keep trying to do the work later.  This also cleans up roots if we
5867          * don't have it in the radix (like when we recover after a power fail
5868          * or unmount) so we don't leak memory.
5869          */
5870         if (!for_reloc && !root_dropped)
5871                 btrfs_add_dead_root(root);
5872         return err;
5873 }
5874
5875 /*
5876  * drop subtree rooted at tree block 'node'.
5877  *
5878  * NOTE: this function will unlock and release tree block 'node'
5879  * only used by relocation code
5880  */
5881 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
5882                         struct btrfs_root *root,
5883                         struct extent_buffer *node,
5884                         struct extent_buffer *parent)
5885 {
5886         struct btrfs_fs_info *fs_info = root->fs_info;
5887         struct btrfs_path *path;
5888         struct walk_control *wc;
5889         int level;
5890         int parent_level;
5891         int ret = 0;
5892         int wret;
5893
5894         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
5895
5896         path = btrfs_alloc_path();
5897         if (!path)
5898                 return -ENOMEM;
5899
5900         wc = kzalloc(sizeof(*wc), GFP_NOFS);
5901         if (!wc) {
5902                 btrfs_free_path(path);
5903                 return -ENOMEM;
5904         }
5905
5906         btrfs_assert_tree_write_locked(parent);
5907         parent_level = btrfs_header_level(parent);
5908         atomic_inc(&parent->refs);
5909         path->nodes[parent_level] = parent;
5910         path->slots[parent_level] = btrfs_header_nritems(parent);
5911
5912         btrfs_assert_tree_write_locked(node);
5913         level = btrfs_header_level(node);
5914         path->nodes[level] = node;
5915         path->slots[level] = 0;
5916         path->locks[level] = BTRFS_WRITE_LOCK;
5917
5918         wc->refs[parent_level] = 1;
5919         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
5920         wc->level = level;
5921         wc->shared_level = -1;
5922         wc->stage = DROP_REFERENCE;
5923         wc->update_ref = 0;
5924         wc->keep_locks = 1;
5925         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
5926
5927         while (1) {
5928                 wret = walk_down_tree(trans, root, path, wc);
5929                 if (wret < 0) {
5930                         ret = wret;
5931                         break;
5932                 }
5933
5934                 wret = walk_up_tree(trans, root, path, wc, parent_level);
5935                 if (wret < 0)
5936                         ret = wret;
5937                 if (wret != 0)
5938                         break;
5939         }
5940
5941         kfree(wc);
5942         btrfs_free_path(path);
5943         return ret;
5944 }
5945
5946 int btrfs_error_unpin_extent_range(struct btrfs_fs_info *fs_info,
5947                                    u64 start, u64 end)
5948 {
5949         return unpin_extent_range(fs_info, start, end, false);
5950 }
5951
5952 /*
5953  * It used to be that old block groups would be left around forever.
5954  * Iterating over them would be enough to trim unused space.  Since we
5955  * now automatically remove them, we also need to iterate over unallocated
5956  * space.
5957  *
5958  * We don't want a transaction for this since the discard may take a
5959  * substantial amount of time.  We don't require that a transaction be
5960  * running, but we do need to take a running transaction into account
5961  * to ensure that we're not discarding chunks that were released or
5962  * allocated in the current transaction.
5963  *
5964  * Holding the chunks lock will prevent other threads from allocating
5965  * or releasing chunks, but it won't prevent a running transaction
5966  * from committing and releasing the memory that the pending chunks
5967  * list head uses.  For that, we need to take a reference to the
5968  * transaction and hold the commit root sem.  We only need to hold
5969  * it while performing the free space search since we have already
5970  * held back allocations.
5971  */
5972 static int btrfs_trim_free_extents(struct btrfs_device *device, u64 *trimmed)
5973 {
5974         u64 start = BTRFS_DEVICE_RANGE_RESERVED, len = 0, end = 0;
5975         int ret;
5976
5977         *trimmed = 0;
5978
5979         /* Discard not supported = nothing to do. */
5980         if (!bdev_max_discard_sectors(device->bdev))
5981                 return 0;
5982
5983         /* Not writable = nothing to do. */
5984         if (!test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state))
5985                 return 0;
5986
5987         /* No free space = nothing to do. */
5988         if (device->total_bytes <= device->bytes_used)
5989                 return 0;
5990
5991         ret = 0;
5992
5993         while (1) {
5994                 struct btrfs_fs_info *fs_info = device->fs_info;
5995                 u64 bytes;
5996
5997                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
5998                 if (ret)
5999                         break;
6000
6001                 find_first_clear_extent_bit(&device->alloc_state, start,
6002                                             &start, &end,
6003                                             CHUNK_TRIMMED | CHUNK_ALLOCATED);
6004
6005                 /* Check if there are any CHUNK_* bits left */
6006                 if (start > device->total_bytes) {
6007                         WARN_ON(IS_ENABLED(CONFIG_BTRFS_DEBUG));
6008                         btrfs_warn_in_rcu(fs_info,
6009 "ignoring attempt to trim beyond device size: offset %llu length %llu device %s device size %llu",
6010                                           start, end - start + 1,
6011                                           btrfs_dev_name(device),
6012                                           device->total_bytes);
6013                         mutex_unlock(&fs_info->chunk_mutex);
6014                         ret = 0;
6015                         break;
6016                 }
6017
6018                 /* Ensure we skip the reserved space on each device. */
6019                 start = max_t(u64, start, BTRFS_DEVICE_RANGE_RESERVED);
6020
6021                 /*
6022                  * If find_first_clear_extent_bit find a range that spans the
6023                  * end of the device it will set end to -1, in this case it's up
6024                  * to the caller to trim the value to the size of the device.
6025                  */
6026                 end = min(end, device->total_bytes - 1);
6027
6028                 len = end - start + 1;
6029
6030                 /* We didn't find any extents */
6031                 if (!len) {
6032                         mutex_unlock(&fs_info->chunk_mutex);
6033                         ret = 0;
6034                         break;
6035                 }
6036
6037                 ret = btrfs_issue_discard(device->bdev, start, len,
6038                                           &bytes);
6039                 if (!ret)
6040                         set_extent_bit(&device->alloc_state, start,
6041                                        start + bytes - 1, CHUNK_TRIMMED, NULL);
6042                 mutex_unlock(&fs_info->chunk_mutex);
6043
6044                 if (ret)
6045                         break;
6046
6047                 start += len;
6048                 *trimmed += bytes;
6049
6050                 if (fatal_signal_pending(current)) {
6051                         ret = -ERESTARTSYS;
6052                         break;
6053                 }
6054
6055                 cond_resched();
6056         }
6057
6058         return ret;
6059 }
6060
6061 /*
6062  * Trim the whole filesystem by:
6063  * 1) trimming the free space in each block group
6064  * 2) trimming the unallocated space on each device
6065  *
6066  * This will also continue trimming even if a block group or device encounters
6067  * an error.  The return value will be the last error, or 0 if nothing bad
6068  * happens.
6069  */
6070 int btrfs_trim_fs(struct btrfs_fs_info *fs_info, struct fstrim_range *range)
6071 {
6072         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
6073         struct btrfs_block_group *cache = NULL;
6074         struct btrfs_device *device;
6075         u64 group_trimmed;
6076         u64 range_end = U64_MAX;
6077         u64 start;
6078         u64 end;
6079         u64 trimmed = 0;
6080         u64 bg_failed = 0;
6081         u64 dev_failed = 0;
6082         int bg_ret = 0;
6083         int dev_ret = 0;
6084         int ret = 0;
6085
6086         if (range->start == U64_MAX)
6087                 return -EINVAL;
6088
6089         /*
6090          * Check range overflow if range->len is set.
6091          * The default range->len is U64_MAX.
6092          */
6093         if (range->len != U64_MAX &&
6094             check_add_overflow(range->start, range->len, &range_end))
6095                 return -EINVAL;
6096
6097         cache = btrfs_lookup_first_block_group(fs_info, range->start);
6098         for (; cache; cache = btrfs_next_block_group(cache)) {
6099                 if (cache->start >= range_end) {
6100                         btrfs_put_block_group(cache);
6101                         break;
6102                 }
6103
6104                 start = max(range->start, cache->start);
6105                 end = min(range_end, cache->start + cache->length);
6106
6107                 if (end - start >= range->minlen) {
6108                         if (!btrfs_block_group_done(cache)) {
6109                                 ret = btrfs_cache_block_group(cache, true);
6110                                 if (ret) {
6111                                         bg_failed++;
6112                                         bg_ret = ret;
6113                                         continue;
6114                                 }
6115                         }
6116                         ret = btrfs_trim_block_group(cache,
6117                                                      &group_trimmed,
6118                                                      start,
6119                                                      end,
6120                                                      range->minlen);
6121
6122                         trimmed += group_trimmed;
6123                         if (ret) {
6124                                 bg_failed++;
6125                                 bg_ret = ret;
6126                                 continue;
6127                         }
6128                 }
6129         }
6130
6131         if (bg_failed)
6132                 btrfs_warn(fs_info,
6133                         "failed to trim %llu block group(s), last error %d",
6134                         bg_failed, bg_ret);
6135
6136         mutex_lock(&fs_devices->device_list_mutex);
6137         list_for_each_entry(device, &fs_devices->devices, dev_list) {
6138                 if (test_bit(BTRFS_DEV_STATE_MISSING, &device->dev_state))
6139                         continue;
6140
6141                 ret = btrfs_trim_free_extents(device, &group_trimmed);
6142                 if (ret) {
6143                         dev_failed++;
6144                         dev_ret = ret;
6145                         break;
6146                 }
6147
6148                 trimmed += group_trimmed;
6149         }
6150         mutex_unlock(&fs_devices->device_list_mutex);
6151
6152         if (dev_failed)
6153                 btrfs_warn(fs_info,
6154                         "failed to trim %llu device(s), last error %d",
6155                         dev_failed, dev_ret);
6156         range->len = trimmed;
6157         if (bg_ret)
6158                 return bg_ret;
6159         return dev_ret;
6160 }