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 "transaction.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 "qgroup.h"
  30 #include "ref-verify.h"
  31 #include "space-info.h"
  32 #include "block-rsv.h"
  33 #include "discard.h"
  34 #include "zoned.h"
  35 #include "dev-replace.h"
  36 #include "fs.h"
  37 #include "accessors.h"
  38 #include "root-tree.h"
  39 #include "file-item.h"
  40 #include "orphan.h"
  41 #include "tree-checker.h"
  42 #include "raid-stripe-tree.h"
  43
  44 #undef SCRAMBLE_DELAYED_REFS
  45
  46
  47 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  48                                struct btrfs_delayed_ref_head *href,
  49                                struct btrfs_delayed_ref_node *node, u64 parent,
  50                                u64 root_objectid, u64 owner_objectid,
  51                                u64 owner_offset,
  52                                struct btrfs_delayed_extent_op *extra_op);
  53 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  54                                     struct extent_buffer *leaf,
  55                                     struct btrfs_extent_item *ei);
  56 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  57                                       u64 parent, u64 root_objectid,
  58                                       u64 flags, u64 owner, u64 offset,
  59                                       struct btrfs_key *ins, int ref_mod, u64 oref_root);
  60 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  61                                      struct btrfs_delayed_ref_node *node,
  62                                      struct btrfs_delayed_extent_op *extent_op);
  63 static int find_next_key(struct btrfs_path *path, int level,
  64                          struct btrfs_key *key);
  65
  66 static int block_group_bits(struct btrfs_block_group *cache, u64 bits)
  67 {
  68         return (cache->flags & bits) == bits;
  69 }
  70
  71 /* simple helper to search for an existing data extent at a given offset */
  72 int btrfs_lookup_data_extent(struct btrfs_fs_info *fs_info, u64 start, u64 len)
  73 {
  74         struct btrfs_root *root = btrfs_extent_root(fs_info, start);
  75         int ret;
  76         struct btrfs_key key;
  77         struct btrfs_path *path;
  78
  79         path = btrfs_alloc_path();
  80         if (!path)
  81                 return -ENOMEM;
  82
  83         key.objectid = start;
  84         key.offset = len;
  85         key.type = BTRFS_EXTENT_ITEM_KEY;
  86         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
  87         btrfs_free_path(path);
  88         return ret;
  89 }
  90
  91 /*
  92  * helper function to lookup reference count and flags of a tree block.
  93  *
  94  * the head node for delayed ref is used to store the sum of all the
  95  * reference count modifications queued up in the rbtree. the head
  96  * node may also store the extent flags to set. This way you can check
  97  * to see what the reference count and extent flags would be if all of
  98  * the delayed refs are not processed.
  99  */
 100 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 101                              struct btrfs_fs_info *fs_info, u64 bytenr,
 102                              u64 offset, int metadata, u64 *refs, u64 *flags,
 103                              u64 *owning_root)
 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         u64 owner = 0;
 116         int ret;
 117
 118         /*
 119          * If we don't have skinny metadata, don't bother doing anything
 120          * different
 121          */
 122         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA)) {
 123                 offset = fs_info->nodesize;
 124                 metadata = 0;
 125         }
 126
 127         path = btrfs_alloc_path();
 128         if (!path)
 129                 return -ENOMEM;
 130
 131         if (!trans) {
 132                 path->skip_locking = 1;
 133                 path->search_commit_root = 1;
 134         }
 135
 136 search_again:
 137         key.objectid = bytenr;
 138         key.offset = offset;
 139         if (metadata)
 140                 key.type = BTRFS_METADATA_ITEM_KEY;
 141         else
 142                 key.type = BTRFS_EXTENT_ITEM_KEY;
 143
 144         extent_root = btrfs_extent_root(fs_info, bytenr);
 145         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 146         if (ret < 0)
 147                 goto out_free;
 148
 149         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 150                 if (path->slots[0]) {
 151                         path->slots[0]--;
 152                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 153                                               path->slots[0]);
 154                         if (key.objectid == bytenr &&
 155                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 156                             key.offset == fs_info->nodesize)
 157                                 ret = 0;
 158                 }
 159         }
 160
 161         if (ret == 0) {
 162                 leaf = path->nodes[0];
 163                 item_size = btrfs_item_size(leaf, path->slots[0]);
 164                 if (item_size >= sizeof(*ei)) {
 165                         ei = btrfs_item_ptr(leaf, path->slots[0],
 166                                             struct btrfs_extent_item);
 167                         num_refs = btrfs_extent_refs(leaf, ei);
 168                         extent_flags = btrfs_extent_flags(leaf, ei);
 169                         owner = btrfs_get_extent_owner_root(fs_info, leaf,
 170                                                             path->slots[0]);
 171                 } else {
 172                         ret = -EUCLEAN;
 173                         btrfs_err(fs_info,
 174                         "unexpected extent item size, has %u expect >= %zu",
 175                                   item_size, sizeof(*ei));
 176                         if (trans)
 177                                 btrfs_abort_transaction(trans, ret);
 178                         else
 179                                 btrfs_handle_fs_error(fs_info, ret, NULL);
 180
 181                         goto out_free;
 182                 }
 183
 184                 BUG_ON(num_refs == 0);
 185         } else {
 186                 num_refs = 0;
 187                 extent_flags = 0;
 188                 ret = 0;
 189         }
 190
 191         if (!trans)
 192                 goto out;
 193
 194         delayed_refs = &trans->transaction->delayed_refs;
 195         spin_lock(&delayed_refs->lock);
 196         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
 197         if (head) {
 198                 if (!mutex_trylock(&head->mutex)) {
 199                         refcount_inc(&head->refs);
 200                         spin_unlock(&delayed_refs->lock);
 201
 202                         btrfs_release_path(path);
 203
 204                         /*
 205                          * Mutex was contended, block until it's released and try
 206                          * again
 207                          */
 208                         mutex_lock(&head->mutex);
 209                         mutex_unlock(&head->mutex);
 210                         btrfs_put_delayed_ref_head(head);
 211                         goto search_again;
 212                 }
 213                 spin_lock(&head->lock);
 214                 if (head->extent_op && head->extent_op->update_flags)
 215                         extent_flags |= head->extent_op->flags_to_set;
 216                 else
 217                         BUG_ON(num_refs == 0);
 218
 219                 num_refs += head->ref_mod;
 220                 spin_unlock(&head->lock);
 221                 mutex_unlock(&head->mutex);
 222         }
 223         spin_unlock(&delayed_refs->lock);
 224 out:
 225         WARN_ON(num_refs == 0);
 226         if (refs)
 227                 *refs = num_refs;
 228         if (flags)
 229                 *flags = extent_flags;
 230         if (owning_root)
 231                 *owning_root = owner;
 232 out_free:
 233         btrfs_free_path(path);
 234         return ret;
 235 }
 236
 237 /*
 238  * Back reference rules.  Back refs have three main goals:
 239  *
 240  * 1) differentiate between all holders of references to an extent so that
 241  *    when a reference is dropped we can make sure it was a valid reference
 242  *    before freeing the extent.
 243  *
 244  * 2) Provide enough information to quickly find the holders of an extent
 245  *    if we notice a given block is corrupted or bad.
 246  *
 247  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 248  *    maintenance.  This is actually the same as #2, but with a slightly
 249  *    different use case.
 250  *
 251  * There are two kinds of back refs. The implicit back refs is optimized
 252  * for pointers in non-shared tree blocks. For a given pointer in a block,
 253  * back refs of this kind provide information about the block's owner tree
 254  * and the pointer's key. These information allow us to find the block by
 255  * b-tree searching. The full back refs is for pointers in tree blocks not
 256  * referenced by their owner trees. The location of tree block is recorded
 257  * in the back refs. Actually the full back refs is generic, and can be
 258  * used in all cases the implicit back refs is used. The major shortcoming
 259  * of the full back refs is its overhead. Every time a tree block gets
 260  * COWed, we have to update back refs entry for all pointers in it.
 261  *
 262  * For a newly allocated tree block, we use implicit back refs for
 263  * pointers in it. This means most tree related operations only involve
 264  * implicit back refs. For a tree block created in old transaction, the
 265  * only way to drop a reference to it is COW it. So we can detect the
 266  * event that tree block loses its owner tree's reference and do the
 267  * back refs conversion.
 268  *
 269  * When a tree block is COWed through a tree, there are four cases:
 270  *
 271  * The reference count of the block is one and the tree is the block's
 272  * owner tree. Nothing to do in this case.
 273  *
 274  * The reference count of the block is one and the tree is not the
 275  * block's owner tree. In this case, full back refs is used for pointers
 276  * in the block. Remove these full back refs, add implicit back refs for
 277  * every pointers in the new block.
 278  *
 279  * The reference count of the block is greater than one and the tree is
 280  * the block's owner tree. In this case, implicit back refs is used for
 281  * pointers in the block. Add full back refs for every pointers in the
 282  * block, increase lower level extents' reference counts. The original
 283  * implicit back refs are entailed to the new block.
 284  *
 285  * The reference count of the block is greater than one and the tree is
 286  * not the block's owner tree. Add implicit back refs for every pointer in
 287  * the new block, increase lower level extents' reference count.
 288  *
 289  * Back Reference Key composing:
 290  *
 291  * The key objectid corresponds to the first byte in the extent,
 292  * The key type is used to differentiate between types of back refs.
 293  * There are different meanings of the key offset for different types
 294  * of back refs.
 295  *
 296  * File extents can be referenced by:
 297  *
 298  * - multiple snapshots, subvolumes, or different generations in one subvol
 299  * - different files inside a single subvolume
 300  * - different offsets inside a file (bookend extents in file.c)
 301  *
 302  * The extent ref structure for the implicit back refs has fields for:
 303  *
 304  * - Objectid of the subvolume root
 305  * - objectid of the file holding the reference
 306  * - original offset in the file
 307  * - how many bookend extents
 308  *
 309  * The key offset for the implicit back refs is hash of the first
 310  * three fields.
 311  *
 312  * The extent ref structure for the full back refs has field for:
 313  *
 314  * - number of pointers in the tree leaf
 315  *
 316  * The key offset for the implicit back refs is the first byte of
 317  * the tree leaf
 318  *
 319  * When a file extent is allocated, The implicit back refs is used.
 320  * the fields are filled in:
 321  *
 322  *     (root_key.objectid, inode objectid, offset in file, 1)
 323  *
 324  * When a file extent is removed file truncation, we find the
 325  * corresponding implicit back refs and check the following fields:
 326  *
 327  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
 328  *
 329  * Btree extents can be referenced by:
 330  *
 331  * - Different subvolumes
 332  *
 333  * Both the implicit back refs and the full back refs for tree blocks
 334  * only consist of key. The key offset for the implicit back refs is
 335  * objectid of block's owner tree. The key offset for the full back refs
 336  * is the first byte of parent block.
 337  *
 338  * When implicit back refs is used, information about the lowest key and
 339  * level of the tree block are required. These information are stored in
 340  * tree block info structure.
 341  */
 342
 343 /*
 344  * is_data == BTRFS_REF_TYPE_BLOCK, tree block type is required,
 345  * is_data == BTRFS_REF_TYPE_DATA, data type is requiried,
 346  * is_data == BTRFS_REF_TYPE_ANY, either type is OK.
 347  */
 348 int btrfs_get_extent_inline_ref_type(const struct extent_buffer *eb,
 349                                      struct btrfs_extent_inline_ref *iref,
 350                                      enum btrfs_inline_ref_type is_data)
 351 {
 352         struct btrfs_fs_info *fs_info = eb->fs_info;
 353         int type = btrfs_extent_inline_ref_type(eb, iref);
 354         u64 offset = btrfs_extent_inline_ref_offset(eb, iref);
 355
 356         if (type == BTRFS_EXTENT_OWNER_REF_KEY) {
 357                 ASSERT(btrfs_fs_incompat(fs_info, SIMPLE_QUOTA));
 358                 return type;
 359         }
 360
 361         if (type == BTRFS_TREE_BLOCK_REF_KEY ||
 362             type == BTRFS_SHARED_BLOCK_REF_KEY ||
 363             type == BTRFS_SHARED_DATA_REF_KEY ||
 364             type == BTRFS_EXTENT_DATA_REF_KEY) {
 365                 if (is_data == BTRFS_REF_TYPE_BLOCK) {
 366                         if (type == BTRFS_TREE_BLOCK_REF_KEY)
 367                                 return type;
 368                         if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
 369                                 ASSERT(fs_info);
 370                                 /*
 371                                  * Every shared one has parent tree block,
 372                                  * which must be aligned to sector size.
 373                                  */
 374                                 if (offset && IS_ALIGNED(offset, fs_info->sectorsize))
 375                                         return type;
 376                         }
 377                 } else if (is_data == BTRFS_REF_TYPE_DATA) {
 378                         if (type == BTRFS_EXTENT_DATA_REF_KEY)
 379                                 return type;
 380                         if (type == BTRFS_SHARED_DATA_REF_KEY) {
 381                                 ASSERT(fs_info);
 382                                 /*
 383                                  * Every shared one has parent tree block,
 384                                  * which must be aligned to sector size.
 385                                  */
 386                                 if (offset &&
 387                                     IS_ALIGNED(offset, fs_info->sectorsize))
 388                                         return type;
 389                         }
 390                 } else {
 391                         ASSERT(is_data == BTRFS_REF_TYPE_ANY);
 392                         return type;
 393                 }
 394         }
 395
 396         WARN_ON(1);
 397         btrfs_print_leaf(eb);
 398         btrfs_err(fs_info,
 399                   "eb %llu iref 0x%lx invalid extent inline ref type %d",
 400                   eb->start, (unsigned long)iref, type);
 401
 402         return BTRFS_REF_TYPE_INVALID;
 403 }
 404
 405 u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
 406 {
 407         u32 high_crc = ~(u32)0;
 408         u32 low_crc = ~(u32)0;
 409         __le64 lenum;
 410
 411         lenum = cpu_to_le64(root_objectid);
 412         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
 413         lenum = cpu_to_le64(owner);
 414         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
 415         lenum = cpu_to_le64(offset);
 416         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
 417
 418         return ((u64)high_crc << 31) ^ (u64)low_crc;
 419 }
 420
 421 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
 422                                      struct btrfs_extent_data_ref *ref)
 423 {
 424         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
 425                                     btrfs_extent_data_ref_objectid(leaf, ref),
 426                                     btrfs_extent_data_ref_offset(leaf, ref));
 427 }
 428
 429 static int match_extent_data_ref(struct extent_buffer *leaf,
 430                                  struct btrfs_extent_data_ref *ref,
 431                                  u64 root_objectid, u64 owner, u64 offset)
 432 {
 433         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
 434             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
 435             btrfs_extent_data_ref_offset(leaf, ref) != offset)
 436                 return 0;
 437         return 1;
 438 }
 439
 440 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
 441                                            struct btrfs_path *path,
 442                                            u64 bytenr, u64 parent,
 443                                            u64 root_objectid,
 444                                            u64 owner, u64 offset)
 445 {
 446         struct btrfs_root *root = btrfs_extent_root(trans->fs_info, bytenr);
 447         struct btrfs_key key;
 448         struct btrfs_extent_data_ref *ref;
 449         struct extent_buffer *leaf;
 450         u32 nritems;
 451         int ret;
 452         int recow;
 453         int err = -ENOENT;
 454
 455         key.objectid = bytenr;
 456         if (parent) {
 457                 key.type = BTRFS_SHARED_DATA_REF_KEY;
 458                 key.offset = parent;
 459         } else {
 460                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
 461                 key.offset = hash_extent_data_ref(root_objectid,
 462                                                   owner, offset);
 463         }
 464 again:
 465         recow = 0;
 466         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
 467         if (ret < 0) {
 468                 err = ret;
 469                 goto fail;
 470         }
 471
 472         if (parent) {
 473                 if (!ret)
 474                         return 0;
 475                 goto fail;
 476         }
 477
 478         leaf = path->nodes[0];
 479         nritems = btrfs_header_nritems(leaf);
 480         while (1) {
 481                 if (path->slots[0] >= nritems) {
 482                         ret = btrfs_next_leaf(root, path);
 483                         if (ret < 0)
 484                                 err = ret;
 485                         if (ret)
 486                                 goto fail;
 487
 488                         leaf = path->nodes[0];
 489                         nritems = btrfs_header_nritems(leaf);
 490                         recow = 1;
 491                 }
 492
 493                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 494                 if (key.objectid != bytenr ||
 495                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
 496                         goto fail;
 497
 498                 ref = btrfs_item_ptr(leaf, path->slots[0],
 499                                      struct btrfs_extent_data_ref);
 500
 501                 if (match_extent_data_ref(leaf, ref, root_objectid,
 502                                           owner, offset)) {
 503                         if (recow) {
 504                                 btrfs_release_path(path);
 505                                 goto again;
 506                         }
 507                         err = 0;
 508                         break;
 509                 }
 510                 path->slots[0]++;
 511         }
 512 fail:
 513         return err;
 514 }
 515
 516 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
 517                                            struct btrfs_path *path,
 518                                            u64 bytenr, u64 parent,
 519                                            u64 root_objectid, u64 owner,
 520                                            u64 offset, int refs_to_add)
 521 {
 522         struct btrfs_root *root = btrfs_extent_root(trans->fs_info, bytenr);
 523         struct btrfs_key key;
 524         struct extent_buffer *leaf;
 525         u32 size;
 526         u32 num_refs;
 527         int ret;
 528
 529         key.objectid = bytenr;
 530         if (parent) {
 531                 key.type = BTRFS_SHARED_DATA_REF_KEY;
 532                 key.offset = parent;
 533                 size = sizeof(struct btrfs_shared_data_ref);
 534         } else {
 535                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
 536                 key.offset = hash_extent_data_ref(root_objectid,
 537                                                   owner, offset);
 538                 size = sizeof(struct btrfs_extent_data_ref);
 539         }
 540
 541         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
 542         if (ret && ret != -EEXIST)
 543                 goto fail;
 544
 545         leaf = path->nodes[0];
 546         if (parent) {
 547                 struct btrfs_shared_data_ref *ref;
 548                 ref = btrfs_item_ptr(leaf, path->slots[0],
 549                                      struct btrfs_shared_data_ref);
 550                 if (ret == 0) {
 551                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
 552                 } else {
 553                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
 554                         num_refs += refs_to_add;
 555                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
 556                 }
 557         } else {
 558                 struct btrfs_extent_data_ref *ref;
 559                 while (ret == -EEXIST) {
 560                         ref = btrfs_item_ptr(leaf, path->slots[0],
 561                                              struct btrfs_extent_data_ref);
 562                         if (match_extent_data_ref(leaf, ref, root_objectid,
 563                                                   owner, offset))
 564                                 break;
 565                         btrfs_release_path(path);
 566                         key.offset++;
 567                         ret = btrfs_insert_empty_item(trans, root, path, &key,
 568                                                       size);
 569                         if (ret && ret != -EEXIST)
 570                                 goto fail;
 571
 572                         leaf = path->nodes[0];
 573                 }
 574                 ref = btrfs_item_ptr(leaf, path->slots[0],
 575                                      struct btrfs_extent_data_ref);
 576                 if (ret == 0) {
 577                         btrfs_set_extent_data_ref_root(leaf, ref,
 578                                                        root_objectid);
 579                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
 580                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
 581                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
 582                 } else {
 583                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
 584                         num_refs += refs_to_add;
 585                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
 586                 }
 587         }
 588         btrfs_mark_buffer_dirty(trans, leaf);
 589         ret = 0;
 590 fail:
 591         btrfs_release_path(path);
 592         return ret;
 593 }
 594
 595 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
 596                                            struct btrfs_root *root,
 597                                            struct btrfs_path *path,
 598                                            int refs_to_drop)
 599 {
 600         struct btrfs_key key;
 601         struct btrfs_extent_data_ref *ref1 = NULL;
 602         struct btrfs_shared_data_ref *ref2 = NULL;
 603         struct extent_buffer *leaf;
 604         u32 num_refs = 0;
 605         int ret = 0;
 606
 607         leaf = path->nodes[0];
 608         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 609
 610         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
 611                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
 612                                       struct btrfs_extent_data_ref);
 613                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
 614         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
 615                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
 616                                       struct btrfs_shared_data_ref);
 617                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
 618         } else {
 619                 btrfs_err(trans->fs_info,
 620                           "unrecognized backref key (%llu %u %llu)",
 621                           key.objectid, key.type, key.offset);
 622                 btrfs_abort_transaction(trans, -EUCLEAN);
 623                 return -EUCLEAN;
 624         }
 625
 626         BUG_ON(num_refs < refs_to_drop);
 627         num_refs -= refs_to_drop;
 628
 629         if (num_refs == 0) {
 630                 ret = btrfs_del_item(trans, root, path);
 631         } else {
 632                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
 633                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
 634                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
 635                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
 636                 btrfs_mark_buffer_dirty(trans, leaf);
 637         }
 638         return ret;
 639 }
 640
 641 static noinline u32 extent_data_ref_count(struct btrfs_path *path,
 642                                           struct btrfs_extent_inline_ref *iref)
 643 {
 644         struct btrfs_key key;
 645         struct extent_buffer *leaf;
 646         struct btrfs_extent_data_ref *ref1;
 647         struct btrfs_shared_data_ref *ref2;
 648         u32 num_refs = 0;
 649         int type;
 650
 651         leaf = path->nodes[0];
 652         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 653
 654         if (iref) {
 655                 /*
 656                  * If type is invalid, we should have bailed out earlier than
 657                  * this call.
 658                  */
 659                 type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
 660                 ASSERT(type != BTRFS_REF_TYPE_INVALID);
 661                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
 662                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
 663                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
 664                 } else {
 665                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
 666                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
 667                 }
 668         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
 669                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
 670                                       struct btrfs_extent_data_ref);
 671                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
 672         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
 673                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
 674                                       struct btrfs_shared_data_ref);
 675                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
 676         } else {
 677                 WARN_ON(1);
 678         }
 679         return num_refs;
 680 }
 681
 682 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
 683                                           struct btrfs_path *path,
 684                                           u64 bytenr, u64 parent,
 685                                           u64 root_objectid)
 686 {
 687         struct btrfs_root *root = btrfs_extent_root(trans->fs_info, bytenr);
 688         struct btrfs_key key;
 689         int ret;
 690
 691         key.objectid = bytenr;
 692         if (parent) {
 693                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
 694                 key.offset = parent;
 695         } else {
 696                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
 697                 key.offset = root_objectid;
 698         }
 699
 700         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
 701         if (ret > 0)
 702                 ret = -ENOENT;
 703         return ret;
 704 }
 705
 706 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
 707                                           struct btrfs_path *path,
 708                                           u64 bytenr, u64 parent,
 709                                           u64 root_objectid)
 710 {
 711         struct btrfs_root *root = btrfs_extent_root(trans->fs_info, bytenr);
 712         struct btrfs_key key;
 713         int ret;
 714
 715         key.objectid = bytenr;
 716         if (parent) {
 717                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
 718                 key.offset = parent;
 719         } else {
 720                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
 721                 key.offset = root_objectid;
 722         }
 723
 724         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
 725         btrfs_release_path(path);
 726         return ret;
 727 }
 728
 729 static inline int extent_ref_type(u64 parent, u64 owner)
 730 {
 731         int type;
 732         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
 733                 if (parent > 0)
 734                         type = BTRFS_SHARED_BLOCK_REF_KEY;
 735                 else
 736                         type = BTRFS_TREE_BLOCK_REF_KEY;
 737         } else {
 738                 if (parent > 0)
 739                         type = BTRFS_SHARED_DATA_REF_KEY;
 740                 else
 741                         type = BTRFS_EXTENT_DATA_REF_KEY;
 742         }
 743         return type;
 744 }
 745
 746 static int find_next_key(struct btrfs_path *path, int level,
 747                          struct btrfs_key *key)
 748
 749 {
 750         for (; level < BTRFS_MAX_LEVEL; level++) {
 751                 if (!path->nodes[level])
 752                         break;
 753                 if (path->slots[level] + 1 >=
 754                     btrfs_header_nritems(path->nodes[level]))
 755                         continue;
 756                 if (level == 0)
 757                         btrfs_item_key_to_cpu(path->nodes[level], key,
 758                                               path->slots[level] + 1);
 759                 else
 760                         btrfs_node_key_to_cpu(path->nodes[level], key,
 761                                               path->slots[level] + 1);
 762                 return 0;
 763         }
 764         return 1;
 765 }
 766
 767 /*
 768  * look for inline back ref. if back ref is found, *ref_ret is set
 769  * to the address of inline back ref, and 0 is returned.
 770  *
 771  * if back ref isn't found, *ref_ret is set to the address where it
 772  * should be inserted, and -ENOENT is returned.
 773  *
 774  * if insert is true and there are too many inline back refs, the path
 775  * points to the extent item, and -EAGAIN is returned.
 776  *
 777  * NOTE: inline back refs are ordered in the same way that back ref
 778  *       items in the tree are ordered.
 779  */
 780 static noinline_for_stack
 781 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
 782                                  struct btrfs_path *path,
 783                                  struct btrfs_extent_inline_ref **ref_ret,
 784                                  u64 bytenr, u64 num_bytes,
 785                                  u64 parent, u64 root_objectid,
 786                                  u64 owner, u64 offset, int insert)
 787 {
 788         struct btrfs_fs_info *fs_info = trans->fs_info;
 789         struct btrfs_root *root = btrfs_extent_root(fs_info, bytenr);
 790         struct btrfs_key key;
 791         struct extent_buffer *leaf;
 792         struct btrfs_extent_item *ei;
 793         struct btrfs_extent_inline_ref *iref;
 794         u64 flags;
 795         u64 item_size;
 796         unsigned long ptr;
 797         unsigned long end;
 798         int extra_size;
 799         int type;
 800         int want;
 801         int ret;
 802         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
 803         int needed;
 804
 805         key.objectid = bytenr;
 806         key.type = BTRFS_EXTENT_ITEM_KEY;
 807         key.offset = num_bytes;
 808
 809         want = extent_ref_type(parent, owner);
 810         if (insert) {
 811                 extra_size = btrfs_extent_inline_ref_size(want);
 812                 path->search_for_extension = 1;
 813                 path->keep_locks = 1;
 814         } else
 815                 extra_size = -1;
 816
 817         /*
 818          * Owner is our level, so we can just add one to get the level for the
 819          * block we are interested in.
 820          */
 821         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
 822                 key.type = BTRFS_METADATA_ITEM_KEY;
 823                 key.offset = owner;
 824         }
 825
 826 again:
 827         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
 828         if (ret < 0)
 829                 goto out;
 830
 831         /*
 832          * We may be a newly converted file system which still has the old fat
 833          * extent entries for metadata, so try and see if we have one of those.
 834          */
 835         if (ret > 0 && skinny_metadata) {
 836                 skinny_metadata = false;
 837                 if (path->slots[0]) {
 838                         path->slots[0]--;
 839                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 840                                               path->slots[0]);
 841                         if (key.objectid == bytenr &&
 842                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 843                             key.offset == num_bytes)
 844                                 ret = 0;
 845                 }
 846                 if (ret) {
 847                         key.objectid = bytenr;
 848                         key.type = BTRFS_EXTENT_ITEM_KEY;
 849                         key.offset = num_bytes;
 850                         btrfs_release_path(path);
 851                         goto again;
 852                 }
 853         }
 854
 855         if (ret && !insert) {
 856                 ret = -ENOENT;
 857                 goto out;
 858         } else if (WARN_ON(ret)) {
 859                 btrfs_print_leaf(path->nodes[0]);
 860                 btrfs_err(fs_info,
 861 "extent item not found for insert, bytenr %llu num_bytes %llu parent %llu root_objectid %llu owner %llu offset %llu",
 862                           bytenr, num_bytes, parent, root_objectid, owner,
 863                           offset);
 864                 ret = -EUCLEAN;
 865                 goto out;
 866         }
 867
 868         leaf = path->nodes[0];
 869         item_size = btrfs_item_size(leaf, path->slots[0]);
 870         if (unlikely(item_size < sizeof(*ei))) {
 871                 ret = -EUCLEAN;
 872                 btrfs_err(fs_info,
 873                           "unexpected extent item size, has %llu expect >= %zu",
 874                           item_size, sizeof(*ei));
 875                 btrfs_abort_transaction(trans, ret);
 876                 goto out;
 877         }
 878
 879         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
 880         flags = btrfs_extent_flags(leaf, ei);
 881
 882         ptr = (unsigned long)(ei + 1);
 883         end = (unsigned long)ei + item_size;
 884
 885         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
 886                 ptr += sizeof(struct btrfs_tree_block_info);
 887                 BUG_ON(ptr > end);
 888         }
 889
 890         if (owner >= BTRFS_FIRST_FREE_OBJECTID)
 891                 needed = BTRFS_REF_TYPE_DATA;
 892         else
 893                 needed = BTRFS_REF_TYPE_BLOCK;
 894
 895         ret = -ENOENT;
 896         while (ptr < end) {
 897                 iref = (struct btrfs_extent_inline_ref *)ptr;
 898                 type = btrfs_get_extent_inline_ref_type(leaf, iref, needed);
 899                 if (type == BTRFS_EXTENT_OWNER_REF_KEY) {
 900                         ASSERT(btrfs_fs_incompat(fs_info, SIMPLE_QUOTA));
 901                         ptr += btrfs_extent_inline_ref_size(type);
 902                         continue;
 903                 }
 904                 if (type == BTRFS_REF_TYPE_INVALID) {
 905                         ret = -EUCLEAN;
 906                         goto out;
 907                 }
 908
 909                 if (want < type)
 910                         break;
 911                 if (want > type) {
 912                         ptr += btrfs_extent_inline_ref_size(type);
 913                         continue;
 914                 }
 915
 916                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
 917                         struct btrfs_extent_data_ref *dref;
 918                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
 919                         if (match_extent_data_ref(leaf, dref, root_objectid,
 920                                                   owner, offset)) {
 921                                 ret = 0;
 922                                 break;
 923                         }
 924                         if (hash_extent_data_ref_item(leaf, dref) <
 925                             hash_extent_data_ref(root_objectid, owner, offset))
 926                                 break;
 927                 } else {
 928                         u64 ref_offset;
 929                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
 930                         if (parent > 0) {
 931                                 if (parent == ref_offset) {
 932                                         ret = 0;
 933                                         break;
 934                                 }
 935                                 if (ref_offset < parent)
 936                                         break;
 937                         } else {
 938                                 if (root_objectid == ref_offset) {
 939                                         ret = 0;
 940                                         break;
 941                                 }
 942                                 if (ref_offset < root_objectid)
 943                                         break;
 944                         }
 945                 }
 946                 ptr += btrfs_extent_inline_ref_size(type);
 947         }
 948
 949         if (unlikely(ptr > end)) {
 950                 ret = -EUCLEAN;
 951                 btrfs_print_leaf(path->nodes[0]);
 952                 btrfs_crit(fs_info,
 953 "overrun extent record at slot %d while looking for inline extent for root %llu owner %llu offset %llu parent %llu",
 954                            path->slots[0], root_objectid, owner, offset, parent);
 955                 goto out;
 956         }
 957
 958         if (ret == -ENOENT && insert) {
 959                 if (item_size + extra_size >=
 960                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
 961                         ret = -EAGAIN;
 962                         goto out;
 963                 }
 964                 /*
 965                  * To add new inline back ref, we have to make sure
 966                  * there is no corresponding back ref item.
 967                  * For simplicity, we just do not add new inline back
 968                  * ref if there is any kind of item for this block
 969                  */
 970                 if (find_next_key(path, 0, &key) == 0 &&
 971                     key.objectid == bytenr &&
 972                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
 973                         ret = -EAGAIN;
 974                         goto out;
 975                 }
 976         }
 977         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
 978 out:
 979         if (insert) {
 980                 path->keep_locks = 0;
 981                 path->search_for_extension = 0;
 982                 btrfs_unlock_up_safe(path, 1);
 983         }
 984         return ret;
 985 }
 986
 987 /*
 988  * helper to add new inline back ref
 989  */
 990 static noinline_for_stack
 991 void setup_inline_extent_backref(struct btrfs_trans_handle *trans,
 992                                  struct btrfs_path *path,
 993                                  struct btrfs_extent_inline_ref *iref,
 994                                  u64 parent, u64 root_objectid,
 995                                  u64 owner, u64 offset, int refs_to_add,
 996                                  struct btrfs_delayed_extent_op *extent_op)
 997 {
 998         struct extent_buffer *leaf;
 999         struct btrfs_extent_item *ei;
1000         unsigned long ptr;
1001         unsigned long end;
1002         unsigned long item_offset;
1003         u64 refs;
1004         int size;
1005         int type;
1006
1007         leaf = path->nodes[0];
1008         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1009         item_offset = (unsigned long)iref - (unsigned long)ei;
1010
1011         type = extent_ref_type(parent, owner);
1012         size = btrfs_extent_inline_ref_size(type);
1013
1014         btrfs_extend_item(trans, path, size);
1015
1016         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1017         refs = btrfs_extent_refs(leaf, ei);
1018         refs += refs_to_add;
1019         btrfs_set_extent_refs(leaf, ei, refs);
1020         if (extent_op)
1021                 __run_delayed_extent_op(extent_op, leaf, ei);
1022
1023         ptr = (unsigned long)ei + item_offset;
1024         end = (unsigned long)ei + btrfs_item_size(leaf, path->slots[0]);
1025         if (ptr < end - size)
1026                 memmove_extent_buffer(leaf, ptr + size, ptr,
1027                                       end - size - ptr);
1028
1029         iref = (struct btrfs_extent_inline_ref *)ptr;
1030         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1031         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1032                 struct btrfs_extent_data_ref *dref;
1033                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1034                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1035                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1036                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1037                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1038         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1039                 struct btrfs_shared_data_ref *sref;
1040                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1041                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1042                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1043         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1044                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1045         } else {
1046                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1047         }
1048         btrfs_mark_buffer_dirty(trans, leaf);
1049 }
1050
1051 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1052                                  struct btrfs_path *path,
1053                                  struct btrfs_extent_inline_ref **ref_ret,
1054                                  u64 bytenr, u64 num_bytes, u64 parent,
1055                                  u64 root_objectid, u64 owner, u64 offset)
1056 {
1057         int ret;
1058
1059         ret = lookup_inline_extent_backref(trans, path, ref_ret, bytenr,
1060                                            num_bytes, parent, root_objectid,
1061                                            owner, offset, 0);
1062         if (ret != -ENOENT)
1063                 return ret;
1064
1065         btrfs_release_path(path);
1066         *ref_ret = NULL;
1067
1068         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1069                 ret = lookup_tree_block_ref(trans, path, bytenr, parent,
1070                                             root_objectid);
1071         } else {
1072                 ret = lookup_extent_data_ref(trans, path, bytenr, parent,
1073                                              root_objectid, owner, offset);
1074         }
1075         return ret;
1076 }
1077
1078 /*
1079  * helper to update/remove inline back ref
1080  */
1081 static noinline_for_stack int update_inline_extent_backref(
1082                                   struct btrfs_trans_handle *trans,
1083                                   struct btrfs_path *path,
1084                                   struct btrfs_extent_inline_ref *iref,
1085                                   int refs_to_mod,
1086                                   struct btrfs_delayed_extent_op *extent_op)
1087 {
1088         struct extent_buffer *leaf = path->nodes[0];
1089         struct btrfs_fs_info *fs_info = leaf->fs_info;
1090         struct btrfs_extent_item *ei;
1091         struct btrfs_extent_data_ref *dref = NULL;
1092         struct btrfs_shared_data_ref *sref = NULL;
1093         unsigned long ptr;
1094         unsigned long end;
1095         u32 item_size;
1096         int size;
1097         int type;
1098         u64 refs;
1099
1100         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1101         refs = btrfs_extent_refs(leaf, ei);
1102         if (unlikely(refs_to_mod < 0 && refs + refs_to_mod <= 0)) {
1103                 struct btrfs_key key;
1104                 u32 extent_size;
1105
1106                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1107                 if (key.type == BTRFS_METADATA_ITEM_KEY)
1108                         extent_size = fs_info->nodesize;
1109                 else
1110                         extent_size = key.offset;
1111                 btrfs_print_leaf(leaf);
1112                 btrfs_err(fs_info,
1113         "invalid refs_to_mod for extent %llu num_bytes %u, has %d expect >= -%llu",
1114                           key.objectid, extent_size, refs_to_mod, refs);
1115                 return -EUCLEAN;
1116         }
1117         refs += refs_to_mod;
1118         btrfs_set_extent_refs(leaf, ei, refs);
1119         if (extent_op)
1120                 __run_delayed_extent_op(extent_op, leaf, ei);
1121
1122         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_ANY);
1123         /*
1124          * Function btrfs_get_extent_inline_ref_type() has already printed
1125          * error messages.
1126          */
1127         if (unlikely(type == BTRFS_REF_TYPE_INVALID))
1128                 return -EUCLEAN;
1129
1130         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1131                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1132                 refs = btrfs_extent_data_ref_count(leaf, dref);
1133         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1134                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1135                 refs = btrfs_shared_data_ref_count(leaf, sref);
1136         } else {
1137                 refs = 1;
1138                 /*
1139                  * For tree blocks we can only drop one ref for it, and tree
1140                  * blocks should not have refs > 1.
1141                  *
1142                  * Furthermore if we're inserting a new inline backref, we
1143                  * won't reach this path either. That would be
1144                  * setup_inline_extent_backref().
1145                  */
1146                 if (unlikely(refs_to_mod != -1)) {
1147                         struct btrfs_key key;
1148
1149                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1150
1151                         btrfs_print_leaf(leaf);
1152                         btrfs_err(fs_info,
1153                         "invalid refs_to_mod for tree block %llu, has %d expect -1",
1154                                   key.objectid, refs_to_mod);
1155                         return -EUCLEAN;
1156                 }
1157         }
1158
1159         if (unlikely(refs_to_mod < 0 && refs < -refs_to_mod)) {
1160                 struct btrfs_key key;
1161                 u32 extent_size;
1162
1163                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1164                 if (key.type == BTRFS_METADATA_ITEM_KEY)
1165                         extent_size = fs_info->nodesize;
1166                 else
1167                         extent_size = key.offset;
1168                 btrfs_print_leaf(leaf);
1169                 btrfs_err(fs_info,
1170 "invalid refs_to_mod for backref entry, iref %lu extent %llu num_bytes %u, has %d expect >= -%llu",
1171                           (unsigned long)iref, key.objectid, extent_size,
1172                           refs_to_mod, refs);
1173                 return -EUCLEAN;
1174         }
1175         refs += refs_to_mod;
1176
1177         if (refs > 0) {
1178                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1179                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1180                 else
1181                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1182         } else {
1183                 size =  btrfs_extent_inline_ref_size(type);
1184                 item_size = btrfs_item_size(leaf, path->slots[0]);
1185                 ptr = (unsigned long)iref;
1186                 end = (unsigned long)ei + item_size;
1187                 if (ptr + size < end)
1188                         memmove_extent_buffer(leaf, ptr, ptr + size,
1189                                               end - ptr - size);
1190                 item_size -= size;
1191                 btrfs_truncate_item(trans, path, item_size, 1);
1192         }
1193         btrfs_mark_buffer_dirty(trans, leaf);
1194         return 0;
1195 }
1196
1197 static noinline_for_stack
1198 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1199                                  struct btrfs_path *path,
1200                                  u64 bytenr, u64 num_bytes, u64 parent,
1201                                  u64 root_objectid, u64 owner,
1202                                  u64 offset, int refs_to_add,
1203                                  struct btrfs_delayed_extent_op *extent_op)
1204 {
1205         struct btrfs_extent_inline_ref *iref;
1206         int ret;
1207
1208         ret = lookup_inline_extent_backref(trans, path, &iref, bytenr,
1209                                            num_bytes, parent, root_objectid,
1210                                            owner, offset, 1);
1211         if (ret == 0) {
1212                 /*
1213                  * We're adding refs to a tree block we already own, this
1214                  * should not happen at all.
1215                  */
1216                 if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1217                         btrfs_print_leaf(path->nodes[0]);
1218                         btrfs_crit(trans->fs_info,
1219 "adding refs to an existing tree ref, bytenr %llu num_bytes %llu root_objectid %llu slot %u",
1220                                    bytenr, num_bytes, root_objectid, path->slots[0]);
1221                         return -EUCLEAN;
1222                 }
1223                 ret = update_inline_extent_backref(trans, path, iref,
1224                                                    refs_to_add, extent_op);
1225         } else if (ret == -ENOENT) {
1226                 setup_inline_extent_backref(trans, path, iref, parent,
1227                                             root_objectid, owner, offset,
1228                                             refs_to_add, extent_op);
1229                 ret = 0;
1230         }
1231         return ret;
1232 }
1233
1234 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1235                                  struct btrfs_root *root,
1236                                  struct btrfs_path *path,
1237                                  struct btrfs_extent_inline_ref *iref,
1238                                  int refs_to_drop, int is_data)
1239 {
1240         int ret = 0;
1241
1242         BUG_ON(!is_data && refs_to_drop != 1);
1243         if (iref)
1244                 ret = update_inline_extent_backref(trans, path, iref,
1245                                                    -refs_to_drop, NULL);
1246         else if (is_data)
1247                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop);
1248         else
1249                 ret = btrfs_del_item(trans, root, path);
1250         return ret;
1251 }
1252
1253 static int btrfs_issue_discard(struct block_device *bdev, u64 start, u64 len,
1254                                u64 *discarded_bytes)
1255 {
1256         int j, ret = 0;
1257         u64 bytes_left, end;
1258         u64 aligned_start = ALIGN(start, 1 << SECTOR_SHIFT);
1259
1260         /* Adjust the range to be aligned to 512B sectors if necessary. */
1261         if (start != aligned_start) {
1262                 len -= aligned_start - start;
1263                 len = round_down(len, 1 << SECTOR_SHIFT);
1264                 start = aligned_start;
1265         }
1266
1267         *discarded_bytes = 0;
1268
1269         if (!len)
1270                 return 0;
1271
1272         end = start + len;
1273         bytes_left = len;
1274
1275         /* Skip any superblocks on this device. */
1276         for (j = 0; j < BTRFS_SUPER_MIRROR_MAX; j++) {
1277                 u64 sb_start = btrfs_sb_offset(j);
1278                 u64 sb_end = sb_start + BTRFS_SUPER_INFO_SIZE;
1279                 u64 size = sb_start - start;
1280
1281                 if (!in_range(sb_start, start, bytes_left) &&
1282                     !in_range(sb_end, start, bytes_left) &&
1283                     !in_range(start, sb_start, BTRFS_SUPER_INFO_SIZE))
1284                         continue;
1285
1286                 /*
1287                  * Superblock spans beginning of range.  Adjust start and
1288                  * try again.
1289                  */
1290                 if (sb_start <= start) {
1291                         start += sb_end - start;
1292                         if (start > end) {
1293                                 bytes_left = 0;
1294                                 break;
1295                         }
1296                         bytes_left = end - start;
1297                         continue;
1298                 }
1299
1300                 if (size) {
1301                         ret = blkdev_issue_discard(bdev, start >> SECTOR_SHIFT,
1302                                                    size >> SECTOR_SHIFT,
1303                                                    GFP_NOFS);
1304                         if (!ret)
1305                                 *discarded_bytes += size;
1306                         else if (ret != -EOPNOTSUPP)
1307                                 return ret;
1308                 }
1309
1310                 start = sb_end;
1311                 if (start > end) {
1312                         bytes_left = 0;
1313                         break;
1314                 }
1315                 bytes_left = end - start;
1316         }
1317
1318         if (bytes_left) {
1319                 ret = blkdev_issue_discard(bdev, start >> SECTOR_SHIFT,
1320                                            bytes_left >> SECTOR_SHIFT,
1321                                            GFP_NOFS);
1322                 if (!ret)
1323                         *discarded_bytes += bytes_left;
1324         }
1325         return ret;
1326 }
1327
1328 static int do_discard_extent(struct btrfs_discard_stripe *stripe, u64 *bytes)
1329 {
1330         struct btrfs_device *dev = stripe->dev;
1331         struct btrfs_fs_info *fs_info = dev->fs_info;
1332         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1333         u64 phys = stripe->physical;
1334         u64 len = stripe->length;
1335         u64 discarded = 0;
1336         int ret = 0;
1337
1338         /* Zone reset on a zoned filesystem */
1339         if (btrfs_can_zone_reset(dev, phys, len)) {
1340                 u64 src_disc;
1341
1342                 ret = btrfs_reset_device_zone(dev, phys, len, &discarded);
1343                 if (ret)
1344                         goto out;
1345
1346                 if (!btrfs_dev_replace_is_ongoing(dev_replace) ||
1347                     dev != dev_replace->srcdev)
1348                         goto out;
1349
1350                 src_disc = discarded;
1351
1352                 /* Send to replace target as well */
1353                 ret = btrfs_reset_device_zone(dev_replace->tgtdev, phys, len,
1354                                               &discarded);
1355                 discarded += src_disc;
1356         } else if (bdev_max_discard_sectors(stripe->dev->bdev)) {
1357                 ret = btrfs_issue_discard(dev->bdev, phys, len, &discarded);
1358         } else {
1359                 ret = 0;
1360                 *bytes = 0;
1361         }
1362
1363 out:
1364         *bytes = discarded;
1365         return ret;
1366 }
1367
1368 int btrfs_discard_extent(struct btrfs_fs_info *fs_info, u64 bytenr,
1369                          u64 num_bytes, u64 *actual_bytes)
1370 {
1371         int ret = 0;
1372         u64 discarded_bytes = 0;
1373         u64 end = bytenr + num_bytes;
1374         u64 cur = bytenr;
1375
1376         /*
1377          * Avoid races with device replace and make sure the devices in the
1378          * stripes don't go away while we are discarding.
1379          */
1380         btrfs_bio_counter_inc_blocked(fs_info);
1381         while (cur < end) {
1382                 struct btrfs_discard_stripe *stripes;
1383                 unsigned int num_stripes;
1384                 int i;
1385
1386                 num_bytes = end - cur;
1387                 stripes = btrfs_map_discard(fs_info, cur, &num_bytes, &num_stripes);
1388                 if (IS_ERR(stripes)) {
1389                         ret = PTR_ERR(stripes);
1390                         if (ret == -EOPNOTSUPP)
1391                                 ret = 0;
1392                         break;
1393                 }
1394
1395                 for (i = 0; i < num_stripes; i++) {
1396                         struct btrfs_discard_stripe *stripe = stripes + i;
1397                         u64 bytes;
1398
1399                         if (!stripe->dev->bdev) {
1400                                 ASSERT(btrfs_test_opt(fs_info, DEGRADED));
1401                                 continue;
1402                         }
1403
1404                         if (!test_bit(BTRFS_DEV_STATE_WRITEABLE,
1405                                         &stripe->dev->dev_state))
1406                                 continue;
1407
1408                         ret = do_discard_extent(stripe, &bytes);
1409                         if (ret) {
1410                                 /*
1411                                  * Keep going if discard is not supported by the
1412                                  * device.
1413                                  */
1414                                 if (ret != -EOPNOTSUPP)
1415                                         break;
1416                                 ret = 0;
1417                         } else {
1418                                 discarded_bytes += bytes;
1419                         }
1420                 }
1421                 kfree(stripes);
1422                 if (ret)
1423                         break;
1424                 cur += num_bytes;
1425         }
1426         btrfs_bio_counter_dec(fs_info);
1427         if (actual_bytes)
1428                 *actual_bytes = discarded_bytes;
1429         return ret;
1430 }
1431
1432 /* Can return -ENOMEM */
1433 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1434                          struct btrfs_ref *generic_ref)
1435 {
1436         struct btrfs_fs_info *fs_info = trans->fs_info;
1437         int ret;
1438
1439         ASSERT(generic_ref->type != BTRFS_REF_NOT_SET &&
1440                generic_ref->action);
1441         BUG_ON(generic_ref->type == BTRFS_REF_METADATA &&
1442                generic_ref->tree_ref.ref_root == BTRFS_TREE_LOG_OBJECTID);
1443
1444         if (generic_ref->type == BTRFS_REF_METADATA)
1445                 ret = btrfs_add_delayed_tree_ref(trans, generic_ref, NULL);
1446         else
1447                 ret = btrfs_add_delayed_data_ref(trans, generic_ref, 0);
1448
1449         btrfs_ref_tree_mod(fs_info, generic_ref);
1450
1451         return ret;
1452 }
1453
1454 /*
1455  * Insert backreference for a given extent.
1456  *
1457  * The counterpart is in __btrfs_free_extent(), with examples and more details
1458  * how it works.
1459  *
1460  * @trans:          Handle of transaction
1461  *
1462  * @node:           The delayed ref node used to get the bytenr/length for
1463  *                  extent whose references are incremented.
1464  *
1465  * @parent:         If this is a shared extent (BTRFS_SHARED_DATA_REF_KEY/
1466  *                  BTRFS_SHARED_BLOCK_REF_KEY) then it holds the logical
1467  *                  bytenr of the parent block. Since new extents are always
1468  *                  created with indirect references, this will only be the case
1469  *                  when relocating a shared extent. In that case, root_objectid
1470  *                  will be BTRFS_TREE_RELOC_OBJECTID. Otherwise, parent must
1471  *                  be 0
1472  *
1473  * @root_objectid:  The id of the root where this modification has originated,
1474  *                  this can be either one of the well-known metadata trees or
1475  *                  the subvolume id which references this extent.
1476  *
1477  * @owner:          For data extents it is the inode number of the owning file.
1478  *                  For metadata extents this parameter holds the level in the
1479  *                  tree of the extent.
1480  *
1481  * @offset:         For metadata extents the offset is ignored and is currently
1482  *                  always passed as 0. For data extents it is the fileoffset
1483  *                  this extent belongs to.
1484  *
1485  * @extent_op       Pointer to a structure, holding information necessary when
1486  *                  updating a tree block's flags
1487  *
1488  */
1489 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1490                                   struct btrfs_delayed_ref_node *node,
1491                                   u64 parent, u64 root_objectid,
1492                                   u64 owner, u64 offset,
1493                                   struct btrfs_delayed_extent_op *extent_op)
1494 {
1495         struct btrfs_path *path;
1496         struct extent_buffer *leaf;
1497         struct btrfs_extent_item *item;
1498         struct btrfs_key key;
1499         u64 bytenr = node->bytenr;
1500         u64 num_bytes = node->num_bytes;
1501         u64 refs;
1502         int refs_to_add = node->ref_mod;
1503         int ret;
1504
1505         path = btrfs_alloc_path();
1506         if (!path)
1507                 return -ENOMEM;
1508
1509         /* this will setup the path even if it fails to insert the back ref */
1510         ret = insert_inline_extent_backref(trans, path, bytenr, num_bytes,
1511                                            parent, root_objectid, owner,
1512                                            offset, refs_to_add, extent_op);
1513         if ((ret < 0 && ret != -EAGAIN) || !ret)
1514                 goto out;
1515
1516         /*
1517          * Ok we had -EAGAIN which means we didn't have space to insert and
1518          * inline extent ref, so just update the reference count and add a
1519          * normal backref.
1520          */
1521         leaf = path->nodes[0];
1522         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1523         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1524         refs = btrfs_extent_refs(leaf, item);
1525         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
1526         if (extent_op)
1527                 __run_delayed_extent_op(extent_op, leaf, item);
1528
1529         btrfs_mark_buffer_dirty(trans, leaf);
1530         btrfs_release_path(path);
1531
1532         /* now insert the actual backref */
1533         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1534                 ret = insert_tree_block_ref(trans, path, bytenr, parent,
1535                                             root_objectid);
1536         else
1537                 ret = insert_extent_data_ref(trans, path, bytenr, parent,
1538                                              root_objectid, owner, offset,
1539                                              refs_to_add);
1540
1541         if (ret)
1542                 btrfs_abort_transaction(trans, ret);
1543 out:
1544         btrfs_free_path(path);
1545         return ret;
1546 }
1547
1548 static void free_head_ref_squota_rsv(struct btrfs_fs_info *fs_info,
1549                                      struct btrfs_delayed_ref_head *href)
1550 {
1551         u64 root = href->owning_root;
1552
1553         /*
1554          * Don't check must_insert_reserved, as this is called from contexts
1555          * where it has already been unset.
1556          */
1557         if (btrfs_qgroup_mode(fs_info) != BTRFS_QGROUP_MODE_SIMPLE ||
1558             !href->is_data || !is_fstree(root))
1559                 return;
1560
1561         btrfs_qgroup_free_refroot(fs_info, root, href->reserved_bytes,
1562                                   BTRFS_QGROUP_RSV_DATA);
1563 }
1564
1565 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
1566                                 struct btrfs_delayed_ref_head *href,
1567                                 struct btrfs_delayed_ref_node *node,
1568                                 struct btrfs_delayed_extent_op *extent_op,
1569                                 bool insert_reserved)
1570 {
1571         int ret = 0;
1572         struct btrfs_delayed_data_ref *ref;
1573         u64 parent = 0;
1574         u64 flags = 0;
1575
1576         ref = btrfs_delayed_node_to_data_ref(node);
1577         trace_run_delayed_data_ref(trans->fs_info, node, ref, node->action);
1578
1579         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
1580                 parent = ref->parent;
1581
1582         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1583                 struct btrfs_key key;
1584                 struct btrfs_squota_delta delta = {
1585                         .root = href->owning_root,
1586                         .num_bytes = node->num_bytes,
1587                         .is_data = true,
1588                         .is_inc = true,
1589                         .generation = trans->transid,
1590                 };
1591
1592                 if (extent_op)
1593                         flags |= extent_op->flags_to_set;
1594
1595                 key.objectid = node->bytenr;
1596                 key.type = BTRFS_EXTENT_ITEM_KEY;
1597                 key.offset = node->num_bytes;
1598
1599                 ret = alloc_reserved_file_extent(trans, parent, ref->root,
1600                                                  flags, ref->objectid,
1601                                                  ref->offset, &key,
1602                                                  node->ref_mod, href->owning_root);
1603                 free_head_ref_squota_rsv(trans->fs_info, href);
1604                 if (!ret)
1605                         ret = btrfs_record_squota_delta(trans->fs_info, &delta);
1606         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1607                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref->root,
1608                                              ref->objectid, ref->offset,
1609                                              extent_op);
1610         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1611                 ret = __btrfs_free_extent(trans, href, node, parent,
1612                                           ref->root, ref->objectid,
1613                                           ref->offset, extent_op);
1614         } else {
1615                 BUG();
1616         }
1617         return ret;
1618 }
1619
1620 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
1621                                     struct extent_buffer *leaf,
1622                                     struct btrfs_extent_item *ei)
1623 {
1624         u64 flags = btrfs_extent_flags(leaf, ei);
1625         if (extent_op->update_flags) {
1626                 flags |= extent_op->flags_to_set;
1627                 btrfs_set_extent_flags(leaf, ei, flags);
1628         }
1629
1630         if (extent_op->update_key) {
1631                 struct btrfs_tree_block_info *bi;
1632                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
1633                 bi = (struct btrfs_tree_block_info *)(ei + 1);
1634                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
1635         }
1636 }
1637
1638 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
1639                                  struct btrfs_delayed_ref_head *head,
1640                                  struct btrfs_delayed_extent_op *extent_op)
1641 {
1642         struct btrfs_fs_info *fs_info = trans->fs_info;
1643         struct btrfs_root *root;
1644         struct btrfs_key key;
1645         struct btrfs_path *path;
1646         struct btrfs_extent_item *ei;
1647         struct extent_buffer *leaf;
1648         u32 item_size;
1649         int ret;
1650         int metadata = 1;
1651
1652         if (TRANS_ABORTED(trans))
1653                 return 0;
1654
1655         if (!btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1656                 metadata = 0;
1657
1658         path = btrfs_alloc_path();
1659         if (!path)
1660                 return -ENOMEM;
1661
1662         key.objectid = head->bytenr;
1663
1664         if (metadata) {
1665                 key.type = BTRFS_METADATA_ITEM_KEY;
1666                 key.offset = extent_op->level;
1667         } else {
1668                 key.type = BTRFS_EXTENT_ITEM_KEY;
1669                 key.offset = head->num_bytes;
1670         }
1671
1672         root = btrfs_extent_root(fs_info, key.objectid);
1673 again:
1674         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
1675         if (ret < 0) {
1676                 goto out;
1677         } else if (ret > 0) {
1678                 if (metadata) {
1679                         if (path->slots[0] > 0) {
1680                                 path->slots[0]--;
1681                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
1682                                                       path->slots[0]);
1683                                 if (key.objectid == head->bytenr &&
1684                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
1685                                     key.offset == head->num_bytes)
1686                                         ret = 0;
1687                         }
1688                         if (ret > 0) {
1689                                 btrfs_release_path(path);
1690                                 metadata = 0;
1691
1692                                 key.objectid = head->bytenr;
1693                                 key.offset = head->num_bytes;
1694                                 key.type = BTRFS_EXTENT_ITEM_KEY;
1695                                 goto again;
1696                         }
1697                 } else {
1698                         ret = -EUCLEAN;
1699                         btrfs_err(fs_info,
1700                   "missing extent item for extent %llu num_bytes %llu level %d",
1701                                   head->bytenr, head->num_bytes, extent_op->level);
1702                         goto out;
1703                 }
1704         }
1705
1706         leaf = path->nodes[0];
1707         item_size = btrfs_item_size(leaf, path->slots[0]);
1708
1709         if (unlikely(item_size < sizeof(*ei))) {
1710                 ret = -EUCLEAN;
1711                 btrfs_err(fs_info,
1712                           "unexpected extent item size, has %u expect >= %zu",
1713                           item_size, sizeof(*ei));
1714                 btrfs_abort_transaction(trans, ret);
1715                 goto out;
1716         }
1717
1718         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1719         __run_delayed_extent_op(extent_op, leaf, ei);
1720
1721         btrfs_mark_buffer_dirty(trans, leaf);
1722 out:
1723         btrfs_free_path(path);
1724         return ret;
1725 }
1726
1727 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
1728                                 struct btrfs_delayed_ref_head *href,
1729                                 struct btrfs_delayed_ref_node *node,
1730                                 struct btrfs_delayed_extent_op *extent_op,
1731                                 bool insert_reserved)
1732 {
1733         int ret = 0;
1734         struct btrfs_fs_info *fs_info = trans->fs_info;
1735         struct btrfs_delayed_tree_ref *ref;
1736         u64 parent = 0;
1737         u64 ref_root = 0;
1738
1739         ref = btrfs_delayed_node_to_tree_ref(node);
1740         trace_run_delayed_tree_ref(trans->fs_info, node, ref, node->action);
1741
1742         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
1743                 parent = ref->parent;
1744         ref_root = ref->root;
1745
1746         if (unlikely(node->ref_mod != 1)) {
1747                 btrfs_err(trans->fs_info,
1748         "btree block %llu has %d references rather than 1: action %d ref_root %llu parent %llu",
1749                           node->bytenr, node->ref_mod, node->action, ref_root,
1750                           parent);
1751                 return -EUCLEAN;
1752         }
1753         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1754                 struct btrfs_squota_delta delta = {
1755                         .root = href->owning_root,
1756                         .num_bytes = fs_info->nodesize,
1757                         .is_data = false,
1758                         .is_inc = true,
1759                         .generation = trans->transid,
1760                 };
1761
1762                 BUG_ON(!extent_op || !extent_op->update_flags);
1763                 ret = alloc_reserved_tree_block(trans, node, extent_op);
1764                 if (!ret)
1765                         btrfs_record_squota_delta(fs_info, &delta);
1766         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1767                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
1768                                              ref->level, 0, extent_op);
1769         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1770                 ret = __btrfs_free_extent(trans, href, node, parent, ref_root,
1771                                           ref->level, 0, extent_op);
1772         } else {
1773                 BUG();
1774         }
1775         return ret;
1776 }
1777
1778 /* helper function to actually process a single delayed ref entry */
1779 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
1780                                struct btrfs_delayed_ref_head *href,
1781                                struct btrfs_delayed_ref_node *node,
1782                                struct btrfs_delayed_extent_op *extent_op,
1783                                bool insert_reserved)
1784 {
1785         int ret = 0;
1786
1787         if (TRANS_ABORTED(trans)) {
1788                 if (insert_reserved) {
1789                         btrfs_pin_extent(trans, node->bytenr, node->num_bytes, 1);
1790                         free_head_ref_squota_rsv(trans->fs_info, href);
1791                 }
1792                 return 0;
1793         }
1794
1795         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
1796             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
1797                 ret = run_delayed_tree_ref(trans, href, node, extent_op,
1798                                            insert_reserved);
1799         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
1800                  node->type == BTRFS_SHARED_DATA_REF_KEY)
1801                 ret = run_delayed_data_ref(trans, href, node, extent_op,
1802                                            insert_reserved);
1803         else if (node->type == BTRFS_EXTENT_OWNER_REF_KEY)
1804                 ret = 0;
1805         else
1806                 BUG();
1807         if (ret && insert_reserved)
1808                 btrfs_pin_extent(trans, node->bytenr, node->num_bytes, 1);
1809         if (ret < 0)
1810                 btrfs_err(trans->fs_info,
1811 "failed to run delayed ref for logical %llu num_bytes %llu type %u action %u ref_mod %d: %d",
1812                           node->bytenr, node->num_bytes, node->type,
1813                           node->action, node->ref_mod, ret);
1814         return ret;
1815 }
1816
1817 static inline struct btrfs_delayed_ref_node *
1818 select_delayed_ref(struct btrfs_delayed_ref_head *head)
1819 {
1820         struct btrfs_delayed_ref_node *ref;
1821
1822         if (RB_EMPTY_ROOT(&head->ref_tree.rb_root))
1823                 return NULL;
1824
1825         /*
1826          * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
1827          * This is to prevent a ref count from going down to zero, which deletes
1828          * the extent item from the extent tree, when there still are references
1829          * to add, which would fail because they would not find the extent item.
1830          */
1831         if (!list_empty(&head->ref_add_list))
1832                 return list_first_entry(&head->ref_add_list,
1833                                 struct btrfs_delayed_ref_node, add_list);
1834
1835         ref = rb_entry(rb_first_cached(&head->ref_tree),
1836                        struct btrfs_delayed_ref_node, ref_node);
1837         ASSERT(list_empty(&ref->add_list));
1838         return ref;
1839 }
1840
1841 static void unselect_delayed_ref_head(struct btrfs_delayed_ref_root *delayed_refs,
1842                                       struct btrfs_delayed_ref_head *head)
1843 {
1844         spin_lock(&delayed_refs->lock);
1845         head->processing = false;
1846         delayed_refs->num_heads_ready++;
1847         spin_unlock(&delayed_refs->lock);
1848         btrfs_delayed_ref_unlock(head);
1849 }
1850
1851 static struct btrfs_delayed_extent_op *cleanup_extent_op(
1852                                 struct btrfs_delayed_ref_head *head)
1853 {
1854         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
1855
1856         if (!extent_op)
1857                 return NULL;
1858
1859         if (head->must_insert_reserved) {
1860                 head->extent_op = NULL;
1861                 btrfs_free_delayed_extent_op(extent_op);
1862                 return NULL;
1863         }
1864         return extent_op;
1865 }
1866
1867 static int run_and_cleanup_extent_op(struct btrfs_trans_handle *trans,
1868                                      struct btrfs_delayed_ref_head *head)
1869 {
1870         struct btrfs_delayed_extent_op *extent_op;
1871         int ret;
1872
1873         extent_op = cleanup_extent_op(head);
1874         if (!extent_op)
1875                 return 0;
1876         head->extent_op = NULL;
1877         spin_unlock(&head->lock);
1878         ret = run_delayed_extent_op(trans, head, extent_op);
1879         btrfs_free_delayed_extent_op(extent_op);
1880         return ret ? ret : 1;
1881 }
1882
1883 u64 btrfs_cleanup_ref_head_accounting(struct btrfs_fs_info *fs_info,
1884                                   struct btrfs_delayed_ref_root *delayed_refs,
1885                                   struct btrfs_delayed_ref_head *head)
1886 {
1887         u64 ret = 0;
1888
1889         /*
1890          * We had csum deletions accounted for in our delayed refs rsv, we need
1891          * to drop the csum leaves for this update from our delayed_refs_rsv.
1892          */
1893         if (head->total_ref_mod < 0 && head->is_data) {
1894                 int nr_csums;
1895
1896                 spin_lock(&delayed_refs->lock);
1897                 delayed_refs->pending_csums -= head->num_bytes;
1898                 spin_unlock(&delayed_refs->lock);
1899                 nr_csums = btrfs_csum_bytes_to_leaves(fs_info, head->num_bytes);
1900
1901                 btrfs_delayed_refs_rsv_release(fs_info, 0, nr_csums);
1902
1903                 ret = btrfs_calc_delayed_ref_csum_bytes(fs_info, nr_csums);
1904         }
1905         /* must_insert_reserved can be set only if we didn't run the head ref. */
1906         if (head->must_insert_reserved)
1907                 free_head_ref_squota_rsv(fs_info, head);
1908
1909         return ret;
1910 }
1911
1912 static int cleanup_ref_head(struct btrfs_trans_handle *trans,
1913                             struct btrfs_delayed_ref_head *head,
1914                             u64 *bytes_released)
1915 {
1916
1917         struct btrfs_fs_info *fs_info = trans->fs_info;
1918         struct btrfs_delayed_ref_root *delayed_refs;
1919         int ret;
1920
1921         delayed_refs = &trans->transaction->delayed_refs;
1922
1923         ret = run_and_cleanup_extent_op(trans, head);
1924         if (ret < 0) {
1925                 unselect_delayed_ref_head(delayed_refs, head);
1926                 btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
1927                 return ret;
1928         } else if (ret) {
1929                 return ret;
1930         }
1931
1932         /*
1933          * Need to drop our head ref lock and re-acquire the delayed ref lock
1934          * and then re-check to make sure nobody got added.
1935          */
1936         spin_unlock(&head->lock);
1937         spin_lock(&delayed_refs->lock);
1938         spin_lock(&head->lock);
1939         if (!RB_EMPTY_ROOT(&head->ref_tree.rb_root) || head->extent_op) {
1940                 spin_unlock(&head->lock);
1941                 spin_unlock(&delayed_refs->lock);
1942                 return 1;
1943         }
1944         btrfs_delete_ref_head(delayed_refs, head);
1945         spin_unlock(&head->lock);
1946         spin_unlock(&delayed_refs->lock);
1947
1948         if (head->must_insert_reserved) {
1949                 btrfs_pin_extent(trans, head->bytenr, head->num_bytes, 1);
1950                 if (head->is_data) {
1951                         struct btrfs_root *csum_root;
1952
1953                         csum_root = btrfs_csum_root(fs_info, head->bytenr);
1954                         ret = btrfs_del_csums(trans, csum_root, head->bytenr,
1955                                               head->num_bytes);
1956                 }
1957         }
1958
1959         *bytes_released += btrfs_cleanup_ref_head_accounting(fs_info, delayed_refs, head);
1960
1961         trace_run_delayed_ref_head(fs_info, head, 0);
1962         btrfs_delayed_ref_unlock(head);
1963         btrfs_put_delayed_ref_head(head);
1964         return ret;
1965 }
1966
1967 static struct btrfs_delayed_ref_head *btrfs_obtain_ref_head(
1968                                         struct btrfs_trans_handle *trans)
1969 {
1970         struct btrfs_delayed_ref_root *delayed_refs =
1971                 &trans->transaction->delayed_refs;
1972         struct btrfs_delayed_ref_head *head = NULL;
1973         int ret;
1974
1975         spin_lock(&delayed_refs->lock);
1976         head = btrfs_select_ref_head(delayed_refs);
1977         if (!head) {
1978                 spin_unlock(&delayed_refs->lock);
1979                 return head;
1980         }
1981
1982         /*
1983          * Grab the lock that says we are going to process all the refs for
1984          * this head
1985          */
1986         ret = btrfs_delayed_ref_lock(delayed_refs, head);
1987         spin_unlock(&delayed_refs->lock);
1988
1989         /*
1990          * We may have dropped the spin lock to get the head mutex lock, and
1991          * that might have given someone else time to free the head.  If that's
1992          * true, it has been removed from our list and we can move on.
1993          */
1994         if (ret == -EAGAIN)
1995                 head = ERR_PTR(-EAGAIN);
1996
1997         return head;
1998 }
1999
2000 static int btrfs_run_delayed_refs_for_head(struct btrfs_trans_handle *trans,
2001                                            struct btrfs_delayed_ref_head *locked_ref,
2002                                            u64 *bytes_released)
2003 {
2004         struct btrfs_fs_info *fs_info = trans->fs_info;
2005         struct btrfs_delayed_ref_root *delayed_refs;
2006         struct btrfs_delayed_extent_op *extent_op;
2007         struct btrfs_delayed_ref_node *ref;
2008         bool must_insert_reserved;
2009         int ret;
2010
2011         delayed_refs = &trans->transaction->delayed_refs;
2012
2013         lockdep_assert_held(&locked_ref->mutex);
2014         lockdep_assert_held(&locked_ref->lock);
2015
2016         while ((ref = select_delayed_ref(locked_ref))) {
2017                 if (ref->seq &&
2018                     btrfs_check_delayed_seq(fs_info, ref->seq)) {
2019                         spin_unlock(&locked_ref->lock);
2020                         unselect_delayed_ref_head(delayed_refs, locked_ref);
2021                         return -EAGAIN;
2022                 }
2023
2024                 rb_erase_cached(&ref->ref_node, &locked_ref->ref_tree);
2025                 RB_CLEAR_NODE(&ref->ref_node);
2026                 if (!list_empty(&ref->add_list))
2027                         list_del(&ref->add_list);
2028                 /*
2029                  * When we play the delayed ref, also correct the ref_mod on
2030                  * head
2031                  */
2032                 switch (ref->action) {
2033                 case BTRFS_ADD_DELAYED_REF:
2034                 case BTRFS_ADD_DELAYED_EXTENT:
2035                         locked_ref->ref_mod -= ref->ref_mod;
2036                         break;
2037                 case BTRFS_DROP_DELAYED_REF:
2038                         locked_ref->ref_mod += ref->ref_mod;
2039                         break;
2040                 default:
2041                         WARN_ON(1);
2042                 }
2043                 atomic_dec(&delayed_refs->num_entries);
2044
2045                 /*
2046                  * Record the must_insert_reserved flag before we drop the
2047                  * spin lock.
2048                  */
2049                 must_insert_reserved = locked_ref->must_insert_reserved;
2050                 /*
2051                  * Unsetting this on the head ref relinquishes ownership of
2052                  * the rsv_bytes, so it is critical that every possible code
2053                  * path from here forward frees all reserves including qgroup
2054                  * reserve.
2055                  */
2056                 locked_ref->must_insert_reserved = false;
2057
2058                 extent_op = locked_ref->extent_op;
2059                 locked_ref->extent_op = NULL;
2060                 spin_unlock(&locked_ref->lock);
2061
2062                 ret = run_one_delayed_ref(trans, locked_ref, ref, extent_op,
2063                                           must_insert_reserved);
2064                 btrfs_delayed_refs_rsv_release(fs_info, 1, 0);
2065                 *bytes_released += btrfs_calc_delayed_ref_bytes(fs_info, 1);
2066
2067                 btrfs_free_delayed_extent_op(extent_op);
2068                 if (ret) {
2069                         unselect_delayed_ref_head(delayed_refs, locked_ref);
2070                         btrfs_put_delayed_ref(ref);
2071                         return ret;
2072                 }
2073
2074                 btrfs_put_delayed_ref(ref);
2075                 cond_resched();
2076
2077                 spin_lock(&locked_ref->lock);
2078                 btrfs_merge_delayed_refs(fs_info, delayed_refs, locked_ref);
2079         }
2080
2081         return 0;
2082 }
2083
2084 /*
2085  * Returns 0 on success or if called with an already aborted transaction.
2086  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2087  */
2088 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2089                                              u64 min_bytes)
2090 {
2091         struct btrfs_fs_info *fs_info = trans->fs_info;
2092         struct btrfs_delayed_ref_root *delayed_refs;
2093         struct btrfs_delayed_ref_head *locked_ref = NULL;
2094         int ret;
2095         unsigned long count = 0;
2096         unsigned long max_count = 0;
2097         u64 bytes_processed = 0;
2098
2099         delayed_refs = &trans->transaction->delayed_refs;
2100         if (min_bytes == 0) {
2101                 max_count = delayed_refs->num_heads_ready;
2102                 min_bytes = U64_MAX;
2103         }
2104
2105         do {
2106                 if (!locked_ref) {
2107                         locked_ref = btrfs_obtain_ref_head(trans);
2108                         if (IS_ERR_OR_NULL(locked_ref)) {
2109                                 if (PTR_ERR(locked_ref) == -EAGAIN) {
2110                                         continue;
2111                                 } else {
2112                                         break;
2113                                 }
2114                         }
2115                         count++;
2116                 }
2117                 /*
2118                  * We need to try and merge add/drops of the same ref since we
2119                  * can run into issues with relocate dropping the implicit ref
2120                  * and then it being added back again before the drop can
2121                  * finish.  If we merged anything we need to re-loop so we can
2122                  * get a good ref.
2123                  * Or we can get node references of the same type that weren't
2124                  * merged when created due to bumps in the tree mod seq, and
2125                  * we need to merge them to prevent adding an inline extent
2126                  * backref before dropping it (triggering a BUG_ON at
2127                  * insert_inline_extent_backref()).
2128                  */
2129                 spin_lock(&locked_ref->lock);
2130                 btrfs_merge_delayed_refs(fs_info, delayed_refs, locked_ref);
2131
2132                 ret = btrfs_run_delayed_refs_for_head(trans, locked_ref, &bytes_processed);
2133                 if (ret < 0 && ret != -EAGAIN) {
2134                         /*
2135                          * Error, btrfs_run_delayed_refs_for_head already
2136                          * unlocked everything so just bail out
2137                          */
2138                         return ret;
2139                 } else if (!ret) {
2140                         /*
2141                          * Success, perform the usual cleanup of a processed
2142                          * head
2143                          */
2144                         ret = cleanup_ref_head(trans, locked_ref, &bytes_processed);
2145                         if (ret > 0 ) {
2146                                 /* We dropped our lock, we need to loop. */
2147                                 ret = 0;
2148                                 continue;
2149                         } else if (ret) {
2150                                 return ret;
2151                         }
2152                 }
2153
2154                 /*
2155                  * Either success case or btrfs_run_delayed_refs_for_head
2156                  * returned -EAGAIN, meaning we need to select another head
2157                  */
2158
2159                 locked_ref = NULL;
2160                 cond_resched();
2161         } while ((min_bytes != U64_MAX && bytes_processed < min_bytes) ||
2162                  (max_count > 0 && count < max_count) ||
2163                  locked_ref);
2164
2165         return 0;
2166 }
2167
2168 #ifdef SCRAMBLE_DELAYED_REFS
2169 /*
2170  * Normally delayed refs get processed in ascending bytenr order. This
2171  * correlates in most cases to the order added. To expose dependencies on this
2172  * order, we start to process the tree in the middle instead of the beginning
2173  */
2174 static u64 find_middle(struct rb_root *root)
2175 {
2176         struct rb_node *n = root->rb_node;
2177         struct btrfs_delayed_ref_node *entry;
2178         int alt = 1;
2179         u64 middle;
2180         u64 first = 0, last = 0;
2181
2182         n = rb_first(root);
2183         if (n) {
2184                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2185                 first = entry->bytenr;
2186         }
2187         n = rb_last(root);
2188         if (n) {
2189                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2190                 last = entry->bytenr;
2191         }
2192         n = root->rb_node;
2193
2194         while (n) {
2195                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2196                 WARN_ON(!entry->in_tree);
2197
2198                 middle = entry->bytenr;
2199
2200                 if (alt)
2201                         n = n->rb_left;
2202                 else
2203                         n = n->rb_right;
2204
2205                 alt = 1 - alt;
2206         }
2207         return middle;
2208 }
2209 #endif
2210
2211 /*
2212  * Start processing the delayed reference count updates and extent insertions
2213  * we have queued up so far.
2214  *
2215  * @trans:      Transaction handle.
2216  * @min_bytes:  How many bytes of delayed references to process. After this
2217  *              many bytes we stop processing delayed references if there are
2218  *              any more. If 0 it means to run all existing delayed references,
2219  *              but not new ones added after running all existing ones.
2220  *              Use (u64)-1 (U64_MAX) to run all existing delayed references
2221  *              plus any new ones that are added.
2222  *
2223  * Returns 0 on success or if called with an aborted transaction
2224  * Returns <0 on error and aborts the transaction
2225  */
2226 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans, u64 min_bytes)
2227 {
2228         struct btrfs_fs_info *fs_info = trans->fs_info;
2229         struct btrfs_delayed_ref_root *delayed_refs;
2230         int ret;
2231
2232         /* We'll clean this up in btrfs_cleanup_transaction */
2233         if (TRANS_ABORTED(trans))
2234                 return 0;
2235
2236         if (test_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags))
2237                 return 0;
2238
2239         delayed_refs = &trans->transaction->delayed_refs;
2240 again:
2241 #ifdef SCRAMBLE_DELAYED_REFS
2242         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2243 #endif
2244         ret = __btrfs_run_delayed_refs(trans, min_bytes);
2245         if (ret < 0) {
2246                 btrfs_abort_transaction(trans, ret);
2247                 return ret;
2248         }
2249
2250         if (min_bytes == U64_MAX) {
2251                 btrfs_create_pending_block_groups(trans);
2252
2253                 spin_lock(&delayed_refs->lock);
2254                 if (RB_EMPTY_ROOT(&delayed_refs->href_root.rb_root)) {
2255                         spin_unlock(&delayed_refs->lock);
2256                         return 0;
2257                 }
2258                 spin_unlock(&delayed_refs->lock);
2259
2260                 cond_resched();
2261                 goto again;
2262         }
2263
2264         return 0;
2265 }
2266
2267 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2268                                 struct extent_buffer *eb, u64 flags)
2269 {
2270         struct btrfs_delayed_extent_op *extent_op;
2271         int level = btrfs_header_level(eb);
2272         int ret;
2273
2274         extent_op = btrfs_alloc_delayed_extent_op();
2275         if (!extent_op)
2276                 return -ENOMEM;
2277
2278         extent_op->flags_to_set = flags;
2279         extent_op->update_flags = true;
2280         extent_op->update_key = false;
2281         extent_op->level = level;
2282
2283         ret = btrfs_add_delayed_extent_op(trans, eb->start, eb->len, extent_op);
2284         if (ret)
2285                 btrfs_free_delayed_extent_op(extent_op);
2286         return ret;
2287 }
2288
2289 static noinline int check_delayed_ref(struct btrfs_root *root,
2290                                       struct btrfs_path *path,
2291                                       u64 objectid, u64 offset, u64 bytenr)
2292 {
2293         struct btrfs_delayed_ref_head *head;
2294         struct btrfs_delayed_ref_node *ref;
2295         struct btrfs_delayed_data_ref *data_ref;
2296         struct btrfs_delayed_ref_root *delayed_refs;
2297         struct btrfs_transaction *cur_trans;
2298         struct rb_node *node;
2299         int ret = 0;
2300
2301         spin_lock(&root->fs_info->trans_lock);
2302         cur_trans = root->fs_info->running_transaction;
2303         if (cur_trans)
2304                 refcount_inc(&cur_trans->use_count);
2305         spin_unlock(&root->fs_info->trans_lock);
2306         if (!cur_trans)
2307                 return 0;
2308
2309         delayed_refs = &cur_trans->delayed_refs;
2310         spin_lock(&delayed_refs->lock);
2311         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
2312         if (!head) {
2313                 spin_unlock(&delayed_refs->lock);
2314                 btrfs_put_transaction(cur_trans);
2315                 return 0;
2316         }
2317
2318         if (!mutex_trylock(&head->mutex)) {
2319                 if (path->nowait) {
2320                         spin_unlock(&delayed_refs->lock);
2321                         btrfs_put_transaction(cur_trans);
2322                         return -EAGAIN;
2323                 }
2324
2325                 refcount_inc(&head->refs);
2326                 spin_unlock(&delayed_refs->lock);
2327
2328                 btrfs_release_path(path);
2329
2330                 /*
2331                  * Mutex was contended, block until it's released and let
2332                  * caller try again
2333                  */
2334                 mutex_lock(&head->mutex);
2335                 mutex_unlock(&head->mutex);
2336                 btrfs_put_delayed_ref_head(head);
2337                 btrfs_put_transaction(cur_trans);
2338                 return -EAGAIN;
2339         }
2340         spin_unlock(&delayed_refs->lock);
2341
2342         spin_lock(&head->lock);
2343         /*
2344          * XXX: We should replace this with a proper search function in the
2345          * future.
2346          */
2347         for (node = rb_first_cached(&head->ref_tree); node;
2348              node = rb_next(node)) {
2349                 ref = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
2350                 /* If it's a shared ref we know a cross reference exists */
2351                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
2352                         ret = 1;
2353                         break;
2354                 }
2355
2356                 data_ref = btrfs_delayed_node_to_data_ref(ref);
2357
2358                 /*
2359                  * If our ref doesn't match the one we're currently looking at
2360                  * then we have a cross reference.
2361                  */
2362                 if (data_ref->root != root->root_key.objectid ||
2363                     data_ref->objectid != objectid ||
2364                     data_ref->offset != offset) {
2365                         ret = 1;
2366                         break;
2367                 }
2368         }
2369         spin_unlock(&head->lock);
2370         mutex_unlock(&head->mutex);
2371         btrfs_put_transaction(cur_trans);
2372         return ret;
2373 }
2374
2375 static noinline int check_committed_ref(struct btrfs_root *root,
2376                                         struct btrfs_path *path,
2377                                         u64 objectid, u64 offset, u64 bytenr,
2378                                         bool strict)
2379 {
2380         struct btrfs_fs_info *fs_info = root->fs_info;
2381         struct btrfs_root *extent_root = btrfs_extent_root(fs_info, bytenr);
2382         struct extent_buffer *leaf;
2383         struct btrfs_extent_data_ref *ref;
2384         struct btrfs_extent_inline_ref *iref;
2385         struct btrfs_extent_item *ei;
2386         struct btrfs_key key;
2387         u32 item_size;
2388         u32 expected_size;
2389         int type;
2390         int ret;
2391
2392         key.objectid = bytenr;
2393         key.offset = (u64)-1;
2394         key.type = BTRFS_EXTENT_ITEM_KEY;
2395
2396         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2397         if (ret < 0)
2398                 goto out;
2399         if (ret == 0) {
2400                 /*
2401                  * Key with offset -1 found, there would have to exist an extent
2402                  * item with such offset, but this is out of the valid range.
2403                  */
2404                 ret = -EUCLEAN;
2405                 goto out;
2406         }
2407
2408         ret = -ENOENT;
2409         if (path->slots[0] == 0)
2410                 goto out;
2411
2412         path->slots[0]--;
2413         leaf = path->nodes[0];
2414         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2415
2416         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2417                 goto out;
2418
2419         ret = 1;
2420         item_size = btrfs_item_size(leaf, path->slots[0]);
2421         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2422         expected_size = sizeof(*ei) + btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY);
2423
2424         /* No inline refs; we need to bail before checking for owner ref. */
2425         if (item_size == sizeof(*ei))
2426                 goto out;
2427
2428         /* Check for an owner ref; skip over it to the real inline refs. */
2429         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2430         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
2431         if (btrfs_fs_incompat(fs_info, SIMPLE_QUOTA) && type == BTRFS_EXTENT_OWNER_REF_KEY) {
2432                 expected_size += btrfs_extent_inline_ref_size(BTRFS_EXTENT_OWNER_REF_KEY);
2433                 iref = (struct btrfs_extent_inline_ref *)(iref + 1);
2434         }
2435
2436         /* If extent item has more than 1 inline ref then it's shared */
2437         if (item_size != expected_size)
2438                 goto out;
2439
2440         /*
2441          * If extent created before last snapshot => it's shared unless the
2442          * snapshot has been deleted. Use the heuristic if strict is false.
2443          */
2444         if (!strict &&
2445             (btrfs_extent_generation(leaf, ei) <=
2446              btrfs_root_last_snapshot(&root->root_item)))
2447                 goto out;
2448
2449         /* If this extent has SHARED_DATA_REF then it's shared */
2450         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
2451         if (type != BTRFS_EXTENT_DATA_REF_KEY)
2452                 goto out;
2453
2454         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2455         if (btrfs_extent_refs(leaf, ei) !=
2456             btrfs_extent_data_ref_count(leaf, ref) ||
2457             btrfs_extent_data_ref_root(leaf, ref) !=
2458             root->root_key.objectid ||
2459             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2460             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2461                 goto out;
2462
2463         ret = 0;
2464 out:
2465         return ret;
2466 }
2467
2468 int btrfs_cross_ref_exist(struct btrfs_root *root, u64 objectid, u64 offset,
2469                           u64 bytenr, bool strict, struct btrfs_path *path)
2470 {
2471         int ret;
2472
2473         do {
2474                 ret = check_committed_ref(root, path, objectid,
2475                                           offset, bytenr, strict);
2476                 if (ret && ret != -ENOENT)
2477                         goto out;
2478
2479                 ret = check_delayed_ref(root, path, objectid, offset, bytenr);
2480         } while (ret == -EAGAIN);
2481
2482 out:
2483         btrfs_release_path(path);
2484         if (btrfs_is_data_reloc_root(root))
2485                 WARN_ON(ret > 0);
2486         return ret;
2487 }
2488
2489 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2490                            struct btrfs_root *root,
2491                            struct extent_buffer *buf,
2492                            int full_backref, int inc)
2493 {
2494         struct btrfs_fs_info *fs_info = root->fs_info;
2495         u64 bytenr;
2496         u64 num_bytes;
2497         u64 parent;
2498         u64 ref_root;
2499         u32 nritems;
2500         struct btrfs_key key;
2501         struct btrfs_file_extent_item *fi;
2502         struct btrfs_ref generic_ref = { 0 };
2503         bool for_reloc = btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC);
2504         int i;
2505         int action;
2506         int level;
2507         int ret = 0;
2508
2509         if (btrfs_is_testing(fs_info))
2510                 return 0;
2511
2512         ref_root = btrfs_header_owner(buf);
2513         nritems = btrfs_header_nritems(buf);
2514         level = btrfs_header_level(buf);
2515
2516         if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state) && level == 0)
2517                 return 0;
2518
2519         if (full_backref)
2520                 parent = buf->start;
2521         else
2522                 parent = 0;
2523         if (inc)
2524                 action = BTRFS_ADD_DELAYED_REF;
2525         else
2526                 action = BTRFS_DROP_DELAYED_REF;
2527
2528         for (i = 0; i < nritems; i++) {
2529                 if (level == 0) {
2530                         btrfs_item_key_to_cpu(buf, &key, i);
2531                         if (key.type != BTRFS_EXTENT_DATA_KEY)
2532                                 continue;
2533                         fi = btrfs_item_ptr(buf, i,
2534                                             struct btrfs_file_extent_item);
2535                         if (btrfs_file_extent_type(buf, fi) ==
2536                             BTRFS_FILE_EXTENT_INLINE)
2537                                 continue;
2538                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2539                         if (bytenr == 0)
2540                                 continue;
2541
2542                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
2543                         key.offset -= btrfs_file_extent_offset(buf, fi);
2544                         btrfs_init_generic_ref(&generic_ref, action, bytenr,
2545                                                num_bytes, parent, ref_root);
2546                         btrfs_init_data_ref(&generic_ref, ref_root, key.objectid,
2547                                             key.offset, root->root_key.objectid,
2548                                             for_reloc);
2549                         if (inc)
2550                                 ret = btrfs_inc_extent_ref(trans, &generic_ref);
2551                         else
2552                                 ret = btrfs_free_extent(trans, &generic_ref);
2553                         if (ret)
2554                                 goto fail;
2555                 } else {
2556                         bytenr = btrfs_node_blockptr(buf, i);
2557                         num_bytes = fs_info->nodesize;
2558                         /* We don't know the owning_root, use 0. */
2559                         btrfs_init_generic_ref(&generic_ref, action, bytenr,
2560                                                num_bytes, parent, 0);
2561                         btrfs_init_tree_ref(&generic_ref, level - 1, ref_root,
2562                                             root->root_key.objectid, for_reloc);
2563                         if (inc)
2564                                 ret = btrfs_inc_extent_ref(trans, &generic_ref);
2565                         else
2566                                 ret = btrfs_free_extent(trans, &generic_ref);
2567                         if (ret)
2568                                 goto fail;
2569                 }
2570         }
2571         return 0;
2572 fail:
2573         return ret;
2574 }
2575
2576 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2577                   struct extent_buffer *buf, int full_backref)
2578 {
2579         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
2580 }
2581
2582 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2583                   struct extent_buffer *buf, int full_backref)
2584 {
2585         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
2586 }
2587
2588 static u64 get_alloc_profile_by_root(struct btrfs_root *root, int data)
2589 {
2590         struct btrfs_fs_info *fs_info = root->fs_info;
2591         u64 flags;
2592         u64 ret;
2593
2594         if (data)
2595                 flags = BTRFS_BLOCK_GROUP_DATA;
2596         else if (root == fs_info->chunk_root)
2597                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
2598         else
2599                 flags = BTRFS_BLOCK_GROUP_METADATA;
2600
2601         ret = btrfs_get_alloc_profile(fs_info, flags);
2602         return ret;
2603 }
2604
2605 static u64 first_logical_byte(struct btrfs_fs_info *fs_info)
2606 {
2607         struct rb_node *leftmost;
2608         u64 bytenr = 0;
2609
2610         read_lock(&fs_info->block_group_cache_lock);
2611         /* Get the block group with the lowest logical start address. */
2612         leftmost = rb_first_cached(&fs_info->block_group_cache_tree);
2613         if (leftmost) {
2614                 struct btrfs_block_group *bg;
2615
2616                 bg = rb_entry(leftmost, struct btrfs_block_group, cache_node);
2617                 bytenr = bg->start;
2618         }
2619         read_unlock(&fs_info->block_group_cache_lock);
2620
2621         return bytenr;
2622 }
2623
2624 static int pin_down_extent(struct btrfs_trans_handle *trans,
2625                            struct btrfs_block_group *cache,
2626                            u64 bytenr, u64 num_bytes, int reserved)
2627 {
2628         struct btrfs_fs_info *fs_info = cache->fs_info;
2629
2630         spin_lock(&cache->space_info->lock);
2631         spin_lock(&cache->lock);
2632         cache->pinned += num_bytes;
2633         btrfs_space_info_update_bytes_pinned(fs_info, cache->space_info,
2634                                              num_bytes);
2635         if (reserved) {
2636                 cache->reserved -= num_bytes;
2637                 cache->space_info->bytes_reserved -= num_bytes;
2638         }
2639         spin_unlock(&cache->lock);
2640         spin_unlock(&cache->space_info->lock);
2641
2642         set_extent_bit(&trans->transaction->pinned_extents, bytenr,
2643                        bytenr + num_bytes - 1, EXTENT_DIRTY, NULL);
2644         return 0;
2645 }
2646
2647 int btrfs_pin_extent(struct btrfs_trans_handle *trans,
2648                      u64 bytenr, u64 num_bytes, int reserved)
2649 {
2650         struct btrfs_block_group *cache;
2651
2652         cache = btrfs_lookup_block_group(trans->fs_info, bytenr);
2653         BUG_ON(!cache); /* Logic error */
2654
2655         pin_down_extent(trans, cache, bytenr, num_bytes, reserved);
2656
2657         btrfs_put_block_group(cache);
2658         return 0;
2659 }
2660
2661 int btrfs_pin_extent_for_log_replay(struct btrfs_trans_handle *trans,
2662                                     const struct extent_buffer *eb)
2663 {
2664         struct btrfs_block_group *cache;
2665         int ret;
2666
2667         cache = btrfs_lookup_block_group(trans->fs_info, eb->start);
2668         if (!cache)
2669                 return -EINVAL;
2670
2671         /*
2672          * Fully cache the free space first so that our pin removes the free space
2673          * from the cache.
2674          */
2675         ret = btrfs_cache_block_group(cache, true);
2676         if (ret)
2677                 goto out;
2678
2679         pin_down_extent(trans, cache, eb->start, eb->len, 0);
2680
2681         /* remove us from the free space cache (if we're there at all) */
2682         ret = btrfs_remove_free_space(cache, eb->start, eb->len);
2683 out:
2684         btrfs_put_block_group(cache);
2685         return ret;
2686 }
2687
2688 static int __exclude_logged_extent(struct btrfs_fs_info *fs_info,
2689                                    u64 start, u64 num_bytes)
2690 {
2691         int ret;
2692         struct btrfs_block_group *block_group;
2693
2694         block_group = btrfs_lookup_block_group(fs_info, start);
2695         if (!block_group)
2696                 return -EINVAL;
2697
2698         ret = btrfs_cache_block_group(block_group, true);
2699         if (ret)
2700                 goto out;
2701
2702         ret = btrfs_remove_free_space(block_group, start, num_bytes);
2703 out:
2704         btrfs_put_block_group(block_group);
2705         return ret;
2706 }
2707
2708 int btrfs_exclude_logged_extents(struct extent_buffer *eb)
2709 {
2710         struct btrfs_fs_info *fs_info = eb->fs_info;
2711         struct btrfs_file_extent_item *item;
2712         struct btrfs_key key;
2713         int found_type;
2714         int i;
2715         int ret = 0;
2716
2717         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS))
2718                 return 0;
2719
2720         for (i = 0; i < btrfs_header_nritems(eb); i++) {
2721                 btrfs_item_key_to_cpu(eb, &key, i);
2722                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2723                         continue;
2724                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
2725                 found_type = btrfs_file_extent_type(eb, item);
2726                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
2727                         continue;
2728                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
2729                         continue;
2730                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
2731                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
2732                 ret = __exclude_logged_extent(fs_info, key.objectid, key.offset);
2733                 if (ret)
2734                         break;
2735         }
2736
2737         return ret;
2738 }
2739
2740 static void
2741 btrfs_inc_block_group_reservations(struct btrfs_block_group *bg)
2742 {
2743         atomic_inc(&bg->reservations);
2744 }
2745
2746 /*
2747  * Returns the free cluster for the given space info and sets empty_cluster to
2748  * what it should be based on the mount options.
2749  */
2750 static struct btrfs_free_cluster *
2751 fetch_cluster_info(struct btrfs_fs_info *fs_info,
2752                    struct btrfs_space_info *space_info, u64 *empty_cluster)
2753 {
2754         struct btrfs_free_cluster *ret = NULL;
2755
2756         *empty_cluster = 0;
2757         if (btrfs_mixed_space_info(space_info))
2758                 return ret;
2759
2760         if (space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
2761                 ret = &fs_info->meta_alloc_cluster;
2762                 if (btrfs_test_opt(fs_info, SSD))
2763                         *empty_cluster = SZ_2M;
2764                 else
2765                         *empty_cluster = SZ_64K;
2766         } else if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) &&
2767                    btrfs_test_opt(fs_info, SSD_SPREAD)) {
2768                 *empty_cluster = SZ_2M;
2769                 ret = &fs_info->data_alloc_cluster;
2770         }
2771
2772         return ret;
2773 }
2774
2775 static int unpin_extent_range(struct btrfs_fs_info *fs_info,
2776                               u64 start, u64 end,
2777                               const bool return_free_space)
2778 {
2779         struct btrfs_block_group *cache = NULL;
2780         struct btrfs_space_info *space_info;
2781         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
2782         struct btrfs_free_cluster *cluster = NULL;
2783         u64 len;
2784         u64 total_unpinned = 0;
2785         u64 empty_cluster = 0;
2786         bool readonly;
2787         int ret = 0;
2788
2789         while (start <= end) {
2790                 readonly = false;
2791                 if (!cache ||
2792                     start >= cache->start + cache->length) {
2793                         if (cache)
2794                                 btrfs_put_block_group(cache);
2795                         total_unpinned = 0;
2796                         cache = btrfs_lookup_block_group(fs_info, start);
2797                         if (cache == NULL) {
2798                                 /* Logic error, something removed the block group. */
2799                                 ret = -EUCLEAN;
2800                                 goto out;
2801                         }
2802
2803                         cluster = fetch_cluster_info(fs_info,
2804                                                      cache->space_info,
2805                                                      &empty_cluster);
2806                         empty_cluster <<= 1;
2807                 }
2808
2809                 len = cache->start + cache->length - start;
2810                 len = min(len, end + 1 - start);
2811
2812                 if (return_free_space)
2813                         btrfs_add_free_space(cache, start, len);
2814
2815                 start += len;
2816                 total_unpinned += len;
2817                 space_info = cache->space_info;
2818
2819                 /*
2820                  * If this space cluster has been marked as fragmented and we've
2821                  * unpinned enough in this block group to potentially allow a
2822                  * cluster to be created inside of it go ahead and clear the
2823                  * fragmented check.
2824                  */
2825                 if (cluster && cluster->fragmented &&
2826                     total_unpinned > empty_cluster) {
2827                         spin_lock(&cluster->lock);
2828                         cluster->fragmented = 0;
2829                         spin_unlock(&cluster->lock);
2830                 }
2831
2832                 spin_lock(&space_info->lock);
2833                 spin_lock(&cache->lock);
2834                 cache->pinned -= len;
2835                 btrfs_space_info_update_bytes_pinned(fs_info, space_info, -len);
2836                 space_info->max_extent_size = 0;
2837                 if (cache->ro) {
2838                         space_info->bytes_readonly += len;
2839                         readonly = true;
2840                 } else if (btrfs_is_zoned(fs_info)) {
2841                         /* Need reset before reusing in a zoned block group */
2842                         space_info->bytes_zone_unusable += len;
2843                         readonly = true;
2844                 }
2845                 spin_unlock(&cache->lock);
2846                 if (!readonly && return_free_space &&
2847                     global_rsv->space_info == space_info) {
2848                         spin_lock(&global_rsv->lock);
2849                         if (!global_rsv->full) {
2850                                 u64 to_add = min(len, global_rsv->size -
2851                                                       global_rsv->reserved);
2852
2853                                 global_rsv->reserved += to_add;
2854                                 btrfs_space_info_update_bytes_may_use(fs_info,
2855                                                 space_info, to_add);
2856                                 if (global_rsv->reserved >= global_rsv->size)
2857                                         global_rsv->full = 1;
2858                                 len -= to_add;
2859                         }
2860                         spin_unlock(&global_rsv->lock);
2861                 }
2862                 /* Add to any tickets we may have */
2863                 if (!readonly && return_free_space && len)
2864                         btrfs_try_granting_tickets(fs_info, space_info);
2865                 spin_unlock(&space_info->lock);
2866         }
2867
2868         if (cache)
2869                 btrfs_put_block_group(cache);
2870 out:
2871         return ret;
2872 }
2873
2874 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans)
2875 {
2876         struct btrfs_fs_info *fs_info = trans->fs_info;
2877         struct btrfs_block_group *block_group, *tmp;
2878         struct list_head *deleted_bgs;
2879         struct extent_io_tree *unpin;
2880         u64 start;
2881         u64 end;
2882         int ret;
2883
2884         unpin = &trans->transaction->pinned_extents;
2885
2886         while (!TRANS_ABORTED(trans)) {
2887                 struct extent_state *cached_state = NULL;
2888
2889                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
2890                 if (!find_first_extent_bit(unpin, 0, &start, &end,
2891                                            EXTENT_DIRTY, &cached_state)) {
2892                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
2893                         break;
2894                 }
2895
2896                 if (btrfs_test_opt(fs_info, DISCARD_SYNC))
2897                         ret = btrfs_discard_extent(fs_info, start,
2898                                                    end + 1 - start, NULL);
2899
2900                 clear_extent_dirty(unpin, start, end, &cached_state);
2901                 ret = unpin_extent_range(fs_info, start, end, true);
2902                 BUG_ON(ret);
2903                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
2904                 free_extent_state(cached_state);
2905                 cond_resched();
2906         }
2907
2908         if (btrfs_test_opt(fs_info, DISCARD_ASYNC)) {
2909                 btrfs_discard_calc_delay(&fs_info->discard_ctl);
2910                 btrfs_discard_schedule_work(&fs_info->discard_ctl, true);
2911         }
2912
2913         /*
2914          * Transaction is finished.  We don't need the lock anymore.  We
2915          * do need to clean up the block groups in case of a transaction
2916          * abort.
2917          */
2918         deleted_bgs = &trans->transaction->deleted_bgs;
2919         list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
2920                 u64 trimmed = 0;
2921
2922                 ret = -EROFS;
2923                 if (!TRANS_ABORTED(trans))
2924                         ret = btrfs_discard_extent(fs_info,
2925                                                    block_group->start,
2926                                                    block_group->length,
2927                                                    &trimmed);
2928
2929                 list_del_init(&block_group->bg_list);
2930                 btrfs_unfreeze_block_group(block_group);
2931                 btrfs_put_block_group(block_group);
2932
2933                 if (ret) {
2934                         const char *errstr = btrfs_decode_error(ret);
2935                         btrfs_warn(fs_info,
2936                            "discard failed while removing blockgroup: errno=%d %s",
2937                                    ret, errstr);
2938                 }
2939         }
2940
2941         return 0;
2942 }
2943
2944 /*
2945  * Parse an extent item's inline extents looking for a simple quotas owner ref.
2946  *
2947  * @fs_info:    the btrfs_fs_info for this mount
2948  * @leaf:       a leaf in the extent tree containing the extent item
2949  * @slot:       the slot in the leaf where the extent item is found
2950  *
2951  * Returns the objectid of the root that originally allocated the extent item
2952  * if the inline owner ref is expected and present, otherwise 0.
2953  *
2954  * If an extent item has an owner ref item, it will be the first inline ref
2955  * item. Therefore the logic is to check whether there are any inline ref
2956  * items, then check the type of the first one.
2957  */
2958 u64 btrfs_get_extent_owner_root(struct btrfs_fs_info *fs_info,
2959                                 struct extent_buffer *leaf, int slot)
2960 {
2961         struct btrfs_extent_item *ei;
2962         struct btrfs_extent_inline_ref *iref;
2963         struct btrfs_extent_owner_ref *oref;
2964         unsigned long ptr;
2965         unsigned long end;
2966         int type;
2967
2968         if (!btrfs_fs_incompat(fs_info, SIMPLE_QUOTA))
2969                 return 0;
2970
2971         ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
2972         ptr = (unsigned long)(ei + 1);
2973         end = (unsigned long)ei + btrfs_item_size(leaf, slot);
2974
2975         /* No inline ref items of any kind, can't check type. */
2976         if (ptr == end)
2977                 return 0;
2978
2979         iref = (struct btrfs_extent_inline_ref *)ptr;
2980         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_ANY);
2981
2982         /* We found an owner ref, get the root out of it. */
2983         if (type == BTRFS_EXTENT_OWNER_REF_KEY) {
2984                 oref = (struct btrfs_extent_owner_ref *)(&iref->offset);
2985                 return btrfs_extent_owner_ref_root_id(leaf, oref);
2986         }
2987
2988         /* We have inline refs, but not an owner ref. */
2989         return 0;
2990 }
2991
2992 static int do_free_extent_accounting(struct btrfs_trans_handle *trans,
2993                                      u64 bytenr, struct btrfs_squota_delta *delta)
2994 {
2995         int ret;
2996         u64 num_bytes = delta->num_bytes;
2997
2998         if (delta->is_data) {
2999                 struct btrfs_root *csum_root;
3000
3001                 csum_root = btrfs_csum_root(trans->fs_info, bytenr);
3002                 ret = btrfs_del_csums(trans, csum_root, bytenr, num_bytes);
3003                 if (ret) {
3004                         btrfs_abort_transaction(trans, ret);
3005                         return ret;
3006                 }
3007
3008                 ret = btrfs_delete_raid_extent(trans, bytenr, num_bytes);
3009                 if (ret) {
3010                         btrfs_abort_transaction(trans, ret);
3011                         return ret;
3012                 }
3013         }
3014
3015         ret = btrfs_record_squota_delta(trans->fs_info, delta);
3016         if (ret) {
3017                 btrfs_abort_transaction(trans, ret);
3018                 return ret;
3019         }
3020
3021         ret = add_to_free_space_tree(trans, bytenr, num_bytes);
3022         if (ret) {
3023                 btrfs_abort_transaction(trans, ret);
3024                 return ret;
3025         }
3026
3027         ret = btrfs_update_block_group(trans, bytenr, num_bytes, false);
3028         if (ret)
3029                 btrfs_abort_transaction(trans, ret);
3030
3031         return ret;
3032 }
3033
3034 #define abort_and_dump(trans, path, fmt, args...)       \
3035 ({                                                      \
3036         btrfs_abort_transaction(trans, -EUCLEAN);       \
3037         btrfs_print_leaf(path->nodes[0]);               \
3038         btrfs_crit(trans->fs_info, fmt, ##args);        \
3039 })
3040
3041 /*
3042  * Drop one or more refs of @node.
3043  *
3044  * 1. Locate the extent refs.
3045  *    It's either inline in EXTENT/METADATA_ITEM or in keyed SHARED_* item.
3046  *    Locate it, then reduce the refs number or remove the ref line completely.
3047  *
3048  * 2. Update the refs count in EXTENT/METADATA_ITEM
3049  *
3050  * Inline backref case:
3051  *
3052  * in extent tree we have:
3053  *
3054  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 16201 itemsize 82
3055  *              refs 2 gen 6 flags DATA
3056  *              extent data backref root FS_TREE objectid 258 offset 0 count 1
3057  *              extent data backref root FS_TREE objectid 257 offset 0 count 1
3058  *
3059  * This function gets called with:
3060  *
3061  *    node->bytenr = 13631488
3062  *    node->num_bytes = 1048576
3063  *    root_objectid = FS_TREE
3064  *    owner_objectid = 257
3065  *    owner_offset = 0
3066  *    refs_to_drop = 1
3067  *
3068  * Then we should get some like:
3069  *
3070  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 16201 itemsize 82
3071  *              refs 1 gen 6 flags DATA
3072  *              extent data backref root FS_TREE objectid 258 offset 0 count 1
3073  *
3074  * Keyed backref case:
3075  *
3076  * in extent tree we have:
3077  *
3078  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 3971 itemsize 24
3079  *              refs 754 gen 6 flags DATA
3080  *      [...]
3081  *      item 2 key (13631488 EXTENT_DATA_REF <HASH>) itemoff 3915 itemsize 28
3082  *              extent data backref root FS_TREE objectid 866 offset 0 count 1
3083  *
3084  * This function get called with:
3085  *
3086  *    node->bytenr = 13631488
3087  *    node->num_bytes = 1048576
3088  *    root_objectid = FS_TREE
3089  *    owner_objectid = 866
3090  *    owner_offset = 0
3091  *    refs_to_drop = 1
3092  *
3093  * Then we should get some like:
3094  *
3095  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 3971 itemsize 24
3096  *              refs 753 gen 6 flags DATA
3097  *
3098  * And that (13631488 EXTENT_DATA_REF <HASH>) gets removed.
3099  */
3100 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
3101                                struct btrfs_delayed_ref_head *href,
3102                                struct btrfs_delayed_ref_node *node, u64 parent,
3103                                u64 root_objectid, u64 owner_objectid,
3104                                u64 owner_offset,
3105                                struct btrfs_delayed_extent_op *extent_op)
3106 {
3107         struct btrfs_fs_info *info = trans->fs_info;
3108         struct btrfs_key key;
3109         struct btrfs_path *path;
3110         struct btrfs_root *extent_root;
3111         struct extent_buffer *leaf;
3112         struct btrfs_extent_item *ei;
3113         struct btrfs_extent_inline_ref *iref;
3114         int ret;
3115         int is_data;
3116         int extent_slot = 0;
3117         int found_extent = 0;
3118         int num_to_del = 1;
3119         int refs_to_drop = node->ref_mod;
3120         u32 item_size;
3121         u64 refs;
3122         u64 bytenr = node->bytenr;
3123         u64 num_bytes = node->num_bytes;
3124         bool skinny_metadata = btrfs_fs_incompat(info, SKINNY_METADATA);
3125         u64 delayed_ref_root = href->owning_root;
3126
3127         extent_root = btrfs_extent_root(info, bytenr);
3128         ASSERT(extent_root);
3129
3130         path = btrfs_alloc_path();
3131         if (!path)
3132                 return -ENOMEM;
3133
3134         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
3135
3136         if (!is_data && refs_to_drop != 1) {
3137                 btrfs_crit(info,
3138 "invalid refs_to_drop, dropping more than 1 refs for tree block %llu refs_to_drop %u",
3139                            node->bytenr, refs_to_drop);
3140                 ret = -EINVAL;
3141                 btrfs_abort_transaction(trans, ret);
3142                 goto out;
3143         }
3144
3145         if (is_data)
3146                 skinny_metadata = false;
3147
3148         ret = lookup_extent_backref(trans, path, &iref, bytenr, num_bytes,
3149                                     parent, root_objectid, owner_objectid,
3150                                     owner_offset);
3151         if (ret == 0) {
3152                 /*
3153                  * Either the inline backref or the SHARED_DATA_REF/
3154                  * SHARED_BLOCK_REF is found
3155                  *
3156                  * Here is a quick path to locate EXTENT/METADATA_ITEM.
3157                  * It's possible the EXTENT/METADATA_ITEM is near current slot.
3158                  */
3159                 extent_slot = path->slots[0];
3160                 while (extent_slot >= 0) {
3161                         btrfs_item_key_to_cpu(path->nodes[0], &key,
3162                                               extent_slot);
3163                         if (key.objectid != bytenr)
3164                                 break;
3165                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
3166                             key.offset == num_bytes) {
3167                                 found_extent = 1;
3168                                 break;
3169                         }
3170                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
3171                             key.offset == owner_objectid) {
3172                                 found_extent = 1;
3173                                 break;
3174                         }
3175
3176                         /* Quick path didn't find the EXTEMT/METADATA_ITEM */
3177                         if (path->slots[0] - extent_slot > 5)
3178                                 break;
3179                         extent_slot--;
3180                 }
3181
3182                 if (!found_extent) {
3183                         if (iref) {
3184                                 abort_and_dump(trans, path,
3185 "invalid iref slot %u, no EXTENT/METADATA_ITEM found but has inline extent ref",
3186                                            path->slots[0]);
3187                                 ret = -EUCLEAN;
3188                                 goto out;
3189                         }
3190                         /* Must be SHARED_* item, remove the backref first */
3191                         ret = remove_extent_backref(trans, extent_root, path,
3192                                                     NULL, refs_to_drop, is_data);
3193                         if (ret) {
3194                                 btrfs_abort_transaction(trans, ret);
3195                                 goto out;
3196                         }
3197                         btrfs_release_path(path);
3198
3199                         /* Slow path to locate EXTENT/METADATA_ITEM */
3200                         key.objectid = bytenr;
3201                         key.type = BTRFS_EXTENT_ITEM_KEY;
3202                         key.offset = num_bytes;
3203
3204                         if (!is_data && skinny_metadata) {
3205                                 key.type = BTRFS_METADATA_ITEM_KEY;
3206                                 key.offset = owner_objectid;
3207                         }
3208
3209                         ret = btrfs_search_slot(trans, extent_root,
3210                                                 &key, path, -1, 1);
3211                         if (ret > 0 && skinny_metadata && path->slots[0]) {
3212                                 /*
3213                                  * Couldn't find our skinny metadata item,
3214                                  * see if we have ye olde extent item.
3215                                  */
3216                                 path->slots[0]--;
3217                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
3218                                                       path->slots[0]);
3219                                 if (key.objectid == bytenr &&
3220                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
3221                                     key.offset == num_bytes)
3222                                         ret = 0;
3223                         }
3224
3225                         if (ret > 0 && skinny_metadata) {
3226                                 skinny_metadata = false;
3227                                 key.objectid = bytenr;
3228                                 key.type = BTRFS_EXTENT_ITEM_KEY;
3229                                 key.offset = num_bytes;
3230                                 btrfs_release_path(path);
3231                                 ret = btrfs_search_slot(trans, extent_root,
3232                                                         &key, path, -1, 1);
3233                         }
3234
3235                         if (ret) {
3236                                 if (ret > 0)
3237                                         btrfs_print_leaf(path->nodes[0]);
3238                                 btrfs_err(info,
3239                         "umm, got %d back from search, was looking for %llu, slot %d",
3240                                           ret, bytenr, path->slots[0]);
3241                         }
3242                         if (ret < 0) {
3243                                 btrfs_abort_transaction(trans, ret);
3244                                 goto out;
3245                         }
3246                         extent_slot = path->slots[0];
3247                 }
3248         } else if (WARN_ON(ret == -ENOENT)) {
3249                 abort_and_dump(trans, path,
3250 "unable to find ref byte nr %llu parent %llu root %llu owner %llu offset %llu slot %d",
3251                                bytenr, parent, root_objectid, owner_objectid,
3252                                owner_offset, path->slots[0]);
3253                 goto out;
3254         } else {
3255                 btrfs_abort_transaction(trans, ret);
3256                 goto out;
3257         }
3258
3259         leaf = path->nodes[0];
3260         item_size = btrfs_item_size(leaf, extent_slot);
3261         if (unlikely(item_size < sizeof(*ei))) {
3262                 ret = -EUCLEAN;
3263                 btrfs_err(trans->fs_info,
3264                           "unexpected extent item size, has %u expect >= %zu",
3265                           item_size, sizeof(*ei));
3266                 btrfs_abort_transaction(trans, ret);
3267                 goto out;
3268         }
3269         ei = btrfs_item_ptr(leaf, extent_slot,
3270                             struct btrfs_extent_item);
3271         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
3272             key.type == BTRFS_EXTENT_ITEM_KEY) {
3273                 struct btrfs_tree_block_info *bi;
3274
3275                 if (item_size < sizeof(*ei) + sizeof(*bi)) {
3276                         abort_and_dump(trans, path,
3277 "invalid extent item size for key (%llu, %u, %llu) slot %u owner %llu, has %u expect >= %zu",
3278                                        key.objectid, key.type, key.offset,
3279                                        path->slots[0], owner_objectid, item_size,
3280                                        sizeof(*ei) + sizeof(*bi));
3281                         ret = -EUCLEAN;
3282                         goto out;
3283                 }
3284                 bi = (struct btrfs_tree_block_info *)(ei + 1);
3285                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
3286         }
3287
3288         refs = btrfs_extent_refs(leaf, ei);
3289         if (refs < refs_to_drop) {
3290                 abort_and_dump(trans, path,
3291                 "trying to drop %d refs but we only have %llu for bytenr %llu slot %u",
3292                                refs_to_drop, refs, bytenr, path->slots[0]);
3293                 ret = -EUCLEAN;
3294                 goto out;
3295         }
3296         refs -= refs_to_drop;
3297
3298         if (refs > 0) {
3299                 if (extent_op)
3300                         __run_delayed_extent_op(extent_op, leaf, ei);
3301                 /*
3302                  * In the case of inline back ref, reference count will
3303                  * be updated by remove_extent_backref
3304                  */
3305                 if (iref) {
3306                         if (!found_extent) {
3307                                 abort_and_dump(trans, path,
3308 "invalid iref, got inlined extent ref but no EXTENT/METADATA_ITEM found, slot %u",
3309                                                path->slots[0]);
3310                                 ret = -EUCLEAN;
3311                                 goto out;
3312                         }
3313                 } else {
3314                         btrfs_set_extent_refs(leaf, ei, refs);
3315                         btrfs_mark_buffer_dirty(trans, leaf);
3316                 }
3317                 if (found_extent) {
3318                         ret = remove_extent_backref(trans, extent_root, path,
3319                                                     iref, refs_to_drop, is_data);
3320                         if (ret) {
3321                                 btrfs_abort_transaction(trans, ret);
3322                                 goto out;
3323                         }
3324                 }
3325         } else {
3326                 struct btrfs_squota_delta delta = {
3327                         .root = delayed_ref_root,
3328                         .num_bytes = num_bytes,
3329                         .is_data = is_data,
3330                         .is_inc = false,
3331                         .generation = btrfs_extent_generation(leaf, ei),
3332                 };
3333
3334                 /* In this branch refs == 1 */
3335                 if (found_extent) {
3336                         if (is_data && refs_to_drop !=
3337                             extent_data_ref_count(path, iref)) {
3338                                 abort_and_dump(trans, path,
3339                 "invalid refs_to_drop, current refs %u refs_to_drop %u slot %u",
3340                                                extent_data_ref_count(path, iref),
3341                                                refs_to_drop, path->slots[0]);
3342                                 ret = -EUCLEAN;
3343                                 goto out;
3344                         }
3345                         if (iref) {
3346                                 if (path->slots[0] != extent_slot) {
3347                                         abort_and_dump(trans, path,
3348 "invalid iref, extent item key (%llu %u %llu) slot %u doesn't have wanted iref",
3349                                                        key.objectid, key.type,
3350                                                        key.offset, path->slots[0]);
3351                                         ret = -EUCLEAN;
3352                                         goto out;
3353                                 }
3354                         } else {
3355                                 /*
3356                                  * No inline ref, we must be at SHARED_* item,
3357                                  * And it's single ref, it must be:
3358                                  * |    extent_slot       ||extent_slot + 1|
3359                                  * [ EXTENT/METADATA_ITEM ][ SHARED_* ITEM ]
3360                                  */
3361                                 if (path->slots[0] != extent_slot + 1) {
3362                                         abort_and_dump(trans, path,
3363         "invalid SHARED_* item slot %u, previous item is not EXTENT/METADATA_ITEM",
3364                                                        path->slots[0]);
3365                                         ret = -EUCLEAN;
3366                                         goto out;
3367                                 }
3368                                 path->slots[0] = extent_slot;
3369                                 num_to_del = 2;
3370                         }
3371                 }
3372                 /*
3373                  * We can't infer the data owner from the delayed ref, so we need
3374                  * to try to get it from the owning ref item.
3375                  *
3376                  * If it is not present, then that extent was not written under
3377                  * simple quotas mode, so we don't need to account for its deletion.
3378                  */
3379                 if (is_data)
3380                         delta.root = btrfs_get_extent_owner_root(trans->fs_info,
3381                                                                  leaf, extent_slot);
3382
3383                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
3384                                       num_to_del);
3385                 if (ret) {
3386                         btrfs_abort_transaction(trans, ret);
3387                         goto out;
3388                 }
3389                 btrfs_release_path(path);
3390
3391                 ret = do_free_extent_accounting(trans, bytenr, &delta);
3392         }
3393         btrfs_release_path(path);
3394
3395 out:
3396         btrfs_free_path(path);
3397         return ret;
3398 }
3399
3400 /*
3401  * when we free an block, it is possible (and likely) that we free the last
3402  * delayed ref for that extent as well.  This searches the delayed ref tree for
3403  * a given extent, and if there are no other delayed refs to be processed, it
3404  * removes it from the tree.
3405  */
3406 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
3407                                       u64 bytenr)
3408 {
3409         struct btrfs_delayed_ref_head *head;
3410         struct btrfs_delayed_ref_root *delayed_refs;
3411         int ret = 0;
3412
3413         delayed_refs = &trans->transaction->delayed_refs;
3414         spin_lock(&delayed_refs->lock);
3415         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
3416         if (!head)
3417                 goto out_delayed_unlock;
3418
3419         spin_lock(&head->lock);
3420         if (!RB_EMPTY_ROOT(&head->ref_tree.rb_root))
3421                 goto out;
3422
3423         if (cleanup_extent_op(head) != NULL)
3424                 goto out;
3425
3426         /*
3427          * waiting for the lock here would deadlock.  If someone else has it
3428          * locked they are already in the process of dropping it anyway
3429          */
3430         if (!mutex_trylock(&head->mutex))
3431                 goto out;
3432
3433         btrfs_delete_ref_head(delayed_refs, head);
3434         head->processing = false;
3435
3436         spin_unlock(&head->lock);
3437         spin_unlock(&delayed_refs->lock);
3438
3439         BUG_ON(head->extent_op);
3440         if (head->must_insert_reserved)
3441                 ret = 1;
3442
3443         btrfs_cleanup_ref_head_accounting(trans->fs_info, delayed_refs, head);
3444         mutex_unlock(&head->mutex);
3445         btrfs_put_delayed_ref_head(head);
3446         return ret;
3447 out:
3448         spin_unlock(&head->lock);
3449
3450 out_delayed_unlock:
3451         spin_unlock(&delayed_refs->lock);
3452         return 0;
3453 }
3454
3455 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
3456                            u64 root_id,
3457                            struct extent_buffer *buf,
3458                            u64 parent, int last_ref)
3459 {
3460         struct btrfs_fs_info *fs_info = trans->fs_info;
3461         struct btrfs_block_group *bg;
3462         int ret;
3463
3464         if (root_id != BTRFS_TREE_LOG_OBJECTID) {
3465                 struct btrfs_ref generic_ref = { 0 };
3466
3467                 /*
3468                  * Assert that the extent buffer is not cleared due to
3469                  * EXTENT_BUFFER_ZONED_ZEROOUT. Please refer
3470                  * btrfs_clear_buffer_dirty() and btree_csum_one_bio() for
3471                  * detail.
3472                  */
3473                 ASSERT(btrfs_header_bytenr(buf) != 0);
3474
3475                 btrfs_init_generic_ref(&generic_ref, BTRFS_DROP_DELAYED_REF,
3476                                        buf->start, buf->len, parent,
3477                                        btrfs_header_owner(buf));
3478                 btrfs_init_tree_ref(&generic_ref, btrfs_header_level(buf),
3479                                     root_id, 0, false);
3480                 btrfs_ref_tree_mod(fs_info, &generic_ref);
3481                 ret = btrfs_add_delayed_tree_ref(trans, &generic_ref, NULL);
3482                 BUG_ON(ret); /* -ENOMEM */
3483         }
3484
3485         if (!last_ref)
3486                 return;
3487
3488         if (btrfs_header_generation(buf) != trans->transid)
3489                 goto out;
3490
3491         if (root_id != BTRFS_TREE_LOG_OBJECTID) {
3492                 ret = check_ref_cleanup(trans, buf->start);
3493                 if (!ret)
3494                         goto out;
3495         }
3496
3497         bg = btrfs_lookup_block_group(fs_info, buf->start);
3498
3499         if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
3500                 pin_down_extent(trans, bg, buf->start, buf->len, 1);
3501                 btrfs_put_block_group(bg);
3502                 goto out;
3503         }
3504
3505         /*
3506          * If there are tree mod log users we may have recorded mod log
3507          * operations for this node.  If we re-allocate this node we
3508          * could replay operations on this node that happened when it
3509          * existed in a completely different root.  For example if it
3510          * was part of root A, then was reallocated to root B, and we
3511          * are doing a btrfs_old_search_slot(root b), we could replay
3512          * operations that happened when the block was part of root A,
3513          * giving us an inconsistent view of the btree.
3514          *
3515          * We are safe from races here because at this point no other
3516          * node or root points to this extent buffer, so if after this
3517          * check a new tree mod log user joins we will not have an
3518          * existing log of operations on this node that we have to
3519          * contend with.
3520          */
3521
3522         if (test_bit(BTRFS_FS_TREE_MOD_LOG_USERS, &fs_info->flags)
3523                      || btrfs_is_zoned(fs_info)) {
3524                 pin_down_extent(trans, bg, buf->start, buf->len, 1);
3525                 btrfs_put_block_group(bg);
3526                 goto out;
3527         }
3528
3529         WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
3530
3531         btrfs_add_free_space(bg, buf->start, buf->len);
3532         btrfs_free_reserved_bytes(bg, buf->len, 0);
3533         btrfs_put_block_group(bg);
3534         trace_btrfs_reserved_extent_free(fs_info, buf->start, buf->len);
3535
3536 out:
3537
3538         /*
3539          * Deleting the buffer, clear the corrupt flag since it doesn't
3540          * matter anymore.
3541          */
3542         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
3543 }
3544
3545 /* Can return -ENOMEM */
3546 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_ref *ref)
3547 {
3548         struct btrfs_fs_info *fs_info = trans->fs_info;
3549         int ret;
3550
3551         if (btrfs_is_testing(fs_info))
3552                 return 0;
3553
3554         /*
3555          * tree log blocks never actually go into the extent allocation
3556          * tree, just update pinning info and exit early.
3557          */
3558         if ((ref->type == BTRFS_REF_METADATA &&
3559              ref->tree_ref.ref_root == BTRFS_TREE_LOG_OBJECTID) ||
3560             (ref->type == BTRFS_REF_DATA &&
3561              ref->data_ref.ref_root == BTRFS_TREE_LOG_OBJECTID)) {
3562                 btrfs_pin_extent(trans, ref->bytenr, ref->len, 1);
3563                 ret = 0;
3564         } else if (ref->type == BTRFS_REF_METADATA) {
3565                 ret = btrfs_add_delayed_tree_ref(trans, ref, NULL);
3566         } else {
3567                 ret = btrfs_add_delayed_data_ref(trans, ref, 0);
3568         }
3569
3570         if (!((ref->type == BTRFS_REF_METADATA &&
3571                ref->tree_ref.ref_root == BTRFS_TREE_LOG_OBJECTID) ||
3572               (ref->type == BTRFS_REF_DATA &&
3573                ref->data_ref.ref_root == BTRFS_TREE_LOG_OBJECTID)))
3574                 btrfs_ref_tree_mod(fs_info, ref);
3575
3576         return ret;
3577 }
3578
3579 enum btrfs_loop_type {
3580         /*
3581          * Start caching block groups but do not wait for progress or for them
3582          * to be done.
3583          */
3584         LOOP_CACHING_NOWAIT,
3585
3586         /*
3587          * Wait for the block group free_space >= the space we're waiting for if
3588          * the block group isn't cached.
3589          */
3590         LOOP_CACHING_WAIT,
3591
3592         /*
3593          * Allow allocations to happen from block groups that do not yet have a
3594          * size classification.
3595          */
3596         LOOP_UNSET_SIZE_CLASS,
3597
3598         /*
3599          * Allocate a chunk and then retry the allocation.
3600          */
3601         LOOP_ALLOC_CHUNK,
3602
3603         /*
3604          * Ignore the size class restrictions for this allocation.
3605          */
3606         LOOP_WRONG_SIZE_CLASS,
3607
3608         /*
3609          * Ignore the empty size, only try to allocate the number of bytes
3610          * needed for this allocation.
3611          */
3612         LOOP_NO_EMPTY_SIZE,
3613 };
3614
3615 static inline void
3616 btrfs_lock_block_group(struct btrfs_block_group *cache,
3617                        int delalloc)
3618 {
3619         if (delalloc)
3620                 down_read(&cache->data_rwsem);
3621 }
3622
3623 static inline void btrfs_grab_block_group(struct btrfs_block_group *cache,
3624                        int delalloc)
3625 {
3626         btrfs_get_block_group(cache);
3627         if (delalloc)
3628                 down_read(&cache->data_rwsem);
3629 }
3630
3631 static struct btrfs_block_group *btrfs_lock_cluster(
3632                    struct btrfs_block_group *block_group,
3633                    struct btrfs_free_cluster *cluster,
3634                    int delalloc)
3635         __acquires(&cluster->refill_lock)
3636 {
3637         struct btrfs_block_group *used_bg = NULL;
3638
3639         spin_lock(&cluster->refill_lock);
3640         while (1) {
3641                 used_bg = cluster->block_group;
3642                 if (!used_bg)
3643                         return NULL;
3644
3645                 if (used_bg == block_group)
3646                         return used_bg;
3647
3648                 btrfs_get_block_group(used_bg);
3649
3650                 if (!delalloc)
3651                         return used_bg;
3652
3653                 if (down_read_trylock(&used_bg->data_rwsem))
3654                         return used_bg;
3655
3656                 spin_unlock(&cluster->refill_lock);
3657
3658                 /* We should only have one-level nested. */
3659                 down_read_nested(&used_bg->data_rwsem, SINGLE_DEPTH_NESTING);
3660
3661                 spin_lock(&cluster->refill_lock);
3662                 if (used_bg == cluster->block_group)
3663                         return used_bg;
3664
3665                 up_read(&used_bg->data_rwsem);
3666                 btrfs_put_block_group(used_bg);
3667         }
3668 }
3669
3670 static inline void
3671 btrfs_release_block_group(struct btrfs_block_group *cache,
3672                          int delalloc)
3673 {
3674         if (delalloc)
3675                 up_read(&cache->data_rwsem);
3676         btrfs_put_block_group(cache);
3677 }
3678
3679 /*
3680  * Helper function for find_free_extent().
3681  *
3682  * Return -ENOENT to inform caller that we need fallback to unclustered mode.
3683  * Return >0 to inform caller that we find nothing
3684  * Return 0 means we have found a location and set ffe_ctl->found_offset.
3685  */
3686 static int find_free_extent_clustered(struct btrfs_block_group *bg,
3687                                       struct find_free_extent_ctl *ffe_ctl,
3688                                       struct btrfs_block_group **cluster_bg_ret)
3689 {
3690         struct btrfs_block_group *cluster_bg;
3691         struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
3692         u64 aligned_cluster;
3693         u64 offset;
3694         int ret;
3695
3696         cluster_bg = btrfs_lock_cluster(bg, last_ptr, ffe_ctl->delalloc);
3697         if (!cluster_bg)
3698                 goto refill_cluster;
3699         if (cluster_bg != bg && (cluster_bg->ro ||
3700             !block_group_bits(cluster_bg, ffe_ctl->flags)))
3701                 goto release_cluster;
3702
3703         offset = btrfs_alloc_from_cluster(cluster_bg, last_ptr,
3704                         ffe_ctl->num_bytes, cluster_bg->start,
3705                         &ffe_ctl->max_extent_size);
3706         if (offset) {
3707                 /* We have a block, we're done */
3708                 spin_unlock(&last_ptr->refill_lock);
3709                 trace_btrfs_reserve_extent_cluster(cluster_bg, ffe_ctl);
3710                 *cluster_bg_ret = cluster_bg;
3711                 ffe_ctl->found_offset = offset;
3712                 return 0;
3713         }
3714         WARN_ON(last_ptr->block_group != cluster_bg);
3715
3716 release_cluster:
3717         /*
3718          * If we are on LOOP_NO_EMPTY_SIZE, we can't set up a new clusters, so
3719          * lets just skip it and let the allocator find whatever block it can
3720          * find. If we reach this point, we will have tried the cluster
3721          * allocator plenty of times and not have found anything, so we are
3722          * likely way too fragmented for the clustering stuff to find anything.
3723          *
3724          * However, if the cluster is taken from the current block group,
3725          * release the cluster first, so that we stand a better chance of
3726          * succeeding in the unclustered allocation.
3727          */
3728         if (ffe_ctl->loop >= LOOP_NO_EMPTY_SIZE && cluster_bg != bg) {
3729                 spin_unlock(&last_ptr->refill_lock);
3730                 btrfs_release_block_group(cluster_bg, ffe_ctl->delalloc);
3731                 return -ENOENT;
3732         }
3733
3734         /* This cluster didn't work out, free it and start over */
3735         btrfs_return_cluster_to_free_space(NULL, last_ptr);
3736
3737         if (cluster_bg != bg)
3738                 btrfs_release_block_group(cluster_bg, ffe_ctl->delalloc);
3739
3740 refill_cluster:
3741         if (ffe_ctl->loop >= LOOP_NO_EMPTY_SIZE) {
3742                 spin_unlock(&last_ptr->refill_lock);
3743                 return -ENOENT;
3744         }
3745
3746         aligned_cluster = max_t(u64,
3747                         ffe_ctl->empty_cluster + ffe_ctl->empty_size,
3748                         bg->full_stripe_len);
3749         ret = btrfs_find_space_cluster(bg, last_ptr, ffe_ctl->search_start,
3750                         ffe_ctl->num_bytes, aligned_cluster);
3751         if (ret == 0) {
3752                 /* Now pull our allocation out of this cluster */
3753                 offset = btrfs_alloc_from_cluster(bg, last_ptr,
3754                                 ffe_ctl->num_bytes, ffe_ctl->search_start,
3755                                 &ffe_ctl->max_extent_size);
3756                 if (offset) {
3757                         /* We found one, proceed */
3758                         spin_unlock(&last_ptr->refill_lock);
3759                         ffe_ctl->found_offset = offset;
3760                         trace_btrfs_reserve_extent_cluster(bg, ffe_ctl);
3761                         return 0;
3762                 }
3763         }
3764         /*
3765          * At this point we either didn't find a cluster or we weren't able to
3766          * allocate a block from our cluster.  Free the cluster we've been
3767          * trying to use, and go to the next block group.
3768          */
3769         btrfs_return_cluster_to_free_space(NULL, last_ptr);
3770         spin_unlock(&last_ptr->refill_lock);
3771         return 1;
3772 }
3773
3774 /*
3775  * Return >0 to inform caller that we find nothing
3776  * Return 0 when we found an free extent and set ffe_ctrl->found_offset
3777  */
3778 static int find_free_extent_unclustered(struct btrfs_block_group *bg,
3779                                         struct find_free_extent_ctl *ffe_ctl)
3780 {
3781         struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
3782         u64 offset;
3783
3784         /*
3785          * We are doing an unclustered allocation, set the fragmented flag so
3786          * we don't bother trying to setup a cluster again until we get more
3787          * space.
3788          */
3789         if (unlikely(last_ptr)) {
3790                 spin_lock(&last_ptr->lock);
3791                 last_ptr->fragmented = 1;
3792                 spin_unlock(&last_ptr->lock);
3793         }
3794         if (ffe_ctl->cached) {
3795                 struct btrfs_free_space_ctl *free_space_ctl;
3796
3797                 free_space_ctl = bg->free_space_ctl;
3798                 spin_lock(&free_space_ctl->tree_lock);
3799                 if (free_space_ctl->free_space <
3800                     ffe_ctl->num_bytes + ffe_ctl->empty_cluster +
3801                     ffe_ctl->empty_size) {
3802                         ffe_ctl->total_free_space = max_t(u64,
3803                                         ffe_ctl->total_free_space,
3804                                         free_space_ctl->free_space);
3805                         spin_unlock(&free_space_ctl->tree_lock);
3806                         return 1;
3807                 }
3808                 spin_unlock(&free_space_ctl->tree_lock);
3809         }
3810
3811         offset = btrfs_find_space_for_alloc(bg, ffe_ctl->search_start,
3812                         ffe_ctl->num_bytes, ffe_ctl->empty_size,
3813                         &ffe_ctl->max_extent_size);
3814         if (!offset)
3815                 return 1;
3816         ffe_ctl->found_offset = offset;
3817         return 0;
3818 }
3819
3820 static int do_allocation_clustered(struct btrfs_block_group *block_group,
3821                                    struct find_free_extent_ctl *ffe_ctl,
3822                                    struct btrfs_block_group **bg_ret)
3823 {
3824         int ret;
3825
3826         /* We want to try and use the cluster allocator, so lets look there */
3827         if (ffe_ctl->last_ptr && ffe_ctl->use_cluster) {
3828                 ret = find_free_extent_clustered(block_group, ffe_ctl, bg_ret);
3829                 if (ret >= 0)
3830                         return ret;
3831                 /* ret == -ENOENT case falls through */
3832         }
3833
3834         return find_free_extent_unclustered(block_group, ffe_ctl);
3835 }
3836
3837 /*
3838  * Tree-log block group locking
3839  * ============================
3840  *
3841  * fs_info::treelog_bg_lock protects the fs_info::treelog_bg which
3842  * indicates the starting address of a block group, which is reserved only
3843  * for tree-log metadata.
3844  *
3845  * Lock nesting
3846  * ============
3847  *
3848  * space_info::lock
3849  *   block_group::lock
3850  *     fs_info::treelog_bg_lock
3851  */
3852
3853 /*
3854  * Simple allocator for sequential-only block group. It only allows sequential
3855  * allocation. No need to play with trees. This function also reserves the
3856  * bytes as in btrfs_add_reserved_bytes.
3857  */
3858 static int do_allocation_zoned(struct btrfs_block_group *block_group,
3859                                struct find_free_extent_ctl *ffe_ctl,
3860                                struct btrfs_block_group **bg_ret)
3861 {
3862         struct btrfs_fs_info *fs_info = block_group->fs_info;
3863         struct btrfs_space_info *space_info = block_group->space_info;
3864         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3865         u64 start = block_group->start;
3866         u64 num_bytes = ffe_ctl->num_bytes;
3867         u64 avail;
3868         u64 bytenr = block_group->start;
3869         u64 log_bytenr;
3870         u64 data_reloc_bytenr;
3871         int ret = 0;
3872         bool skip = false;
3873
3874         ASSERT(btrfs_is_zoned(block_group->fs_info));
3875
3876         /*
3877          * Do not allow non-tree-log blocks in the dedicated tree-log block
3878          * group, and vice versa.
3879          */
3880         spin_lock(&fs_info->treelog_bg_lock);
3881         log_bytenr = fs_info->treelog_bg;
3882         if (log_bytenr && ((ffe_ctl->for_treelog && bytenr != log_bytenr) ||
3883                            (!ffe_ctl->for_treelog && bytenr == log_bytenr)))
3884                 skip = true;
3885         spin_unlock(&fs_info->treelog_bg_lock);
3886         if (skip)
3887                 return 1;
3888
3889         /*
3890          * Do not allow non-relocation blocks in the dedicated relocation block
3891          * group, and vice versa.
3892          */
3893         spin_lock(&fs_info->relocation_bg_lock);
3894         data_reloc_bytenr = fs_info->data_reloc_bg;
3895         if (data_reloc_bytenr &&
3896             ((ffe_ctl->for_data_reloc && bytenr != data_reloc_bytenr) ||
3897              (!ffe_ctl->for_data_reloc && bytenr == data_reloc_bytenr)))
3898                 skip = true;
3899         spin_unlock(&fs_info->relocation_bg_lock);
3900         if (skip)
3901                 return 1;
3902
3903         /* Check RO and no space case before trying to activate it */
3904         spin_lock(&block_group->lock);
3905         if (block_group->ro || btrfs_zoned_bg_is_full(block_group)) {
3906                 ret = 1;
3907                 /*
3908                  * May need to clear fs_info->{treelog,data_reloc}_bg.
3909                  * Return the error after taking the locks.
3910                  */
3911         }
3912         spin_unlock(&block_group->lock);
3913
3914         /* Metadata block group is activated at write time. */
3915         if (!ret && (block_group->flags & BTRFS_BLOCK_GROUP_DATA) &&
3916             !btrfs_zone_activate(block_group)) {
3917                 ret = 1;
3918                 /*
3919                  * May need to clear fs_info->{treelog,data_reloc}_bg.
3920                  * Return the error after taking the locks.
3921                  */
3922         }
3923
3924         spin_lock(&space_info->lock);
3925         spin_lock(&block_group->lock);
3926         spin_lock(&fs_info->treelog_bg_lock);
3927         spin_lock(&fs_info->relocation_bg_lock);
3928
3929         if (ret)
3930                 goto out;
3931
3932         ASSERT(!ffe_ctl->for_treelog ||
3933                block_group->start == fs_info->treelog_bg ||
3934                fs_info->treelog_bg == 0);
3935         ASSERT(!ffe_ctl->for_data_reloc ||
3936                block_group->start == fs_info->data_reloc_bg ||
3937                fs_info->data_reloc_bg == 0);
3938
3939         if (block_group->ro ||
3940             (!ffe_ctl->for_data_reloc &&
3941              test_bit(BLOCK_GROUP_FLAG_ZONED_DATA_RELOC, &block_group->runtime_flags))) {
3942                 ret = 1;
3943                 goto out;
3944         }
3945
3946         /*
3947          * Do not allow currently using block group to be tree-log dedicated
3948          * block group.
3949          */
3950         if (ffe_ctl->for_treelog && !fs_info->treelog_bg &&
3951             (block_group->used || block_group->reserved)) {
3952                 ret = 1;
3953                 goto out;
3954         }
3955
3956         /*
3957          * Do not allow currently used block group to be the data relocation
3958          * dedicated block group.
3959          */
3960         if (ffe_ctl->for_data_reloc && !fs_info->data_reloc_bg &&
3961             (block_group->used || block_group->reserved)) {
3962                 ret = 1;
3963                 goto out;
3964         }
3965
3966         WARN_ON_ONCE(block_group->alloc_offset > block_group->zone_capacity);
3967         avail = block_group->zone_capacity - block_group->alloc_offset;
3968         if (avail < num_bytes) {
3969                 if (ffe_ctl->max_extent_size < avail) {
3970                         /*
3971                          * With sequential allocator, free space is always
3972                          * contiguous
3973                          */
3974                         ffe_ctl->max_extent_size = avail;
3975                         ffe_ctl->total_free_space = avail;
3976                 }
3977                 ret = 1;
3978                 goto out;
3979         }
3980
3981         if (ffe_ctl->for_treelog && !fs_info->treelog_bg)
3982                 fs_info->treelog_bg = block_group->start;
3983
3984         if (ffe_ctl->for_data_reloc) {
3985                 if (!fs_info->data_reloc_bg)
3986                         fs_info->data_reloc_bg = block_group->start;
3987                 /*
3988                  * Do not allow allocations from this block group, unless it is
3989                  * for data relocation. Compared to increasing the ->ro, setting
3990                  * the ->zoned_data_reloc_ongoing flag still allows nocow
3991                  * writers to come in. See btrfs_inc_nocow_writers().
3992                  *
3993                  * We need to disable an allocation to avoid an allocation of
3994                  * regular (non-relocation data) extent. With mix of relocation
3995                  * extents and regular extents, we can dispatch WRITE commands
3996                  * (for relocation extents) and ZONE APPEND commands (for
3997                  * regular extents) at the same time to the same zone, which
3998                  * easily break the write pointer.
3999                  *
4000                  * Also, this flag avoids this block group to be zone finished.
4001                  */
4002                 set_bit(BLOCK_GROUP_FLAG_ZONED_DATA_RELOC, &block_group->runtime_flags);
4003         }
4004
4005         ffe_ctl->found_offset = start + block_group->alloc_offset;
4006         block_group->alloc_offset += num_bytes;
4007         spin_lock(&ctl->tree_lock);
4008         ctl->free_space -= num_bytes;
4009         spin_unlock(&ctl->tree_lock);
4010
4011         /*
4012          * We do not check if found_offset is aligned to stripesize. The
4013          * address is anyway rewritten when using zone append writing.
4014          */
4015
4016         ffe_ctl->search_start = ffe_ctl->found_offset;
4017
4018 out:
4019         if (ret && ffe_ctl->for_treelog)
4020                 fs_info->treelog_bg = 0;
4021         if (ret && ffe_ctl->for_data_reloc)
4022                 fs_info->data_reloc_bg = 0;
4023         spin_unlock(&fs_info->relocation_bg_lock);
4024         spin_unlock(&fs_info->treelog_bg_lock);
4025         spin_unlock(&block_group->lock);
4026         spin_unlock(&space_info->lock);
4027         return ret;
4028 }
4029
4030 static int do_allocation(struct btrfs_block_group *block_group,
4031                          struct find_free_extent_ctl *ffe_ctl,
4032                          struct btrfs_block_group **bg_ret)
4033 {
4034         switch (ffe_ctl->policy) {
4035         case BTRFS_EXTENT_ALLOC_CLUSTERED:
4036                 return do_allocation_clustered(block_group, ffe_ctl, bg_ret);
4037         case BTRFS_EXTENT_ALLOC_ZONED:
4038                 return do_allocation_zoned(block_group, ffe_ctl, bg_ret);
4039         default:
4040                 BUG();
4041         }
4042 }
4043
4044 static void release_block_group(struct btrfs_block_group *block_group,
4045                                 struct find_free_extent_ctl *ffe_ctl,
4046                                 int delalloc)
4047 {
4048         switch (ffe_ctl->policy) {
4049         case BTRFS_EXTENT_ALLOC_CLUSTERED:
4050                 ffe_ctl->retry_uncached = false;
4051                 break;
4052         case BTRFS_EXTENT_ALLOC_ZONED:
4053                 /* Nothing to do */
4054                 break;
4055         default:
4056                 BUG();
4057         }
4058
4059         BUG_ON(btrfs_bg_flags_to_raid_index(block_group->flags) !=
4060                ffe_ctl->index);
4061         btrfs_release_block_group(block_group, delalloc);
4062 }
4063
4064 static void found_extent_clustered(struct find_free_extent_ctl *ffe_ctl,
4065                                    struct btrfs_key *ins)
4066 {
4067         struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
4068
4069         if (!ffe_ctl->use_cluster && last_ptr) {
4070                 spin_lock(&last_ptr->lock);
4071                 last_ptr->window_start = ins->objectid;
4072                 spin_unlock(&last_ptr->lock);
4073         }
4074 }
4075
4076 static void found_extent(struct find_free_extent_ctl *ffe_ctl,
4077                          struct btrfs_key *ins)
4078 {
4079         switch (ffe_ctl->policy) {
4080         case BTRFS_EXTENT_ALLOC_CLUSTERED:
4081                 found_extent_clustered(ffe_ctl, ins);
4082                 break;
4083         case BTRFS_EXTENT_ALLOC_ZONED:
4084                 /* Nothing to do */
4085                 break;
4086         default:
4087                 BUG();
4088         }
4089 }
4090
4091 static int can_allocate_chunk_zoned(struct btrfs_fs_info *fs_info,
4092                                     struct find_free_extent_ctl *ffe_ctl)
4093 {
4094         /* Block group's activeness is not a requirement for METADATA block groups. */
4095         if (!(ffe_ctl->flags & BTRFS_BLOCK_GROUP_DATA))
4096                 return 0;
4097
4098         /* If we can activate new zone, just allocate a chunk and use it */
4099         if (btrfs_can_activate_zone(fs_info->fs_devices, ffe_ctl->flags))
4100                 return 0;
4101
4102         /*
4103          * We already reached the max active zones. Try to finish one block
4104          * group to make a room for a new block group. This is only possible
4105          * for a data block group because btrfs_zone_finish() may need to wait
4106          * for a running transaction which can cause a deadlock for metadata
4107          * allocation.
4108          */
4109         if (ffe_ctl->flags & BTRFS_BLOCK_GROUP_DATA) {
4110                 int ret = btrfs_zone_finish_one_bg(fs_info);
4111
4112                 if (ret == 1)
4113                         return 0;
4114                 else if (ret < 0)
4115                         return ret;
4116         }
4117
4118         /*
4119          * If we have enough free space left in an already active block group
4120          * and we can't activate any other zone now, do not allow allocating a
4121          * new chunk and let find_free_extent() retry with a smaller size.
4122          */
4123         if (ffe_ctl->max_extent_size >= ffe_ctl->min_alloc_size)
4124                 return -ENOSPC;
4125
4126         /*
4127          * Even min_alloc_size is not left in any block groups. Since we cannot
4128          * activate a new block group, allocating it may not help. Let's tell a
4129          * caller to try again and hope it progress something by writing some
4130          * parts of the region. That is only possible for data block groups,
4131          * where a part of the region can be written.
4132          */
4133         if (ffe_ctl->flags & BTRFS_BLOCK_GROUP_DATA)
4134                 return -EAGAIN;
4135
4136         /*
4137          * We cannot activate a new block group and no enough space left in any
4138          * block groups. So, allocating a new block group may not help. But,
4139          * there is nothing to do anyway, so let's go with it.
4140          */
4141         return 0;
4142 }
4143
4144 static int can_allocate_chunk(struct btrfs_fs_info *fs_info,
4145                               struct find_free_extent_ctl *ffe_ctl)
4146 {
4147         switch (ffe_ctl->policy) {
4148         case BTRFS_EXTENT_ALLOC_CLUSTERED:
4149                 return 0;
4150         case BTRFS_EXTENT_ALLOC_ZONED:
4151                 return can_allocate_chunk_zoned(fs_info, ffe_ctl);
4152         default:
4153                 BUG();
4154         }
4155 }
4156
4157 /*
4158  * Return >0 means caller needs to re-search for free extent
4159  * Return 0 means we have the needed free extent.
4160  * Return <0 means we failed to locate any free extent.
4161  */
4162 static int find_free_extent_update_loop(struct btrfs_fs_info *fs_info,
4163                                         struct btrfs_key *ins,
4164                                         struct find_free_extent_ctl *ffe_ctl,
4165                                         bool full_search)
4166 {
4167         struct btrfs_root *root = fs_info->chunk_root;
4168         int ret;
4169
4170         if ((ffe_ctl->loop == LOOP_CACHING_NOWAIT) &&
4171             ffe_ctl->have_caching_bg && !ffe_ctl->orig_have_caching_bg)
4172                 ffe_ctl->orig_have_caching_bg = true;
4173
4174         if (ins->objectid) {
4175                 found_extent(ffe_ctl, ins);
4176                 return 0;
4177         }
4178
4179         if (ffe_ctl->loop >= LOOP_CACHING_WAIT && ffe_ctl->have_caching_bg)
4180                 return 1;
4181
4182         ffe_ctl->index++;
4183         if (ffe_ctl->index < BTRFS_NR_RAID_TYPES)
4184                 return 1;
4185
4186         /* See the comments for btrfs_loop_type for an explanation of the phases. */
4187         if (ffe_ctl->loop < LOOP_NO_EMPTY_SIZE) {
4188                 ffe_ctl->index = 0;
4189                 /*
4190                  * We want to skip the LOOP_CACHING_WAIT step if we don't have
4191                  * any uncached bgs and we've already done a full search
4192                  * through.
4193                  */
4194                 if (ffe_ctl->loop == LOOP_CACHING_NOWAIT &&
4195                     (!ffe_ctl->orig_have_caching_bg && full_search))
4196                         ffe_ctl->loop++;
4197                 ffe_ctl->loop++;
4198
4199                 if (ffe_ctl->loop == LOOP_ALLOC_CHUNK) {
4200                         struct btrfs_trans_handle *trans;
4201                         int exist = 0;
4202
4203                         /* Check if allocation policy allows to create a new chunk */
4204                         ret = can_allocate_chunk(fs_info, ffe_ctl);
4205                         if (ret)
4206                                 return ret;
4207
4208                         trans = current->journal_info;
4209                         if (trans)
4210                                 exist = 1;
4211                         else
4212                                 trans = btrfs_join_transaction(root);
4213
4214                         if (IS_ERR(trans)) {
4215                                 ret = PTR_ERR(trans);
4216                                 return ret;
4217                         }
4218
4219                         ret = btrfs_chunk_alloc(trans, ffe_ctl->flags,
4220                                                 CHUNK_ALLOC_FORCE_FOR_EXTENT);
4221
4222                         /* Do not bail out on ENOSPC since we can do more. */
4223                         if (ret == -ENOSPC) {
4224                                 ret = 0;
4225                                 ffe_ctl->loop++;
4226                         }
4227                         else if (ret < 0)
4228                                 btrfs_abort_transaction(trans, ret);
4229                         else
4230                                 ret = 0;
4231                         if (!exist)
4232                                 btrfs_end_transaction(trans);
4233                         if (ret)
4234                                 return ret;
4235                 }
4236
4237                 if (ffe_ctl->loop == LOOP_NO_EMPTY_SIZE) {
4238                         if (ffe_ctl->policy != BTRFS_EXTENT_ALLOC_CLUSTERED)
4239                                 return -ENOSPC;
4240
4241                         /*
4242                          * Don't loop again if we already have no empty_size and
4243                          * no empty_cluster.
4244                          */
4245                         if (ffe_ctl->empty_size == 0 &&
4246                             ffe_ctl->empty_cluster == 0)
4247                                 return -ENOSPC;
4248                         ffe_ctl->empty_size = 0;
4249                         ffe_ctl->empty_cluster = 0;
4250                 }
4251                 return 1;
4252         }
4253         return -ENOSPC;
4254 }
4255
4256 static bool find_free_extent_check_size_class(struct find_free_extent_ctl *ffe_ctl,
4257                                               struct btrfs_block_group *bg)
4258 {
4259         if (ffe_ctl->policy == BTRFS_EXTENT_ALLOC_ZONED)
4260                 return true;
4261         if (!btrfs_block_group_should_use_size_class(bg))
4262                 return true;
4263         if (ffe_ctl->loop >= LOOP_WRONG_SIZE_CLASS)
4264                 return true;
4265         if (ffe_ctl->loop >= LOOP_UNSET_SIZE_CLASS &&
4266             bg->size_class == BTRFS_BG_SZ_NONE)
4267                 return true;
4268         return ffe_ctl->size_class == bg->size_class;
4269 }
4270
4271 static int prepare_allocation_clustered(struct btrfs_fs_info *fs_info,
4272                                         struct find_free_extent_ctl *ffe_ctl,
4273                                         struct btrfs_space_info *space_info,
4274                                         struct btrfs_key *ins)
4275 {
4276         /*
4277          * If our free space is heavily fragmented we may not be able to make
4278          * big contiguous allocations, so instead of doing the expensive search
4279          * for free space, simply return ENOSPC with our max_extent_size so we
4280          * can go ahead and search for a more manageable chunk.
4281          *
4282          * If our max_extent_size is large enough for our allocation simply
4283          * disable clustering since we will likely not be able to find enough
4284          * space to create a cluster and induce latency trying.
4285          */
4286         if (space_info->max_extent_size) {
4287                 spin_lock(&space_info->lock);
4288                 if (space_info->max_extent_size &&
4289                     ffe_ctl->num_bytes > space_info->max_extent_size) {
4290                         ins->offset = space_info->max_extent_size;
4291                         spin_unlock(&space_info->lock);
4292                         return -ENOSPC;
4293                 } else if (space_info->max_extent_size) {
4294                         ffe_ctl->use_cluster = false;
4295                 }
4296                 spin_unlock(&space_info->lock);
4297         }
4298
4299         ffe_ctl->last_ptr = fetch_cluster_info(fs_info, space_info,
4300                                                &ffe_ctl->empty_cluster);
4301         if (ffe_ctl->last_ptr) {
4302                 struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
4303
4304                 spin_lock(&last_ptr->lock);
4305                 if (last_ptr->block_group)
4306                         ffe_ctl->hint_byte = last_ptr->window_start;
4307                 if (last_ptr->fragmented) {
4308                         /*
4309                          * We still set window_start so we can keep track of the
4310                          * last place we found an allocation to try and save
4311                          * some time.
4312                          */
4313                         ffe_ctl->hint_byte = last_ptr->window_start;
4314                         ffe_ctl->use_cluster = false;
4315                 }
4316                 spin_unlock(&last_ptr->lock);
4317         }
4318
4319         return 0;
4320 }
4321
4322 static int prepare_allocation_zoned(struct btrfs_fs_info *fs_info,
4323                                     struct find_free_extent_ctl *ffe_ctl)
4324 {
4325         if (ffe_ctl->for_treelog) {
4326                 spin_lock(&fs_info->treelog_bg_lock);
4327                 if (fs_info->treelog_bg)
4328                         ffe_ctl->hint_byte = fs_info->treelog_bg;
4329                 spin_unlock(&fs_info->treelog_bg_lock);
4330         } else if (ffe_ctl->for_data_reloc) {
4331                 spin_lock(&fs_info->relocation_bg_lock);
4332                 if (fs_info->data_reloc_bg)
4333                         ffe_ctl->hint_byte = fs_info->data_reloc_bg;
4334                 spin_unlock(&fs_info->relocation_bg_lock);
4335         } else if (ffe_ctl->flags & BTRFS_BLOCK_GROUP_DATA) {
4336                 struct btrfs_block_group *block_group;
4337
4338                 spin_lock(&fs_info->zone_active_bgs_lock);
4339                 list_for_each_entry(block_group, &fs_info->zone_active_bgs, active_bg_list) {
4340                         /*
4341                          * No lock is OK here because avail is monotinically
4342                          * decreasing, and this is just a hint.
4343                          */
4344                         u64 avail = block_group->zone_capacity - block_group->alloc_offset;
4345
4346                         if (block_group_bits(block_group, ffe_ctl->flags) &&
4347                             avail >= ffe_ctl->num_bytes) {
4348                                 ffe_ctl->hint_byte = block_group->start;
4349                                 break;
4350                         }
4351                 }
4352                 spin_unlock(&fs_info->zone_active_bgs_lock);
4353         }
4354
4355         return 0;
4356 }
4357
4358 static int prepare_allocation(struct btrfs_fs_info *fs_info,
4359                               struct find_free_extent_ctl *ffe_ctl,
4360                               struct btrfs_space_info *space_info,
4361                               struct btrfs_key *ins)
4362 {
4363         switch (ffe_ctl->policy) {
4364         case BTRFS_EXTENT_ALLOC_CLUSTERED:
4365                 return prepare_allocation_clustered(fs_info, ffe_ctl,
4366                                                     space_info, ins);
4367         case BTRFS_EXTENT_ALLOC_ZONED:
4368                 return prepare_allocation_zoned(fs_info, ffe_ctl);
4369         default:
4370                 BUG();
4371         }
4372 }
4373
4374 /*
4375  * walks the btree of allocated extents and find a hole of a given size.
4376  * The key ins is changed to record the hole:
4377  * ins->objectid == start position
4378  * ins->flags = BTRFS_EXTENT_ITEM_KEY
4379  * ins->offset == the size of the hole.
4380  * Any available blocks before search_start are skipped.
4381  *
4382  * If there is no suitable free space, we will record the max size of
4383  * the free space extent currently.
4384  *
4385  * The overall logic and call chain:
4386  *
4387  * find_free_extent()
4388  * |- Iterate through all block groups
4389  * |  |- Get a valid block group
4390  * |  |- Try to do clustered allocation in that block group
4391  * |  |- Try to do unclustered allocation in that block group
4392  * |  |- Check if the result is valid
4393  * |  |  |- If valid, then exit
4394  * |  |- Jump to next block group
4395  * |
4396  * |- Push harder to find free extents
4397  *    |- If not found, re-iterate all block groups
4398  */
4399 static noinline int find_free_extent(struct btrfs_root *root,
4400                                      struct btrfs_key *ins,
4401                                      struct find_free_extent_ctl *ffe_ctl)
4402 {
4403         struct btrfs_fs_info *fs_info = root->fs_info;
4404         int ret = 0;
4405         int cache_block_group_error = 0;
4406         struct btrfs_block_group *block_group = NULL;
4407         struct btrfs_space_info *space_info;
4408         bool full_search = false;
4409
4410         WARN_ON(ffe_ctl->num_bytes < fs_info->sectorsize);
4411
4412         ffe_ctl->search_start = 0;
4413         /* For clustered allocation */
4414         ffe_ctl->empty_cluster = 0;
4415         ffe_ctl->last_ptr = NULL;
4416         ffe_ctl->use_cluster = true;
4417         ffe_ctl->have_caching_bg = false;
4418         ffe_ctl->orig_have_caching_bg = false;
4419         ffe_ctl->index = btrfs_bg_flags_to_raid_index(ffe_ctl->flags);
4420         ffe_ctl->loop = 0;
4421         ffe_ctl->retry_uncached = false;
4422         ffe_ctl->cached = 0;
4423         ffe_ctl->max_extent_size = 0;
4424         ffe_ctl->total_free_space = 0;
4425         ffe_ctl->found_offset = 0;
4426         ffe_ctl->policy = BTRFS_EXTENT_ALLOC_CLUSTERED;
4427         ffe_ctl->size_class = btrfs_calc_block_group_size_class(ffe_ctl->num_bytes);
4428
4429         if (btrfs_is_zoned(fs_info))
4430                 ffe_ctl->policy = BTRFS_EXTENT_ALLOC_ZONED;
4431
4432         ins->type = BTRFS_EXTENT_ITEM_KEY;
4433         ins->objectid = 0;
4434         ins->offset = 0;
4435
4436         trace_find_free_extent(root, ffe_ctl);
4437
4438         space_info = btrfs_find_space_info(fs_info, ffe_ctl->flags);
4439         if (!space_info) {
4440                 btrfs_err(fs_info, "No space info for %llu", ffe_ctl->flags);
4441                 return -ENOSPC;
4442         }
4443
4444         ret = prepare_allocation(fs_info, ffe_ctl, space_info, ins);
4445         if (ret < 0)
4446                 return ret;
4447
4448         ffe_ctl->search_start = max(ffe_ctl->search_start,
4449                                     first_logical_byte(fs_info));
4450         ffe_ctl->search_start = max(ffe_ctl->search_start, ffe_ctl->hint_byte);
4451         if (ffe_ctl->search_start == ffe_ctl->hint_byte) {
4452                 block_group = btrfs_lookup_block_group(fs_info,
4453                                                        ffe_ctl->search_start);
4454                 /*
4455                  * we don't want to use the block group if it doesn't match our
4456                  * allocation bits, or if its not cached.
4457                  *
4458                  * However if we are re-searching with an ideal block group
4459                  * picked out then we don't care that the block group is cached.
4460                  */
4461                 if (block_group && block_group_bits(block_group, ffe_ctl->flags) &&
4462                     block_group->cached != BTRFS_CACHE_NO) {
4463                         down_read(&space_info->groups_sem);
4464                         if (list_empty(&block_group->list) ||
4465                             block_group->ro) {
4466                                 /*
4467                                  * someone is removing this block group,
4468                                  * we can't jump into the have_block_group
4469                                  * target because our list pointers are not
4470                                  * valid
4471                                  */
4472                                 btrfs_put_block_group(block_group);
4473                                 up_read(&space_info->groups_sem);
4474                         } else {
4475                                 ffe_ctl->index = btrfs_bg_flags_to_raid_index(
4476                                                         block_group->flags);
4477                                 btrfs_lock_block_group(block_group,
4478                                                        ffe_ctl->delalloc);
4479                                 ffe_ctl->hinted = true;
4480                                 goto have_block_group;
4481                         }
4482                 } else if (block_group) {
4483                         btrfs_put_block_group(block_group);
4484                 }
4485         }
4486 search:
4487         trace_find_free_extent_search_loop(root, ffe_ctl);
4488         ffe_ctl->have_caching_bg = false;
4489         if (ffe_ctl->index == btrfs_bg_flags_to_raid_index(ffe_ctl->flags) ||
4490             ffe_ctl->index == 0)
4491                 full_search = true;
4492         down_read(&space_info->groups_sem);
4493         list_for_each_entry(block_group,
4494                             &space_info->block_groups[ffe_ctl->index], list) {
4495                 struct btrfs_block_group *bg_ret;
4496
4497                 ffe_ctl->hinted = false;
4498                 /* If the block group is read-only, we can skip it entirely. */
4499                 if (unlikely(block_group->ro)) {
4500                         if (ffe_ctl->for_treelog)
4501                                 btrfs_clear_treelog_bg(block_group);
4502                         if (ffe_ctl->for_data_reloc)
4503                                 btrfs_clear_data_reloc_bg(block_group);
4504                         continue;
4505                 }
4506
4507                 btrfs_grab_block_group(block_group, ffe_ctl->delalloc);
4508                 ffe_ctl->search_start = block_group->start;
4509
4510                 /*
4511                  * this can happen if we end up cycling through all the
4512                  * raid types, but we want to make sure we only allocate
4513                  * for the proper type.
4514                  */
4515                 if (!block_group_bits(block_group, ffe_ctl->flags)) {
4516                         u64 extra = BTRFS_BLOCK_GROUP_DUP |
4517                                 BTRFS_BLOCK_GROUP_RAID1_MASK |
4518                                 BTRFS_BLOCK_GROUP_RAID56_MASK |
4519                                 BTRFS_BLOCK_GROUP_RAID10;
4520
4521                         /*
4522                          * if they asked for extra copies and this block group
4523                          * doesn't provide them, bail.  This does allow us to
4524                          * fill raid0 from raid1.
4525                          */
4526                         if ((ffe_ctl->flags & extra) && !(block_group->flags & extra))
4527                                 goto loop;
4528
4529                         /*
4530                          * This block group has different flags than we want.
4531                          * It's possible that we have MIXED_GROUP flag but no
4532                          * block group is mixed.  Just skip such block group.
4533                          */
4534                         btrfs_release_block_group(block_group, ffe_ctl->delalloc);
4535                         continue;
4536                 }
4537
4538 have_block_group:
4539                 trace_find_free_extent_have_block_group(root, ffe_ctl, block_group);
4540                 ffe_ctl->cached = btrfs_block_group_done(block_group);
4541                 if (unlikely(!ffe_ctl->cached)) {
4542                         ffe_ctl->have_caching_bg = true;
4543                         ret = btrfs_cache_block_group(block_group, false);
4544
4545                         /*
4546                          * If we get ENOMEM here or something else we want to
4547                          * try other block groups, because it may not be fatal.
4548                          * However if we can't find anything else we need to
4549                          * save our return here so that we return the actual
4550                          * error that caused problems, not ENOSPC.
4551                          */
4552                         if (ret < 0) {
4553                                 if (!cache_block_group_error)
4554                                         cache_block_group_error = ret;
4555                                 ret = 0;
4556                                 goto loop;
4557                         }
4558                         ret = 0;
4559                 }
4560
4561                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR)) {
4562                         if (!cache_block_group_error)
4563                                 cache_block_group_error = -EIO;
4564                         goto loop;
4565                 }
4566
4567                 if (!find_free_extent_check_size_class(ffe_ctl, block_group))
4568                         goto loop;
4569
4570                 bg_ret = NULL;
4571                 ret = do_allocation(block_group, ffe_ctl, &bg_ret);
4572                 if (ret > 0)
4573                         goto loop;
4574
4575                 if (bg_ret && bg_ret != block_group) {
4576                         btrfs_release_block_group(block_group, ffe_ctl->delalloc);
4577                         block_group = bg_ret;
4578                 }
4579
4580                 /* Checks */
4581                 ffe_ctl->search_start = round_up(ffe_ctl->found_offset,
4582                                                  fs_info->stripesize);
4583
4584                 /* move on to the next group */
4585                 if (ffe_ctl->search_start + ffe_ctl->num_bytes >
4586                     block_group->start + block_group->length) {
4587                         btrfs_add_free_space_unused(block_group,
4588                                             ffe_ctl->found_offset,
4589                                             ffe_ctl->num_bytes);
4590                         goto loop;
4591                 }
4592
4593                 if (ffe_ctl->found_offset < ffe_ctl->search_start)
4594                         btrfs_add_free_space_unused(block_group,
4595                                         ffe_ctl->found_offset,
4596                                         ffe_ctl->search_start - ffe_ctl->found_offset);
4597
4598                 ret = btrfs_add_reserved_bytes(block_group, ffe_ctl->ram_bytes,
4599                                                ffe_ctl->num_bytes,
4600                                                ffe_ctl->delalloc,
4601                                                ffe_ctl->loop >= LOOP_WRONG_SIZE_CLASS);
4602                 if (ret == -EAGAIN) {
4603                         btrfs_add_free_space_unused(block_group,
4604                                         ffe_ctl->found_offset,
4605                                         ffe_ctl->num_bytes);
4606                         goto loop;
4607                 }
4608                 btrfs_inc_block_group_reservations(block_group);
4609
4610                 /* we are all good, lets return */
4611                 ins->objectid = ffe_ctl->search_start;
4612                 ins->offset = ffe_ctl->num_bytes;
4613
4614                 trace_btrfs_reserve_extent(block_group, ffe_ctl);
4615                 btrfs_release_block_group(block_group, ffe_ctl->delalloc);
4616                 break;
4617 loop:
4618                 if (!ffe_ctl->cached && ffe_ctl->loop > LOOP_CACHING_NOWAIT &&
4619                     !ffe_ctl->retry_uncached) {
4620                         ffe_ctl->retry_uncached = true;
4621                         btrfs_wait_block_group_cache_progress(block_group,
4622                                                 ffe_ctl->num_bytes +
4623                                                 ffe_ctl->empty_cluster +
4624                                                 ffe_ctl->empty_size);
4625                         goto have_block_group;
4626                 }
4627                 release_block_group(block_group, ffe_ctl, ffe_ctl->delalloc);
4628                 cond_resched();
4629         }
4630         up_read(&space_info->groups_sem);
4631
4632         ret = find_free_extent_update_loop(fs_info, ins, ffe_ctl, full_search);
4633         if (ret > 0)
4634                 goto search;
4635
4636         if (ret == -ENOSPC && !cache_block_group_error) {
4637                 /*
4638                  * Use ffe_ctl->total_free_space as fallback if we can't find
4639                  * any contiguous hole.
4640                  */
4641                 if (!ffe_ctl->max_extent_size)
4642                         ffe_ctl->max_extent_size = ffe_ctl->total_free_space;
4643                 spin_lock(&space_info->lock);
4644                 space_info->max_extent_size = ffe_ctl->max_extent_size;
4645                 spin_unlock(&space_info->lock);
4646                 ins->offset = ffe_ctl->max_extent_size;
4647         } else if (ret == -ENOSPC) {
4648                 ret = cache_block_group_error;
4649         }
4650         return ret;
4651 }
4652
4653 /*
4654  * Entry point to the extent allocator. Tries to find a hole that is at least
4655  * as big as @num_bytes.
4656  *
4657  * @root           -    The root that will contain this extent
4658  *
4659  * @ram_bytes      -    The amount of space in ram that @num_bytes take. This
4660  *                      is used for accounting purposes. This value differs
4661  *                      from @num_bytes only in the case of compressed extents.
4662  *
4663  * @num_bytes      -    Number of bytes to allocate on-disk.
4664  *
4665  * @min_alloc_size -    Indicates the minimum amount of space that the
4666  *                      allocator should try to satisfy. In some cases
4667  *                      @num_bytes may be larger than what is required and if
4668  *                      the filesystem is fragmented then allocation fails.
4669  *                      However, the presence of @min_alloc_size gives a
4670  *                      chance to try and satisfy the smaller allocation.
4671  *
4672  * @empty_size     -    A hint that you plan on doing more COW. This is the
4673  *                      size in bytes the allocator should try to find free
4674  *                      next to the block it returns.  This is just a hint and
4675  *                      may be ignored by the allocator.
4676  *
4677  * @hint_byte      -    Hint to the allocator to start searching above the byte
4678  *                      address passed. It might be ignored.
4679  *
4680  * @ins            -    This key is modified to record the found hole. It will
4681  *                      have the following values:
4682  *                      ins->objectid == start position
4683  *                      ins->flags = BTRFS_EXTENT_ITEM_KEY
4684  *                      ins->offset == the size of the hole.
4685  *
4686  * @is_data        -    Boolean flag indicating whether an extent is
4687  *                      allocated for data (true) or metadata (false)
4688  *
4689  * @delalloc       -    Boolean flag indicating whether this allocation is for
4690  *                      delalloc or not. If 'true' data_rwsem of block groups
4691  *                      is going to be acquired.
4692  *
4693  *
4694  * Returns 0 when an allocation succeeded or < 0 when an error occurred. In
4695  * case -ENOSPC is returned then @ins->offset will contain the size of the
4696  * largest available hole the allocator managed to find.
4697  */
4698 int btrfs_reserve_extent(struct btrfs_root *root, u64 ram_bytes,
4699                          u64 num_bytes, u64 min_alloc_size,
4700                          u64 empty_size, u64 hint_byte,
4701                          struct btrfs_key *ins, int is_data, int delalloc)
4702 {
4703         struct btrfs_fs_info *fs_info = root->fs_info;
4704         struct find_free_extent_ctl ffe_ctl = {};
4705         bool final_tried = num_bytes == min_alloc_size;
4706         u64 flags;
4707         int ret;
4708         bool for_treelog = (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
4709         bool for_data_reloc = (btrfs_is_data_reloc_root(root) && is_data);
4710
4711         flags = get_alloc_profile_by_root(root, is_data);
4712 again:
4713         WARN_ON(num_bytes < fs_info->sectorsize);
4714
4715         ffe_ctl.ram_bytes = ram_bytes;
4716         ffe_ctl.num_bytes = num_bytes;
4717         ffe_ctl.min_alloc_size = min_alloc_size;
4718         ffe_ctl.empty_size = empty_size;
4719         ffe_ctl.flags = flags;
4720         ffe_ctl.delalloc = delalloc;
4721         ffe_ctl.hint_byte = hint_byte;
4722         ffe_ctl.for_treelog = for_treelog;
4723         ffe_ctl.for_data_reloc = for_data_reloc;
4724
4725         ret = find_free_extent(root, ins, &ffe_ctl);
4726         if (!ret && !is_data) {
4727                 btrfs_dec_block_group_reservations(fs_info, ins->objectid);
4728         } else if (ret == -ENOSPC) {
4729                 if (!final_tried && ins->offset) {
4730                         num_bytes = min(num_bytes >> 1, ins->offset);
4731                         num_bytes = round_down(num_bytes,
4732                                                fs_info->sectorsize);
4733                         num_bytes = max(num_bytes, min_alloc_size);
4734                         ram_bytes = num_bytes;
4735                         if (num_bytes == min_alloc_size)
4736                                 final_tried = true;
4737                         goto again;
4738                 } else if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
4739                         struct btrfs_space_info *sinfo;
4740
4741                         sinfo = btrfs_find_space_info(fs_info, flags);
4742                         btrfs_err(fs_info,
4743         "allocation failed flags %llu, wanted %llu tree-log %d, relocation: %d",
4744                                   flags, num_bytes, for_treelog, for_data_reloc);
4745                         if (sinfo)
4746                                 btrfs_dump_space_info(fs_info, sinfo,
4747                                                       num_bytes, 1);
4748                 }
4749         }
4750
4751         return ret;
4752 }
4753
4754 int btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
4755                                u64 start, u64 len, int delalloc)
4756 {
4757         struct btrfs_block_group *cache;
4758
4759         cache = btrfs_lookup_block_group(fs_info, start);
4760         if (!cache) {
4761                 btrfs_err(fs_info, "Unable to find block group for %llu",
4762                           start);
4763                 return -ENOSPC;
4764         }
4765
4766         btrfs_add_free_space(cache, start, len);
4767         btrfs_free_reserved_bytes(cache, len, delalloc);
4768         trace_btrfs_reserved_extent_free(fs_info, start, len);
4769
4770         btrfs_put_block_group(cache);
4771         return 0;
4772 }
4773
4774 int btrfs_pin_reserved_extent(struct btrfs_trans_handle *trans,
4775                               const struct extent_buffer *eb)
4776 {
4777         struct btrfs_block_group *cache;
4778         int ret = 0;
4779
4780         cache = btrfs_lookup_block_group(trans->fs_info, eb->start);
4781         if (!cache) {
4782                 btrfs_err(trans->fs_info, "unable to find block group for %llu",
4783                           eb->start);
4784                 return -ENOSPC;
4785         }
4786
4787         ret = pin_down_extent(trans, cache, eb->start, eb->len, 1);
4788         btrfs_put_block_group(cache);
4789         return ret;
4790 }
4791
4792 static int alloc_reserved_extent(struct btrfs_trans_handle *trans, u64 bytenr,
4793                                  u64 num_bytes)
4794 {
4795         struct btrfs_fs_info *fs_info = trans->fs_info;
4796         int ret;
4797
4798         ret = remove_from_free_space_tree(trans, bytenr, num_bytes);
4799         if (ret)
4800                 return ret;
4801
4802         ret = btrfs_update_block_group(trans, bytenr, num_bytes, true);
4803         if (ret) {
4804                 ASSERT(!ret);
4805                 btrfs_err(fs_info, "update block group failed for %llu %llu",
4806                           bytenr, num_bytes);
4807                 return ret;
4808         }
4809
4810         trace_btrfs_reserved_extent_alloc(fs_info, bytenr, num_bytes);
4811         return 0;
4812 }
4813
4814 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
4815                                       u64 parent, u64 root_objectid,
4816                                       u64 flags, u64 owner, u64 offset,
4817                                       struct btrfs_key *ins, int ref_mod, u64 oref_root)
4818 {
4819         struct btrfs_fs_info *fs_info = trans->fs_info;
4820         struct btrfs_root *extent_root;
4821         int ret;
4822         struct btrfs_extent_item *extent_item;
4823         struct btrfs_extent_owner_ref *oref;
4824         struct btrfs_extent_inline_ref *iref;
4825         struct btrfs_path *path;
4826         struct extent_buffer *leaf;
4827         int type;
4828         u32 size;
4829         const bool simple_quota = (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_SIMPLE);
4830
4831         if (parent > 0)
4832                 type = BTRFS_SHARED_DATA_REF_KEY;
4833         else
4834                 type = BTRFS_EXTENT_DATA_REF_KEY;
4835
4836         size = sizeof(*extent_item);
4837         if (simple_quota)
4838                 size += btrfs_extent_inline_ref_size(BTRFS_EXTENT_OWNER_REF_KEY);
4839         size += btrfs_extent_inline_ref_size(type);
4840
4841         path = btrfs_alloc_path();
4842         if (!path)
4843                 return -ENOMEM;
4844
4845         extent_root = btrfs_extent_root(fs_info, ins->objectid);
4846         ret = btrfs_insert_empty_item(trans, extent_root, path, ins, size);
4847         if (ret) {
4848                 btrfs_free_path(path);
4849                 return ret;
4850         }
4851
4852         leaf = path->nodes[0];
4853         extent_item = btrfs_item_ptr(leaf, path->slots[0],
4854                                      struct btrfs_extent_item);
4855         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
4856         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
4857         btrfs_set_extent_flags(leaf, extent_item,
4858                                flags | BTRFS_EXTENT_FLAG_DATA);
4859
4860         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
4861         if (simple_quota) {
4862                 btrfs_set_extent_inline_ref_type(leaf, iref, BTRFS_EXTENT_OWNER_REF_KEY);
4863                 oref = (struct btrfs_extent_owner_ref *)(&iref->offset);
4864                 btrfs_set_extent_owner_ref_root_id(leaf, oref, oref_root);
4865                 iref = (struct btrfs_extent_inline_ref *)(oref + 1);
4866         }
4867         btrfs_set_extent_inline_ref_type(leaf, iref, type);
4868
4869         if (parent > 0) {
4870                 struct btrfs_shared_data_ref *ref;
4871                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
4872                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
4873                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
4874         } else {
4875                 struct btrfs_extent_data_ref *ref;
4876                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
4877                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
4878                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
4879                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
4880                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
4881         }
4882
4883         btrfs_mark_buffer_dirty(trans, path->nodes[0]);
4884         btrfs_free_path(path);
4885
4886         return alloc_reserved_extent(trans, ins->objectid, ins->offset);
4887 }
4888
4889 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
4890                                      struct btrfs_delayed_ref_node *node,
4891                                      struct btrfs_delayed_extent_op *extent_op)
4892 {
4893         struct btrfs_fs_info *fs_info = trans->fs_info;
4894         struct btrfs_root *extent_root;
4895         int ret;
4896         struct btrfs_extent_item *extent_item;
4897         struct btrfs_key extent_key;
4898         struct btrfs_tree_block_info *block_info;
4899         struct btrfs_extent_inline_ref *iref;
4900         struct btrfs_path *path;
4901         struct extent_buffer *leaf;
4902         struct btrfs_delayed_tree_ref *ref;
4903         u32 size = sizeof(*extent_item) + sizeof(*iref);
4904         u64 flags = extent_op->flags_to_set;
4905         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
4906
4907         ref = btrfs_delayed_node_to_tree_ref(node);
4908
4909         extent_key.objectid = node->bytenr;
4910         if (skinny_metadata) {
4911                 extent_key.offset = ref->level;
4912                 extent_key.type = BTRFS_METADATA_ITEM_KEY;
4913         } else {
4914                 extent_key.offset = node->num_bytes;
4915                 extent_key.type = BTRFS_EXTENT_ITEM_KEY;
4916                 size += sizeof(*block_info);
4917         }
4918
4919         path = btrfs_alloc_path();
4920         if (!path)
4921                 return -ENOMEM;
4922
4923         extent_root = btrfs_extent_root(fs_info, extent_key.objectid);
4924         ret = btrfs_insert_empty_item(trans, extent_root, path, &extent_key,
4925                                       size);
4926         if (ret) {
4927                 btrfs_free_path(path);
4928                 return ret;
4929         }
4930
4931         leaf = path->nodes[0];
4932         extent_item = btrfs_item_ptr(leaf, path->slots[0],
4933                                      struct btrfs_extent_item);
4934         btrfs_set_extent_refs(leaf, extent_item, 1);
4935         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
4936         btrfs_set_extent_flags(leaf, extent_item,
4937                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
4938
4939         if (skinny_metadata) {
4940                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
4941         } else {
4942                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
4943                 btrfs_set_tree_block_key(leaf, block_info, &extent_op->key);
4944                 btrfs_set_tree_block_level(leaf, block_info, ref->level);
4945                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
4946         }
4947
4948         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY) {
4949                 btrfs_set_extent_inline_ref_type(leaf, iref,
4950                                                  BTRFS_SHARED_BLOCK_REF_KEY);
4951                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->parent);
4952         } else {
4953                 btrfs_set_extent_inline_ref_type(leaf, iref,
4954                                                  BTRFS_TREE_BLOCK_REF_KEY);
4955                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->root);
4956         }
4957
4958         btrfs_mark_buffer_dirty(trans, leaf);
4959         btrfs_free_path(path);
4960
4961         return alloc_reserved_extent(trans, node->bytenr, fs_info->nodesize);
4962 }
4963
4964 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
4965                                      struct btrfs_root *root, u64 owner,
4966                                      u64 offset, u64 ram_bytes,
4967                                      struct btrfs_key *ins)
4968 {
4969         struct btrfs_ref generic_ref = { 0 };
4970         u64 root_objectid = root->root_key.objectid;
4971         u64 owning_root = root_objectid;
4972
4973         ASSERT(root_objectid != BTRFS_TREE_LOG_OBJECTID);
4974
4975         if (btrfs_is_data_reloc_root(root) && is_fstree(root->relocation_src_root))
4976                 owning_root = root->relocation_src_root;
4977
4978         btrfs_init_generic_ref(&generic_ref, BTRFS_ADD_DELAYED_EXTENT,
4979                                ins->objectid, ins->offset, 0, owning_root);
4980         btrfs_init_data_ref(&generic_ref, root_objectid, owner,
4981                             offset, 0, false);
4982         btrfs_ref_tree_mod(root->fs_info, &generic_ref);
4983
4984         return btrfs_add_delayed_data_ref(trans, &generic_ref, ram_bytes);
4985 }
4986
4987 /*
4988  * this is used by the tree logging recovery code.  It records that
4989  * an extent has been allocated and makes sure to clear the free
4990  * space cache bits as well
4991  */
4992 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
4993                                    u64 root_objectid, u64 owner, u64 offset,
4994                                    struct btrfs_key *ins)
4995 {
4996         struct btrfs_fs_info *fs_info = trans->fs_info;
4997         int ret;
4998         struct btrfs_block_group *block_group;
4999         struct btrfs_space_info *space_info;
5000         struct btrfs_squota_delta delta = {
5001                 .root = root_objectid,
5002                 .num_bytes = ins->offset,
5003                 .generation = trans->transid,
5004                 .is_data = true,
5005                 .is_inc = true,
5006         };
5007
5008         /*
5009          * Mixed block groups will exclude before processing the log so we only
5010          * need to do the exclude dance if this fs isn't mixed.
5011          */
5012         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS)) {
5013                 ret = __exclude_logged_extent(fs_info, ins->objectid,
5014                                               ins->offset);
5015                 if (ret)
5016                         return ret;
5017         }
5018
5019         block_group = btrfs_lookup_block_group(fs_info, ins->objectid);
5020         if (!block_group)
5021                 return -EINVAL;
5022
5023         space_info = block_group->space_info;
5024         spin_lock(&space_info->lock);
5025         spin_lock(&block_group->lock);
5026         space_info->bytes_reserved += ins->offset;
5027         block_group->reserved += ins->offset;
5028         spin_unlock(&block_group->lock);
5029         spin_unlock(&space_info->lock);
5030
5031         ret = alloc_reserved_file_extent(trans, 0, root_objectid, 0, owner,
5032                                          offset, ins, 1, root_objectid);
5033         if (ret)
5034                 btrfs_pin_extent(trans, ins->objectid, ins->offset, 1);
5035         ret = btrfs_record_squota_delta(fs_info, &delta);
5036         btrfs_put_block_group(block_group);
5037         return ret;
5038 }
5039
5040 #ifdef CONFIG_BTRFS_DEBUG
5041 /*
5042  * Extra safety check in case the extent tree is corrupted and extent allocator
5043  * chooses to use a tree block which is already used and locked.
5044  */
5045 static bool check_eb_lock_owner(const struct extent_buffer *eb)
5046 {
5047         if (eb->lock_owner == current->pid) {
5048                 btrfs_err_rl(eb->fs_info,
5049 "tree block %llu owner %llu already locked by pid=%d, extent tree corruption detected",
5050                              eb->start, btrfs_header_owner(eb), current->pid);
5051                 return true;
5052         }
5053         return false;
5054 }
5055 #else
5056 static bool check_eb_lock_owner(struct extent_buffer *eb)
5057 {
5058         return false;
5059 }
5060 #endif
5061
5062 static struct extent_buffer *
5063 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
5064                       u64 bytenr, int level, u64 owner,
5065                       enum btrfs_lock_nesting nest)
5066 {
5067         struct btrfs_fs_info *fs_info = root->fs_info;
5068         struct extent_buffer *buf;
5069         u64 lockdep_owner = owner;
5070
5071         buf = btrfs_find_create_tree_block(fs_info, bytenr, owner, level);
5072         if (IS_ERR(buf))
5073                 return buf;
5074
5075         if (check_eb_lock_owner(buf)) {
5076                 free_extent_buffer(buf);
5077                 return ERR_PTR(-EUCLEAN);
5078         }
5079
5080         /*
5081          * The reloc trees are just snapshots, so we need them to appear to be
5082          * just like any other fs tree WRT lockdep.
5083          *
5084          * The exception however is in replace_path() in relocation, where we
5085          * hold the lock on the original fs root and then search for the reloc
5086          * root.  At that point we need to make sure any reloc root buffers are
5087          * set to the BTRFS_TREE_RELOC_OBJECTID lockdep class in order to make
5088          * lockdep happy.
5089          */
5090         if (lockdep_owner == BTRFS_TREE_RELOC_OBJECTID &&
5091             !test_bit(BTRFS_ROOT_RESET_LOCKDEP_CLASS, &root->state))
5092                 lockdep_owner = BTRFS_FS_TREE_OBJECTID;
5093
5094         /* btrfs_clear_buffer_dirty() accesses generation field. */
5095         btrfs_set_header_generation(buf, trans->transid);
5096
5097         /*
5098          * This needs to stay, because we could allocate a freed block from an
5099          * old tree into a new tree, so we need to make sure this new block is
5100          * set to the appropriate level and owner.
5101          */
5102         btrfs_set_buffer_lockdep_class(lockdep_owner, buf, level);
5103
5104         __btrfs_tree_lock(buf, nest);
5105         btrfs_clear_buffer_dirty(trans, buf);
5106         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
5107         clear_bit(EXTENT_BUFFER_ZONED_ZEROOUT, &buf->bflags);
5108
5109         set_extent_buffer_uptodate(buf);
5110
5111         memzero_extent_buffer(buf, 0, sizeof(struct btrfs_header));
5112         btrfs_set_header_level(buf, level);
5113         btrfs_set_header_bytenr(buf, buf->start);
5114         btrfs_set_header_generation(buf, trans->transid);
5115         btrfs_set_header_backref_rev(buf, BTRFS_MIXED_BACKREF_REV);
5116         btrfs_set_header_owner(buf, owner);
5117         write_extent_buffer_fsid(buf, fs_info->fs_devices->metadata_uuid);
5118         write_extent_buffer_chunk_tree_uuid(buf, fs_info->chunk_tree_uuid);
5119         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
5120                 buf->log_index = root->log_transid % 2;
5121                 /*
5122                  * we allow two log transactions at a time, use different
5123                  * EXTENT bit to differentiate dirty pages.
5124                  */
5125                 if (buf->log_index == 0)
5126                         set_extent_bit(&root->dirty_log_pages, buf->start,
5127                                        buf->start + buf->len - 1,
5128                                        EXTENT_DIRTY, NULL);
5129                 else
5130                         set_extent_bit(&root->dirty_log_pages, buf->start,
5131                                        buf->start + buf->len - 1,
5132                                        EXTENT_NEW, NULL);
5133         } else {
5134                 buf->log_index = -1;
5135                 set_extent_bit(&trans->transaction->dirty_pages, buf->start,
5136                                buf->start + buf->len - 1, EXTENT_DIRTY, NULL);
5137         }
5138         /* this returns a buffer locked for blocking */
5139         return buf;
5140 }
5141
5142 /*
5143  * finds a free extent and does all the dirty work required for allocation
5144  * returns the tree buffer or an ERR_PTR on error.
5145  */
5146 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
5147                                              struct btrfs_root *root,
5148                                              u64 parent, u64 root_objectid,
5149                                              const struct btrfs_disk_key *key,
5150                                              int level, u64 hint,
5151                                              u64 empty_size,
5152                                              u64 reloc_src_root,
5153                                              enum btrfs_lock_nesting nest)
5154 {
5155         struct btrfs_fs_info *fs_info = root->fs_info;
5156         struct btrfs_key ins;
5157         struct btrfs_block_rsv *block_rsv;
5158         struct extent_buffer *buf;
5159         struct btrfs_delayed_extent_op *extent_op;
5160         struct btrfs_ref generic_ref = { 0 };
5161         u64 flags = 0;
5162         int ret;
5163         u32 blocksize = fs_info->nodesize;
5164         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
5165         u64 owning_root;
5166
5167 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
5168         if (btrfs_is_testing(fs_info)) {
5169                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
5170                                             level, root_objectid, nest);
5171                 if (!IS_ERR(buf))
5172                         root->alloc_bytenr += blocksize;
5173                 return buf;
5174         }
5175 #endif
5176
5177         block_rsv = btrfs_use_block_rsv(trans, root, blocksize);
5178         if (IS_ERR(block_rsv))
5179                 return ERR_CAST(block_rsv);
5180
5181         ret = btrfs_reserve_extent(root, blocksize, blocksize, blocksize,
5182                                    empty_size, hint, &ins, 0, 0);
5183         if (ret)
5184                 goto out_unuse;
5185
5186         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level,
5187                                     root_objectid, nest);
5188         if (IS_ERR(buf)) {
5189                 ret = PTR_ERR(buf);
5190                 goto out_free_reserved;
5191         }
5192         owning_root = btrfs_header_owner(buf);
5193
5194         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
5195                 if (parent == 0)
5196                         parent = ins.objectid;
5197                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
5198                 owning_root = reloc_src_root;
5199         } else
5200                 BUG_ON(parent > 0);
5201
5202         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
5203                 extent_op = btrfs_alloc_delayed_extent_op();
5204                 if (!extent_op) {
5205                         ret = -ENOMEM;
5206                         goto out_free_buf;
5207                 }
5208                 if (key)
5209                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
5210                 else
5211                         memset(&extent_op->key, 0, sizeof(extent_op->key));
5212                 extent_op->flags_to_set = flags;
5213                 extent_op->update_key = skinny_metadata ? false : true;
5214                 extent_op->update_flags = true;
5215                 extent_op->level = level;
5216
5217                 btrfs_init_generic_ref(&generic_ref, BTRFS_ADD_DELAYED_EXTENT,
5218                                        ins.objectid, ins.offset, parent, owning_root);
5219                 btrfs_init_tree_ref(&generic_ref, level, root_objectid,
5220                                     root->root_key.objectid, false);
5221                 btrfs_ref_tree_mod(fs_info, &generic_ref);
5222                 ret = btrfs_add_delayed_tree_ref(trans, &generic_ref, extent_op);
5223                 if (ret)
5224                         goto out_free_delayed;
5225         }
5226         return buf;
5227
5228 out_free_delayed:
5229         btrfs_free_delayed_extent_op(extent_op);
5230 out_free_buf:
5231         btrfs_tree_unlock(buf);
5232         free_extent_buffer(buf);
5233 out_free_reserved:
5234         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 0);
5235 out_unuse:
5236         btrfs_unuse_block_rsv(fs_info, block_rsv, blocksize);
5237         return ERR_PTR(ret);
5238 }
5239
5240 struct walk_control {
5241         u64 refs[BTRFS_MAX_LEVEL];
5242         u64 flags[BTRFS_MAX_LEVEL];
5243         struct btrfs_key update_progress;
5244         struct btrfs_key drop_progress;
5245         int drop_level;
5246         int stage;
5247         int level;
5248         int shared_level;
5249         int update_ref;
5250         int keep_locks;
5251         int reada_slot;
5252         int reada_count;
5253         int restarted;
5254 };
5255
5256 #define DROP_REFERENCE  1
5257 #define UPDATE_BACKREF  2
5258
5259 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
5260                                      struct btrfs_root *root,
5261                                      struct walk_control *wc,
5262                                      struct btrfs_path *path)
5263 {
5264         struct btrfs_fs_info *fs_info = root->fs_info;
5265         u64 bytenr;
5266         u64 generation;
5267         u64 refs;
5268         u64 flags;
5269         u32 nritems;
5270         struct btrfs_key key;
5271         struct extent_buffer *eb;
5272         int ret;
5273         int slot;
5274         int nread = 0;
5275
5276         if (path->slots[wc->level] < wc->reada_slot) {
5277                 wc->reada_count = wc->reada_count * 2 / 3;
5278                 wc->reada_count = max(wc->reada_count, 2);
5279         } else {
5280                 wc->reada_count = wc->reada_count * 3 / 2;
5281                 wc->reada_count = min_t(int, wc->reada_count,
5282                                         BTRFS_NODEPTRS_PER_BLOCK(fs_info));
5283         }
5284
5285         eb = path->nodes[wc->level];
5286         nritems = btrfs_header_nritems(eb);
5287
5288         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
5289                 if (nread >= wc->reada_count)
5290                         break;
5291
5292                 cond_resched();
5293                 bytenr = btrfs_node_blockptr(eb, slot);
5294                 generation = btrfs_node_ptr_generation(eb, slot);
5295
5296                 if (slot == path->slots[wc->level])
5297                         goto reada;
5298
5299                 if (wc->stage == UPDATE_BACKREF &&
5300                     generation <= root->root_key.offset)
5301                         continue;
5302
5303                 /* We don't lock the tree block, it's OK to be racy here */
5304                 ret = btrfs_lookup_extent_info(trans, fs_info, bytenr,
5305                                                wc->level - 1, 1, &refs,
5306                                                &flags, NULL);
5307                 /* We don't care about errors in readahead. */
5308                 if (ret < 0)
5309                         continue;
5310                 BUG_ON(refs == 0);
5311
5312                 if (wc->stage == DROP_REFERENCE) {
5313                         if (refs == 1)
5314                                 goto reada;
5315
5316                         if (wc->level == 1 &&
5317                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5318                                 continue;
5319                         if (!wc->update_ref ||
5320                             generation <= root->root_key.offset)
5321                                 continue;
5322                         btrfs_node_key_to_cpu(eb, &key, slot);
5323                         ret = btrfs_comp_cpu_keys(&key,
5324                                                   &wc->update_progress);
5325                         if (ret < 0)
5326                                 continue;
5327                 } else {
5328                         if (wc->level == 1 &&
5329                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5330                                 continue;
5331                 }
5332 reada:
5333                 btrfs_readahead_node_child(eb, slot);
5334                 nread++;
5335         }
5336         wc->reada_slot = slot;
5337 }
5338
5339 /*
5340  * helper to process tree block while walking down the tree.
5341  *
5342  * when wc->stage == UPDATE_BACKREF, this function updates
5343  * back refs for pointers in the block.
5344  *
5345  * NOTE: return value 1 means we should stop walking down.
5346  */
5347 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
5348                                    struct btrfs_root *root,
5349                                    struct btrfs_path *path,
5350                                    struct walk_control *wc, int lookup_info)
5351 {
5352         struct btrfs_fs_info *fs_info = root->fs_info;
5353         int level = wc->level;
5354         struct extent_buffer *eb = path->nodes[level];
5355         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
5356         int ret;
5357
5358         if (wc->stage == UPDATE_BACKREF &&
5359             btrfs_header_owner(eb) != root->root_key.objectid)
5360                 return 1;
5361
5362         /*
5363          * when reference count of tree block is 1, it won't increase
5364          * again. once full backref flag is set, we never clear it.
5365          */
5366         if (lookup_info &&
5367             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
5368              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
5369                 BUG_ON(!path->locks[level]);
5370                 ret = btrfs_lookup_extent_info(trans, fs_info,
5371                                                eb->start, level, 1,
5372                                                &wc->refs[level],
5373                                                &wc->flags[level],
5374                                                NULL);
5375                 BUG_ON(ret == -ENOMEM);
5376                 if (ret)
5377                         return ret;
5378                 BUG_ON(wc->refs[level] == 0);
5379         }
5380
5381         if (wc->stage == DROP_REFERENCE) {
5382                 if (wc->refs[level] > 1)
5383                         return 1;
5384
5385                 if (path->locks[level] && !wc->keep_locks) {
5386                         btrfs_tree_unlock_rw(eb, path->locks[level]);
5387                         path->locks[level] = 0;
5388                 }
5389                 return 0;
5390         }
5391
5392         /* wc->stage == UPDATE_BACKREF */
5393         if (!(wc->flags[level] & flag)) {
5394                 BUG_ON(!path->locks[level]);
5395                 ret = btrfs_inc_ref(trans, root, eb, 1);
5396                 BUG_ON(ret); /* -ENOMEM */
5397                 ret = btrfs_dec_ref(trans, root, eb, 0);
5398                 BUG_ON(ret); /* -ENOMEM */
5399                 ret = btrfs_set_disk_extent_flags(trans, eb, flag);
5400                 BUG_ON(ret); /* -ENOMEM */
5401                 wc->flags[level] |= flag;
5402         }
5403
5404         /*
5405          * the block is shared by multiple trees, so it's not good to
5406          * keep the tree lock
5407          */
5408         if (path->locks[level] && level > 0) {
5409                 btrfs_tree_unlock_rw(eb, path->locks[level]);
5410                 path->locks[level] = 0;
5411         }
5412         return 0;
5413 }
5414
5415 /*
5416  * This is used to verify a ref exists for this root to deal with a bug where we
5417  * would have a drop_progress key that hadn't been updated properly.
5418  */
5419 static int check_ref_exists(struct btrfs_trans_handle *trans,
5420                             struct btrfs_root *root, u64 bytenr, u64 parent,
5421                             int level)
5422 {
5423         struct btrfs_path *path;
5424         struct btrfs_extent_inline_ref *iref;
5425         int ret;
5426
5427         path = btrfs_alloc_path();
5428         if (!path)
5429                 return -ENOMEM;
5430
5431         ret = lookup_extent_backref(trans, path, &iref, bytenr,
5432                                     root->fs_info->nodesize, parent,
5433                                     root->root_key.objectid, level, 0);
5434         btrfs_free_path(path);
5435         if (ret == -ENOENT)
5436                 return 0;
5437         if (ret < 0)
5438                 return ret;
5439         return 1;
5440 }
5441
5442 /*
5443  * helper to process tree block pointer.
5444  *
5445  * when wc->stage == DROP_REFERENCE, this function checks
5446  * reference count of the block pointed to. if the block
5447  * is shared and we need update back refs for the subtree
5448  * rooted at the block, this function changes wc->stage to
5449  * UPDATE_BACKREF. if the block is shared and there is no
5450  * need to update back, this function drops the reference
5451  * to the block.
5452  *
5453  * NOTE: return value 1 means we should stop walking down.
5454  */
5455 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
5456                                  struct btrfs_root *root,
5457                                  struct btrfs_path *path,
5458                                  struct walk_control *wc, int *lookup_info)
5459 {
5460         struct btrfs_fs_info *fs_info = root->fs_info;
5461         u64 bytenr;
5462         u64 generation;
5463         u64 parent;
5464         u64 owner_root = 0;
5465         struct btrfs_tree_parent_check check = { 0 };
5466         struct btrfs_key key;
5467         struct btrfs_ref ref = { 0 };
5468         struct extent_buffer *next;
5469         int level = wc->level;
5470         int reada = 0;
5471         int ret = 0;
5472         bool need_account = false;
5473
5474         generation = btrfs_node_ptr_generation(path->nodes[level],
5475                                                path->slots[level]);
5476         /*
5477          * if the lower level block was created before the snapshot
5478          * was created, we know there is no need to update back refs
5479          * for the subtree
5480          */
5481         if (wc->stage == UPDATE_BACKREF &&
5482             generation <= root->root_key.offset) {
5483                 *lookup_info = 1;
5484                 return 1;
5485         }
5486
5487         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
5488
5489         check.level = level - 1;
5490         check.transid = generation;
5491         check.owner_root = root->root_key.objectid;
5492         check.has_first_key = true;
5493         btrfs_node_key_to_cpu(path->nodes[level], &check.first_key,
5494                               path->slots[level]);
5495
5496         next = find_extent_buffer(fs_info, bytenr);
5497         if (!next) {
5498                 next = btrfs_find_create_tree_block(fs_info, bytenr,
5499                                 root->root_key.objectid, level - 1);
5500                 if (IS_ERR(next))
5501                         return PTR_ERR(next);
5502                 reada = 1;
5503         }
5504         btrfs_tree_lock(next);
5505
5506         ret = btrfs_lookup_extent_info(trans, fs_info, bytenr, level - 1, 1,
5507                                        &wc->refs[level - 1],
5508                                        &wc->flags[level - 1],
5509                                        &owner_root);
5510         if (ret < 0)
5511                 goto out_unlock;
5512
5513         if (unlikely(wc->refs[level - 1] == 0)) {
5514                 btrfs_err(fs_info, "Missing references.");
5515                 ret = -EIO;
5516                 goto out_unlock;
5517         }
5518         *lookup_info = 0;
5519
5520         if (wc->stage == DROP_REFERENCE) {
5521                 if (wc->refs[level - 1] > 1) {
5522                         need_account = true;
5523                         if (level == 1 &&
5524                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5525                                 goto skip;
5526
5527                         if (!wc->update_ref ||
5528                             generation <= root->root_key.offset)
5529                                 goto skip;
5530
5531                         btrfs_node_key_to_cpu(path->nodes[level], &key,
5532                                               path->slots[level]);
5533                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
5534                         if (ret < 0)
5535                                 goto skip;
5536
5537                         wc->stage = UPDATE_BACKREF;
5538                         wc->shared_level = level - 1;
5539                 }
5540         } else {
5541                 if (level == 1 &&
5542                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5543                         goto skip;
5544         }
5545
5546         if (!btrfs_buffer_uptodate(next, generation, 0)) {
5547                 btrfs_tree_unlock(next);
5548                 free_extent_buffer(next);
5549                 next = NULL;
5550                 *lookup_info = 1;
5551         }
5552
5553         if (!next) {
5554                 if (reada && level == 1)
5555                         reada_walk_down(trans, root, wc, path);
5556                 next = read_tree_block(fs_info, bytenr, &check);
5557                 if (IS_ERR(next)) {
5558                         return PTR_ERR(next);
5559                 } else if (!extent_buffer_uptodate(next)) {
5560                         free_extent_buffer(next);
5561                         return -EIO;
5562                 }
5563                 btrfs_tree_lock(next);
5564         }
5565
5566         level--;
5567         ASSERT(level == btrfs_header_level(next));
5568         if (level != btrfs_header_level(next)) {
5569                 btrfs_err(root->fs_info, "mismatched level");
5570                 ret = -EIO;
5571                 goto out_unlock;
5572         }
5573         path->nodes[level] = next;
5574         path->slots[level] = 0;
5575         path->locks[level] = BTRFS_WRITE_LOCK;
5576         wc->level = level;
5577         if (wc->level == 1)
5578                 wc->reada_slot = 0;
5579         return 0;
5580 skip:
5581         wc->refs[level - 1] = 0;
5582         wc->flags[level - 1] = 0;
5583         if (wc->stage == DROP_REFERENCE) {
5584                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
5585                         parent = path->nodes[level]->start;
5586                 } else {
5587                         ASSERT(root->root_key.objectid ==
5588                                btrfs_header_owner(path->nodes[level]));
5589                         if (root->root_key.objectid !=
5590                             btrfs_header_owner(path->nodes[level])) {
5591                                 btrfs_err(root->fs_info,
5592                                                 "mismatched block owner");
5593                                 ret = -EIO;
5594                                 goto out_unlock;
5595                         }
5596                         parent = 0;
5597                 }
5598
5599                 /*
5600                  * If we had a drop_progress we need to verify the refs are set
5601                  * as expected.  If we find our ref then we know that from here
5602                  * on out everything should be correct, and we can clear the
5603                  * ->restarted flag.
5604                  */
5605                 if (wc->restarted) {
5606                         ret = check_ref_exists(trans, root, bytenr, parent,
5607                                                level - 1);
5608                         if (ret < 0)
5609                                 goto out_unlock;
5610                         if (ret == 0)
5611                                 goto no_delete;
5612                         ret = 0;
5613                         wc->restarted = 0;
5614                 }
5615
5616                 /*
5617                  * Reloc tree doesn't contribute to qgroup numbers, and we have
5618                  * already accounted them at merge time (replace_path),
5619                  * thus we could skip expensive subtree trace here.
5620                  */
5621                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
5622                     need_account) {
5623                         ret = btrfs_qgroup_trace_subtree(trans, next,
5624                                                          generation, level - 1);
5625                         if (ret) {
5626                                 btrfs_err_rl(fs_info,
5627                                              "Error %d accounting shared subtree. Quota is out of sync, rescan required.",
5628                                              ret);
5629                         }
5630                 }
5631
5632                 /*
5633                  * We need to update the next key in our walk control so we can
5634                  * update the drop_progress key accordingly.  We don't care if
5635                  * find_next_key doesn't find a key because that means we're at
5636                  * the end and are going to clean up now.
5637                  */
5638                 wc->drop_level = level;
5639                 find_next_key(path, level, &wc->drop_progress);
5640
5641                 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
5642                                        fs_info->nodesize, parent, owner_root);
5643                 btrfs_init_tree_ref(&ref, level - 1, root->root_key.objectid,
5644                                     0, false);
5645                 ret = btrfs_free_extent(trans, &ref);
5646                 if (ret)
5647                         goto out_unlock;
5648         }
5649 no_delete:
5650         *lookup_info = 1;
5651         ret = 1;
5652
5653 out_unlock:
5654         btrfs_tree_unlock(next);
5655         free_extent_buffer(next);
5656
5657         return ret;
5658 }
5659
5660 /*
5661  * helper to process tree block while walking up the tree.
5662  *
5663  * when wc->stage == DROP_REFERENCE, this function drops
5664  * reference count on the block.
5665  *
5666  * when wc->stage == UPDATE_BACKREF, this function changes
5667  * wc->stage back to DROP_REFERENCE if we changed wc->stage
5668  * to UPDATE_BACKREF previously while processing the block.
5669  *
5670  * NOTE: return value 1 means we should stop walking up.
5671  */
5672 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
5673                                  struct btrfs_root *root,
5674                                  struct btrfs_path *path,
5675                                  struct walk_control *wc)
5676 {
5677         struct btrfs_fs_info *fs_info = root->fs_info;
5678         int ret;
5679         int level = wc->level;
5680         struct extent_buffer *eb = path->nodes[level];
5681         u64 parent = 0;
5682
5683         if (wc->stage == UPDATE_BACKREF) {
5684                 BUG_ON(wc->shared_level < level);
5685                 if (level < wc->shared_level)
5686                         goto out;
5687
5688                 ret = find_next_key(path, level + 1, &wc->update_progress);
5689                 if (ret > 0)
5690                         wc->update_ref = 0;
5691
5692                 wc->stage = DROP_REFERENCE;
5693                 wc->shared_level = -1;
5694                 path->slots[level] = 0;
5695
5696                 /*
5697                  * check reference count again if the block isn't locked.
5698                  * we should start walking down the tree again if reference
5699                  * count is one.
5700                  */
5701                 if (!path->locks[level]) {
5702                         BUG_ON(level == 0);
5703                         btrfs_tree_lock(eb);
5704                         path->locks[level] = BTRFS_WRITE_LOCK;
5705
5706                         ret = btrfs_lookup_extent_info(trans, fs_info,
5707                                                        eb->start, level, 1,
5708                                                        &wc->refs[level],
5709                                                        &wc->flags[level],
5710                                                        NULL);
5711                         if (ret < 0) {
5712                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
5713                                 path->locks[level] = 0;
5714                                 return ret;
5715                         }
5716                         BUG_ON(wc->refs[level] == 0);
5717                         if (wc->refs[level] == 1) {
5718                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
5719                                 path->locks[level] = 0;
5720                                 return 1;
5721                         }
5722                 }
5723         }
5724
5725         /* wc->stage == DROP_REFERENCE */
5726         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
5727
5728         if (wc->refs[level] == 1) {
5729                 if (level == 0) {
5730                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
5731                                 ret = btrfs_dec_ref(trans, root, eb, 1);
5732                         else
5733                                 ret = btrfs_dec_ref(trans, root, eb, 0);
5734                         BUG_ON(ret); /* -ENOMEM */
5735                         if (is_fstree(root->root_key.objectid)) {
5736                                 ret = btrfs_qgroup_trace_leaf_items(trans, eb);
5737                                 if (ret) {
5738                                         btrfs_err_rl(fs_info,
5739         "error %d accounting leaf items, quota is out of sync, rescan required",
5740                                              ret);
5741                                 }
5742                         }
5743                 }
5744                 /* Make block locked assertion in btrfs_clear_buffer_dirty happy. */
5745                 if (!path->locks[level]) {
5746                         btrfs_tree_lock(eb);
5747                         path->locks[level] = BTRFS_WRITE_LOCK;
5748                 }
5749                 btrfs_clear_buffer_dirty(trans, eb);
5750         }
5751
5752         if (eb == root->node) {
5753                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
5754                         parent = eb->start;
5755                 else if (root->root_key.objectid != btrfs_header_owner(eb))
5756                         goto owner_mismatch;
5757         } else {
5758                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
5759                         parent = path->nodes[level + 1]->start;
5760                 else if (root->root_key.objectid !=
5761                          btrfs_header_owner(path->nodes[level + 1]))
5762                         goto owner_mismatch;
5763         }
5764
5765         btrfs_free_tree_block(trans, btrfs_root_id(root), eb, parent,
5766                               wc->refs[level] == 1);
5767 out:
5768         wc->refs[level] = 0;
5769         wc->flags[level] = 0;
5770         return 0;
5771
5772 owner_mismatch:
5773         btrfs_err_rl(fs_info, "unexpected tree owner, have %llu expect %llu",
5774                      btrfs_header_owner(eb), root->root_key.objectid);
5775         return -EUCLEAN;
5776 }
5777
5778 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
5779                                    struct btrfs_root *root,
5780                                    struct btrfs_path *path,
5781                                    struct walk_control *wc)
5782 {
5783         int level = wc->level;
5784         int lookup_info = 1;
5785         int ret = 0;
5786
5787         while (level >= 0) {
5788                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
5789                 if (ret)
5790                         break;
5791
5792                 if (level == 0)
5793                         break;
5794
5795                 if (path->slots[level] >=
5796                     btrfs_header_nritems(path->nodes[level]))
5797                         break;
5798
5799                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
5800                 if (ret > 0) {
5801                         path->slots[level]++;
5802                         continue;
5803                 } else if (ret < 0)
5804                         break;
5805                 level = wc->level;
5806         }
5807         return (ret == 1) ? 0 : ret;
5808 }
5809
5810 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
5811                                  struct btrfs_root *root,
5812                                  struct btrfs_path *path,
5813                                  struct walk_control *wc, int max_level)
5814 {
5815         int level = wc->level;
5816         int ret;
5817
5818         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
5819         while (level < max_level && path->nodes[level]) {
5820                 wc->level = level;
5821                 if (path->slots[level] + 1 <
5822                     btrfs_header_nritems(path->nodes[level])) {
5823                         path->slots[level]++;
5824                         return 0;
5825                 } else {
5826                         ret = walk_up_proc(trans, root, path, wc);
5827                         if (ret > 0)
5828                                 return 0;
5829                         if (ret < 0)
5830                                 return ret;
5831
5832                         if (path->locks[level]) {
5833                                 btrfs_tree_unlock_rw(path->nodes[level],
5834                                                      path->locks[level]);
5835                                 path->locks[level] = 0;
5836                         }
5837                         free_extent_buffer(path->nodes[level]);
5838                         path->nodes[level] = NULL;
5839                         level++;
5840                 }
5841         }
5842         return 1;
5843 }
5844
5845 /*
5846  * drop a subvolume tree.
5847  *
5848  * this function traverses the tree freeing any blocks that only
5849  * referenced by the tree.
5850  *
5851  * when a shared tree block is found. this function decreases its
5852  * reference count by one. if update_ref is true, this function
5853  * also make sure backrefs for the shared block and all lower level
5854  * blocks are properly updated.
5855  *
5856  * If called with for_reloc == 0, may exit early with -EAGAIN
5857  */
5858 int btrfs_drop_snapshot(struct btrfs_root *root, int update_ref, int for_reloc)
5859 {
5860         const bool is_reloc_root = (root->root_key.objectid ==
5861                                     BTRFS_TREE_RELOC_OBJECTID);
5862         struct btrfs_fs_info *fs_info = root->fs_info;
5863         struct btrfs_path *path;
5864         struct btrfs_trans_handle *trans;
5865         struct btrfs_root *tree_root = fs_info->tree_root;
5866         struct btrfs_root_item *root_item = &root->root_item;
5867         struct walk_control *wc;
5868         struct btrfs_key key;
5869         int err = 0;
5870         int ret;
5871         int level;
5872         bool root_dropped = false;
5873         bool unfinished_drop = false;
5874
5875         btrfs_debug(fs_info, "Drop subvolume %llu", root->root_key.objectid);
5876
5877         path = btrfs_alloc_path();
5878         if (!path) {
5879                 err = -ENOMEM;
5880                 goto out;
5881         }
5882
5883         wc = kzalloc(sizeof(*wc), GFP_NOFS);
5884         if (!wc) {
5885                 btrfs_free_path(path);
5886                 err = -ENOMEM;
5887                 goto out;
5888         }
5889
5890         /*
5891          * Use join to avoid potential EINTR from transaction start. See
5892          * wait_reserve_ticket and the whole reservation callchain.
5893          */
5894         if (for_reloc)
5895                 trans = btrfs_join_transaction(tree_root);
5896         else
5897                 trans = btrfs_start_transaction(tree_root, 0);
5898         if (IS_ERR(trans)) {
5899                 err = PTR_ERR(trans);
5900                 goto out_free;
5901         }
5902
5903         err = btrfs_run_delayed_items(trans);
5904         if (err)
5905                 goto out_end_trans;
5906
5907         /*
5908          * This will help us catch people modifying the fs tree while we're
5909          * dropping it.  It is unsafe to mess with the fs tree while it's being
5910          * dropped as we unlock the root node and parent nodes as we walk down
5911          * the tree, assuming nothing will change.  If something does change
5912          * then we'll have stale information and drop references to blocks we've
5913          * already dropped.
5914          */
5915         set_bit(BTRFS_ROOT_DELETING, &root->state);
5916         unfinished_drop = test_bit(BTRFS_ROOT_UNFINISHED_DROP, &root->state);
5917
5918         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
5919                 level = btrfs_header_level(root->node);
5920                 path->nodes[level] = btrfs_lock_root_node(root);
5921                 path->slots[level] = 0;
5922                 path->locks[level] = BTRFS_WRITE_LOCK;
5923                 memset(&wc->update_progress, 0,
5924                        sizeof(wc->update_progress));
5925         } else {
5926                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
5927                 memcpy(&wc->update_progress, &key,
5928                        sizeof(wc->update_progress));
5929
5930                 level = btrfs_root_drop_level(root_item);
5931                 BUG_ON(level == 0);
5932                 path->lowest_level = level;
5933                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5934                 path->lowest_level = 0;
5935                 if (ret < 0) {
5936                         err = ret;
5937                         goto out_end_trans;
5938                 }
5939                 WARN_ON(ret > 0);
5940
5941                 /*
5942                  * unlock our path, this is safe because only this
5943                  * function is allowed to delete this snapshot
5944                  */
5945                 btrfs_unlock_up_safe(path, 0);
5946
5947                 level = btrfs_header_level(root->node);
5948                 while (1) {
5949                         btrfs_tree_lock(path->nodes[level]);
5950                         path->locks[level] = BTRFS_WRITE_LOCK;
5951
5952                         ret = btrfs_lookup_extent_info(trans, fs_info,
5953                                                 path->nodes[level]->start,
5954                                                 level, 1, &wc->refs[level],
5955                                                 &wc->flags[level], NULL);
5956                         if (ret < 0) {
5957                                 err = ret;
5958                                 goto out_end_trans;
5959                         }
5960                         BUG_ON(wc->refs[level] == 0);
5961
5962                         if (level == btrfs_root_drop_level(root_item))
5963                                 break;
5964
5965                         btrfs_tree_unlock(path->nodes[level]);
5966                         path->locks[level] = 0;
5967                         WARN_ON(wc->refs[level] != 1);
5968                         level--;
5969                 }
5970         }
5971
5972         wc->restarted = test_bit(BTRFS_ROOT_DEAD_TREE, &root->state);
5973         wc->level = level;
5974         wc->shared_level = -1;
5975         wc->stage = DROP_REFERENCE;
5976         wc->update_ref = update_ref;
5977         wc->keep_locks = 0;
5978         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
5979
5980         while (1) {
5981
5982                 ret = walk_down_tree(trans, root, path, wc);
5983                 if (ret < 0) {
5984                         btrfs_abort_transaction(trans, ret);
5985                         err = ret;
5986                         break;
5987                 }
5988
5989                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
5990                 if (ret < 0) {
5991                         btrfs_abort_transaction(trans, ret);
5992                         err = ret;
5993                         break;
5994                 }
5995
5996                 if (ret > 0) {
5997                         BUG_ON(wc->stage != DROP_REFERENCE);
5998                         break;
5999                 }
6000
6001                 if (wc->stage == DROP_REFERENCE) {
6002                         wc->drop_level = wc->level;
6003                         btrfs_node_key_to_cpu(path->nodes[wc->drop_level],
6004                                               &wc->drop_progress,
6005                                               path->slots[wc->drop_level]);
6006                 }
6007                 btrfs_cpu_key_to_disk(&root_item->drop_progress,
6008                                       &wc->drop_progress);
6009                 btrfs_set_root_drop_level(root_item, wc->drop_level);
6010
6011                 BUG_ON(wc->level == 0);
6012                 if (btrfs_should_end_transaction(trans) ||
6013                     (!for_reloc && btrfs_need_cleaner_sleep(fs_info))) {
6014                         ret = btrfs_update_root(trans, tree_root,
6015                                                 &root->root_key,
6016                                                 root_item);
6017                         if (ret) {
6018                                 btrfs_abort_transaction(trans, ret);
6019                                 err = ret;
6020                                 goto out_end_trans;
6021                         }
6022
6023                         if (!is_reloc_root)
6024                                 btrfs_set_last_root_drop_gen(fs_info, trans->transid);
6025
6026                         btrfs_end_transaction_throttle(trans);
6027                         if (!for_reloc && btrfs_need_cleaner_sleep(fs_info)) {
6028                                 btrfs_debug(fs_info,
6029                                             "drop snapshot early exit");
6030                                 err = -EAGAIN;
6031                                 goto out_free;
6032                         }
6033
6034                        /*
6035                         * Use join to avoid potential EINTR from transaction
6036                         * start. See wait_reserve_ticket and the whole
6037                         * reservation callchain.
6038                         */
6039                         if (for_reloc)
6040                                 trans = btrfs_join_transaction(tree_root);
6041                         else
6042                                 trans = btrfs_start_transaction(tree_root, 0);
6043                         if (IS_ERR(trans)) {
6044                                 err = PTR_ERR(trans);
6045                                 goto out_free;
6046                         }
6047                 }
6048         }
6049         btrfs_release_path(path);
6050         if (err)
6051                 goto out_end_trans;
6052
6053         ret = btrfs_del_root(trans, &root->root_key);
6054         if (ret) {
6055                 btrfs_abort_transaction(trans, ret);
6056                 err = ret;
6057                 goto out_end_trans;
6058         }
6059
6060         if (!is_reloc_root) {
6061                 ret = btrfs_find_root(tree_root, &root->root_key, path,
6062                                       NULL, NULL);
6063                 if (ret < 0) {
6064                         btrfs_abort_transaction(trans, ret);
6065                         err = ret;
6066                         goto out_end_trans;
6067                 } else if (ret > 0) {
6068                         /* if we fail to delete the orphan item this time
6069                          * around, it'll get picked up the next time.
6070                          *
6071                          * The most common failure here is just -ENOENT.
6072                          */
6073                         btrfs_del_orphan_item(trans, tree_root,
6074                                               root->root_key.objectid);
6075                 }
6076         }
6077
6078         /*
6079          * This subvolume is going to be completely dropped, and won't be
6080          * recorded as dirty roots, thus pertrans meta rsv will not be freed at
6081          * commit transaction time.  So free it here manually.
6082          */
6083         btrfs_qgroup_convert_reserved_meta(root, INT_MAX);
6084         btrfs_qgroup_free_meta_all_pertrans(root);
6085
6086         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state))
6087                 btrfs_add_dropped_root(trans, root);
6088         else
6089                 btrfs_put_root(root);
6090         root_dropped = true;
6091 out_end_trans:
6092         if (!is_reloc_root)
6093                 btrfs_set_last_root_drop_gen(fs_info, trans->transid);
6094
6095         btrfs_end_transaction_throttle(trans);
6096 out_free:
6097         kfree(wc);
6098         btrfs_free_path(path);
6099 out:
6100         /*
6101          * We were an unfinished drop root, check to see if there are any
6102          * pending, and if not clear and wake up any waiters.
6103          */
6104         if (!err && unfinished_drop)
6105                 btrfs_maybe_wake_unfinished_drop(fs_info);
6106
6107         /*
6108          * So if we need to stop dropping the snapshot for whatever reason we
6109          * need to make sure to add it back to the dead root list so that we
6110          * keep trying to do the work later.  This also cleans up roots if we
6111          * don't have it in the radix (like when we recover after a power fail
6112          * or unmount) so we don't leak memory.
6113          */
6114         if (!for_reloc && !root_dropped)
6115                 btrfs_add_dead_root(root);
6116         return err;
6117 }
6118
6119 /*
6120  * drop subtree rooted at tree block 'node'.
6121  *
6122  * NOTE: this function will unlock and release tree block 'node'
6123  * only used by relocation code
6124  */
6125 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
6126                         struct btrfs_root *root,
6127                         struct extent_buffer *node,
6128                         struct extent_buffer *parent)
6129 {
6130         struct btrfs_fs_info *fs_info = root->fs_info;
6131         struct btrfs_path *path;
6132         struct walk_control *wc;
6133         int level;
6134         int parent_level;
6135         int ret = 0;
6136         int wret;
6137
6138         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
6139
6140         path = btrfs_alloc_path();
6141         if (!path)
6142                 return -ENOMEM;
6143
6144         wc = kzalloc(sizeof(*wc), GFP_NOFS);
6145         if (!wc) {
6146                 btrfs_free_path(path);
6147                 return -ENOMEM;
6148         }
6149
6150         btrfs_assert_tree_write_locked(parent);
6151         parent_level = btrfs_header_level(parent);
6152         atomic_inc(&parent->refs);
6153         path->nodes[parent_level] = parent;
6154         path->slots[parent_level] = btrfs_header_nritems(parent);
6155
6156         btrfs_assert_tree_write_locked(node);
6157         level = btrfs_header_level(node);
6158         path->nodes[level] = node;
6159         path->slots[level] = 0;
6160         path->locks[level] = BTRFS_WRITE_LOCK;
6161
6162         wc->refs[parent_level] = 1;
6163         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
6164         wc->level = level;
6165         wc->shared_level = -1;
6166         wc->stage = DROP_REFERENCE;
6167         wc->update_ref = 0;
6168         wc->keep_locks = 1;
6169         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
6170
6171         while (1) {
6172                 wret = walk_down_tree(trans, root, path, wc);
6173                 if (wret < 0) {
6174                         ret = wret;
6175                         break;
6176                 }
6177
6178                 wret = walk_up_tree(trans, root, path, wc, parent_level);
6179                 if (wret < 0)
6180                         ret = wret;
6181                 if (wret != 0)
6182                         break;
6183         }
6184
6185         kfree(wc);
6186         btrfs_free_path(path);
6187         return ret;
6188 }
6189
6190 /*
6191  * Unpin the extent range in an error context and don't add the space back.
6192  * Errors are not propagated further.
6193  */
6194 void btrfs_error_unpin_extent_range(struct btrfs_fs_info *fs_info, u64 start, u64 end)
6195 {
6196         unpin_extent_range(fs_info, start, end, false);
6197 }
6198
6199 /*
6200  * It used to be that old block groups would be left around forever.
6201  * Iterating over them would be enough to trim unused space.  Since we
6202  * now automatically remove them, we also need to iterate over unallocated
6203  * space.
6204  *
6205  * We don't want a transaction for this since the discard may take a
6206  * substantial amount of time.  We don't require that a transaction be
6207  * running, but we do need to take a running transaction into account
6208  * to ensure that we're not discarding chunks that were released or
6209  * allocated in the current transaction.
6210  *
6211  * Holding the chunks lock will prevent other threads from allocating
6212  * or releasing chunks, but it won't prevent a running transaction
6213  * from committing and releasing the memory that the pending chunks
6214  * list head uses.  For that, we need to take a reference to the
6215  * transaction and hold the commit root sem.  We only need to hold
6216  * it while performing the free space search since we have already
6217  * held back allocations.
6218  */
6219 static int btrfs_trim_free_extents(struct btrfs_device *device, u64 *trimmed)
6220 {
6221         u64 start = BTRFS_DEVICE_RANGE_RESERVED, len = 0, end = 0;
6222         int ret;
6223
6224         *trimmed = 0;
6225
6226         /* Discard not supported = nothing to do. */
6227         if (!bdev_max_discard_sectors(device->bdev))
6228                 return 0;
6229
6230         /* Not writable = nothing to do. */
6231         if (!test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state))
6232                 return 0;
6233
6234         /* No free space = nothing to do. */
6235         if (device->total_bytes <= device->bytes_used)
6236                 return 0;
6237
6238         ret = 0;
6239
6240         while (1) {
6241                 struct btrfs_fs_info *fs_info = device->fs_info;
6242                 u64 bytes;
6243
6244                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
6245                 if (ret)
6246                         break;
6247
6248                 find_first_clear_extent_bit(&device->alloc_state, start,
6249                                             &start, &end,
6250                                             CHUNK_TRIMMED | CHUNK_ALLOCATED);
6251
6252                 /* Check if there are any CHUNK_* bits left */
6253                 if (start > device->total_bytes) {
6254                         WARN_ON(IS_ENABLED(CONFIG_BTRFS_DEBUG));
6255                         btrfs_warn_in_rcu(fs_info,
6256 "ignoring attempt to trim beyond device size: offset %llu length %llu device %s device size %llu",
6257                                           start, end - start + 1,
6258                                           btrfs_dev_name(device),
6259                                           device->total_bytes);
6260                         mutex_unlock(&fs_info->chunk_mutex);
6261                         ret = 0;
6262                         break;
6263                 }
6264
6265                 /* Ensure we skip the reserved space on each device. */
6266                 start = max_t(u64, start, BTRFS_DEVICE_RANGE_RESERVED);
6267
6268                 /*
6269                  * If find_first_clear_extent_bit find a range that spans the
6270                  * end of the device it will set end to -1, in this case it's up
6271                  * to the caller to trim the value to the size of the device.
6272                  */
6273                 end = min(end, device->total_bytes - 1);
6274
6275                 len = end - start + 1;
6276
6277                 /* We didn't find any extents */
6278                 if (!len) {
6279                         mutex_unlock(&fs_info->chunk_mutex);
6280                         ret = 0;
6281                         break;
6282                 }
6283
6284                 ret = btrfs_issue_discard(device->bdev, start, len,
6285                                           &bytes);
6286                 if (!ret)
6287                         set_extent_bit(&device->alloc_state, start,
6288                                        start + bytes - 1, CHUNK_TRIMMED, NULL);
6289                 mutex_unlock(&fs_info->chunk_mutex);
6290
6291                 if (ret)
6292                         break;
6293
6294                 start += len;
6295                 *trimmed += bytes;
6296
6297                 if (fatal_signal_pending(current)) {
6298                         ret = -ERESTARTSYS;
6299                         break;
6300                 }
6301
6302                 cond_resched();
6303         }
6304
6305         return ret;
6306 }
6307
6308 /*
6309  * Trim the whole filesystem by:
6310  * 1) trimming the free space in each block group
6311  * 2) trimming the unallocated space on each device
6312  *
6313  * This will also continue trimming even if a block group or device encounters
6314  * an error.  The return value will be the last error, or 0 if nothing bad
6315  * happens.
6316  */
6317 int btrfs_trim_fs(struct btrfs_fs_info *fs_info, struct fstrim_range *range)
6318 {
6319         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
6320         struct btrfs_block_group *cache = NULL;
6321         struct btrfs_device *device;
6322         u64 group_trimmed;
6323         u64 range_end = U64_MAX;
6324         u64 start;
6325         u64 end;
6326         u64 trimmed = 0;
6327         u64 bg_failed = 0;
6328         u64 dev_failed = 0;
6329         int bg_ret = 0;
6330         int dev_ret = 0;
6331         int ret = 0;
6332
6333         if (range->start == U64_MAX)
6334                 return -EINVAL;
6335
6336         /*
6337          * Check range overflow if range->len is set.
6338          * The default range->len is U64_MAX.
6339          */
6340         if (range->len != U64_MAX &&
6341             check_add_overflow(range->start, range->len, &range_end))
6342                 return -EINVAL;
6343
6344         cache = btrfs_lookup_first_block_group(fs_info, range->start);
6345         for (; cache; cache = btrfs_next_block_group(cache)) {
6346                 if (cache->start >= range_end) {
6347                         btrfs_put_block_group(cache);
6348                         break;
6349                 }
6350
6351                 start = max(range->start, cache->start);
6352                 end = min(range_end, cache->start + cache->length);
6353
6354                 if (end - start >= range->minlen) {
6355                         if (!btrfs_block_group_done(cache)) {
6356                                 ret = btrfs_cache_block_group(cache, true);
6357                                 if (ret) {
6358                                         bg_failed++;
6359                                         bg_ret = ret;
6360                                         continue;
6361                                 }
6362                         }
6363                         ret = btrfs_trim_block_group(cache,
6364                                                      &group_trimmed,
6365                                                      start,
6366                                                      end,
6367                                                      range->minlen);
6368
6369                         trimmed += group_trimmed;
6370                         if (ret) {
6371                                 bg_failed++;
6372                                 bg_ret = ret;
6373                                 continue;
6374                         }
6375                 }
6376         }
6377
6378         if (bg_failed)
6379                 btrfs_warn(fs_info,
6380                         "failed to trim %llu block group(s), last error %d",
6381                         bg_failed, bg_ret);
6382
6383         mutex_lock(&fs_devices->device_list_mutex);
6384         list_for_each_entry(device, &fs_devices->devices, dev_list) {
6385                 if (test_bit(BTRFS_DEV_STATE_MISSING, &device->dev_state))
6386                         continue;
6387
6388                 ret = btrfs_trim_free_extents(device, &group_trimmed);
6389                 if (ret) {
6390                         dev_failed++;
6391                         dev_ret = ret;
6392                         break;
6393                 }
6394
6395                 trimmed += group_trimmed;
6396         }
6397         mutex_unlock(&fs_devices->device_list_mutex);
6398
6399         if (dev_failed)
6400                 btrfs_warn(fs_info,
6401                         "failed to trim %llu device(s), last error %d",
6402                         dev_failed, dev_ret);
6403         range->len = trimmed;
6404         if (bg_ret)
6405                 return bg_ret;
6406         return dev_ret;
6407 }