fs/btrfs/extent-tree.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2007 Oracle.  All rights reserved.
   4  */
   5
   6 #include <linux/sched.h>
   7 #include <linux/sched/signal.h>
   8 #include <linux/pagemap.h>
   9 #include <linux/writeback.h>
  10 #include <linux/blkdev.h>
  11 #include <linux/sort.h>
  12 #include <linux/rcupdate.h>
  13 #include <linux/kthread.h>
  14 #include <linux/slab.h>
  15 #include <linux/ratelimit.h>
  16 #include <linux/percpu_counter.h>
  17 #include <linux/lockdep.h>
  18 #include <linux/crc32c.h>
  19 #include "ctree.h"
  20 #include "extent-tree.h"
  21 #include "tree-log.h"
  22 #include "disk-io.h"
  23 #include "print-tree.h"
  24 #include "volumes.h"
  25 #include "raid56.h"
  26 #include "locking.h"
  27 #include "free-space-cache.h"
  28 #include "free-space-tree.h"
  29 #include "sysfs.h"
  30 #include "qgroup.h"
  31 #include "ref-verify.h"
  32 #include "space-info.h"
  33 #include "block-rsv.h"
  34 #include "delalloc-space.h"
  35 #include "discard.h"
  36 #include "rcu-string.h"
  37 #include "zoned.h"
  38 #include "dev-replace.h"
  39 #include "fs.h"
  40 #include "accessors.h"
  41 #include "root-tree.h"
  42 #include "file-item.h"
  43 #include "orphan.h"
  44 #include "tree-checker.h"
  45
  46 #undef SCRAMBLE_DELAYED_REFS
  47
  48
  49 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  50                                struct btrfs_delayed_ref_node *node, u64 parent,
  51                                u64 root_objectid, u64 owner_objectid,
  52                                u64 owner_offset, int refs_to_drop,
  53                                struct btrfs_delayed_extent_op *extra_op);
  54 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  55                                     struct extent_buffer *leaf,
  56                                     struct btrfs_extent_item *ei);
  57 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  58                                       u64 parent, u64 root_objectid,
  59                                       u64 flags, u64 owner, u64 offset,
  60                                       struct btrfs_key *ins, int ref_mod);
  61 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  62                                      struct btrfs_delayed_ref_node *node,
  63                                      struct btrfs_delayed_extent_op *extent_op);
  64 static int find_next_key(struct btrfs_path *path, int level,
  65                          struct btrfs_key *key);
  66
  67 static int block_group_bits(struct btrfs_block_group *cache, u64 bits)
  68 {
  69         return (cache->flags & bits) == bits;
  70 }
  71
  72 /* simple helper to search for an existing data extent at a given offset */
  73 int btrfs_lookup_data_extent(struct btrfs_fs_info *fs_info, u64 start, u64 len)
  74 {
  75         struct btrfs_root *root = btrfs_extent_root(fs_info, start);
  76         int ret;
  77         struct btrfs_key key;
  78         struct btrfs_path *path;
  79
  80         path = btrfs_alloc_path();
  81         if (!path)
  82                 return -ENOMEM;
  83
  84         key.objectid = start;
  85         key.offset = len;
  86         key.type = BTRFS_EXTENT_ITEM_KEY;
  87         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
  88         btrfs_free_path(path);
  89         return ret;
  90 }
  91
  92 /*
  93  * helper function to lookup reference count and flags of a tree block.
  94  *
  95  * the head node for delayed ref is used to store the sum of all the
  96  * reference count modifications queued up in the rbtree. the head
  97  * node may also store the extent flags to set. This way you can check
  98  * to see what the reference count and extent flags would be if all of
  99  * the delayed refs are not processed.
 100  */
 101 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 102                              struct btrfs_fs_info *fs_info, u64 bytenr,
 103                              u64 offset, int metadata, u64 *refs, u64 *flags)
 104 {
 105         struct btrfs_root *extent_root;
 106         struct btrfs_delayed_ref_head *head;
 107         struct btrfs_delayed_ref_root *delayed_refs;
 108         struct btrfs_path *path;
 109         struct btrfs_extent_item *ei;
 110         struct extent_buffer *leaf;
 111         struct btrfs_key key;
 112         u32 item_size;
 113         u64 num_refs;
 114         u64 extent_flags;
 115         int ret;
 116
 117         /*
 118          * If we don't have skinny metadata, don't bother doing anything
 119          * different
 120          */
 121         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA)) {
 122                 offset = fs_info->nodesize;
 123                 metadata = 0;
 124         }
 125
 126         path = btrfs_alloc_path();
 127         if (!path)
 128                 return -ENOMEM;
 129
 130         if (!trans) {
 131                 path->skip_locking = 1;
 132                 path->search_commit_root = 1;
 133         }
 134
 135 search_again:
 136         key.objectid = bytenr;
 137         key.offset = offset;
 138         if (metadata)
 139                 key.type = BTRFS_METADATA_ITEM_KEY;
 140         else
 141                 key.type = BTRFS_EXTENT_ITEM_KEY;
 142
 143         extent_root = btrfs_extent_root(fs_info, bytenr);
 144         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 145         if (ret < 0)
 146                 goto out_free;
 147
 148         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 149                 if (path->slots[0]) {
 150                         path->slots[0]--;
 151                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 152                                               path->slots[0]);
 153                         if (key.objectid == bytenr &&
 154                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 155                             key.offset == fs_info->nodesize)
 156                                 ret = 0;
 157                 }
 158         }
 159
 160         if (ret == 0) {
 161                 leaf = path->nodes[0];
 162                 item_size = btrfs_item_size(leaf, path->slots[0]);
 163                 if (item_size >= sizeof(*ei)) {
 164                         ei = btrfs_item_ptr(leaf, path->slots[0],
 165                                             struct btrfs_extent_item);
 166                         num_refs = btrfs_extent_refs(leaf, ei);
 167                         extent_flags = btrfs_extent_flags(leaf, ei);
 168                 } else {
 169                         ret = -EUCLEAN;
 170                         btrfs_err(fs_info,
 171                         "unexpected extent item size, has %u expect >= %zu",
 172                                   item_size, sizeof(*ei));
 173                         if (trans)
 174                                 btrfs_abort_transaction(trans, ret);
 175                         else
 176                                 btrfs_handle_fs_error(fs_info, ret, NULL);
 177
 178                         goto out_free;
 179                 }
 180
 181                 BUG_ON(num_refs == 0);
 182         } else {
 183                 num_refs = 0;
 184                 extent_flags = 0;
 185                 ret = 0;
 186         }
 187
 188         if (!trans)
 189                 goto out;
 190
 191         delayed_refs = &trans->transaction->delayed_refs;
 192         spin_lock(&delayed_refs->lock);
 193         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
 194         if (head) {
 195                 if (!mutex_trylock(&head->mutex)) {
 196                         refcount_inc(&head->refs);
 197                         spin_unlock(&delayed_refs->lock);
 198
 199                         btrfs_release_path(path);
 200
 201                         /*
 202                          * Mutex was contended, block until it's released and try
 203                          * again
 204                          */
 205                         mutex_lock(&head->mutex);
 206                         mutex_unlock(&head->mutex);
 207                         btrfs_put_delayed_ref_head(head);
 208                         goto search_again;
 209                 }
 210                 spin_lock(&head->lock);
 211                 if (head->extent_op && head->extent_op->update_flags)
 212                         extent_flags |= head->extent_op->flags_to_set;
 213                 else
 214                         BUG_ON(num_refs == 0);
 215
 216                 num_refs += head->ref_mod;
 217                 spin_unlock(&head->lock);
 218                 mutex_unlock(&head->mutex);
 219         }
 220         spin_unlock(&delayed_refs->lock);
 221 out:
 222         WARN_ON(num_refs == 0);
 223         if (refs)
 224                 *refs = num_refs;
 225         if (flags)
 226                 *flags = extent_flags;
 227 out_free:
 228         btrfs_free_path(path);
 229         return ret;
 230 }
 231
 232 /*
 233  * Back reference rules.  Back refs have three main goals:
 234  *
 235  * 1) differentiate between all holders of references to an extent so that
 236  *    when a reference is dropped we can make sure it was a valid reference
 237  *    before freeing the extent.
 238  *
 239  * 2) Provide enough information to quickly find the holders of an extent
 240  *    if we notice a given block is corrupted or bad.
 241  *
 242  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 243  *    maintenance.  This is actually the same as #2, but with a slightly
 244  *    different use case.
 245  *
 246  * There are two kinds of back refs. The implicit back refs is optimized
 247  * for pointers in non-shared tree blocks. For a given pointer in a block,
 248  * back refs of this kind provide information about the block's owner tree
 249  * and the pointer's key. These information allow us to find the block by
 250  * b-tree searching. The full back refs is for pointers in tree blocks not
 251  * referenced by their owner trees. The location of tree block is recorded
 252  * in the back refs. Actually the full back refs is generic, and can be
 253  * used in all cases the implicit back refs is used. The major shortcoming
 254  * of the full back refs is its overhead. Every time a tree block gets
 255  * COWed, we have to update back refs entry for all pointers in it.
 256  *
 257  * For a newly allocated tree block, we use implicit back refs for
 258  * pointers in it. This means most tree related operations only involve
 259  * implicit back refs. For a tree block created in old transaction, the
 260  * only way to drop a reference to it is COW it. So we can detect the
 261  * event that tree block loses its owner tree's reference and do the
 262  * back refs conversion.
 263  *
 264  * When a tree block is COWed through a tree, there are four cases:
 265  *
 266  * The reference count of the block is one and the tree is the block's
 267  * owner tree. Nothing to do in this case.
 268  *
 269  * The reference count of the block is one and the tree is not the
 270  * block's owner tree. In this case, full back refs is used for pointers
 271  * in the block. Remove these full back refs, add implicit back refs for
 272  * every pointers in the new block.
 273  *
 274  * The reference count of the block is greater than one and the tree is
 275  * the block's owner tree. In this case, implicit back refs is used for
 276  * pointers in the block. Add full back refs for every pointers in the
 277  * block, increase lower level extents' reference counts. The original
 278  * implicit back refs are entailed to the new block.
 279  *
 280  * The reference count of the block is greater than one and the tree is
 281  * not the block's owner tree. Add implicit back refs for every pointer in
 282  * the new block, increase lower level extents' reference count.
 283  *
 284  * Back Reference Key composing:
 285  *
 286  * The key objectid corresponds to the first byte in the extent,
 287  * The key type is used to differentiate between types of back refs.
 288  * There are different meanings of the key offset for different types
 289  * of back refs.
 290  *
 291  * File extents can be referenced by:
 292  *
 293  * - multiple snapshots, subvolumes, or different generations in one subvol
 294  * - different files inside a single subvolume
 295  * - different offsets inside a file (bookend extents in file.c)
 296  *
 297  * The extent ref structure for the implicit back refs has fields for:
 298  *
 299  * - Objectid of the subvolume root
 300  * - objectid of the file holding the reference
 301  * - original offset in the file
 302  * - how many bookend extents
 303  *
 304  * The key offset for the implicit back refs is hash of the first
 305  * three fields.
 306  *
 307  * The extent ref structure for the full back refs has field for:
 308  *
 309  * - number of pointers in the tree leaf
 310  *
 311  * The key offset for the implicit back refs is the first byte of
 312  * the tree leaf
 313  *
 314  * When a file extent is allocated, The implicit back refs is used.
 315  * the fields are filled in:
 316  *
 317  *     (root_key.objectid, inode objectid, offset in file, 1)
 318  *
 319  * When a file extent is removed file truncation, we find the
 320  * corresponding implicit back refs and check the following fields:
 321  *
 322  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
 323  *
 324  * Btree extents can be referenced by:
 325  *
 326  * - Different subvolumes
 327  *
 328  * Both the implicit back refs and the full back refs for tree blocks
 329  * only consist of key. The key offset for the implicit back refs is
 330  * objectid of block's owner tree. The key offset for the full back refs
 331  * is the first byte of parent block.
 332  *
 333  * When implicit back refs is used, information about the lowest key and
 334  * level of the tree block are required. These information are stored in
 335  * tree block info structure.
 336  */
 337
 338 /*
 339  * is_data == BTRFS_REF_TYPE_BLOCK, tree block type is required,
 340  * is_data == BTRFS_REF_TYPE_DATA, data type is requiried,
 341  * is_data == BTRFS_REF_TYPE_ANY, either type is OK.
 342  */
 343 int btrfs_get_extent_inline_ref_type(const struct extent_buffer *eb,
 344                                      struct btrfs_extent_inline_ref *iref,
 345                                      enum btrfs_inline_ref_type is_data)
 346 {
 347         int type = btrfs_extent_inline_ref_type(eb, iref);
 348         u64 offset = btrfs_extent_inline_ref_offset(eb, iref);
 349
 350         if (type == BTRFS_TREE_BLOCK_REF_KEY ||
 351             type == BTRFS_SHARED_BLOCK_REF_KEY ||
 352             type == BTRFS_SHARED_DATA_REF_KEY ||
 353             type == BTRFS_EXTENT_DATA_REF_KEY) {
 354                 if (is_data == BTRFS_REF_TYPE_BLOCK) {
 355                         if (type == BTRFS_TREE_BLOCK_REF_KEY)
 356                                 return type;
 357                         if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
 358                                 ASSERT(eb->fs_info);
 359                                 /*
 360                                  * Every shared one has parent tree block,
 361                                  * which must be aligned to sector size.
 362                                  */
 363                                 if (offset &&
 364                                     IS_ALIGNED(offset, eb->fs_info->sectorsize))
 365                                         return type;
 366                         }
 367                 } else if (is_data == BTRFS_REF_TYPE_DATA) {
 368                         if (type == BTRFS_EXTENT_DATA_REF_KEY)
 369                                 return type;
 370                         if (type == BTRFS_SHARED_DATA_REF_KEY) {
 371                                 ASSERT(eb->fs_info);
 372                                 /*
 373                                  * Every shared one has parent tree block,
 374                                  * which must be aligned to sector size.
 375                                  */
 376                                 if (offset &&
 377                                     IS_ALIGNED(offset, eb->fs_info->sectorsize))
 378                                         return type;
 379                         }
 380                 } else {
 381                         ASSERT(is_data == BTRFS_REF_TYPE_ANY);
 382                         return type;
 383                 }
 384         }
 385
 386         WARN_ON(1);
 387         btrfs_print_leaf(eb);
 388         btrfs_err(eb->fs_info,
 389                   "eb %llu iref 0x%lx invalid extent inline ref type %d",
 390                   eb->start, (unsigned long)iref, type);
 391
 392         return BTRFS_REF_TYPE_INVALID;
 393 }
 394
 395 u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
 396 {
 397         u32 high_crc = ~(u32)0;
 398         u32 low_crc = ~(u32)0;
 399         __le64 lenum;
 400
 401         lenum = cpu_to_le64(root_objectid);
 402         high_crc = btrfs_crc32c(high_crc, &lenum, sizeof(lenum));
 403         lenum = cpu_to_le64(owner);
 404         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
 405         lenum = cpu_to_le64(offset);
 406         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
 407
 408         return ((u64)high_crc << 31) ^ (u64)low_crc;
 409 }
 410
 411 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
 412                                      struct btrfs_extent_data_ref *ref)
 413 {
 414         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
 415                                     btrfs_extent_data_ref_objectid(leaf, ref),
 416                                     btrfs_extent_data_ref_offset(leaf, ref));
 417 }
 418
 419 static int match_extent_data_ref(struct extent_buffer *leaf,
 420                                  struct btrfs_extent_data_ref *ref,
 421                                  u64 root_objectid, u64 owner, u64 offset)
 422 {
 423         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
 424             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
 425             btrfs_extent_data_ref_offset(leaf, ref) != offset)
 426                 return 0;
 427         return 1;
 428 }
 429
 430 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
 431                                            struct btrfs_path *path,
 432                                            u64 bytenr, u64 parent,
 433                                            u64 root_objectid,
 434                                            u64 owner, u64 offset)
 435 {
 436         struct btrfs_root *root = btrfs_extent_root(trans->fs_info, bytenr);
 437         struct btrfs_key key;
 438         struct btrfs_extent_data_ref *ref;
 439         struct extent_buffer *leaf;
 440         u32 nritems;
 441         int ret;
 442         int recow;
 443         int err = -ENOENT;
 444
 445         key.objectid = bytenr;
 446         if (parent) {
 447                 key.type = BTRFS_SHARED_DATA_REF_KEY;
 448                 key.offset = parent;
 449         } else {
 450                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
 451                 key.offset = hash_extent_data_ref(root_objectid,
 452                                                   owner, offset);
 453         }
 454 again:
 455         recow = 0;
 456         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
 457         if (ret < 0) {
 458                 err = ret;
 459                 goto fail;
 460         }
 461
 462         if (parent) {
 463                 if (!ret)
 464                         return 0;
 465                 goto fail;
 466         }
 467
 468         leaf = path->nodes[0];
 469         nritems = btrfs_header_nritems(leaf);
 470         while (1) {
 471                 if (path->slots[0] >= nritems) {
 472                         ret = btrfs_next_leaf(root, path);
 473                         if (ret < 0)
 474                                 err = ret;
 475                         if (ret)
 476                                 goto fail;
 477
 478                         leaf = path->nodes[0];
 479                         nritems = btrfs_header_nritems(leaf);
 480                         recow = 1;
 481                 }
 482
 483                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 484                 if (key.objectid != bytenr ||
 485                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
 486                         goto fail;
 487
 488                 ref = btrfs_item_ptr(leaf, path->slots[0],
 489                                      struct btrfs_extent_data_ref);
 490
 491                 if (match_extent_data_ref(leaf, ref, root_objectid,
 492                                           owner, offset)) {
 493                         if (recow) {
 494                                 btrfs_release_path(path);
 495                                 goto again;
 496                         }
 497                         err = 0;
 498                         break;
 499                 }
 500                 path->slots[0]++;
 501         }
 502 fail:
 503         return err;
 504 }
 505
 506 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
 507                                            struct btrfs_path *path,
 508                                            u64 bytenr, u64 parent,
 509                                            u64 root_objectid, u64 owner,
 510                                            u64 offset, int refs_to_add)
 511 {
 512         struct btrfs_root *root = btrfs_extent_root(trans->fs_info, bytenr);
 513         struct btrfs_key key;
 514         struct extent_buffer *leaf;
 515         u32 size;
 516         u32 num_refs;
 517         int ret;
 518
 519         key.objectid = bytenr;
 520         if (parent) {
 521                 key.type = BTRFS_SHARED_DATA_REF_KEY;
 522                 key.offset = parent;
 523                 size = sizeof(struct btrfs_shared_data_ref);
 524         } else {
 525                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
 526                 key.offset = hash_extent_data_ref(root_objectid,
 527                                                   owner, offset);
 528                 size = sizeof(struct btrfs_extent_data_ref);
 529         }
 530
 531         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
 532         if (ret && ret != -EEXIST)
 533                 goto fail;
 534
 535         leaf = path->nodes[0];
 536         if (parent) {
 537                 struct btrfs_shared_data_ref *ref;
 538                 ref = btrfs_item_ptr(leaf, path->slots[0],
 539                                      struct btrfs_shared_data_ref);
 540                 if (ret == 0) {
 541                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
 542                 } else {
 543                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
 544                         num_refs += refs_to_add;
 545                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
 546                 }
 547         } else {
 548                 struct btrfs_extent_data_ref *ref;
 549                 while (ret == -EEXIST) {
 550                         ref = btrfs_item_ptr(leaf, path->slots[0],
 551                                              struct btrfs_extent_data_ref);
 552                         if (match_extent_data_ref(leaf, ref, root_objectid,
 553                                                   owner, offset))
 554                                 break;
 555                         btrfs_release_path(path);
 556                         key.offset++;
 557                         ret = btrfs_insert_empty_item(trans, root, path, &key,
 558                                                       size);
 559                         if (ret && ret != -EEXIST)
 560                                 goto fail;
 561
 562                         leaf = path->nodes[0];
 563                 }
 564                 ref = btrfs_item_ptr(leaf, path->slots[0],
 565                                      struct btrfs_extent_data_ref);
 566                 if (ret == 0) {
 567                         btrfs_set_extent_data_ref_root(leaf, ref,
 568                                                        root_objectid);
 569                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
 570                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
 571                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
 572                 } else {
 573                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
 574                         num_refs += refs_to_add;
 575                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
 576                 }
 577         }
 578         btrfs_mark_buffer_dirty(trans, leaf);
 579         ret = 0;
 580 fail:
 581         btrfs_release_path(path);
 582         return ret;
 583 }
 584
 585 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
 586                                            struct btrfs_root *root,
 587                                            struct btrfs_path *path,
 588                                            int refs_to_drop)
 589 {
 590         struct btrfs_key key;
 591         struct btrfs_extent_data_ref *ref1 = NULL;
 592         struct btrfs_shared_data_ref *ref2 = NULL;
 593         struct extent_buffer *leaf;
 594         u32 num_refs = 0;
 595         int ret = 0;
 596
 597         leaf = path->nodes[0];
 598         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 599
 600         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
 601                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
 602                                       struct btrfs_extent_data_ref);
 603                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
 604         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
 605                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
 606                                       struct btrfs_shared_data_ref);
 607                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
 608         } else {
 609                 btrfs_err(trans->fs_info,
 610                           "unrecognized backref key (%llu %u %llu)",
 611                           key.objectid, key.type, key.offset);
 612                 btrfs_abort_transaction(trans, -EUCLEAN);
 613                 return -EUCLEAN;
 614         }
 615
 616         BUG_ON(num_refs < refs_to_drop);
 617         num_refs -= refs_to_drop;
 618
 619         if (num_refs == 0) {
 620                 ret = btrfs_del_item(trans, root, path);
 621         } else {
 622                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
 623                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
 624                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
 625                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
 626                 btrfs_mark_buffer_dirty(trans, leaf);
 627         }
 628         return ret;
 629 }
 630
 631 static noinline u32 extent_data_ref_count(struct btrfs_path *path,
 632                                           struct btrfs_extent_inline_ref *iref)
 633 {
 634         struct btrfs_key key;
 635         struct extent_buffer *leaf;
 636         struct btrfs_extent_data_ref *ref1;
 637         struct btrfs_shared_data_ref *ref2;
 638         u32 num_refs = 0;
 639         int type;
 640
 641         leaf = path->nodes[0];
 642         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 643
 644         if (iref) {
 645                 /*
 646                  * If type is invalid, we should have bailed out earlier than
 647                  * this call.
 648                  */
 649                 type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
 650                 ASSERT(type != BTRFS_REF_TYPE_INVALID);
 651                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
 652                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
 653                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
 654                 } else {
 655                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
 656                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
 657                 }
 658         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
 659                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
 660                                       struct btrfs_extent_data_ref);
 661                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
 662         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
 663                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
 664                                       struct btrfs_shared_data_ref);
 665                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
 666         } else {
 667                 WARN_ON(1);
 668         }
 669         return num_refs;
 670 }
 671
 672 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
 673                                           struct btrfs_path *path,
 674                                           u64 bytenr, u64 parent,
 675                                           u64 root_objectid)
 676 {
 677         struct btrfs_root *root = btrfs_extent_root(trans->fs_info, bytenr);
 678         struct btrfs_key key;
 679         int ret;
 680
 681         key.objectid = bytenr;
 682         if (parent) {
 683                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
 684                 key.offset = parent;
 685         } else {
 686                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
 687                 key.offset = root_objectid;
 688         }
 689
 690         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
 691         if (ret > 0)
 692                 ret = -ENOENT;
 693         return ret;
 694 }
 695
 696 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
 697                                           struct btrfs_path *path,
 698                                           u64 bytenr, u64 parent,
 699                                           u64 root_objectid)
 700 {
 701         struct btrfs_root *root = btrfs_extent_root(trans->fs_info, bytenr);
 702         struct btrfs_key key;
 703         int ret;
 704
 705         key.objectid = bytenr;
 706         if (parent) {
 707                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
 708                 key.offset = parent;
 709         } else {
 710                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
 711                 key.offset = root_objectid;
 712         }
 713
 714         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
 715         btrfs_release_path(path);
 716         return ret;
 717 }
 718
 719 static inline int extent_ref_type(u64 parent, u64 owner)
 720 {
 721         int type;
 722         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
 723                 if (parent > 0)
 724                         type = BTRFS_SHARED_BLOCK_REF_KEY;
 725                 else
 726                         type = BTRFS_TREE_BLOCK_REF_KEY;
 727         } else {
 728                 if (parent > 0)
 729                         type = BTRFS_SHARED_DATA_REF_KEY;
 730                 else
 731                         type = BTRFS_EXTENT_DATA_REF_KEY;
 732         }
 733         return type;
 734 }
 735
 736 static int find_next_key(struct btrfs_path *path, int level,
 737                          struct btrfs_key *key)
 738
 739 {
 740         for (; level < BTRFS_MAX_LEVEL; level++) {
 741                 if (!path->nodes[level])
 742                         break;
 743                 if (path->slots[level] + 1 >=
 744                     btrfs_header_nritems(path->nodes[level]))
 745                         continue;
 746                 if (level == 0)
 747                         btrfs_item_key_to_cpu(path->nodes[level], key,
 748                                               path->slots[level] + 1);
 749                 else
 750                         btrfs_node_key_to_cpu(path->nodes[level], key,
 751                                               path->slots[level] + 1);
 752                 return 0;
 753         }
 754         return 1;
 755 }
 756
 757 /*
 758  * look for inline back ref. if back ref is found, *ref_ret is set
 759  * to the address of inline back ref, and 0 is returned.
 760  *
 761  * if back ref isn't found, *ref_ret is set to the address where it
 762  * should be inserted, and -ENOENT is returned.
 763  *
 764  * if insert is true and there are too many inline back refs, the path
 765  * points to the extent item, and -EAGAIN is returned.
 766  *
 767  * NOTE: inline back refs are ordered in the same way that back ref
 768  *       items in the tree are ordered.
 769  */
 770 static noinline_for_stack
 771 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
 772                                  struct btrfs_path *path,
 773                                  struct btrfs_extent_inline_ref **ref_ret,
 774                                  u64 bytenr, u64 num_bytes,
 775                                  u64 parent, u64 root_objectid,
 776                                  u64 owner, u64 offset, int insert)
 777 {
 778         struct btrfs_fs_info *fs_info = trans->fs_info;
 779         struct btrfs_root *root = btrfs_extent_root(fs_info, bytenr);
 780         struct btrfs_key key;
 781         struct extent_buffer *leaf;
 782         struct btrfs_extent_item *ei;
 783         struct btrfs_extent_inline_ref *iref;
 784         u64 flags;
 785         u64 item_size;
 786         unsigned long ptr;
 787         unsigned long end;
 788         int extra_size;
 789         int type;
 790         int want;
 791         int ret;
 792         int err = 0;
 793         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
 794         int needed;
 795
 796         key.objectid = bytenr;
 797         key.type = BTRFS_EXTENT_ITEM_KEY;
 798         key.offset = num_bytes;
 799
 800         want = extent_ref_type(parent, owner);
 801         if (insert) {
 802                 extra_size = btrfs_extent_inline_ref_size(want);
 803                 path->search_for_extension = 1;
 804                 path->keep_locks = 1;
 805         } else
 806                 extra_size = -1;
 807
 808         /*
 809          * Owner is our level, so we can just add one to get the level for the
 810          * block we are interested in.
 811          */
 812         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
 813                 key.type = BTRFS_METADATA_ITEM_KEY;
 814                 key.offset = owner;
 815         }
 816
 817 again:
 818         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
 819         if (ret < 0) {
 820                 err = ret;
 821                 goto out;
 822         }
 823
 824         /*
 825          * We may be a newly converted file system which still has the old fat
 826          * extent entries for metadata, so try and see if we have one of those.
 827          */
 828         if (ret > 0 && skinny_metadata) {
 829                 skinny_metadata = false;
 830                 if (path->slots[0]) {
 831                         path->slots[0]--;
 832                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 833                                               path->slots[0]);
 834                         if (key.objectid == bytenr &&
 835                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 836                             key.offset == num_bytes)
 837                                 ret = 0;
 838                 }
 839                 if (ret) {
 840                         key.objectid = bytenr;
 841                         key.type = BTRFS_EXTENT_ITEM_KEY;
 842                         key.offset = num_bytes;
 843                         btrfs_release_path(path);
 844                         goto again;
 845                 }
 846         }
 847
 848         if (ret && !insert) {
 849                 err = -ENOENT;
 850                 goto out;
 851         } else if (WARN_ON(ret)) {
 852                 btrfs_print_leaf(path->nodes[0]);
 853                 btrfs_err(fs_info,
 854 "extent item not found for insert, bytenr %llu num_bytes %llu parent %llu root_objectid %llu owner %llu offset %llu",
 855                           bytenr, num_bytes, parent, root_objectid, owner,
 856                           offset);
 857                 err = -EIO;
 858                 goto out;
 859         }
 860
 861         leaf = path->nodes[0];
 862         item_size = btrfs_item_size(leaf, path->slots[0]);
 863         if (unlikely(item_size < sizeof(*ei))) {
 864                 err = -EUCLEAN;
 865                 btrfs_err(fs_info,
 866                           "unexpected extent item size, has %llu expect >= %zu",
 867                           item_size, sizeof(*ei));
 868                 btrfs_abort_transaction(trans, err);
 869                 goto out;
 870         }
 871
 872         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
 873         flags = btrfs_extent_flags(leaf, ei);
 874
 875         ptr = (unsigned long)(ei + 1);
 876         end = (unsigned long)ei + item_size;
 877
 878         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
 879                 ptr += sizeof(struct btrfs_tree_block_info);
 880                 BUG_ON(ptr > end);
 881         }
 882
 883         if (owner >= BTRFS_FIRST_FREE_OBJECTID)
 884                 needed = BTRFS_REF_TYPE_DATA;
 885         else
 886                 needed = BTRFS_REF_TYPE_BLOCK;
 887
 888         err = -ENOENT;
 889         while (1) {
 890                 if (ptr >= end) {
 891                         if (ptr > end) {
 892                                 err = -EUCLEAN;
 893                                 btrfs_print_leaf(path->nodes[0]);
 894                                 btrfs_crit(fs_info,
 895 "overrun extent record at slot %d while looking for inline extent for root %llu owner %llu offset %llu parent %llu",
 896                                         path->slots[0], root_objectid, owner, offset, parent);
 897                         }
 898                         break;
 899                 }
 900                 iref = (struct btrfs_extent_inline_ref *)ptr;
 901                 type = btrfs_get_extent_inline_ref_type(leaf, iref, needed);
 902                 if (type == BTRFS_REF_TYPE_INVALID) {
 903                         err = -EUCLEAN;
 904                         goto out;
 905                 }
 906
 907                 if (want < type)
 908                         break;
 909                 if (want > type) {
 910                         ptr += btrfs_extent_inline_ref_size(type);
 911                         continue;
 912                 }
 913
 914                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
 915                         struct btrfs_extent_data_ref *dref;
 916                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
 917                         if (match_extent_data_ref(leaf, dref, root_objectid,
 918                                                   owner, offset)) {
 919                                 err = 0;
 920                                 break;
 921                         }
 922                         if (hash_extent_data_ref_item(leaf, dref) <
 923                             hash_extent_data_ref(root_objectid, owner, offset))
 924                                 break;
 925                 } else {
 926                         u64 ref_offset;
 927                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
 928                         if (parent > 0) {
 929                                 if (parent == ref_offset) {
 930                                         err = 0;
 931                                         break;
 932                                 }
 933                                 if (ref_offset < parent)
 934                                         break;
 935                         } else {
 936                                 if (root_objectid == ref_offset) {
 937                                         err = 0;
 938                                         break;
 939                                 }
 940                                 if (ref_offset < root_objectid)
 941                                         break;
 942                         }
 943                 }
 944                 ptr += btrfs_extent_inline_ref_size(type);
 945         }
 946         if (err == -ENOENT && insert) {
 947                 if (item_size + extra_size >=
 948                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
 949                         err = -EAGAIN;
 950                         goto out;
 951                 }
 952                 /*
 953                  * To add new inline back ref, we have to make sure
 954                  * there is no corresponding back ref item.
 955                  * For simplicity, we just do not add new inline back
 956                  * ref if there is any kind of item for this block
 957                  */
 958                 if (find_next_key(path, 0, &key) == 0 &&
 959                     key.objectid == bytenr &&
 960                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
 961                         err = -EAGAIN;
 962                         goto out;
 963                 }
 964         }
 965         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
 966 out:
 967         if (insert) {
 968                 path->keep_locks = 0;
 969                 path->search_for_extension = 0;
 970                 btrfs_unlock_up_safe(path, 1);
 971         }
 972         return err;
 973 }
 974
 975 /*
 976  * helper to add new inline back ref
 977  */
 978 static noinline_for_stack
 979 void setup_inline_extent_backref(struct btrfs_trans_handle *trans,
 980                                  struct btrfs_path *path,
 981                                  struct btrfs_extent_inline_ref *iref,
 982                                  u64 parent, u64 root_objectid,
 983                                  u64 owner, u64 offset, int refs_to_add,
 984                                  struct btrfs_delayed_extent_op *extent_op)
 985 {
 986         struct extent_buffer *leaf;
 987         struct btrfs_extent_item *ei;
 988         unsigned long ptr;
 989         unsigned long end;
 990         unsigned long item_offset;
 991         u64 refs;
 992         int size;
 993         int type;
 994
 995         leaf = path->nodes[0];
 996         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
 997         item_offset = (unsigned long)iref - (unsigned long)ei;
 998
 999         type = extent_ref_type(parent, owner);
1000         size = btrfs_extent_inline_ref_size(type);
1001
1002         btrfs_extend_item(trans, path, size);
1003
1004         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1005         refs = btrfs_extent_refs(leaf, ei);
1006         refs += refs_to_add;
1007         btrfs_set_extent_refs(leaf, ei, refs);
1008         if (extent_op)
1009                 __run_delayed_extent_op(extent_op, leaf, ei);
1010
1011         ptr = (unsigned long)ei + item_offset;
1012         end = (unsigned long)ei + btrfs_item_size(leaf, path->slots[0]);
1013         if (ptr < end - size)
1014                 memmove_extent_buffer(leaf, ptr + size, ptr,
1015                                       end - size - ptr);
1016
1017         iref = (struct btrfs_extent_inline_ref *)ptr;
1018         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1019         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1020                 struct btrfs_extent_data_ref *dref;
1021                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1022                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1023                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1024                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1025                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1026         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1027                 struct btrfs_shared_data_ref *sref;
1028                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1029                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1030                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1031         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1032                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1033         } else {
1034                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1035         }
1036         btrfs_mark_buffer_dirty(trans, leaf);
1037 }
1038
1039 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1040                                  struct btrfs_path *path,
1041                                  struct btrfs_extent_inline_ref **ref_ret,
1042                                  u64 bytenr, u64 num_bytes, u64 parent,
1043                                  u64 root_objectid, u64 owner, u64 offset)
1044 {
1045         int ret;
1046
1047         ret = lookup_inline_extent_backref(trans, path, ref_ret, bytenr,
1048                                            num_bytes, parent, root_objectid,
1049                                            owner, offset, 0);
1050         if (ret != -ENOENT)
1051                 return ret;
1052
1053         btrfs_release_path(path);
1054         *ref_ret = NULL;
1055
1056         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1057                 ret = lookup_tree_block_ref(trans, path, bytenr, parent,
1058                                             root_objectid);
1059         } else {
1060                 ret = lookup_extent_data_ref(trans, path, bytenr, parent,
1061                                              root_objectid, owner, offset);
1062         }
1063         return ret;
1064 }
1065
1066 /*
1067  * helper to update/remove inline back ref
1068  */
1069 static noinline_for_stack int update_inline_extent_backref(
1070                                   struct btrfs_trans_handle *trans,
1071                                   struct btrfs_path *path,
1072                                   struct btrfs_extent_inline_ref *iref,
1073                                   int refs_to_mod,
1074                                   struct btrfs_delayed_extent_op *extent_op)
1075 {
1076         struct extent_buffer *leaf = path->nodes[0];
1077         struct btrfs_fs_info *fs_info = leaf->fs_info;
1078         struct btrfs_extent_item *ei;
1079         struct btrfs_extent_data_ref *dref = NULL;
1080         struct btrfs_shared_data_ref *sref = NULL;
1081         unsigned long ptr;
1082         unsigned long end;
1083         u32 item_size;
1084         int size;
1085         int type;
1086         u64 refs;
1087
1088         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1089         refs = btrfs_extent_refs(leaf, ei);
1090         if (unlikely(refs_to_mod < 0 && refs + refs_to_mod <= 0)) {
1091                 struct btrfs_key key;
1092                 u32 extent_size;
1093
1094                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1095                 if (key.type == BTRFS_METADATA_ITEM_KEY)
1096                         extent_size = fs_info->nodesize;
1097                 else
1098                         extent_size = key.offset;
1099                 btrfs_print_leaf(leaf);
1100                 btrfs_err(fs_info,
1101         "invalid refs_to_mod for extent %llu num_bytes %u, has %d expect >= -%llu",
1102                           key.objectid, extent_size, refs_to_mod, refs);
1103                 return -EUCLEAN;
1104         }
1105         refs += refs_to_mod;
1106         btrfs_set_extent_refs(leaf, ei, refs);
1107         if (extent_op)
1108                 __run_delayed_extent_op(extent_op, leaf, ei);
1109
1110         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_ANY);
1111         /*
1112          * Function btrfs_get_extent_inline_ref_type() has already printed
1113          * error messages.
1114          */
1115         if (unlikely(type == BTRFS_REF_TYPE_INVALID))
1116                 return -EUCLEAN;
1117
1118         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1119                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1120                 refs = btrfs_extent_data_ref_count(leaf, dref);
1121         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1122                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1123                 refs = btrfs_shared_data_ref_count(leaf, sref);
1124         } else {
1125                 refs = 1;
1126                 /*
1127                  * For tree blocks we can only drop one ref for it, and tree
1128                  * blocks should not have refs > 1.
1129                  *
1130                  * Furthermore if we're inserting a new inline backref, we
1131                  * won't reach this path either. That would be
1132                  * setup_inline_extent_backref().
1133                  */
1134                 if (unlikely(refs_to_mod != -1)) {
1135                         struct btrfs_key key;
1136
1137                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1138
1139                         btrfs_print_leaf(leaf);
1140                         btrfs_err(fs_info,
1141                         "invalid refs_to_mod for tree block %llu, has %d expect -1",
1142                                   key.objectid, refs_to_mod);
1143                         return -EUCLEAN;
1144                 }
1145         }
1146
1147         if (unlikely(refs_to_mod < 0 && refs < -refs_to_mod)) {
1148                 struct btrfs_key key;
1149                 u32 extent_size;
1150
1151                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1152                 if (key.type == BTRFS_METADATA_ITEM_KEY)
1153                         extent_size = fs_info->nodesize;
1154                 else
1155                         extent_size = key.offset;
1156                 btrfs_print_leaf(leaf);
1157                 btrfs_err(fs_info,
1158 "invalid refs_to_mod for backref entry, iref %lu extent %llu num_bytes %u, has %d expect >= -%llu",
1159                           (unsigned long)iref, key.objectid, extent_size,
1160                           refs_to_mod, refs);
1161                 return -EUCLEAN;
1162         }
1163         refs += refs_to_mod;
1164
1165         if (refs > 0) {
1166                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1167                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1168                 else
1169                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1170         } else {
1171                 size =  btrfs_extent_inline_ref_size(type);
1172                 item_size = btrfs_item_size(leaf, path->slots[0]);
1173                 ptr = (unsigned long)iref;
1174                 end = (unsigned long)ei + item_size;
1175                 if (ptr + size < end)
1176                         memmove_extent_buffer(leaf, ptr, ptr + size,
1177                                               end - ptr - size);
1178                 item_size -= size;
1179                 btrfs_truncate_item(trans, path, item_size, 1);
1180         }
1181         btrfs_mark_buffer_dirty(trans, leaf);
1182         return 0;
1183 }
1184
1185 static noinline_for_stack
1186 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1187                                  struct btrfs_path *path,
1188                                  u64 bytenr, u64 num_bytes, u64 parent,
1189                                  u64 root_objectid, u64 owner,
1190                                  u64 offset, int refs_to_add,
1191                                  struct btrfs_delayed_extent_op *extent_op)
1192 {
1193         struct btrfs_extent_inline_ref *iref;
1194         int ret;
1195
1196         ret = lookup_inline_extent_backref(trans, path, &iref, bytenr,
1197                                            num_bytes, parent, root_objectid,
1198                                            owner, offset, 1);
1199         if (ret == 0) {
1200                 /*
1201                  * We're adding refs to a tree block we already own, this
1202                  * should not happen at all.
1203                  */
1204                 if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1205                         btrfs_print_leaf(path->nodes[0]);
1206                         btrfs_crit(trans->fs_info,
1207 "adding refs to an existing tree ref, bytenr %llu num_bytes %llu root_objectid %llu slot %u",
1208                                    bytenr, num_bytes, root_objectid, path->slots[0]);
1209                         return -EUCLEAN;
1210                 }
1211                 ret = update_inline_extent_backref(trans, path, iref,
1212                                                    refs_to_add, extent_op);
1213         } else if (ret == -ENOENT) {
1214                 setup_inline_extent_backref(trans, path, iref, parent,
1215                                             root_objectid, owner, offset,
1216                                             refs_to_add, extent_op);
1217                 ret = 0;
1218         }
1219         return ret;
1220 }
1221
1222 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1223                                  struct btrfs_root *root,
1224                                  struct btrfs_path *path,
1225                                  struct btrfs_extent_inline_ref *iref,
1226                                  int refs_to_drop, int is_data)
1227 {
1228         int ret = 0;
1229
1230         BUG_ON(!is_data && refs_to_drop != 1);
1231         if (iref)
1232                 ret = update_inline_extent_backref(trans, path, iref,
1233                                                    -refs_to_drop, NULL);
1234         else if (is_data)
1235                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop);
1236         else
1237                 ret = btrfs_del_item(trans, root, path);
1238         return ret;
1239 }
1240
1241 static int btrfs_issue_discard(struct block_device *bdev, u64 start, u64 len,
1242                                u64 *discarded_bytes)
1243 {
1244         int j, ret = 0;
1245         u64 bytes_left, end;
1246         u64 aligned_start = ALIGN(start, 1 << SECTOR_SHIFT);
1247
1248         if (WARN_ON(start != aligned_start)) {
1249                 len -= aligned_start - start;
1250                 len = round_down(len, 1 << SECTOR_SHIFT);
1251                 start = aligned_start;
1252         }
1253
1254         *discarded_bytes = 0;
1255
1256         if (!len)
1257                 return 0;
1258
1259         end = start + len;
1260         bytes_left = len;
1261
1262         /* Skip any superblocks on this device. */
1263         for (j = 0; j < BTRFS_SUPER_MIRROR_MAX; j++) {
1264                 u64 sb_start = btrfs_sb_offset(j);
1265                 u64 sb_end = sb_start + BTRFS_SUPER_INFO_SIZE;
1266                 u64 size = sb_start - start;
1267
1268                 if (!in_range(sb_start, start, bytes_left) &&
1269                     !in_range(sb_end, start, bytes_left) &&
1270                     !in_range(start, sb_start, BTRFS_SUPER_INFO_SIZE))
1271                         continue;
1272
1273                 /*
1274                  * Superblock spans beginning of range.  Adjust start and
1275                  * try again.
1276                  */
1277                 if (sb_start <= start) {
1278                         start += sb_end - start;
1279                         if (start > end) {
1280                                 bytes_left = 0;
1281                                 break;
1282                         }
1283                         bytes_left = end - start;
1284                         continue;
1285                 }
1286
1287                 if (size) {
1288                         ret = blkdev_issue_discard(bdev, start >> SECTOR_SHIFT,
1289                                                    size >> SECTOR_SHIFT,
1290                                                    GFP_NOFS);
1291                         if (!ret)
1292                                 *discarded_bytes += size;
1293                         else if (ret != -EOPNOTSUPP)
1294                                 return ret;
1295                 }
1296
1297                 start = sb_end;
1298                 if (start > end) {
1299                         bytes_left = 0;
1300                         break;
1301                 }
1302                 bytes_left = end - start;
1303         }
1304
1305         if (bytes_left) {
1306                 ret = blkdev_issue_discard(bdev, start >> SECTOR_SHIFT,
1307                                            bytes_left >> SECTOR_SHIFT,
1308                                            GFP_NOFS);
1309                 if (!ret)
1310                         *discarded_bytes += bytes_left;
1311         }
1312         return ret;
1313 }
1314
1315 static int do_discard_extent(struct btrfs_discard_stripe *stripe, u64 *bytes)
1316 {
1317         struct btrfs_device *dev = stripe->dev;
1318         struct btrfs_fs_info *fs_info = dev->fs_info;
1319         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1320         u64 phys = stripe->physical;
1321         u64 len = stripe->length;
1322         u64 discarded = 0;
1323         int ret = 0;
1324
1325         /* Zone reset on a zoned filesystem */
1326         if (btrfs_can_zone_reset(dev, phys, len)) {
1327                 u64 src_disc;
1328
1329                 ret = btrfs_reset_device_zone(dev, phys, len, &discarded);
1330                 if (ret)
1331                         goto out;
1332
1333                 if (!btrfs_dev_replace_is_ongoing(dev_replace) ||
1334                     dev != dev_replace->srcdev)
1335                         goto out;
1336
1337                 src_disc = discarded;
1338
1339                 /* Send to replace target as well */
1340                 ret = btrfs_reset_device_zone(dev_replace->tgtdev, phys, len,
1341                                               &discarded);
1342                 discarded += src_disc;
1343         } else if (bdev_max_discard_sectors(stripe->dev->bdev)) {
1344                 ret = btrfs_issue_discard(dev->bdev, phys, len, &discarded);
1345         } else {
1346                 ret = 0;
1347                 *bytes = 0;
1348         }
1349
1350 out:
1351         *bytes = discarded;
1352         return ret;
1353 }
1354
1355 int btrfs_discard_extent(struct btrfs_fs_info *fs_info, u64 bytenr,
1356                          u64 num_bytes, u64 *actual_bytes)
1357 {
1358         int ret = 0;
1359         u64 discarded_bytes = 0;
1360         u64 end = bytenr + num_bytes;
1361         u64 cur = bytenr;
1362
1363         /*
1364          * Avoid races with device replace and make sure the devices in the
1365          * stripes don't go away while we are discarding.
1366          */
1367         btrfs_bio_counter_inc_blocked(fs_info);
1368         while (cur < end) {
1369                 struct btrfs_discard_stripe *stripes;
1370                 unsigned int num_stripes;
1371                 int i;
1372
1373                 num_bytes = end - cur;
1374                 stripes = btrfs_map_discard(fs_info, cur, &num_bytes, &num_stripes);
1375                 if (IS_ERR(stripes)) {
1376                         ret = PTR_ERR(stripes);
1377                         if (ret == -EOPNOTSUPP)
1378                                 ret = 0;
1379                         break;
1380                 }
1381
1382                 for (i = 0; i < num_stripes; i++) {
1383                         struct btrfs_discard_stripe *stripe = stripes + i;
1384                         u64 bytes;
1385
1386                         if (!stripe->dev->bdev) {
1387                                 ASSERT(btrfs_test_opt(fs_info, DEGRADED));
1388                                 continue;
1389                         }
1390
1391                         if (!test_bit(BTRFS_DEV_STATE_WRITEABLE,
1392                                         &stripe->dev->dev_state))
1393                                 continue;
1394
1395                         ret = do_discard_extent(stripe, &bytes);
1396                         if (ret) {
1397                                 /*
1398                                  * Keep going if discard is not supported by the
1399                                  * device.
1400                                  */
1401                                 if (ret != -EOPNOTSUPP)
1402                                         break;
1403                                 ret = 0;
1404                         } else {
1405                                 discarded_bytes += bytes;
1406                         }
1407                 }
1408                 kfree(stripes);
1409                 if (ret)
1410                         break;
1411                 cur += num_bytes;
1412         }
1413         btrfs_bio_counter_dec(fs_info);
1414         if (actual_bytes)
1415                 *actual_bytes = discarded_bytes;
1416         return ret;
1417 }
1418
1419 /* Can return -ENOMEM */
1420 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1421                          struct btrfs_ref *generic_ref)
1422 {
1423         struct btrfs_fs_info *fs_info = trans->fs_info;
1424         int ret;
1425
1426         ASSERT(generic_ref->type != BTRFS_REF_NOT_SET &&
1427                generic_ref->action);
1428         BUG_ON(generic_ref->type == BTRFS_REF_METADATA &&
1429                generic_ref->tree_ref.owning_root == BTRFS_TREE_LOG_OBJECTID);
1430
1431         if (generic_ref->type == BTRFS_REF_METADATA)
1432                 ret = btrfs_add_delayed_tree_ref(trans, generic_ref, NULL);
1433         else
1434                 ret = btrfs_add_delayed_data_ref(trans, generic_ref, 0);
1435
1436         btrfs_ref_tree_mod(fs_info, generic_ref);
1437
1438         return ret;
1439 }
1440
1441 /*
1442  * __btrfs_inc_extent_ref - insert backreference for a given extent
1443  *
1444  * The counterpart is in __btrfs_free_extent(), with examples and more details
1445  * how it works.
1446  *
1447  * @trans:          Handle of transaction
1448  *
1449  * @node:           The delayed ref node used to get the bytenr/length for
1450  *                  extent whose references are incremented.
1451  *
1452  * @parent:         If this is a shared extent (BTRFS_SHARED_DATA_REF_KEY/
1453  *                  BTRFS_SHARED_BLOCK_REF_KEY) then it holds the logical
1454  *                  bytenr of the parent block. Since new extents are always
1455  *                  created with indirect references, this will only be the case
1456  *                  when relocating a shared extent. In that case, root_objectid
1457  *                  will be BTRFS_TREE_RELOC_OBJECTID. Otherwise, parent must
1458  *                  be 0
1459  *
1460  * @root_objectid:  The id of the root where this modification has originated,
1461  *                  this can be either one of the well-known metadata trees or
1462  *                  the subvolume id which references this extent.
1463  *
1464  * @owner:          For data extents it is the inode number of the owning file.
1465  *                  For metadata extents this parameter holds the level in the
1466  *                  tree of the extent.
1467  *
1468  * @offset:         For metadata extents the offset is ignored and is currently
1469  *                  always passed as 0. For data extents it is the fileoffset
1470  *                  this extent belongs to.
1471  *
1472  * @refs_to_add     Number of references to add
1473  *
1474  * @extent_op       Pointer to a structure, holding information necessary when
1475  *                  updating a tree block's flags
1476  *
1477  */
1478 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1479                                   struct btrfs_delayed_ref_node *node,
1480                                   u64 parent, u64 root_objectid,
1481                                   u64 owner, u64 offset, int refs_to_add,
1482                                   struct btrfs_delayed_extent_op *extent_op)
1483 {
1484         struct btrfs_path *path;
1485         struct extent_buffer *leaf;
1486         struct btrfs_extent_item *item;
1487         struct btrfs_key key;
1488         u64 bytenr = node->bytenr;
1489         u64 num_bytes = node->num_bytes;
1490         u64 refs;
1491         int ret;
1492
1493         path = btrfs_alloc_path();
1494         if (!path)
1495                 return -ENOMEM;
1496
1497         /* this will setup the path even if it fails to insert the back ref */
1498         ret = insert_inline_extent_backref(trans, path, bytenr, num_bytes,
1499                                            parent, root_objectid, owner,
1500                                            offset, refs_to_add, extent_op);
1501         if ((ret < 0 && ret != -EAGAIN) || !ret)
1502                 goto out;
1503
1504         /*
1505          * Ok we had -EAGAIN which means we didn't have space to insert and
1506          * inline extent ref, so just update the reference count and add a
1507          * normal backref.
1508          */
1509         leaf = path->nodes[0];
1510         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1511         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1512         refs = btrfs_extent_refs(leaf, item);
1513         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
1514         if (extent_op)
1515                 __run_delayed_extent_op(extent_op, leaf, item);
1516
1517         btrfs_mark_buffer_dirty(trans, leaf);
1518         btrfs_release_path(path);
1519
1520         /* now insert the actual backref */
1521         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1522                 ret = insert_tree_block_ref(trans, path, bytenr, parent,
1523                                             root_objectid);
1524         else
1525                 ret = insert_extent_data_ref(trans, path, bytenr, parent,
1526                                              root_objectid, owner, offset,
1527                                              refs_to_add);
1528
1529         if (ret)
1530                 btrfs_abort_transaction(trans, ret);
1531 out:
1532         btrfs_free_path(path);
1533         return ret;
1534 }
1535
1536 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
1537                                 struct btrfs_delayed_ref_node *node,
1538                                 struct btrfs_delayed_extent_op *extent_op,
1539                                 bool insert_reserved)
1540 {
1541         int ret = 0;
1542         struct btrfs_delayed_data_ref *ref;
1543         struct btrfs_key ins;
1544         u64 parent = 0;
1545         u64 ref_root = 0;
1546         u64 flags = 0;
1547
1548         ins.objectid = node->bytenr;
1549         ins.offset = node->num_bytes;
1550         ins.type = BTRFS_EXTENT_ITEM_KEY;
1551
1552         ref = btrfs_delayed_node_to_data_ref(node);
1553         trace_run_delayed_data_ref(trans->fs_info, node, ref, node->action);
1554
1555         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
1556                 parent = ref->parent;
1557         ref_root = ref->root;
1558
1559         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1560                 if (extent_op)
1561                         flags |= extent_op->flags_to_set;
1562                 ret = alloc_reserved_file_extent(trans, parent, ref_root,
1563                                                  flags, ref->objectid,
1564                                                  ref->offset, &ins,
1565                                                  node->ref_mod);
1566         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1567                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
1568                                              ref->objectid, ref->offset,
1569                                              node->ref_mod, extent_op);
1570         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1571                 ret = __btrfs_free_extent(trans, node, parent,
1572                                           ref_root, ref->objectid,
1573                                           ref->offset, node->ref_mod,
1574                                           extent_op);
1575         } else {
1576                 BUG();
1577         }
1578         return ret;
1579 }
1580
1581 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
1582                                     struct extent_buffer *leaf,
1583                                     struct btrfs_extent_item *ei)
1584 {
1585         u64 flags = btrfs_extent_flags(leaf, ei);
1586         if (extent_op->update_flags) {
1587                 flags |= extent_op->flags_to_set;
1588                 btrfs_set_extent_flags(leaf, ei, flags);
1589         }
1590
1591         if (extent_op->update_key) {
1592                 struct btrfs_tree_block_info *bi;
1593                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
1594                 bi = (struct btrfs_tree_block_info *)(ei + 1);
1595                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
1596         }
1597 }
1598
1599 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
1600                                  struct btrfs_delayed_ref_head *head,
1601                                  struct btrfs_delayed_extent_op *extent_op)
1602 {
1603         struct btrfs_fs_info *fs_info = trans->fs_info;
1604         struct btrfs_root *root;
1605         struct btrfs_key key;
1606         struct btrfs_path *path;
1607         struct btrfs_extent_item *ei;
1608         struct extent_buffer *leaf;
1609         u32 item_size;
1610         int ret;
1611         int err = 0;
1612         int metadata = 1;
1613
1614         if (TRANS_ABORTED(trans))
1615                 return 0;
1616
1617         if (!btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1618                 metadata = 0;
1619
1620         path = btrfs_alloc_path();
1621         if (!path)
1622                 return -ENOMEM;
1623
1624         key.objectid = head->bytenr;
1625
1626         if (metadata) {
1627                 key.type = BTRFS_METADATA_ITEM_KEY;
1628                 key.offset = extent_op->level;
1629         } else {
1630                 key.type = BTRFS_EXTENT_ITEM_KEY;
1631                 key.offset = head->num_bytes;
1632         }
1633
1634         root = btrfs_extent_root(fs_info, key.objectid);
1635 again:
1636         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
1637         if (ret < 0) {
1638                 err = ret;
1639                 goto out;
1640         }
1641         if (ret > 0) {
1642                 if (metadata) {
1643                         if (path->slots[0] > 0) {
1644                                 path->slots[0]--;
1645                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
1646                                                       path->slots[0]);
1647                                 if (key.objectid == head->bytenr &&
1648                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
1649                                     key.offset == head->num_bytes)
1650                                         ret = 0;
1651                         }
1652                         if (ret > 0) {
1653                                 btrfs_release_path(path);
1654                                 metadata = 0;
1655
1656                                 key.objectid = head->bytenr;
1657                                 key.offset = head->num_bytes;
1658                                 key.type = BTRFS_EXTENT_ITEM_KEY;
1659                                 goto again;
1660                         }
1661                 } else {
1662                         err = -EUCLEAN;
1663                         btrfs_err(fs_info,
1664                   "missing extent item for extent %llu num_bytes %llu level %d",
1665                                   head->bytenr, head->num_bytes, extent_op->level);
1666                         goto out;
1667                 }
1668         }
1669
1670         leaf = path->nodes[0];
1671         item_size = btrfs_item_size(leaf, path->slots[0]);
1672
1673         if (unlikely(item_size < sizeof(*ei))) {
1674                 err = -EUCLEAN;
1675                 btrfs_err(fs_info,
1676                           "unexpected extent item size, has %u expect >= %zu",
1677                           item_size, sizeof(*ei));
1678                 btrfs_abort_transaction(trans, err);
1679                 goto out;
1680         }
1681
1682         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1683         __run_delayed_extent_op(extent_op, leaf, ei);
1684
1685         btrfs_mark_buffer_dirty(trans, leaf);
1686 out:
1687         btrfs_free_path(path);
1688         return err;
1689 }
1690
1691 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
1692                                 struct btrfs_delayed_ref_node *node,
1693                                 struct btrfs_delayed_extent_op *extent_op,
1694                                 bool insert_reserved)
1695 {
1696         int ret = 0;
1697         struct btrfs_delayed_tree_ref *ref;
1698         u64 parent = 0;
1699         u64 ref_root = 0;
1700
1701         ref = btrfs_delayed_node_to_tree_ref(node);
1702         trace_run_delayed_tree_ref(trans->fs_info, node, ref, node->action);
1703
1704         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
1705                 parent = ref->parent;
1706         ref_root = ref->root;
1707
1708         if (unlikely(node->ref_mod != 1)) {
1709                 btrfs_err(trans->fs_info,
1710         "btree block %llu has %d references rather than 1: action %d ref_root %llu parent %llu",
1711                           node->bytenr, node->ref_mod, node->action, ref_root,
1712                           parent);
1713                 return -EUCLEAN;
1714         }
1715         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1716                 BUG_ON(!extent_op || !extent_op->update_flags);
1717                 ret = alloc_reserved_tree_block(trans, node, extent_op);
1718         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1719                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
1720                                              ref->level, 0, 1, extent_op);
1721         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1722                 ret = __btrfs_free_extent(trans, node, parent, ref_root,
1723                                           ref->level, 0, 1, extent_op);
1724         } else {
1725                 BUG();
1726         }
1727         return ret;
1728 }
1729
1730 /* helper function to actually process a single delayed ref entry */
1731 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
1732                                struct btrfs_delayed_ref_node *node,
1733                                struct btrfs_delayed_extent_op *extent_op,
1734                                bool insert_reserved)
1735 {
1736         int ret = 0;
1737
1738         if (TRANS_ABORTED(trans)) {
1739                 if (insert_reserved)
1740                         btrfs_pin_extent(trans, node->bytenr, node->num_bytes, 1);
1741                 return 0;
1742         }
1743
1744         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
1745             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
1746                 ret = run_delayed_tree_ref(trans, node, extent_op,
1747                                            insert_reserved);
1748         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
1749                  node->type == BTRFS_SHARED_DATA_REF_KEY)
1750                 ret = run_delayed_data_ref(trans, node, extent_op,
1751                                            insert_reserved);
1752         else
1753                 BUG();
1754         if (ret && insert_reserved)
1755                 btrfs_pin_extent(trans, node->bytenr, node->num_bytes, 1);
1756         if (ret < 0)
1757                 btrfs_err(trans->fs_info,
1758 "failed to run delayed ref for logical %llu num_bytes %llu type %u action %u ref_mod %d: %d",
1759                           node->bytenr, node->num_bytes, node->type,
1760                           node->action, node->ref_mod, ret);
1761         return ret;
1762 }
1763
1764 static inline struct btrfs_delayed_ref_node *
1765 select_delayed_ref(struct btrfs_delayed_ref_head *head)
1766 {
1767         struct btrfs_delayed_ref_node *ref;
1768
1769         if (RB_EMPTY_ROOT(&head->ref_tree.rb_root))
1770                 return NULL;
1771
1772         /*
1773          * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
1774          * This is to prevent a ref count from going down to zero, which deletes
1775          * the extent item from the extent tree, when there still are references
1776          * to add, which would fail because they would not find the extent item.
1777          */
1778         if (!list_empty(&head->ref_add_list))
1779                 return list_first_entry(&head->ref_add_list,
1780                                 struct btrfs_delayed_ref_node, add_list);
1781
1782         ref = rb_entry(rb_first_cached(&head->ref_tree),
1783                        struct btrfs_delayed_ref_node, ref_node);
1784         ASSERT(list_empty(&ref->add_list));
1785         return ref;
1786 }
1787
1788 static void unselect_delayed_ref_head(struct btrfs_delayed_ref_root *delayed_refs,
1789                                       struct btrfs_delayed_ref_head *head)
1790 {
1791         spin_lock(&delayed_refs->lock);
1792         head->processing = false;
1793         delayed_refs->num_heads_ready++;
1794         spin_unlock(&delayed_refs->lock);
1795         btrfs_delayed_ref_unlock(head);
1796 }
1797
1798 static struct btrfs_delayed_extent_op *cleanup_extent_op(
1799                                 struct btrfs_delayed_ref_head *head)
1800 {
1801         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
1802
1803         if (!extent_op)
1804                 return NULL;
1805
1806         if (head->must_insert_reserved) {
1807                 head->extent_op = NULL;
1808                 btrfs_free_delayed_extent_op(extent_op);
1809                 return NULL;
1810         }
1811         return extent_op;
1812 }
1813
1814 static int run_and_cleanup_extent_op(struct btrfs_trans_handle *trans,
1815                                      struct btrfs_delayed_ref_head *head)
1816 {
1817         struct btrfs_delayed_extent_op *extent_op;
1818         int ret;
1819
1820         extent_op = cleanup_extent_op(head);
1821         if (!extent_op)
1822                 return 0;
1823         head->extent_op = NULL;
1824         spin_unlock(&head->lock);
1825         ret = run_delayed_extent_op(trans, head, extent_op);
1826         btrfs_free_delayed_extent_op(extent_op);
1827         return ret ? ret : 1;
1828 }
1829
1830 void btrfs_cleanup_ref_head_accounting(struct btrfs_fs_info *fs_info,
1831                                   struct btrfs_delayed_ref_root *delayed_refs,
1832                                   struct btrfs_delayed_ref_head *head)
1833 {
1834         int nr_items = 1;       /* Dropping this ref head update. */
1835
1836         /*
1837          * We had csum deletions accounted for in our delayed refs rsv, we need
1838          * to drop the csum leaves for this update from our delayed_refs_rsv.
1839          */
1840         if (head->total_ref_mod < 0 && head->is_data) {
1841                 spin_lock(&delayed_refs->lock);
1842                 delayed_refs->pending_csums -= head->num_bytes;
1843                 spin_unlock(&delayed_refs->lock);
1844                 nr_items += btrfs_csum_bytes_to_leaves(fs_info, head->num_bytes);
1845         }
1846
1847         btrfs_delayed_refs_rsv_release(fs_info, nr_items);
1848 }
1849
1850 static int cleanup_ref_head(struct btrfs_trans_handle *trans,
1851                             struct btrfs_delayed_ref_head *head)
1852 {
1853
1854         struct btrfs_fs_info *fs_info = trans->fs_info;
1855         struct btrfs_delayed_ref_root *delayed_refs;
1856         int ret;
1857
1858         delayed_refs = &trans->transaction->delayed_refs;
1859
1860         ret = run_and_cleanup_extent_op(trans, head);
1861         if (ret < 0) {
1862                 unselect_delayed_ref_head(delayed_refs, head);
1863                 btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
1864                 return ret;
1865         } else if (ret) {
1866                 return ret;
1867         }
1868
1869         /*
1870          * Need to drop our head ref lock and re-acquire the delayed ref lock
1871          * and then re-check to make sure nobody got added.
1872          */
1873         spin_unlock(&head->lock);
1874         spin_lock(&delayed_refs->lock);
1875         spin_lock(&head->lock);
1876         if (!RB_EMPTY_ROOT(&head->ref_tree.rb_root) || head->extent_op) {
1877                 spin_unlock(&head->lock);
1878                 spin_unlock(&delayed_refs->lock);
1879                 return 1;
1880         }
1881         btrfs_delete_ref_head(delayed_refs, head);
1882         spin_unlock(&head->lock);
1883         spin_unlock(&delayed_refs->lock);
1884
1885         if (head->must_insert_reserved) {
1886                 btrfs_pin_extent(trans, head->bytenr, head->num_bytes, 1);
1887                 if (head->is_data) {
1888                         struct btrfs_root *csum_root;
1889
1890                         csum_root = btrfs_csum_root(fs_info, head->bytenr);
1891                         ret = btrfs_del_csums(trans, csum_root, head->bytenr,
1892                                               head->num_bytes);
1893                 }
1894         }
1895
1896         btrfs_cleanup_ref_head_accounting(fs_info, delayed_refs, head);
1897
1898         trace_run_delayed_ref_head(fs_info, head, 0);
1899         btrfs_delayed_ref_unlock(head);
1900         btrfs_put_delayed_ref_head(head);
1901         return ret;
1902 }
1903
1904 static struct btrfs_delayed_ref_head *btrfs_obtain_ref_head(
1905                                         struct btrfs_trans_handle *trans)
1906 {
1907         struct btrfs_delayed_ref_root *delayed_refs =
1908                 &trans->transaction->delayed_refs;
1909         struct btrfs_delayed_ref_head *head = NULL;
1910         int ret;
1911
1912         spin_lock(&delayed_refs->lock);
1913         head = btrfs_select_ref_head(delayed_refs);
1914         if (!head) {
1915                 spin_unlock(&delayed_refs->lock);
1916                 return head;
1917         }
1918
1919         /*
1920          * Grab the lock that says we are going to process all the refs for
1921          * this head
1922          */
1923         ret = btrfs_delayed_ref_lock(delayed_refs, head);
1924         spin_unlock(&delayed_refs->lock);
1925
1926         /*
1927          * We may have dropped the spin lock to get the head mutex lock, and
1928          * that might have given someone else time to free the head.  If that's
1929          * true, it has been removed from our list and we can move on.
1930          */
1931         if (ret == -EAGAIN)
1932                 head = ERR_PTR(-EAGAIN);
1933
1934         return head;
1935 }
1936
1937 static int btrfs_run_delayed_refs_for_head(struct btrfs_trans_handle *trans,
1938                                            struct btrfs_delayed_ref_head *locked_ref)
1939 {
1940         struct btrfs_fs_info *fs_info = trans->fs_info;
1941         struct btrfs_delayed_ref_root *delayed_refs;
1942         struct btrfs_delayed_extent_op *extent_op;
1943         struct btrfs_delayed_ref_node *ref;
1944         bool must_insert_reserved;
1945         int ret;
1946
1947         delayed_refs = &trans->transaction->delayed_refs;
1948
1949         lockdep_assert_held(&locked_ref->mutex);
1950         lockdep_assert_held(&locked_ref->lock);
1951
1952         while ((ref = select_delayed_ref(locked_ref))) {
1953                 if (ref->seq &&
1954                     btrfs_check_delayed_seq(fs_info, ref->seq)) {
1955                         spin_unlock(&locked_ref->lock);
1956                         unselect_delayed_ref_head(delayed_refs, locked_ref);
1957                         return -EAGAIN;
1958                 }
1959
1960                 rb_erase_cached(&ref->ref_node, &locked_ref->ref_tree);
1961                 RB_CLEAR_NODE(&ref->ref_node);
1962                 if (!list_empty(&ref->add_list))
1963                         list_del(&ref->add_list);
1964                 /*
1965                  * When we play the delayed ref, also correct the ref_mod on
1966                  * head
1967                  */
1968                 switch (ref->action) {
1969                 case BTRFS_ADD_DELAYED_REF:
1970                 case BTRFS_ADD_DELAYED_EXTENT:
1971                         locked_ref->ref_mod -= ref->ref_mod;
1972                         break;
1973                 case BTRFS_DROP_DELAYED_REF:
1974                         locked_ref->ref_mod += ref->ref_mod;
1975                         break;
1976                 default:
1977                         WARN_ON(1);
1978                 }
1979                 atomic_dec(&delayed_refs->num_entries);
1980
1981                 /*
1982                  * Record the must_insert_reserved flag before we drop the
1983                  * spin lock.
1984                  */
1985                 must_insert_reserved = locked_ref->must_insert_reserved;
1986                 locked_ref->must_insert_reserved = false;
1987
1988                 extent_op = locked_ref->extent_op;
1989                 locked_ref->extent_op = NULL;
1990                 spin_unlock(&locked_ref->lock);
1991
1992                 ret = run_one_delayed_ref(trans, ref, extent_op,
1993                                           must_insert_reserved);
1994
1995                 btrfs_free_delayed_extent_op(extent_op);
1996                 if (ret) {
1997                         unselect_delayed_ref_head(delayed_refs, locked_ref);
1998                         btrfs_put_delayed_ref(ref);
1999                         return ret;
2000                 }
2001
2002                 btrfs_put_delayed_ref(ref);
2003                 cond_resched();
2004
2005                 spin_lock(&locked_ref->lock);
2006                 btrfs_merge_delayed_refs(fs_info, delayed_refs, locked_ref);
2007         }
2008
2009         return 0;
2010 }
2011
2012 /*
2013  * Returns 0 on success or if called with an already aborted transaction.
2014  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2015  */
2016 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2017                                              unsigned long nr)
2018 {
2019         struct btrfs_fs_info *fs_info = trans->fs_info;
2020         struct btrfs_delayed_ref_root *delayed_refs;
2021         struct btrfs_delayed_ref_head *locked_ref = NULL;
2022         int ret;
2023         unsigned long count = 0;
2024
2025         delayed_refs = &trans->transaction->delayed_refs;
2026         do {
2027                 if (!locked_ref) {
2028                         locked_ref = btrfs_obtain_ref_head(trans);
2029                         if (IS_ERR_OR_NULL(locked_ref)) {
2030                                 if (PTR_ERR(locked_ref) == -EAGAIN) {
2031                                         continue;
2032                                 } else {
2033                                         break;
2034                                 }
2035                         }
2036                         count++;
2037                 }
2038                 /*
2039                  * We need to try and merge add/drops of the same ref since we
2040                  * can run into issues with relocate dropping the implicit ref
2041                  * and then it being added back again before the drop can
2042                  * finish.  If we merged anything we need to re-loop so we can
2043                  * get a good ref.
2044                  * Or we can get node references of the same type that weren't
2045                  * merged when created due to bumps in the tree mod seq, and
2046                  * we need to merge them to prevent adding an inline extent
2047                  * backref before dropping it (triggering a BUG_ON at
2048                  * insert_inline_extent_backref()).
2049                  */
2050                 spin_lock(&locked_ref->lock);
2051                 btrfs_merge_delayed_refs(fs_info, delayed_refs, locked_ref);
2052
2053                 ret = btrfs_run_delayed_refs_for_head(trans, locked_ref);
2054                 if (ret < 0 && ret != -EAGAIN) {
2055                         /*
2056                          * Error, btrfs_run_delayed_refs_for_head already
2057                          * unlocked everything so just bail out
2058                          */
2059                         return ret;
2060                 } else if (!ret) {
2061                         /*
2062                          * Success, perform the usual cleanup of a processed
2063                          * head
2064                          */
2065                         ret = cleanup_ref_head(trans, locked_ref);
2066                         if (ret > 0 ) {
2067                                 /* We dropped our lock, we need to loop. */
2068                                 ret = 0;
2069                                 continue;
2070                         } else if (ret) {
2071                                 return ret;
2072                         }
2073                 }
2074
2075                 /*
2076                  * Either success case or btrfs_run_delayed_refs_for_head
2077                  * returned -EAGAIN, meaning we need to select another head
2078                  */
2079
2080                 locked_ref = NULL;
2081                 cond_resched();
2082         } while ((nr != -1 && count < nr) || locked_ref);
2083
2084         return 0;
2085 }
2086
2087 #ifdef SCRAMBLE_DELAYED_REFS
2088 /*
2089  * Normally delayed refs get processed in ascending bytenr order. This
2090  * correlates in most cases to the order added. To expose dependencies on this
2091  * order, we start to process the tree in the middle instead of the beginning
2092  */
2093 static u64 find_middle(struct rb_root *root)
2094 {
2095         struct rb_node *n = root->rb_node;
2096         struct btrfs_delayed_ref_node *entry;
2097         int alt = 1;
2098         u64 middle;
2099         u64 first = 0, last = 0;
2100
2101         n = rb_first(root);
2102         if (n) {
2103                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2104                 first = entry->bytenr;
2105         }
2106         n = rb_last(root);
2107         if (n) {
2108                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2109                 last = entry->bytenr;
2110         }
2111         n = root->rb_node;
2112
2113         while (n) {
2114                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2115                 WARN_ON(!entry->in_tree);
2116
2117                 middle = entry->bytenr;
2118
2119                 if (alt)
2120                         n = n->rb_left;
2121                 else
2122                         n = n->rb_right;
2123
2124                 alt = 1 - alt;
2125         }
2126         return middle;
2127 }
2128 #endif
2129
2130 /*
2131  * this starts processing the delayed reference count updates and
2132  * extent insertions we have queued up so far.  count can be
2133  * 0, which means to process everything in the tree at the start
2134  * of the run (but not newly added entries), or it can be some target
2135  * number you'd like to process.
2136  *
2137  * Returns 0 on success or if called with an aborted transaction
2138  * Returns <0 on error and aborts the transaction
2139  */
2140 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2141                            unsigned long count)
2142 {
2143         struct btrfs_fs_info *fs_info = trans->fs_info;
2144         struct rb_node *node;
2145         struct btrfs_delayed_ref_root *delayed_refs;
2146         struct btrfs_delayed_ref_head *head;
2147         int ret;
2148         int run_all = count == (unsigned long)-1;
2149
2150         /* We'll clean this up in btrfs_cleanup_transaction */
2151         if (TRANS_ABORTED(trans))
2152                 return 0;
2153
2154         if (test_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags))
2155                 return 0;
2156
2157         delayed_refs = &trans->transaction->delayed_refs;
2158         if (count == 0)
2159                 count = delayed_refs->num_heads_ready;
2160
2161 again:
2162 #ifdef SCRAMBLE_DELAYED_REFS
2163         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2164 #endif
2165         ret = __btrfs_run_delayed_refs(trans, count);
2166         if (ret < 0) {
2167                 btrfs_abort_transaction(trans, ret);
2168                 return ret;
2169         }
2170
2171         if (run_all) {
2172                 btrfs_create_pending_block_groups(trans);
2173
2174                 spin_lock(&delayed_refs->lock);
2175                 node = rb_first_cached(&delayed_refs->href_root);
2176                 if (!node) {
2177                         spin_unlock(&delayed_refs->lock);
2178                         goto out;
2179                 }
2180                 head = rb_entry(node, struct btrfs_delayed_ref_head,
2181                                 href_node);
2182                 refcount_inc(&head->refs);
2183                 spin_unlock(&delayed_refs->lock);
2184
2185                 /* Mutex was contended, block until it's released and retry. */
2186                 mutex_lock(&head->mutex);
2187                 mutex_unlock(&head->mutex);
2188
2189                 btrfs_put_delayed_ref_head(head);
2190                 cond_resched();
2191                 goto again;
2192         }
2193 out:
2194         return 0;
2195 }
2196
2197 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2198                                 struct extent_buffer *eb, u64 flags)
2199 {
2200         struct btrfs_delayed_extent_op *extent_op;
2201         int level = btrfs_header_level(eb);
2202         int ret;
2203
2204         extent_op = btrfs_alloc_delayed_extent_op();
2205         if (!extent_op)
2206                 return -ENOMEM;
2207
2208         extent_op->flags_to_set = flags;
2209         extent_op->update_flags = true;
2210         extent_op->update_key = false;
2211         extent_op->level = level;
2212
2213         ret = btrfs_add_delayed_extent_op(trans, eb->start, eb->len, extent_op);
2214         if (ret)
2215                 btrfs_free_delayed_extent_op(extent_op);
2216         return ret;
2217 }
2218
2219 static noinline int check_delayed_ref(struct btrfs_root *root,
2220                                       struct btrfs_path *path,
2221                                       u64 objectid, u64 offset, u64 bytenr)
2222 {
2223         struct btrfs_delayed_ref_head *head;
2224         struct btrfs_delayed_ref_node *ref;
2225         struct btrfs_delayed_data_ref *data_ref;
2226         struct btrfs_delayed_ref_root *delayed_refs;
2227         struct btrfs_transaction *cur_trans;
2228         struct rb_node *node;
2229         int ret = 0;
2230
2231         spin_lock(&root->fs_info->trans_lock);
2232         cur_trans = root->fs_info->running_transaction;
2233         if (cur_trans)
2234                 refcount_inc(&cur_trans->use_count);
2235         spin_unlock(&root->fs_info->trans_lock);
2236         if (!cur_trans)
2237                 return 0;
2238
2239         delayed_refs = &cur_trans->delayed_refs;
2240         spin_lock(&delayed_refs->lock);
2241         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
2242         if (!head) {
2243                 spin_unlock(&delayed_refs->lock);
2244                 btrfs_put_transaction(cur_trans);
2245                 return 0;
2246         }
2247
2248         if (!mutex_trylock(&head->mutex)) {
2249                 if (path->nowait) {
2250                         spin_unlock(&delayed_refs->lock);
2251                         btrfs_put_transaction(cur_trans);
2252                         return -EAGAIN;
2253                 }
2254
2255                 refcount_inc(&head->refs);
2256                 spin_unlock(&delayed_refs->lock);
2257
2258                 btrfs_release_path(path);
2259
2260                 /*
2261                  * Mutex was contended, block until it's released and let
2262                  * caller try again
2263                  */
2264                 mutex_lock(&head->mutex);
2265                 mutex_unlock(&head->mutex);
2266                 btrfs_put_delayed_ref_head(head);
2267                 btrfs_put_transaction(cur_trans);
2268                 return -EAGAIN;
2269         }
2270         spin_unlock(&delayed_refs->lock);
2271
2272         spin_lock(&head->lock);
2273         /*
2274          * XXX: We should replace this with a proper search function in the
2275          * future.
2276          */
2277         for (node = rb_first_cached(&head->ref_tree); node;
2278              node = rb_next(node)) {
2279                 ref = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
2280                 /* If it's a shared ref we know a cross reference exists */
2281                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
2282                         ret = 1;
2283                         break;
2284                 }
2285
2286                 data_ref = btrfs_delayed_node_to_data_ref(ref);
2287
2288                 /*
2289                  * If our ref doesn't match the one we're currently looking at
2290                  * then we have a cross reference.
2291                  */
2292                 if (data_ref->root != root->root_key.objectid ||
2293                     data_ref->objectid != objectid ||
2294                     data_ref->offset != offset) {
2295                         ret = 1;
2296                         break;
2297                 }
2298         }
2299         spin_unlock(&head->lock);
2300         mutex_unlock(&head->mutex);
2301         btrfs_put_transaction(cur_trans);
2302         return ret;
2303 }
2304
2305 static noinline int check_committed_ref(struct btrfs_root *root,
2306                                         struct btrfs_path *path,
2307                                         u64 objectid, u64 offset, u64 bytenr,
2308                                         bool strict)
2309 {
2310         struct btrfs_fs_info *fs_info = root->fs_info;
2311         struct btrfs_root *extent_root = btrfs_extent_root(fs_info, bytenr);
2312         struct extent_buffer *leaf;
2313         struct btrfs_extent_data_ref *ref;
2314         struct btrfs_extent_inline_ref *iref;
2315         struct btrfs_extent_item *ei;
2316         struct btrfs_key key;
2317         u32 item_size;
2318         int type;
2319         int ret;
2320
2321         key.objectid = bytenr;
2322         key.offset = (u64)-1;
2323         key.type = BTRFS_EXTENT_ITEM_KEY;
2324
2325         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2326         if (ret < 0)
2327                 goto out;
2328         BUG_ON(ret == 0); /* Corruption */
2329
2330         ret = -ENOENT;
2331         if (path->slots[0] == 0)
2332                 goto out;
2333
2334         path->slots[0]--;
2335         leaf = path->nodes[0];
2336         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2337
2338         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2339                 goto out;
2340
2341         ret = 1;
2342         item_size = btrfs_item_size(leaf, path->slots[0]);
2343         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2344
2345         /* If extent item has more than 1 inline ref then it's shared */
2346         if (item_size != sizeof(*ei) +
2347             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2348                 goto out;
2349
2350         /*
2351          * If extent created before last snapshot => it's shared unless the
2352          * snapshot has been deleted. Use the heuristic if strict is false.
2353          */
2354         if (!strict &&
2355             (btrfs_extent_generation(leaf, ei) <=
2356              btrfs_root_last_snapshot(&root->root_item)))
2357                 goto out;
2358
2359         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2360
2361         /* If this extent has SHARED_DATA_REF then it's shared */
2362         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
2363         if (type != BTRFS_EXTENT_DATA_REF_KEY)
2364                 goto out;
2365
2366         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2367         if (btrfs_extent_refs(leaf, ei) !=
2368             btrfs_extent_data_ref_count(leaf, ref) ||
2369             btrfs_extent_data_ref_root(leaf, ref) !=
2370             root->root_key.objectid ||
2371             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2372             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2373                 goto out;
2374
2375         ret = 0;
2376 out:
2377         return ret;
2378 }
2379
2380 int btrfs_cross_ref_exist(struct btrfs_root *root, u64 objectid, u64 offset,
2381                           u64 bytenr, bool strict, struct btrfs_path *path)
2382 {
2383         int ret;
2384
2385         do {
2386                 ret = check_committed_ref(root, path, objectid,
2387                                           offset, bytenr, strict);
2388                 if (ret && ret != -ENOENT)
2389                         goto out;
2390
2391                 ret = check_delayed_ref(root, path, objectid, offset, bytenr);
2392         } while (ret == -EAGAIN);
2393
2394 out:
2395         btrfs_release_path(path);
2396         if (btrfs_is_data_reloc_root(root))
2397                 WARN_ON(ret > 0);
2398         return ret;
2399 }
2400
2401 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2402                            struct btrfs_root *root,
2403                            struct extent_buffer *buf,
2404                            int full_backref, int inc)
2405 {
2406         struct btrfs_fs_info *fs_info = root->fs_info;
2407         u64 bytenr;
2408         u64 num_bytes;
2409         u64 parent;
2410         u64 ref_root;
2411         u32 nritems;
2412         struct btrfs_key key;
2413         struct btrfs_file_extent_item *fi;
2414         struct btrfs_ref generic_ref = { 0 };
2415         bool for_reloc = btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC);
2416         int i;
2417         int action;
2418         int level;
2419         int ret = 0;
2420
2421         if (btrfs_is_testing(fs_info))
2422                 return 0;
2423
2424         ref_root = btrfs_header_owner(buf);
2425         nritems = btrfs_header_nritems(buf);
2426         level = btrfs_header_level(buf);
2427
2428         if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state) && level == 0)
2429                 return 0;
2430
2431         if (full_backref)
2432                 parent = buf->start;
2433         else
2434                 parent = 0;
2435         if (inc)
2436                 action = BTRFS_ADD_DELAYED_REF;
2437         else
2438                 action = BTRFS_DROP_DELAYED_REF;
2439
2440         for (i = 0; i < nritems; i++) {
2441                 if (level == 0) {
2442                         btrfs_item_key_to_cpu(buf, &key, i);
2443                         if (key.type != BTRFS_EXTENT_DATA_KEY)
2444                                 continue;
2445                         fi = btrfs_item_ptr(buf, i,
2446                                             struct btrfs_file_extent_item);
2447                         if (btrfs_file_extent_type(buf, fi) ==
2448                             BTRFS_FILE_EXTENT_INLINE)
2449                                 continue;
2450                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2451                         if (bytenr == 0)
2452                                 continue;
2453
2454                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
2455                         key.offset -= btrfs_file_extent_offset(buf, fi);
2456                         btrfs_init_generic_ref(&generic_ref, action, bytenr,
2457                                                num_bytes, parent);
2458                         btrfs_init_data_ref(&generic_ref, ref_root, key.objectid,
2459                                             key.offset, root->root_key.objectid,
2460                                             for_reloc);
2461                         if (inc)
2462                                 ret = btrfs_inc_extent_ref(trans, &generic_ref);
2463                         else
2464                                 ret = btrfs_free_extent(trans, &generic_ref);
2465                         if (ret)
2466                                 goto fail;
2467                 } else {
2468                         bytenr = btrfs_node_blockptr(buf, i);
2469                         num_bytes = fs_info->nodesize;
2470                         btrfs_init_generic_ref(&generic_ref, action, bytenr,
2471                                                num_bytes, parent);
2472                         btrfs_init_tree_ref(&generic_ref, level - 1, ref_root,
2473                                             root->root_key.objectid, for_reloc);
2474                         if (inc)
2475                                 ret = btrfs_inc_extent_ref(trans, &generic_ref);
2476                         else
2477                                 ret = btrfs_free_extent(trans, &generic_ref);
2478                         if (ret)
2479                                 goto fail;
2480                 }
2481         }
2482         return 0;
2483 fail:
2484         return ret;
2485 }
2486
2487 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2488                   struct extent_buffer *buf, int full_backref)
2489 {
2490         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
2491 }
2492
2493 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2494                   struct extent_buffer *buf, int full_backref)
2495 {
2496         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
2497 }
2498
2499 static u64 get_alloc_profile_by_root(struct btrfs_root *root, int data)
2500 {
2501         struct btrfs_fs_info *fs_info = root->fs_info;
2502         u64 flags;
2503         u64 ret;
2504
2505         if (data)
2506                 flags = BTRFS_BLOCK_GROUP_DATA;
2507         else if (root == fs_info->chunk_root)
2508                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
2509         else
2510                 flags = BTRFS_BLOCK_GROUP_METADATA;
2511
2512         ret = btrfs_get_alloc_profile(fs_info, flags);
2513         return ret;
2514 }
2515
2516 static u64 first_logical_byte(struct btrfs_fs_info *fs_info)
2517 {
2518         struct rb_node *leftmost;
2519         u64 bytenr = 0;
2520
2521         read_lock(&fs_info->block_group_cache_lock);
2522         /* Get the block group with the lowest logical start address. */
2523         leftmost = rb_first_cached(&fs_info->block_group_cache_tree);
2524         if (leftmost) {
2525                 struct btrfs_block_group *bg;
2526
2527                 bg = rb_entry(leftmost, struct btrfs_block_group, cache_node);
2528                 bytenr = bg->start;
2529         }
2530         read_unlock(&fs_info->block_group_cache_lock);
2531
2532         return bytenr;
2533 }
2534
2535 static int pin_down_extent(struct btrfs_trans_handle *trans,
2536                            struct btrfs_block_group *cache,
2537                            u64 bytenr, u64 num_bytes, int reserved)
2538 {
2539         struct btrfs_fs_info *fs_info = cache->fs_info;
2540
2541         spin_lock(&cache->space_info->lock);
2542         spin_lock(&cache->lock);
2543         cache->pinned += num_bytes;
2544         btrfs_space_info_update_bytes_pinned(fs_info, cache->space_info,
2545                                              num_bytes);
2546         if (reserved) {
2547                 cache->reserved -= num_bytes;
2548                 cache->space_info->bytes_reserved -= num_bytes;
2549         }
2550         spin_unlock(&cache->lock);
2551         spin_unlock(&cache->space_info->lock);
2552
2553         set_extent_bit(&trans->transaction->pinned_extents, bytenr,
2554                        bytenr + num_bytes - 1, EXTENT_DIRTY, NULL);
2555         return 0;
2556 }
2557
2558 int btrfs_pin_extent(struct btrfs_trans_handle *trans,
2559                      u64 bytenr, u64 num_bytes, int reserved)
2560 {
2561         struct btrfs_block_group *cache;
2562
2563         cache = btrfs_lookup_block_group(trans->fs_info, bytenr);
2564         BUG_ON(!cache); /* Logic error */
2565
2566         pin_down_extent(trans, cache, bytenr, num_bytes, reserved);
2567
2568         btrfs_put_block_group(cache);
2569         return 0;
2570 }
2571
2572 /*
2573  * this function must be called within transaction
2574  */
2575 int btrfs_pin_extent_for_log_replay(struct btrfs_trans_handle *trans,
2576                                     u64 bytenr, u64 num_bytes)
2577 {
2578         struct btrfs_block_group *cache;
2579         int ret;
2580
2581         cache = btrfs_lookup_block_group(trans->fs_info, bytenr);
2582         if (!cache)
2583                 return -EINVAL;
2584
2585         /*
2586          * Fully cache the free space first so that our pin removes the free space
2587          * from the cache.
2588          */
2589         ret = btrfs_cache_block_group(cache, true);
2590         if (ret)
2591                 goto out;
2592
2593         pin_down_extent(trans, cache, bytenr, num_bytes, 0);
2594
2595         /* remove us from the free space cache (if we're there at all) */
2596         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
2597 out:
2598         btrfs_put_block_group(cache);
2599         return ret;
2600 }
2601
2602 static int __exclude_logged_extent(struct btrfs_fs_info *fs_info,
2603                                    u64 start, u64 num_bytes)
2604 {
2605         int ret;
2606         struct btrfs_block_group *block_group;
2607
2608         block_group = btrfs_lookup_block_group(fs_info, start);
2609         if (!block_group)
2610                 return -EINVAL;
2611
2612         ret = btrfs_cache_block_group(block_group, true);
2613         if (ret)
2614                 goto out;
2615
2616         ret = btrfs_remove_free_space(block_group, start, num_bytes);
2617 out:
2618         btrfs_put_block_group(block_group);
2619         return ret;
2620 }
2621
2622 int btrfs_exclude_logged_extents(struct extent_buffer *eb)
2623 {
2624         struct btrfs_fs_info *fs_info = eb->fs_info;
2625         struct btrfs_file_extent_item *item;
2626         struct btrfs_key key;
2627         int found_type;
2628         int i;
2629         int ret = 0;
2630
2631         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS))
2632                 return 0;
2633
2634         for (i = 0; i < btrfs_header_nritems(eb); i++) {
2635                 btrfs_item_key_to_cpu(eb, &key, i);
2636                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2637                         continue;
2638                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
2639                 found_type = btrfs_file_extent_type(eb, item);
2640                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
2641                         continue;
2642                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
2643                         continue;
2644                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
2645                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
2646                 ret = __exclude_logged_extent(fs_info, key.objectid, key.offset);
2647                 if (ret)
2648                         break;
2649         }
2650
2651         return ret;
2652 }
2653
2654 static void
2655 btrfs_inc_block_group_reservations(struct btrfs_block_group *bg)
2656 {
2657         atomic_inc(&bg->reservations);
2658 }
2659
2660 /*
2661  * Returns the free cluster for the given space info and sets empty_cluster to
2662  * what it should be based on the mount options.
2663  */
2664 static struct btrfs_free_cluster *
2665 fetch_cluster_info(struct btrfs_fs_info *fs_info,
2666                    struct btrfs_space_info *space_info, u64 *empty_cluster)
2667 {
2668         struct btrfs_free_cluster *ret = NULL;
2669
2670         *empty_cluster = 0;
2671         if (btrfs_mixed_space_info(space_info))
2672                 return ret;
2673
2674         if (space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
2675                 ret = &fs_info->meta_alloc_cluster;
2676                 if (btrfs_test_opt(fs_info, SSD))
2677                         *empty_cluster = SZ_2M;
2678                 else
2679                         *empty_cluster = SZ_64K;
2680         } else if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) &&
2681                    btrfs_test_opt(fs_info, SSD_SPREAD)) {
2682                 *empty_cluster = SZ_2M;
2683                 ret = &fs_info->data_alloc_cluster;
2684         }
2685
2686         return ret;
2687 }
2688
2689 static int unpin_extent_range(struct btrfs_fs_info *fs_info,
2690                               u64 start, u64 end,
2691                               const bool return_free_space)
2692 {
2693         struct btrfs_block_group *cache = NULL;
2694         struct btrfs_space_info *space_info;
2695         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
2696         struct btrfs_free_cluster *cluster = NULL;
2697         u64 len;
2698         u64 total_unpinned = 0;
2699         u64 empty_cluster = 0;
2700         bool readonly;
2701
2702         while (start <= end) {
2703                 readonly = false;
2704                 if (!cache ||
2705                     start >= cache->start + cache->length) {
2706                         if (cache)
2707                                 btrfs_put_block_group(cache);
2708                         total_unpinned = 0;
2709                         cache = btrfs_lookup_block_group(fs_info, start);
2710                         BUG_ON(!cache); /* Logic error */
2711
2712                         cluster = fetch_cluster_info(fs_info,
2713                                                      cache->space_info,
2714                                                      &empty_cluster);
2715                         empty_cluster <<= 1;
2716                 }
2717
2718                 len = cache->start + cache->length - start;
2719                 len = min(len, end + 1 - start);
2720
2721                 if (return_free_space)
2722                         btrfs_add_free_space(cache, start, len);
2723
2724                 start += len;
2725                 total_unpinned += len;
2726                 space_info = cache->space_info;
2727
2728                 /*
2729                  * If this space cluster has been marked as fragmented and we've
2730                  * unpinned enough in this block group to potentially allow a
2731                  * cluster to be created inside of it go ahead and clear the
2732                  * fragmented check.
2733                  */
2734                 if (cluster && cluster->fragmented &&
2735                     total_unpinned > empty_cluster) {
2736                         spin_lock(&cluster->lock);
2737                         cluster->fragmented = 0;
2738                         spin_unlock(&cluster->lock);
2739                 }
2740
2741                 spin_lock(&space_info->lock);
2742                 spin_lock(&cache->lock);
2743                 cache->pinned -= len;
2744                 btrfs_space_info_update_bytes_pinned(fs_info, space_info, -len);
2745                 space_info->max_extent_size = 0;
2746                 if (cache->ro) {
2747                         space_info->bytes_readonly += len;
2748                         readonly = true;
2749                 } else if (btrfs_is_zoned(fs_info)) {
2750                         /* Need reset before reusing in a zoned block group */
2751                         space_info->bytes_zone_unusable += len;
2752                         readonly = true;
2753                 }
2754                 spin_unlock(&cache->lock);
2755                 if (!readonly && return_free_space &&
2756                     global_rsv->space_info == space_info) {
2757                         spin_lock(&global_rsv->lock);
2758                         if (!global_rsv->full) {
2759                                 u64 to_add = min(len, global_rsv->size -
2760                                                       global_rsv->reserved);
2761
2762                                 global_rsv->reserved += to_add;
2763                                 btrfs_space_info_update_bytes_may_use(fs_info,
2764                                                 space_info, to_add);
2765                                 if (global_rsv->reserved >= global_rsv->size)
2766                                         global_rsv->full = 1;
2767                                 len -= to_add;
2768                         }
2769                         spin_unlock(&global_rsv->lock);
2770                 }
2771                 /* Add to any tickets we may have */
2772                 if (!readonly && return_free_space && len)
2773                         btrfs_try_granting_tickets(fs_info, space_info);
2774                 spin_unlock(&space_info->lock);
2775         }
2776
2777         if (cache)
2778                 btrfs_put_block_group(cache);
2779         return 0;
2780 }
2781
2782 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans)
2783 {
2784         struct btrfs_fs_info *fs_info = trans->fs_info;
2785         struct btrfs_block_group *block_group, *tmp;
2786         struct list_head *deleted_bgs;
2787         struct extent_io_tree *unpin;
2788         u64 start;
2789         u64 end;
2790         int ret;
2791
2792         unpin = &trans->transaction->pinned_extents;
2793
2794         while (!TRANS_ABORTED(trans)) {
2795                 struct extent_state *cached_state = NULL;
2796
2797                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
2798                 if (!find_first_extent_bit(unpin, 0, &start, &end,
2799                                            EXTENT_DIRTY, &cached_state)) {
2800                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
2801                         break;
2802                 }
2803
2804                 if (btrfs_test_opt(fs_info, DISCARD_SYNC))
2805                         ret = btrfs_discard_extent(fs_info, start,
2806                                                    end + 1 - start, NULL);
2807
2808                 clear_extent_dirty(unpin, start, end, &cached_state);
2809                 unpin_extent_range(fs_info, start, end, true);
2810                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
2811                 free_extent_state(cached_state);
2812                 cond_resched();
2813         }
2814
2815         if (btrfs_test_opt(fs_info, DISCARD_ASYNC)) {
2816                 btrfs_discard_calc_delay(&fs_info->discard_ctl);
2817                 btrfs_discard_schedule_work(&fs_info->discard_ctl, true);
2818         }
2819
2820         /*
2821          * Transaction is finished.  We don't need the lock anymore.  We
2822          * do need to clean up the block groups in case of a transaction
2823          * abort.
2824          */
2825         deleted_bgs = &trans->transaction->deleted_bgs;
2826         list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
2827                 u64 trimmed = 0;
2828
2829                 ret = -EROFS;
2830                 if (!TRANS_ABORTED(trans))
2831                         ret = btrfs_discard_extent(fs_info,
2832                                                    block_group->start,
2833                                                    block_group->length,
2834                                                    &trimmed);
2835
2836                 list_del_init(&block_group->bg_list);
2837                 btrfs_unfreeze_block_group(block_group);
2838                 btrfs_put_block_group(block_group);
2839
2840                 if (ret) {
2841                         const char *errstr = btrfs_decode_error(ret);
2842                         btrfs_warn(fs_info,
2843                            "discard failed while removing blockgroup: errno=%d %s",
2844                                    ret, errstr);
2845                 }
2846         }
2847
2848         return 0;
2849 }
2850
2851 static int do_free_extent_accounting(struct btrfs_trans_handle *trans,
2852                                      u64 bytenr, u64 num_bytes, bool is_data)
2853 {
2854         int ret;
2855
2856         if (is_data) {
2857                 struct btrfs_root *csum_root;
2858
2859                 csum_root = btrfs_csum_root(trans->fs_info, bytenr);
2860                 ret = btrfs_del_csums(trans, csum_root, bytenr, num_bytes);
2861                 if (ret) {
2862                         btrfs_abort_transaction(trans, ret);
2863                         return ret;
2864                 }
2865         }
2866
2867         ret = add_to_free_space_tree(trans, bytenr, num_bytes);
2868         if (ret) {
2869                 btrfs_abort_transaction(trans, ret);
2870                 return ret;
2871         }
2872
2873         ret = btrfs_update_block_group(trans, bytenr, num_bytes, false);
2874         if (ret)
2875                 btrfs_abort_transaction(trans, ret);
2876
2877         return ret;
2878 }
2879
2880 #define abort_and_dump(trans, path, fmt, args...)       \
2881 ({                                                      \
2882         btrfs_abort_transaction(trans, -EUCLEAN);       \
2883         btrfs_print_leaf(path->nodes[0]);               \
2884         btrfs_crit(trans->fs_info, fmt, ##args);        \
2885 })
2886
2887 /*
2888  * Drop one or more refs of @node.
2889  *
2890  * 1. Locate the extent refs.
2891  *    It's either inline in EXTENT/METADATA_ITEM or in keyed SHARED_* item.
2892  *    Locate it, then reduce the refs number or remove the ref line completely.
2893  *
2894  * 2. Update the refs count in EXTENT/METADATA_ITEM
2895  *
2896  * Inline backref case:
2897  *
2898  * in extent tree we have:
2899  *
2900  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 16201 itemsize 82
2901  *              refs 2 gen 6 flags DATA
2902  *              extent data backref root FS_TREE objectid 258 offset 0 count 1
2903  *              extent data backref root FS_TREE objectid 257 offset 0 count 1
2904  *
2905  * This function gets called with:
2906  *
2907  *    node->bytenr = 13631488
2908  *    node->num_bytes = 1048576
2909  *    root_objectid = FS_TREE
2910  *    owner_objectid = 257
2911  *    owner_offset = 0
2912  *    refs_to_drop = 1
2913  *
2914  * Then we should get some like:
2915  *
2916  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 16201 itemsize 82
2917  *              refs 1 gen 6 flags DATA
2918  *              extent data backref root FS_TREE objectid 258 offset 0 count 1
2919  *
2920  * Keyed backref case:
2921  *
2922  * in extent tree we have:
2923  *
2924  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 3971 itemsize 24
2925  *              refs 754 gen 6 flags DATA
2926  *      [...]
2927  *      item 2 key (13631488 EXTENT_DATA_REF <HASH>) itemoff 3915 itemsize 28
2928  *              extent data backref root FS_TREE objectid 866 offset 0 count 1
2929  *
2930  * This function get called with:
2931  *
2932  *    node->bytenr = 13631488
2933  *    node->num_bytes = 1048576
2934  *    root_objectid = FS_TREE
2935  *    owner_objectid = 866
2936  *    owner_offset = 0
2937  *    refs_to_drop = 1
2938  *
2939  * Then we should get some like:
2940  *
2941  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 3971 itemsize 24
2942  *              refs 753 gen 6 flags DATA
2943  *
2944  * And that (13631488 EXTENT_DATA_REF <HASH>) gets removed.
2945  */
2946 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
2947                                struct btrfs_delayed_ref_node *node, u64 parent,
2948                                u64 root_objectid, u64 owner_objectid,
2949                                u64 owner_offset, int refs_to_drop,
2950                                struct btrfs_delayed_extent_op *extent_op)
2951 {
2952         struct btrfs_fs_info *info = trans->fs_info;
2953         struct btrfs_key key;
2954         struct btrfs_path *path;
2955         struct btrfs_root *extent_root;
2956         struct extent_buffer *leaf;
2957         struct btrfs_extent_item *ei;
2958         struct btrfs_extent_inline_ref *iref;
2959         int ret;
2960         int is_data;
2961         int extent_slot = 0;
2962         int found_extent = 0;
2963         int num_to_del = 1;
2964         u32 item_size;
2965         u64 refs;
2966         u64 bytenr = node->bytenr;
2967         u64 num_bytes = node->num_bytes;
2968         bool skinny_metadata = btrfs_fs_incompat(info, SKINNY_METADATA);
2969
2970         extent_root = btrfs_extent_root(info, bytenr);
2971         ASSERT(extent_root);
2972
2973         path = btrfs_alloc_path();
2974         if (!path)
2975                 return -ENOMEM;
2976
2977         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
2978
2979         if (!is_data && refs_to_drop != 1) {
2980                 btrfs_crit(info,
2981 "invalid refs_to_drop, dropping more than 1 refs for tree block %llu refs_to_drop %u",
2982                            node->bytenr, refs_to_drop);
2983                 ret = -EINVAL;
2984                 btrfs_abort_transaction(trans, ret);
2985                 goto out;
2986         }
2987
2988         if (is_data)
2989                 skinny_metadata = false;
2990
2991         ret = lookup_extent_backref(trans, path, &iref, bytenr, num_bytes,
2992                                     parent, root_objectid, owner_objectid,
2993                                     owner_offset);
2994         if (ret == 0) {
2995                 /*
2996                  * Either the inline backref or the SHARED_DATA_REF/
2997                  * SHARED_BLOCK_REF is found
2998                  *
2999                  * Here is a quick path to locate EXTENT/METADATA_ITEM.
3000                  * It's possible the EXTENT/METADATA_ITEM is near current slot.
3001                  */
3002                 extent_slot = path->slots[0];
3003                 while (extent_slot >= 0) {
3004                         btrfs_item_key_to_cpu(path->nodes[0], &key,
3005                                               extent_slot);
3006                         if (key.objectid != bytenr)
3007                                 break;
3008                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
3009                             key.offset == num_bytes) {
3010                                 found_extent = 1;
3011                                 break;
3012                         }
3013                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
3014                             key.offset == owner_objectid) {
3015                                 found_extent = 1;
3016                                 break;
3017                         }
3018
3019                         /* Quick path didn't find the EXTEMT/METADATA_ITEM */
3020                         if (path->slots[0] - extent_slot > 5)
3021                                 break;
3022                         extent_slot--;
3023                 }
3024
3025                 if (!found_extent) {
3026                         if (iref) {
3027                                 abort_and_dump(trans, path,
3028 "invalid iref slot %u, no EXTENT/METADATA_ITEM found but has inline extent ref",
3029                                            path->slots[0]);
3030                                 ret = -EUCLEAN;
3031                                 goto out;
3032                         }
3033                         /* Must be SHARED_* item, remove the backref first */
3034                         ret = remove_extent_backref(trans, extent_root, path,
3035                                                     NULL, refs_to_drop, is_data);
3036                         if (ret) {
3037                                 btrfs_abort_transaction(trans, ret);
3038                                 goto out;
3039                         }
3040                         btrfs_release_path(path);
3041
3042                         /* Slow path to locate EXTENT/METADATA_ITEM */
3043                         key.objectid = bytenr;
3044                         key.type = BTRFS_EXTENT_ITEM_KEY;
3045                         key.offset = num_bytes;
3046
3047                         if (!is_data && skinny_metadata) {
3048                                 key.type = BTRFS_METADATA_ITEM_KEY;
3049                                 key.offset = owner_objectid;
3050                         }
3051
3052                         ret = btrfs_search_slot(trans, extent_root,
3053                                                 &key, path, -1, 1);
3054                         if (ret > 0 && skinny_metadata && path->slots[0]) {
3055                                 /*
3056                                  * Couldn't find our skinny metadata item,
3057                                  * see if we have ye olde extent item.
3058                                  */
3059                                 path->slots[0]--;
3060                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
3061                                                       path->slots[0]);
3062                                 if (key.objectid == bytenr &&
3063                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
3064                                     key.offset == num_bytes)
3065                                         ret = 0;
3066                         }
3067
3068                         if (ret > 0 && skinny_metadata) {
3069                                 skinny_metadata = false;
3070                                 key.objectid = bytenr;
3071                                 key.type = BTRFS_EXTENT_ITEM_KEY;
3072                                 key.offset = num_bytes;
3073                                 btrfs_release_path(path);
3074                                 ret = btrfs_search_slot(trans, extent_root,
3075                                                         &key, path, -1, 1);
3076                         }
3077
3078                         if (ret) {
3079                                 if (ret > 0)
3080                                         btrfs_print_leaf(path->nodes[0]);
3081                                 btrfs_err(info,
3082                         "umm, got %d back from search, was looking for %llu, slot %d",
3083                                           ret, bytenr, path->slots[0]);
3084                         }
3085                         if (ret < 0) {
3086                                 btrfs_abort_transaction(trans, ret);
3087                                 goto out;
3088                         }
3089                         extent_slot = path->slots[0];
3090                 }
3091         } else if (WARN_ON(ret == -ENOENT)) {
3092                 abort_and_dump(trans, path,
3093 "unable to find ref byte nr %llu parent %llu root %llu owner %llu offset %llu slot %d",
3094                                bytenr, parent, root_objectid, owner_objectid,
3095                                owner_offset, path->slots[0]);
3096                 goto out;
3097         } else {
3098                 btrfs_abort_transaction(trans, ret);
3099                 goto out;
3100         }
3101
3102         leaf = path->nodes[0];
3103         item_size = btrfs_item_size(leaf, extent_slot);
3104         if (unlikely(item_size < sizeof(*ei))) {
3105                 ret = -EUCLEAN;
3106                 btrfs_err(trans->fs_info,
3107                           "unexpected extent item size, has %u expect >= %zu",
3108                           item_size, sizeof(*ei));
3109                 btrfs_abort_transaction(trans, ret);
3110                 goto out;
3111         }
3112         ei = btrfs_item_ptr(leaf, extent_slot,
3113                             struct btrfs_extent_item);
3114         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
3115             key.type == BTRFS_EXTENT_ITEM_KEY) {
3116                 struct btrfs_tree_block_info *bi;
3117
3118                 if (item_size < sizeof(*ei) + sizeof(*bi)) {
3119                         abort_and_dump(trans, path,
3120 "invalid extent item size for key (%llu, %u, %llu) slot %u owner %llu, has %u expect >= %zu",
3121                                        key.objectid, key.type, key.offset,
3122                                        path->slots[0], owner_objectid, item_size,
3123                                        sizeof(*ei) + sizeof(*bi));
3124                         ret = -EUCLEAN;
3125                         goto out;
3126                 }
3127                 bi = (struct btrfs_tree_block_info *)(ei + 1);
3128                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
3129         }
3130
3131         refs = btrfs_extent_refs(leaf, ei);
3132         if (refs < refs_to_drop) {
3133                 abort_and_dump(trans, path,
3134                 "trying to drop %d refs but we only have %llu for bytenr %llu slot %u",
3135                                refs_to_drop, refs, bytenr, path->slots[0]);
3136                 ret = -EUCLEAN;
3137                 goto out;
3138         }
3139         refs -= refs_to_drop;
3140
3141         if (refs > 0) {
3142                 if (extent_op)
3143                         __run_delayed_extent_op(extent_op, leaf, ei);
3144                 /*
3145                  * In the case of inline back ref, reference count will
3146                  * be updated by remove_extent_backref
3147                  */
3148                 if (iref) {
3149                         if (!found_extent) {
3150                                 abort_and_dump(trans, path,
3151 "invalid iref, got inlined extent ref but no EXTENT/METADATA_ITEM found, slot %u",
3152                                                path->slots[0]);
3153                                 ret = -EUCLEAN;
3154                                 goto out;
3155                         }
3156                 } else {
3157                         btrfs_set_extent_refs(leaf, ei, refs);
3158                         btrfs_mark_buffer_dirty(trans, leaf);
3159                 }
3160                 if (found_extent) {
3161                         ret = remove_extent_backref(trans, extent_root, path,
3162                                                     iref, refs_to_drop, is_data);
3163                         if (ret) {
3164                                 btrfs_abort_transaction(trans, ret);
3165                                 goto out;
3166                         }
3167                 }
3168         } else {
3169                 /* In this branch refs == 1 */
3170                 if (found_extent) {
3171                         if (is_data && refs_to_drop !=
3172                             extent_data_ref_count(path, iref)) {
3173                                 abort_and_dump(trans, path,
3174                 "invalid refs_to_drop, current refs %u refs_to_drop %u slot %u",
3175                                                extent_data_ref_count(path, iref),
3176                                                refs_to_drop, path->slots[0]);
3177                                 ret = -EUCLEAN;
3178                                 goto out;
3179                         }
3180                         if (iref) {
3181                                 if (path->slots[0] != extent_slot) {
3182                                         abort_and_dump(trans, path,
3183 "invalid iref, extent item key (%llu %u %llu) slot %u doesn't have wanted iref",
3184                                                        key.objectid, key.type,
3185                                                        key.offset, path->slots[0]);
3186                                         ret = -EUCLEAN;
3187                                         goto out;
3188                                 }
3189                         } else {
3190                                 /*
3191                                  * No inline ref, we must be at SHARED_* item,
3192                                  * And it's single ref, it must be:
3193                                  * |    extent_slot       ||extent_slot + 1|
3194                                  * [ EXTENT/METADATA_ITEM ][ SHARED_* ITEM ]
3195                                  */
3196                                 if (path->slots[0] != extent_slot + 1) {
3197                                         abort_and_dump(trans, path,
3198         "invalid SHARED_* item slot %u, previous item is not EXTENT/METADATA_ITEM",
3199                                                        path->slots[0]);
3200                                         ret = -EUCLEAN;
3201                                         goto out;
3202                                 }
3203                                 path->slots[0] = extent_slot;
3204                                 num_to_del = 2;
3205                         }
3206                 }
3207
3208                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
3209                                       num_to_del);
3210                 if (ret) {
3211                         btrfs_abort_transaction(trans, ret);
3212                         goto out;
3213                 }
3214                 btrfs_release_path(path);
3215
3216                 ret = do_free_extent_accounting(trans, bytenr, num_bytes, is_data);
3217         }
3218         btrfs_release_path(path);
3219
3220 out:
3221         btrfs_free_path(path);
3222         return ret;
3223 }
3224
3225 /*
3226  * when we free an block, it is possible (and likely) that we free the last
3227  * delayed ref for that extent as well.  This searches the delayed ref tree for
3228  * a given extent, and if there are no other delayed refs to be processed, it
3229  * removes it from the tree.
3230  */
3231 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
3232                                       u64 bytenr)
3233 {
3234         struct btrfs_delayed_ref_head *head;
3235         struct btrfs_delayed_ref_root *delayed_refs;
3236         int ret = 0;
3237
3238         delayed_refs = &trans->transaction->delayed_refs;
3239         spin_lock(&delayed_refs->lock);
3240         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
3241         if (!head)
3242                 goto out_delayed_unlock;
3243
3244         spin_lock(&head->lock);
3245         if (!RB_EMPTY_ROOT(&head->ref_tree.rb_root))
3246                 goto out;
3247
3248         if (cleanup_extent_op(head) != NULL)
3249                 goto out;
3250
3251         /*
3252          * waiting for the lock here would deadlock.  If someone else has it
3253          * locked they are already in the process of dropping it anyway
3254          */
3255         if (!mutex_trylock(&head->mutex))
3256                 goto out;
3257
3258         btrfs_delete_ref_head(delayed_refs, head);
3259         head->processing = false;
3260
3261         spin_unlock(&head->lock);
3262         spin_unlock(&delayed_refs->lock);
3263
3264         BUG_ON(head->extent_op);
3265         if (head->must_insert_reserved)
3266                 ret = 1;
3267
3268         btrfs_cleanup_ref_head_accounting(trans->fs_info, delayed_refs, head);
3269         mutex_unlock(&head->mutex);
3270         btrfs_put_delayed_ref_head(head);
3271         return ret;
3272 out:
3273         spin_unlock(&head->lock);
3274
3275 out_delayed_unlock:
3276         spin_unlock(&delayed_refs->lock);
3277         return 0;
3278 }
3279
3280 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
3281                            u64 root_id,
3282                            struct extent_buffer *buf,
3283                            u64 parent, int last_ref)
3284 {
3285         struct btrfs_fs_info *fs_info = trans->fs_info;
3286         struct btrfs_ref generic_ref = { 0 };
3287         int ret;
3288
3289         btrfs_init_generic_ref(&generic_ref, BTRFS_DROP_DELAYED_REF,
3290                                buf->start, buf->len, parent);
3291         btrfs_init_tree_ref(&generic_ref, btrfs_header_level(buf),
3292                             root_id, 0, false);
3293
3294         if (root_id != BTRFS_TREE_LOG_OBJECTID) {
3295                 btrfs_ref_tree_mod(fs_info, &generic_ref);
3296                 ret = btrfs_add_delayed_tree_ref(trans, &generic_ref, NULL);
3297                 BUG_ON(ret); /* -ENOMEM */
3298         }
3299
3300         if (last_ref && btrfs_header_generation(buf) == trans->transid) {
3301                 struct btrfs_block_group *cache;
3302                 bool must_pin = false;
3303
3304                 if (root_id != BTRFS_TREE_LOG_OBJECTID) {
3305                         ret = check_ref_cleanup(trans, buf->start);
3306                         if (!ret) {
3307                                 btrfs_redirty_list_add(trans->transaction, buf);
3308                                 goto out;
3309                         }
3310                 }
3311
3312                 cache = btrfs_lookup_block_group(fs_info, buf->start);
3313
3314                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
3315                         pin_down_extent(trans, cache, buf->start, buf->len, 1);
3316                         btrfs_put_block_group(cache);
3317                         goto out;
3318                 }
3319
3320                 /*
3321                  * If there are tree mod log users we may have recorded mod log
3322                  * operations for this node.  If we re-allocate this node we
3323                  * could replay operations on this node that happened when it
3324                  * existed in a completely different root.  For example if it
3325                  * was part of root A, then was reallocated to root B, and we
3326                  * are doing a btrfs_old_search_slot(root b), we could replay
3327                  * operations that happened when the block was part of root A,
3328                  * giving us an inconsistent view of the btree.
3329                  *
3330                  * We are safe from races here because at this point no other
3331                  * node or root points to this extent buffer, so if after this
3332                  * check a new tree mod log user joins we will not have an
3333                  * existing log of operations on this node that we have to
3334                  * contend with.
3335                  */
3336                 if (test_bit(BTRFS_FS_TREE_MOD_LOG_USERS, &fs_info->flags))
3337                         must_pin = true;
3338
3339                 if (must_pin || btrfs_is_zoned(fs_info)) {
3340                         btrfs_redirty_list_add(trans->transaction, buf);
3341                         pin_down_extent(trans, cache, buf->start, buf->len, 1);
3342                         btrfs_put_block_group(cache);
3343                         goto out;
3344                 }
3345
3346                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
3347
3348                 btrfs_add_free_space(cache, buf->start, buf->len);
3349                 btrfs_free_reserved_bytes(cache, buf->len, 0);
3350                 btrfs_put_block_group(cache);
3351                 trace_btrfs_reserved_extent_free(fs_info, buf->start, buf->len);
3352         }
3353 out:
3354         if (last_ref) {
3355                 /*
3356                  * Deleting the buffer, clear the corrupt flag since it doesn't
3357                  * matter anymore.
3358                  */
3359                 clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
3360         }
3361 }
3362
3363 /* Can return -ENOMEM */
3364 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_ref *ref)
3365 {
3366         struct btrfs_fs_info *fs_info = trans->fs_info;
3367         int ret;
3368
3369         if (btrfs_is_testing(fs_info))
3370                 return 0;
3371
3372         /*
3373          * tree log blocks never actually go into the extent allocation
3374          * tree, just update pinning info and exit early.
3375          */
3376         if ((ref->type == BTRFS_REF_METADATA &&
3377              ref->tree_ref.owning_root == BTRFS_TREE_LOG_OBJECTID) ||
3378             (ref->type == BTRFS_REF_DATA &&
3379              ref->data_ref.owning_root == BTRFS_TREE_LOG_OBJECTID)) {
3380                 /* unlocks the pinned mutex */
3381                 btrfs_pin_extent(trans, ref->bytenr, ref->len, 1);
3382                 ret = 0;
3383         } else if (ref->type == BTRFS_REF_METADATA) {
3384                 ret = btrfs_add_delayed_tree_ref(trans, ref, NULL);
3385         } else {
3386                 ret = btrfs_add_delayed_data_ref(trans, ref, 0);
3387         }
3388
3389         if (!((ref->type == BTRFS_REF_METADATA &&
3390                ref->tree_ref.owning_root == BTRFS_TREE_LOG_OBJECTID) ||
3391               (ref->type == BTRFS_REF_DATA &&
3392                ref->data_ref.owning_root == BTRFS_TREE_LOG_OBJECTID)))
3393                 btrfs_ref_tree_mod(fs_info, ref);
3394
3395         return ret;
3396 }
3397
3398 enum btrfs_loop_type {
3399         /*
3400          * Start caching block groups but do not wait for progress or for them
3401          * to be done.
3402          */
3403         LOOP_CACHING_NOWAIT,
3404
3405         /*
3406          * Wait for the block group free_space >= the space we're waiting for if
3407          * the block group isn't cached.
3408          */
3409         LOOP_CACHING_WAIT,
3410
3411         /*
3412          * Allow allocations to happen from block groups that do not yet have a
3413          * size classification.
3414          */
3415         LOOP_UNSET_SIZE_CLASS,
3416
3417         /*
3418          * Allocate a chunk and then retry the allocation.
3419          */
3420         LOOP_ALLOC_CHUNK,
3421
3422         /*
3423          * Ignore the size class restrictions for this allocation.
3424          */
3425         LOOP_WRONG_SIZE_CLASS,
3426
3427         /*
3428          * Ignore the empty size, only try to allocate the number of bytes
3429          * needed for this allocation.
3430          */
3431         LOOP_NO_EMPTY_SIZE,
3432 };
3433
3434 static inline void
3435 btrfs_lock_block_group(struct btrfs_block_group *cache,
3436                        int delalloc)
3437 {
3438         if (delalloc)
3439                 down_read(&cache->data_rwsem);
3440 }
3441
3442 static inline void btrfs_grab_block_group(struct btrfs_block_group *cache,
3443                        int delalloc)
3444 {
3445         btrfs_get_block_group(cache);
3446         if (delalloc)
3447                 down_read(&cache->data_rwsem);
3448 }
3449
3450 static struct btrfs_block_group *btrfs_lock_cluster(
3451                    struct btrfs_block_group *block_group,
3452                    struct btrfs_free_cluster *cluster,
3453                    int delalloc)
3454         __acquires(&cluster->refill_lock)
3455 {
3456         struct btrfs_block_group *used_bg = NULL;
3457
3458         spin_lock(&cluster->refill_lock);
3459         while (1) {
3460                 used_bg = cluster->block_group;
3461                 if (!used_bg)
3462                         return NULL;
3463
3464                 if (used_bg == block_group)
3465                         return used_bg;
3466
3467                 btrfs_get_block_group(used_bg);
3468
3469                 if (!delalloc)
3470                         return used_bg;
3471
3472                 if (down_read_trylock(&used_bg->data_rwsem))
3473                         return used_bg;
3474
3475                 spin_unlock(&cluster->refill_lock);
3476
3477                 /* We should only have one-level nested. */
3478                 down_read_nested(&used_bg->data_rwsem, SINGLE_DEPTH_NESTING);
3479
3480                 spin_lock(&cluster->refill_lock);
3481                 if (used_bg == cluster->block_group)
3482                         return used_bg;
3483
3484                 up_read(&used_bg->data_rwsem);
3485                 btrfs_put_block_group(used_bg);
3486         }
3487 }
3488
3489 static inline void
3490 btrfs_release_block_group(struct btrfs_block_group *cache,
3491                          int delalloc)
3492 {
3493         if (delalloc)
3494                 up_read(&cache->data_rwsem);
3495         btrfs_put_block_group(cache);
3496 }
3497
3498 /*
3499  * Helper function for find_free_extent().
3500  *
3501  * Return -ENOENT to inform caller that we need fallback to unclustered mode.
3502  * Return >0 to inform caller that we find nothing
3503  * Return 0 means we have found a location and set ffe_ctl->found_offset.
3504  */
3505 static int find_free_extent_clustered(struct btrfs_block_group *bg,
3506                                       struct find_free_extent_ctl *ffe_ctl,
3507                                       struct btrfs_block_group **cluster_bg_ret)
3508 {
3509         struct btrfs_block_group *cluster_bg;
3510         struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
3511         u64 aligned_cluster;
3512         u64 offset;
3513         int ret;
3514
3515         cluster_bg = btrfs_lock_cluster(bg, last_ptr, ffe_ctl->delalloc);
3516         if (!cluster_bg)
3517                 goto refill_cluster;
3518         if (cluster_bg != bg && (cluster_bg->ro ||
3519             !block_group_bits(cluster_bg, ffe_ctl->flags)))
3520                 goto release_cluster;
3521
3522         offset = btrfs_alloc_from_cluster(cluster_bg, last_ptr,
3523                         ffe_ctl->num_bytes, cluster_bg->start,
3524                         &ffe_ctl->max_extent_size);
3525         if (offset) {
3526                 /* We have a block, we're done */
3527                 spin_unlock(&last_ptr->refill_lock);
3528                 trace_btrfs_reserve_extent_cluster(cluster_bg, ffe_ctl);
3529                 *cluster_bg_ret = cluster_bg;
3530                 ffe_ctl->found_offset = offset;
3531                 return 0;
3532         }
3533         WARN_ON(last_ptr->block_group != cluster_bg);
3534
3535 release_cluster:
3536         /*
3537          * If we are on LOOP_NO_EMPTY_SIZE, we can't set up a new clusters, so
3538          * lets just skip it and let the allocator find whatever block it can
3539          * find. If we reach this point, we will have tried the cluster
3540          * allocator plenty of times and not have found anything, so we are
3541          * likely way too fragmented for the clustering stuff to find anything.
3542          *
3543          * However, if the cluster is taken from the current block group,
3544          * release the cluster first, so that we stand a better chance of
3545          * succeeding in the unclustered allocation.
3546          */
3547         if (ffe_ctl->loop >= LOOP_NO_EMPTY_SIZE && cluster_bg != bg) {
3548                 spin_unlock(&last_ptr->refill_lock);
3549                 btrfs_release_block_group(cluster_bg, ffe_ctl->delalloc);
3550                 return -ENOENT;
3551         }
3552
3553         /* This cluster didn't work out, free it and start over */
3554         btrfs_return_cluster_to_free_space(NULL, last_ptr);
3555
3556         if (cluster_bg != bg)
3557                 btrfs_release_block_group(cluster_bg, ffe_ctl->delalloc);
3558
3559 refill_cluster:
3560         if (ffe_ctl->loop >= LOOP_NO_EMPTY_SIZE) {
3561                 spin_unlock(&last_ptr->refill_lock);
3562                 return -ENOENT;
3563         }
3564
3565         aligned_cluster = max_t(u64,
3566                         ffe_ctl->empty_cluster + ffe_ctl->empty_size,
3567                         bg->full_stripe_len);
3568         ret = btrfs_find_space_cluster(bg, last_ptr, ffe_ctl->search_start,
3569                         ffe_ctl->num_bytes, aligned_cluster);
3570         if (ret == 0) {
3571                 /* Now pull our allocation out of this cluster */
3572                 offset = btrfs_alloc_from_cluster(bg, last_ptr,
3573                                 ffe_ctl->num_bytes, ffe_ctl->search_start,
3574                                 &ffe_ctl->max_extent_size);
3575                 if (offset) {
3576                         /* We found one, proceed */
3577                         spin_unlock(&last_ptr->refill_lock);
3578                         ffe_ctl->found_offset = offset;
3579                         trace_btrfs_reserve_extent_cluster(bg, ffe_ctl);
3580                         return 0;
3581                 }
3582         }
3583         /*
3584          * At this point we either didn't find a cluster or we weren't able to
3585          * allocate a block from our cluster.  Free the cluster we've been
3586          * trying to use, and go to the next block group.
3587          */
3588         btrfs_return_cluster_to_free_space(NULL, last_ptr);
3589         spin_unlock(&last_ptr->refill_lock);
3590         return 1;
3591 }
3592
3593 /*
3594  * Return >0 to inform caller that we find nothing
3595  * Return 0 when we found an free extent and set ffe_ctrl->found_offset
3596  */
3597 static int find_free_extent_unclustered(struct btrfs_block_group *bg,
3598                                         struct find_free_extent_ctl *ffe_ctl)
3599 {
3600         struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
3601         u64 offset;
3602
3603         /*
3604          * We are doing an unclustered allocation, set the fragmented flag so
3605          * we don't bother trying to setup a cluster again until we get more
3606          * space.
3607          */
3608         if (unlikely(last_ptr)) {
3609                 spin_lock(&last_ptr->lock);
3610                 last_ptr->fragmented = 1;
3611                 spin_unlock(&last_ptr->lock);
3612         }
3613         if (ffe_ctl->cached) {
3614                 struct btrfs_free_space_ctl *free_space_ctl;
3615
3616                 free_space_ctl = bg->free_space_ctl;
3617                 spin_lock(&free_space_ctl->tree_lock);
3618                 if (free_space_ctl->free_space <
3619                     ffe_ctl->num_bytes + ffe_ctl->empty_cluster +
3620                     ffe_ctl->empty_size) {
3621                         ffe_ctl->total_free_space = max_t(u64,
3622                                         ffe_ctl->total_free_space,
3623                                         free_space_ctl->free_space);
3624                         spin_unlock(&free_space_ctl->tree_lock);
3625                         return 1;
3626                 }
3627                 spin_unlock(&free_space_ctl->tree_lock);
3628         }
3629
3630         offset = btrfs_find_space_for_alloc(bg, ffe_ctl->search_start,
3631                         ffe_ctl->num_bytes, ffe_ctl->empty_size,
3632                         &ffe_ctl->max_extent_size);
3633         if (!offset)
3634                 return 1;
3635         ffe_ctl->found_offset = offset;
3636         return 0;
3637 }
3638
3639 static int do_allocation_clustered(struct btrfs_block_group *block_group,
3640                                    struct find_free_extent_ctl *ffe_ctl,
3641                                    struct btrfs_block_group **bg_ret)
3642 {
3643         int ret;
3644
3645         /* We want to try and use the cluster allocator, so lets look there */
3646         if (ffe_ctl->last_ptr && ffe_ctl->use_cluster) {
3647                 ret = find_free_extent_clustered(block_group, ffe_ctl, bg_ret);
3648                 if (ret >= 0)
3649                         return ret;
3650                 /* ret == -ENOENT case falls through */
3651         }
3652
3653         return find_free_extent_unclustered(block_group, ffe_ctl);
3654 }
3655
3656 /*
3657  * Tree-log block group locking
3658  * ============================
3659  *
3660  * fs_info::treelog_bg_lock protects the fs_info::treelog_bg which
3661  * indicates the starting address of a block group, which is reserved only
3662  * for tree-log metadata.
3663  *
3664  * Lock nesting
3665  * ============
3666  *
3667  * space_info::lock
3668  *   block_group::lock
3669  *     fs_info::treelog_bg_lock
3670  */
3671
3672 /*
3673  * Simple allocator for sequential-only block group. It only allows sequential
3674  * allocation. No need to play with trees. This function also reserves the
3675  * bytes as in btrfs_add_reserved_bytes.
3676  */
3677 static int do_allocation_zoned(struct btrfs_block_group *block_group,
3678                                struct find_free_extent_ctl *ffe_ctl,
3679                                struct btrfs_block_group **bg_ret)
3680 {
3681         struct btrfs_fs_info *fs_info = block_group->fs_info;
3682         struct btrfs_space_info *space_info = block_group->space_info;
3683         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3684         u64 start = block_group->start;
3685         u64 num_bytes = ffe_ctl->num_bytes;
3686         u64 avail;
3687         u64 bytenr = block_group->start;
3688         u64 log_bytenr;
3689         u64 data_reloc_bytenr;
3690         int ret = 0;
3691         bool skip = false;
3692
3693         ASSERT(btrfs_is_zoned(block_group->fs_info));
3694
3695         /*
3696          * Do not allow non-tree-log blocks in the dedicated tree-log block
3697          * group, and vice versa.
3698          */
3699         spin_lock(&fs_info->treelog_bg_lock);
3700         log_bytenr = fs_info->treelog_bg;
3701         if (log_bytenr && ((ffe_ctl->for_treelog && bytenr != log_bytenr) ||
3702                            (!ffe_ctl->for_treelog && bytenr == log_bytenr)))
3703                 skip = true;
3704         spin_unlock(&fs_info->treelog_bg_lock);
3705         if (skip)
3706                 return 1;
3707
3708         /*
3709          * Do not allow non-relocation blocks in the dedicated relocation block
3710          * group, and vice versa.
3711          */
3712         spin_lock(&fs_info->relocation_bg_lock);
3713         data_reloc_bytenr = fs_info->data_reloc_bg;
3714         if (data_reloc_bytenr &&
3715             ((ffe_ctl->for_data_reloc && bytenr != data_reloc_bytenr) ||
3716              (!ffe_ctl->for_data_reloc && bytenr == data_reloc_bytenr)))
3717                 skip = true;
3718         spin_unlock(&fs_info->relocation_bg_lock);
3719         if (skip)
3720                 return 1;
3721
3722         /* Check RO and no space case before trying to activate it */
3723         spin_lock(&block_group->lock);
3724         if (block_group->ro || btrfs_zoned_bg_is_full(block_group)) {
3725                 ret = 1;
3726                 /*
3727                  * May need to clear fs_info->{treelog,data_reloc}_bg.
3728                  * Return the error after taking the locks.
3729                  */
3730         }
3731         spin_unlock(&block_group->lock);
3732
3733         /* Metadata block group is activated at write time. */
3734         if (!ret && (block_group->flags & BTRFS_BLOCK_GROUP_DATA) &&
3735             !btrfs_zone_activate(block_group)) {
3736                 ret = 1;
3737                 /*
3738                  * May need to clear fs_info->{treelog,data_reloc}_bg.
3739                  * Return the error after taking the locks.
3740                  */
3741         }
3742
3743         spin_lock(&space_info->lock);
3744         spin_lock(&block_group->lock);
3745         spin_lock(&fs_info->treelog_bg_lock);
3746         spin_lock(&fs_info->relocation_bg_lock);
3747
3748         if (ret)
3749                 goto out;
3750
3751         ASSERT(!ffe_ctl->for_treelog ||
3752                block_group->start == fs_info->treelog_bg ||
3753                fs_info->treelog_bg == 0);
3754         ASSERT(!ffe_ctl->for_data_reloc ||
3755                block_group->start == fs_info->data_reloc_bg ||
3756                fs_info->data_reloc_bg == 0);
3757
3758         if (block_group->ro ||
3759             (!ffe_ctl->for_data_reloc &&
3760              test_bit(BLOCK_GROUP_FLAG_ZONED_DATA_RELOC, &block_group->runtime_flags))) {
3761                 ret = 1;
3762                 goto out;
3763         }
3764
3765         /*
3766          * Do not allow currently using block group to be tree-log dedicated
3767          * block group.
3768          */
3769         if (ffe_ctl->for_treelog && !fs_info->treelog_bg &&
3770             (block_group->used || block_group->reserved)) {
3771                 ret = 1;
3772                 goto out;
3773         }
3774
3775         /*
3776          * Do not allow currently used block group to be the data relocation
3777          * dedicated block group.
3778          */
3779         if (ffe_ctl->for_data_reloc && !fs_info->data_reloc_bg &&
3780             (block_group->used || block_group->reserved)) {
3781                 ret = 1;
3782                 goto out;
3783         }
3784
3785         WARN_ON_ONCE(block_group->alloc_offset > block_group->zone_capacity);
3786         avail = block_group->zone_capacity - block_group->alloc_offset;
3787         if (avail < num_bytes) {
3788                 if (ffe_ctl->max_extent_size < avail) {
3789                         /*
3790                          * With sequential allocator, free space is always
3791                          * contiguous
3792                          */
3793                         ffe_ctl->max_extent_size = avail;
3794                         ffe_ctl->total_free_space = avail;
3795                 }
3796                 ret = 1;
3797                 goto out;
3798         }
3799
3800         if (ffe_ctl->for_treelog && !fs_info->treelog_bg)
3801                 fs_info->treelog_bg = block_group->start;
3802
3803         if (ffe_ctl->for_data_reloc) {
3804                 if (!fs_info->data_reloc_bg)
3805                         fs_info->data_reloc_bg = block_group->start;
3806                 /*
3807                  * Do not allow allocations from this block group, unless it is
3808                  * for data relocation. Compared to increasing the ->ro, setting
3809                  * the ->zoned_data_reloc_ongoing flag still allows nocow
3810                  * writers to come in. See btrfs_inc_nocow_writers().
3811                  *
3812                  * We need to disable an allocation to avoid an allocation of
3813                  * regular (non-relocation data) extent. With mix of relocation
3814                  * extents and regular extents, we can dispatch WRITE commands
3815                  * (for relocation extents) and ZONE APPEND commands (for
3816                  * regular extents) at the same time to the same zone, which
3817                  * easily break the write pointer.
3818                  *
3819                  * Also, this flag avoids this block group to be zone finished.
3820                  */
3821                 set_bit(BLOCK_GROUP_FLAG_ZONED_DATA_RELOC, &block_group->runtime_flags);
3822         }
3823
3824         ffe_ctl->found_offset = start + block_group->alloc_offset;
3825         block_group->alloc_offset += num_bytes;
3826         spin_lock(&ctl->tree_lock);
3827         ctl->free_space -= num_bytes;
3828         spin_unlock(&ctl->tree_lock);
3829
3830         /*
3831          * We do not check if found_offset is aligned to stripesize. The
3832          * address is anyway rewritten when using zone append writing.
3833          */
3834
3835         ffe_ctl->search_start = ffe_ctl->found_offset;
3836
3837 out:
3838         if (ret && ffe_ctl->for_treelog)
3839                 fs_info->treelog_bg = 0;
3840         if (ret && ffe_ctl->for_data_reloc)
3841                 fs_info->data_reloc_bg = 0;
3842         spin_unlock(&fs_info->relocation_bg_lock);
3843         spin_unlock(&fs_info->treelog_bg_lock);
3844         spin_unlock(&block_group->lock);
3845         spin_unlock(&space_info->lock);
3846         return ret;
3847 }
3848
3849 static int do_allocation(struct btrfs_block_group *block_group,
3850                          struct find_free_extent_ctl *ffe_ctl,
3851                          struct btrfs_block_group **bg_ret)
3852 {
3853         switch (ffe_ctl->policy) {
3854         case BTRFS_EXTENT_ALLOC_CLUSTERED:
3855                 return do_allocation_clustered(block_group, ffe_ctl, bg_ret);
3856         case BTRFS_EXTENT_ALLOC_ZONED:
3857                 return do_allocation_zoned(block_group, ffe_ctl, bg_ret);
3858         default:
3859                 BUG();
3860         }
3861 }
3862
3863 static void release_block_group(struct btrfs_block_group *block_group,
3864                                 struct find_free_extent_ctl *ffe_ctl,
3865                                 int delalloc)
3866 {
3867         switch (ffe_ctl->policy) {
3868         case BTRFS_EXTENT_ALLOC_CLUSTERED:
3869                 ffe_ctl->retry_uncached = false;
3870                 break;
3871         case BTRFS_EXTENT_ALLOC_ZONED:
3872                 /* Nothing to do */
3873                 break;
3874         default:
3875                 BUG();
3876         }
3877
3878         BUG_ON(btrfs_bg_flags_to_raid_index(block_group->flags) !=
3879                ffe_ctl->index);
3880         btrfs_release_block_group(block_group, delalloc);
3881 }
3882
3883 static void found_extent_clustered(struct find_free_extent_ctl *ffe_ctl,
3884                                    struct btrfs_key *ins)
3885 {
3886         struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
3887
3888         if (!ffe_ctl->use_cluster && last_ptr) {
3889                 spin_lock(&last_ptr->lock);
3890                 last_ptr->window_start = ins->objectid;
3891                 spin_unlock(&last_ptr->lock);
3892         }
3893 }
3894
3895 static void found_extent(struct find_free_extent_ctl *ffe_ctl,
3896                          struct btrfs_key *ins)
3897 {
3898         switch (ffe_ctl->policy) {
3899         case BTRFS_EXTENT_ALLOC_CLUSTERED:
3900                 found_extent_clustered(ffe_ctl, ins);
3901                 break;
3902         case BTRFS_EXTENT_ALLOC_ZONED:
3903                 /* Nothing to do */
3904                 break;
3905         default:
3906                 BUG();
3907         }
3908 }
3909
3910 static int can_allocate_chunk_zoned(struct btrfs_fs_info *fs_info,
3911                                     struct find_free_extent_ctl *ffe_ctl)
3912 {
3913         /* Block group's activeness is not a requirement for METADATA block groups. */
3914         if (!(ffe_ctl->flags & BTRFS_BLOCK_GROUP_DATA))
3915                 return 0;
3916
3917         /* If we can activate new zone, just allocate a chunk and use it */
3918         if (btrfs_can_activate_zone(fs_info->fs_devices, ffe_ctl->flags))
3919                 return 0;
3920
3921         /*
3922          * We already reached the max active zones. Try to finish one block
3923          * group to make a room for a new block group. This is only possible
3924          * for a data block group because btrfs_zone_finish() may need to wait
3925          * for a running transaction which can cause a deadlock for metadata
3926          * allocation.
3927          */
3928         if (ffe_ctl->flags & BTRFS_BLOCK_GROUP_DATA) {
3929                 int ret = btrfs_zone_finish_one_bg(fs_info);
3930
3931                 if (ret == 1)
3932                         return 0;
3933                 else if (ret < 0)
3934                         return ret;
3935         }
3936
3937         /*
3938          * If we have enough free space left in an already active block group
3939          * and we can't activate any other zone now, do not allow allocating a
3940          * new chunk and let find_free_extent() retry with a smaller size.
3941          */
3942         if (ffe_ctl->max_extent_size >= ffe_ctl->min_alloc_size)
3943                 return -ENOSPC;
3944
3945         /*
3946          * Even min_alloc_size is not left in any block groups. Since we cannot
3947          * activate a new block group, allocating it may not help. Let's tell a
3948          * caller to try again and hope it progress something by writing some
3949          * parts of the region. That is only possible for data block groups,
3950          * where a part of the region can be written.
3951          */
3952         if (ffe_ctl->flags & BTRFS_BLOCK_GROUP_DATA)
3953                 return -EAGAIN;
3954
3955         /*
3956          * We cannot activate a new block group and no enough space left in any
3957          * block groups. So, allocating a new block group may not help. But,
3958          * there is nothing to do anyway, so let's go with it.
3959          */
3960         return 0;
3961 }
3962
3963 static int can_allocate_chunk(struct btrfs_fs_info *fs_info,
3964                               struct find_free_extent_ctl *ffe_ctl)
3965 {
3966         switch (ffe_ctl->policy) {
3967         case BTRFS_EXTENT_ALLOC_CLUSTERED:
3968                 return 0;
3969         case BTRFS_EXTENT_ALLOC_ZONED:
3970                 return can_allocate_chunk_zoned(fs_info, ffe_ctl);
3971         default:
3972                 BUG();
3973         }
3974 }
3975
3976 /*
3977  * Return >0 means caller needs to re-search for free extent
3978  * Return 0 means we have the needed free extent.
3979  * Return <0 means we failed to locate any free extent.
3980  */
3981 static int find_free_extent_update_loop(struct btrfs_fs_info *fs_info,
3982                                         struct btrfs_key *ins,
3983                                         struct find_free_extent_ctl *ffe_ctl,
3984                                         bool full_search)
3985 {
3986         struct btrfs_root *root = fs_info->chunk_root;
3987         int ret;
3988
3989         if ((ffe_ctl->loop == LOOP_CACHING_NOWAIT) &&
3990             ffe_ctl->have_caching_bg && !ffe_ctl->orig_have_caching_bg)
3991                 ffe_ctl->orig_have_caching_bg = true;
3992
3993         if (ins->objectid) {
3994                 found_extent(ffe_ctl, ins);
3995                 return 0;
3996         }
3997
3998         if (ffe_ctl->loop >= LOOP_CACHING_WAIT && ffe_ctl->have_caching_bg)
3999                 return 1;
4000
4001         ffe_ctl->index++;
4002         if (ffe_ctl->index < BTRFS_NR_RAID_TYPES)
4003                 return 1;
4004
4005         /* See the comments for btrfs_loop_type for an explanation of the phases. */
4006         if (ffe_ctl->loop < LOOP_NO_EMPTY_SIZE) {
4007                 ffe_ctl->index = 0;
4008                 /*
4009                  * We want to skip the LOOP_CACHING_WAIT step if we don't have
4010                  * any uncached bgs and we've already done a full search
4011                  * through.
4012                  */
4013                 if (ffe_ctl->loop == LOOP_CACHING_NOWAIT &&
4014                     (!ffe_ctl->orig_have_caching_bg && full_search))
4015                         ffe_ctl->loop++;
4016                 ffe_ctl->loop++;
4017
4018                 if (ffe_ctl->loop == LOOP_ALLOC_CHUNK) {
4019                         struct btrfs_trans_handle *trans;
4020                         int exist = 0;
4021
4022                         /* Check if allocation policy allows to create a new chunk */
4023                         ret = can_allocate_chunk(fs_info, ffe_ctl);
4024                         if (ret)
4025                                 return ret;
4026
4027                         trans = current->journal_info;
4028                         if (trans)
4029                                 exist = 1;
4030                         else
4031                                 trans = btrfs_join_transaction(root);
4032
4033                         if (IS_ERR(trans)) {
4034                                 ret = PTR_ERR(trans);
4035                                 return ret;
4036                         }
4037
4038                         ret = btrfs_chunk_alloc(trans, ffe_ctl->flags,
4039                                                 CHUNK_ALLOC_FORCE_FOR_EXTENT);
4040
4041                         /* Do not bail out on ENOSPC since we can do more. */
4042                         if (ret == -ENOSPC) {
4043                                 ret = 0;
4044                                 ffe_ctl->loop++;
4045                         }
4046                         else if (ret < 0)
4047                                 btrfs_abort_transaction(trans, ret);
4048                         else
4049                                 ret = 0;
4050                         if (!exist)
4051                                 btrfs_end_transaction(trans);
4052                         if (ret)
4053                                 return ret;
4054                 }
4055
4056                 if (ffe_ctl->loop == LOOP_NO_EMPTY_SIZE) {
4057                         if (ffe_ctl->policy != BTRFS_EXTENT_ALLOC_CLUSTERED)
4058                                 return -ENOSPC;
4059
4060                         /*
4061                          * Don't loop again if we already have no empty_size and
4062                          * no empty_cluster.
4063                          */
4064                         if (ffe_ctl->empty_size == 0 &&
4065                             ffe_ctl->empty_cluster == 0)
4066                                 return -ENOSPC;
4067                         ffe_ctl->empty_size = 0;
4068                         ffe_ctl->empty_cluster = 0;
4069                 }
4070                 return 1;
4071         }
4072         return -ENOSPC;
4073 }
4074
4075 static bool find_free_extent_check_size_class(struct find_free_extent_ctl *ffe_ctl,
4076                                               struct btrfs_block_group *bg)
4077 {
4078         if (ffe_ctl->policy == BTRFS_EXTENT_ALLOC_ZONED)
4079                 return true;
4080         if (!btrfs_block_group_should_use_size_class(bg))
4081                 return true;
4082         if (ffe_ctl->loop >= LOOP_WRONG_SIZE_CLASS)
4083                 return true;
4084         if (ffe_ctl->loop >= LOOP_UNSET_SIZE_CLASS &&
4085             bg->size_class == BTRFS_BG_SZ_NONE)
4086                 return true;
4087         return ffe_ctl->size_class == bg->size_class;
4088 }
4089
4090 static int prepare_allocation_clustered(struct btrfs_fs_info *fs_info,
4091                                         struct find_free_extent_ctl *ffe_ctl,
4092                                         struct btrfs_space_info *space_info,
4093                                         struct btrfs_key *ins)
4094 {
4095         /*
4096          * If our free space is heavily fragmented we may not be able to make
4097          * big contiguous allocations, so instead of doing the expensive search
4098          * for free space, simply return ENOSPC with our max_extent_size so we
4099          * can go ahead and search for a more manageable chunk.
4100          *
4101          * If our max_extent_size is large enough for our allocation simply
4102          * disable clustering since we will likely not be able to find enough
4103          * space to create a cluster and induce latency trying.
4104          */
4105         if (space_info->max_extent_size) {
4106                 spin_lock(&space_info->lock);
4107                 if (space_info->max_extent_size &&
4108                     ffe_ctl->num_bytes > space_info->max_extent_size) {
4109                         ins->offset = space_info->max_extent_size;
4110                         spin_unlock(&space_info->lock);
4111                         return -ENOSPC;
4112                 } else if (space_info->max_extent_size) {
4113                         ffe_ctl->use_cluster = false;
4114                 }
4115                 spin_unlock(&space_info->lock);
4116         }
4117
4118         ffe_ctl->last_ptr = fetch_cluster_info(fs_info, space_info,
4119                                                &ffe_ctl->empty_cluster);
4120         if (ffe_ctl->last_ptr) {
4121                 struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
4122
4123                 spin_lock(&last_ptr->lock);
4124                 if (last_ptr->block_group)
4125                         ffe_ctl->hint_byte = last_ptr->window_start;
4126                 if (last_ptr->fragmented) {
4127                         /*
4128                          * We still set window_start so we can keep track of the
4129                          * last place we found an allocation to try and save
4130                          * some time.
4131                          */
4132                         ffe_ctl->hint_byte = last_ptr->window_start;
4133                         ffe_ctl->use_cluster = false;
4134                 }
4135                 spin_unlock(&last_ptr->lock);
4136         }
4137
4138         return 0;
4139 }
4140
4141 static int prepare_allocation(struct btrfs_fs_info *fs_info,
4142                               struct find_free_extent_ctl *ffe_ctl,
4143                               struct btrfs_space_info *space_info,
4144                               struct btrfs_key *ins)
4145 {
4146         switch (ffe_ctl->policy) {
4147         case BTRFS_EXTENT_ALLOC_CLUSTERED:
4148                 return prepare_allocation_clustered(fs_info, ffe_ctl,
4149                                                     space_info, ins);
4150         case BTRFS_EXTENT_ALLOC_ZONED:
4151                 if (ffe_ctl->for_treelog) {
4152                         spin_lock(&fs_info->treelog_bg_lock);
4153                         if (fs_info->treelog_bg)
4154                                 ffe_ctl->hint_byte = fs_info->treelog_bg;
4155                         spin_unlock(&fs_info->treelog_bg_lock);
4156                 }
4157                 if (ffe_ctl->for_data_reloc) {
4158                         spin_lock(&fs_info->relocation_bg_lock);
4159                         if (fs_info->data_reloc_bg)
4160                                 ffe_ctl->hint_byte = fs_info->data_reloc_bg;
4161                         spin_unlock(&fs_info->relocation_bg_lock);
4162                 }
4163                 return 0;
4164         default:
4165                 BUG();
4166         }
4167 }
4168
4169 /*
4170  * walks the btree of allocated extents and find a hole of a given size.
4171  * The key ins is changed to record the hole:
4172  * ins->objectid == start position
4173  * ins->flags = BTRFS_EXTENT_ITEM_KEY
4174  * ins->offset == the size of the hole.
4175  * Any available blocks before search_start are skipped.
4176  *
4177  * If there is no suitable free space, we will record the max size of
4178  * the free space extent currently.
4179  *
4180  * The overall logic and call chain:
4181  *
4182  * find_free_extent()
4183  * |- Iterate through all block groups
4184  * |  |- Get a valid block group
4185  * |  |- Try to do clustered allocation in that block group
4186  * |  |- Try to do unclustered allocation in that block group
4187  * |  |- Check if the result is valid
4188  * |  |  |- If valid, then exit
4189  * |  |- Jump to next block group
4190  * |
4191  * |- Push harder to find free extents
4192  *    |- If not found, re-iterate all block groups
4193  */
4194 static noinline int find_free_extent(struct btrfs_root *root,
4195                                      struct btrfs_key *ins,
4196                                      struct find_free_extent_ctl *ffe_ctl)
4197 {
4198         struct btrfs_fs_info *fs_info = root->fs_info;
4199         int ret = 0;
4200         int cache_block_group_error = 0;
4201         struct btrfs_block_group *block_group = NULL;
4202         struct btrfs_space_info *space_info;
4203         bool full_search = false;
4204
4205         WARN_ON(ffe_ctl->num_bytes < fs_info->sectorsize);
4206
4207         ffe_ctl->search_start = 0;
4208         /* For clustered allocation */
4209         ffe_ctl->empty_cluster = 0;
4210         ffe_ctl->last_ptr = NULL;
4211         ffe_ctl->use_cluster = true;
4212         ffe_ctl->have_caching_bg = false;
4213         ffe_ctl->orig_have_caching_bg = false;
4214         ffe_ctl->index = btrfs_bg_flags_to_raid_index(ffe_ctl->flags);
4215         ffe_ctl->loop = 0;
4216         ffe_ctl->retry_uncached = false;
4217         ffe_ctl->cached = 0;
4218         ffe_ctl->max_extent_size = 0;
4219         ffe_ctl->total_free_space = 0;
4220         ffe_ctl->found_offset = 0;
4221         ffe_ctl->policy = BTRFS_EXTENT_ALLOC_CLUSTERED;
4222         ffe_ctl->size_class = btrfs_calc_block_group_size_class(ffe_ctl->num_bytes);
4223
4224         if (btrfs_is_zoned(fs_info))
4225                 ffe_ctl->policy = BTRFS_EXTENT_ALLOC_ZONED;
4226
4227         ins->type = BTRFS_EXTENT_ITEM_KEY;
4228         ins->objectid = 0;
4229         ins->offset = 0;
4230
4231         trace_find_free_extent(root, ffe_ctl);
4232
4233         space_info = btrfs_find_space_info(fs_info, ffe_ctl->flags);
4234         if (!space_info) {
4235                 btrfs_err(fs_info, "No space info for %llu", ffe_ctl->flags);
4236                 return -ENOSPC;
4237         }
4238
4239         ret = prepare_allocation(fs_info, ffe_ctl, space_info, ins);
4240         if (ret < 0)
4241                 return ret;
4242
4243         ffe_ctl->search_start = max(ffe_ctl->search_start,
4244                                     first_logical_byte(fs_info));
4245         ffe_ctl->search_start = max(ffe_ctl->search_start, ffe_ctl->hint_byte);
4246         if (ffe_ctl->search_start == ffe_ctl->hint_byte) {
4247                 block_group = btrfs_lookup_block_group(fs_info,
4248                                                        ffe_ctl->search_start);
4249                 /*
4250                  * we don't want to use the block group if it doesn't match our
4251                  * allocation bits, or if its not cached.
4252                  *
4253                  * However if we are re-searching with an ideal block group
4254                  * picked out then we don't care that the block group is cached.
4255                  */
4256                 if (block_group && block_group_bits(block_group, ffe_ctl->flags) &&
4257                     block_group->cached != BTRFS_CACHE_NO) {
4258                         down_read(&space_info->groups_sem);
4259                         if (list_empty(&block_group->list) ||
4260                             block_group->ro) {
4261                                 /*
4262                                  * someone is removing this block group,
4263                                  * we can't jump into the have_block_group
4264                                  * target because our list pointers are not
4265                                  * valid
4266                                  */
4267                                 btrfs_put_block_group(block_group);
4268                                 up_read(&space_info->groups_sem);
4269                         } else {
4270                                 ffe_ctl->index = btrfs_bg_flags_to_raid_index(
4271                                                         block_group->flags);
4272                                 btrfs_lock_block_group(block_group,
4273                                                        ffe_ctl->delalloc);
4274                                 ffe_ctl->hinted = true;
4275                                 goto have_block_group;
4276                         }
4277                 } else if (block_group) {
4278                         btrfs_put_block_group(block_group);
4279                 }
4280         }
4281 search:
4282         trace_find_free_extent_search_loop(root, ffe_ctl);
4283         ffe_ctl->have_caching_bg = false;
4284         if (ffe_ctl->index == btrfs_bg_flags_to_raid_index(ffe_ctl->flags) ||
4285             ffe_ctl->index == 0)
4286                 full_search = true;
4287         down_read(&space_info->groups_sem);
4288         list_for_each_entry(block_group,
4289                             &space_info->block_groups[ffe_ctl->index], list) {
4290                 struct btrfs_block_group *bg_ret;
4291
4292                 ffe_ctl->hinted = false;
4293                 /* If the block group is read-only, we can skip it entirely. */
4294                 if (unlikely(block_group->ro)) {
4295                         if (ffe_ctl->for_treelog)
4296                                 btrfs_clear_treelog_bg(block_group);
4297                         if (ffe_ctl->for_data_reloc)
4298                                 btrfs_clear_data_reloc_bg(block_group);
4299                         continue;
4300                 }
4301
4302                 btrfs_grab_block_group(block_group, ffe_ctl->delalloc);
4303                 ffe_ctl->search_start = block_group->start;
4304
4305                 /*
4306                  * this can happen if we end up cycling through all the
4307                  * raid types, but we want to make sure we only allocate
4308                  * for the proper type.
4309                  */
4310                 if (!block_group_bits(block_group, ffe_ctl->flags)) {
4311                         u64 extra = BTRFS_BLOCK_GROUP_DUP |
4312                                 BTRFS_BLOCK_GROUP_RAID1_MASK |
4313                                 BTRFS_BLOCK_GROUP_RAID56_MASK |
4314                                 BTRFS_BLOCK_GROUP_RAID10;
4315
4316                         /*
4317                          * if they asked for extra copies and this block group
4318                          * doesn't provide them, bail.  This does allow us to
4319                          * fill raid0 from raid1.
4320                          */
4321                         if ((ffe_ctl->flags & extra) && !(block_group->flags & extra))
4322                                 goto loop;
4323
4324                         /*
4325                          * This block group has different flags than we want.
4326                          * It's possible that we have MIXED_GROUP flag but no
4327                          * block group is mixed.  Just skip such block group.
4328                          */
4329                         btrfs_release_block_group(block_group, ffe_ctl->delalloc);
4330                         continue;
4331                 }
4332
4333 have_block_group:
4334                 trace_find_free_extent_have_block_group(root, ffe_ctl, block_group);
4335                 ffe_ctl->cached = btrfs_block_group_done(block_group);
4336                 if (unlikely(!ffe_ctl->cached)) {
4337                         ffe_ctl->have_caching_bg = true;
4338                         ret = btrfs_cache_block_group(block_group, false);
4339
4340                         /*
4341                          * If we get ENOMEM here or something else we want to
4342                          * try other block groups, because it may not be fatal.
4343                          * However if we can't find anything else we need to
4344                          * save our return here so that we return the actual
4345                          * error that caused problems, not ENOSPC.
4346                          */
4347                         if (ret < 0) {
4348                                 if (!cache_block_group_error)
4349                                         cache_block_group_error = ret;
4350                                 ret = 0;
4351                                 goto loop;
4352                         }
4353                         ret = 0;
4354                 }
4355
4356                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR)) {
4357                         if (!cache_block_group_error)
4358                                 cache_block_group_error = -EIO;
4359                         goto loop;
4360                 }
4361
4362                 if (!find_free_extent_check_size_class(ffe_ctl, block_group))
4363                         goto loop;
4364
4365                 bg_ret = NULL;
4366                 ret = do_allocation(block_group, ffe_ctl, &bg_ret);
4367                 if (ret > 0)
4368                         goto loop;
4369
4370                 if (bg_ret && bg_ret != block_group) {
4371                         btrfs_release_block_group(block_group, ffe_ctl->delalloc);
4372                         block_group = bg_ret;
4373                 }
4374
4375                 /* Checks */
4376                 ffe_ctl->search_start = round_up(ffe_ctl->found_offset,
4377                                                  fs_info->stripesize);
4378
4379                 /* move on to the next group */
4380                 if (ffe_ctl->search_start + ffe_ctl->num_bytes >
4381                     block_group->start + block_group->length) {
4382                         btrfs_add_free_space_unused(block_group,
4383                                             ffe_ctl->found_offset,
4384                                             ffe_ctl->num_bytes);
4385                         goto loop;
4386                 }
4387
4388                 if (ffe_ctl->found_offset < ffe_ctl->search_start)
4389                         btrfs_add_free_space_unused(block_group,
4390                                         ffe_ctl->found_offset,
4391                                         ffe_ctl->search_start - ffe_ctl->found_offset);
4392
4393                 ret = btrfs_add_reserved_bytes(block_group, ffe_ctl->ram_bytes,
4394                                                ffe_ctl->num_bytes,
4395                                                ffe_ctl->delalloc,
4396                                                ffe_ctl->loop >= LOOP_WRONG_SIZE_CLASS);
4397                 if (ret == -EAGAIN) {
4398                         btrfs_add_free_space_unused(block_group,
4399                                         ffe_ctl->found_offset,
4400                                         ffe_ctl->num_bytes);
4401                         goto loop;
4402                 }
4403                 btrfs_inc_block_group_reservations(block_group);
4404
4405                 /* we are all good, lets return */
4406                 ins->objectid = ffe_ctl->search_start;
4407                 ins->offset = ffe_ctl->num_bytes;
4408
4409                 trace_btrfs_reserve_extent(block_group, ffe_ctl);
4410                 btrfs_release_block_group(block_group, ffe_ctl->delalloc);
4411                 break;
4412 loop:
4413                 if (!ffe_ctl->cached && ffe_ctl->loop > LOOP_CACHING_NOWAIT &&
4414                     !ffe_ctl->retry_uncached) {
4415                         ffe_ctl->retry_uncached = true;
4416                         btrfs_wait_block_group_cache_progress(block_group,
4417                                                 ffe_ctl->num_bytes +
4418                                                 ffe_ctl->empty_cluster +
4419                                                 ffe_ctl->empty_size);
4420                         goto have_block_group;
4421                 }
4422                 release_block_group(block_group, ffe_ctl, ffe_ctl->delalloc);
4423                 cond_resched();
4424         }
4425         up_read(&space_info->groups_sem);
4426
4427         ret = find_free_extent_update_loop(fs_info, ins, ffe_ctl, full_search);
4428         if (ret > 0)
4429                 goto search;
4430
4431         if (ret == -ENOSPC && !cache_block_group_error) {
4432                 /*
4433                  * Use ffe_ctl->total_free_space as fallback if we can't find
4434                  * any contiguous hole.
4435                  */
4436                 if (!ffe_ctl->max_extent_size)
4437                         ffe_ctl->max_extent_size = ffe_ctl->total_free_space;
4438                 spin_lock(&space_info->lock);
4439                 space_info->max_extent_size = ffe_ctl->max_extent_size;
4440                 spin_unlock(&space_info->lock);
4441                 ins->offset = ffe_ctl->max_extent_size;
4442         } else if (ret == -ENOSPC) {
4443                 ret = cache_block_group_error;
4444         }
4445         return ret;
4446 }
4447
4448 /*
4449  * btrfs_reserve_extent - entry point to the extent allocator. Tries to find a
4450  *                        hole that is at least as big as @num_bytes.
4451  *
4452  * @root           -    The root that will contain this extent
4453  *
4454  * @ram_bytes      -    The amount of space in ram that @num_bytes take. This
4455  *                      is used for accounting purposes. This value differs
4456  *                      from @num_bytes only in the case of compressed extents.
4457  *
4458  * @num_bytes      -    Number of bytes to allocate on-disk.
4459  *
4460  * @min_alloc_size -    Indicates the minimum amount of space that the
4461  *                      allocator should try to satisfy. In some cases
4462  *                      @num_bytes may be larger than what is required and if
4463  *                      the filesystem is fragmented then allocation fails.
4464  *                      However, the presence of @min_alloc_size gives a
4465  *                      chance to try and satisfy the smaller allocation.
4466  *
4467  * @empty_size     -    A hint that you plan on doing more COW. This is the
4468  *                      size in bytes the allocator should try to find free
4469  *                      next to the block it returns.  This is just a hint and
4470  *                      may be ignored by the allocator.
4471  *
4472  * @hint_byte      -    Hint to the allocator to start searching above the byte
4473  *                      address passed. It might be ignored.
4474  *
4475  * @ins            -    This key is modified to record the found hole. It will
4476  *                      have the following values:
4477  *                      ins->objectid == start position
4478  *                      ins->flags = BTRFS_EXTENT_ITEM_KEY
4479  *                      ins->offset == the size of the hole.
4480  *
4481  * @is_data        -    Boolean flag indicating whether an extent is
4482  *                      allocated for data (true) or metadata (false)
4483  *
4484  * @delalloc       -    Boolean flag indicating whether this allocation is for
4485  *                      delalloc or not. If 'true' data_rwsem of block groups
4486  *                      is going to be acquired.
4487  *
4488  *
4489  * Returns 0 when an allocation succeeded or < 0 when an error occurred. In
4490  * case -ENOSPC is returned then @ins->offset will contain the size of the
4491  * largest available hole the allocator managed to find.
4492  */
4493 int btrfs_reserve_extent(struct btrfs_root *root, u64 ram_bytes,
4494                          u64 num_bytes, u64 min_alloc_size,
4495                          u64 empty_size, u64 hint_byte,
4496                          struct btrfs_key *ins, int is_data, int delalloc)
4497 {
4498         struct btrfs_fs_info *fs_info = root->fs_info;
4499         struct find_free_extent_ctl ffe_ctl = {};
4500         bool final_tried = num_bytes == min_alloc_size;
4501         u64 flags;
4502         int ret;
4503         bool for_treelog = (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
4504         bool for_data_reloc = (btrfs_is_data_reloc_root(root) && is_data);
4505
4506         flags = get_alloc_profile_by_root(root, is_data);
4507 again:
4508         WARN_ON(num_bytes < fs_info->sectorsize);
4509
4510         ffe_ctl.ram_bytes = ram_bytes;
4511         ffe_ctl.num_bytes = num_bytes;
4512         ffe_ctl.min_alloc_size = min_alloc_size;
4513         ffe_ctl.empty_size = empty_size;
4514         ffe_ctl.flags = flags;
4515         ffe_ctl.delalloc = delalloc;
4516         ffe_ctl.hint_byte = hint_byte;
4517         ffe_ctl.for_treelog = for_treelog;
4518         ffe_ctl.for_data_reloc = for_data_reloc;
4519
4520         ret = find_free_extent(root, ins, &ffe_ctl);
4521         if (!ret && !is_data) {
4522                 btrfs_dec_block_group_reservations(fs_info, ins->objectid);
4523         } else if (ret == -ENOSPC) {
4524                 if (!final_tried && ins->offset) {
4525                         num_bytes = min(num_bytes >> 1, ins->offset);
4526                         num_bytes = round_down(num_bytes,
4527                                                fs_info->sectorsize);
4528                         num_bytes = max(num_bytes, min_alloc_size);
4529                         ram_bytes = num_bytes;
4530                         if (num_bytes == min_alloc_size)
4531                                 final_tried = true;
4532                         goto again;
4533                 } else if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
4534                         struct btrfs_space_info *sinfo;
4535
4536                         sinfo = btrfs_find_space_info(fs_info, flags);
4537                         btrfs_err(fs_info,
4538         "allocation failed flags %llu, wanted %llu tree-log %d, relocation: %d",
4539                                   flags, num_bytes, for_treelog, for_data_reloc);
4540                         if (sinfo)
4541                                 btrfs_dump_space_info(fs_info, sinfo,
4542                                                       num_bytes, 1);
4543                 }
4544         }
4545
4546         return ret;
4547 }
4548
4549 int btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
4550                                u64 start, u64 len, int delalloc)
4551 {
4552         struct btrfs_block_group *cache;
4553
4554         cache = btrfs_lookup_block_group(fs_info, start);
4555         if (!cache) {
4556                 btrfs_err(fs_info, "Unable to find block group for %llu",
4557                           start);
4558                 return -ENOSPC;
4559         }
4560
4561         btrfs_add_free_space(cache, start, len);
4562         btrfs_free_reserved_bytes(cache, len, delalloc);
4563         trace_btrfs_reserved_extent_free(fs_info, start, len);
4564
4565         btrfs_put_block_group(cache);
4566         return 0;
4567 }
4568
4569 int btrfs_pin_reserved_extent(struct btrfs_trans_handle *trans, u64 start,
4570                               u64 len)
4571 {
4572         struct btrfs_block_group *cache;
4573         int ret = 0;
4574
4575         cache = btrfs_lookup_block_group(trans->fs_info, start);
4576         if (!cache) {
4577                 btrfs_err(trans->fs_info, "unable to find block group for %llu",
4578                           start);
4579                 return -ENOSPC;
4580         }
4581
4582         ret = pin_down_extent(trans, cache, start, len, 1);
4583         btrfs_put_block_group(cache);
4584         return ret;
4585 }
4586
4587 static int alloc_reserved_extent(struct btrfs_trans_handle *trans, u64 bytenr,
4588                                  u64 num_bytes)
4589 {
4590         struct btrfs_fs_info *fs_info = trans->fs_info;
4591         int ret;
4592
4593         ret = remove_from_free_space_tree(trans, bytenr, num_bytes);
4594         if (ret)
4595                 return ret;
4596
4597         ret = btrfs_update_block_group(trans, bytenr, num_bytes, true);
4598         if (ret) {
4599                 ASSERT(!ret);
4600                 btrfs_err(fs_info, "update block group failed for %llu %llu",
4601                           bytenr, num_bytes);
4602                 return ret;
4603         }
4604
4605         trace_btrfs_reserved_extent_alloc(fs_info, bytenr, num_bytes);
4606         return 0;
4607 }
4608
4609 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
4610                                       u64 parent, u64 root_objectid,
4611                                       u64 flags, u64 owner, u64 offset,
4612                                       struct btrfs_key *ins, int ref_mod)
4613 {
4614         struct btrfs_fs_info *fs_info = trans->fs_info;
4615         struct btrfs_root *extent_root;
4616         int ret;
4617         struct btrfs_extent_item *extent_item;
4618         struct btrfs_extent_inline_ref *iref;
4619         struct btrfs_path *path;
4620         struct extent_buffer *leaf;
4621         int type;
4622         u32 size;
4623
4624         if (parent > 0)
4625                 type = BTRFS_SHARED_DATA_REF_KEY;
4626         else
4627                 type = BTRFS_EXTENT_DATA_REF_KEY;
4628
4629         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
4630
4631         path = btrfs_alloc_path();
4632         if (!path)
4633                 return -ENOMEM;
4634
4635         extent_root = btrfs_extent_root(fs_info, ins->objectid);
4636         ret = btrfs_insert_empty_item(trans, extent_root, path, ins, size);
4637         if (ret) {
4638                 btrfs_free_path(path);
4639                 return ret;
4640         }
4641
4642         leaf = path->nodes[0];
4643         extent_item = btrfs_item_ptr(leaf, path->slots[0],
4644                                      struct btrfs_extent_item);
4645         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
4646         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
4647         btrfs_set_extent_flags(leaf, extent_item,
4648                                flags | BTRFS_EXTENT_FLAG_DATA);
4649
4650         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
4651         btrfs_set_extent_inline_ref_type(leaf, iref, type);
4652         if (parent > 0) {
4653                 struct btrfs_shared_data_ref *ref;
4654                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
4655                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
4656                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
4657         } else {
4658                 struct btrfs_extent_data_ref *ref;
4659                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
4660                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
4661                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
4662                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
4663                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
4664         }
4665
4666         btrfs_mark_buffer_dirty(trans, path->nodes[0]);
4667         btrfs_free_path(path);
4668
4669         return alloc_reserved_extent(trans, ins->objectid, ins->offset);
4670 }
4671
4672 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
4673                                      struct btrfs_delayed_ref_node *node,
4674                                      struct btrfs_delayed_extent_op *extent_op)
4675 {
4676         struct btrfs_fs_info *fs_info = trans->fs_info;
4677         struct btrfs_root *extent_root;
4678         int ret;
4679         struct btrfs_extent_item *extent_item;
4680         struct btrfs_key extent_key;
4681         struct btrfs_tree_block_info *block_info;
4682         struct btrfs_extent_inline_ref *iref;
4683         struct btrfs_path *path;
4684         struct extent_buffer *leaf;
4685         struct btrfs_delayed_tree_ref *ref;
4686         u32 size = sizeof(*extent_item) + sizeof(*iref);
4687         u64 flags = extent_op->flags_to_set;
4688         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
4689
4690         ref = btrfs_delayed_node_to_tree_ref(node);
4691
4692         extent_key.objectid = node->bytenr;
4693         if (skinny_metadata) {
4694                 extent_key.offset = ref->level;
4695                 extent_key.type = BTRFS_METADATA_ITEM_KEY;
4696         } else {
4697                 extent_key.offset = node->num_bytes;
4698                 extent_key.type = BTRFS_EXTENT_ITEM_KEY;
4699                 size += sizeof(*block_info);
4700         }
4701
4702         path = btrfs_alloc_path();
4703         if (!path)
4704                 return -ENOMEM;
4705
4706         extent_root = btrfs_extent_root(fs_info, extent_key.objectid);
4707         ret = btrfs_insert_empty_item(trans, extent_root, path, &extent_key,
4708                                       size);
4709         if (ret) {
4710                 btrfs_free_path(path);
4711                 return ret;
4712         }
4713
4714         leaf = path->nodes[0];
4715         extent_item = btrfs_item_ptr(leaf, path->slots[0],
4716                                      struct btrfs_extent_item);
4717         btrfs_set_extent_refs(leaf, extent_item, 1);
4718         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
4719         btrfs_set_extent_flags(leaf, extent_item,
4720                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
4721
4722         if (skinny_metadata) {
4723                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
4724         } else {
4725                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
4726                 btrfs_set_tree_block_key(leaf, block_info, &extent_op->key);
4727                 btrfs_set_tree_block_level(leaf, block_info, ref->level);
4728                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
4729         }
4730
4731         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY) {
4732                 btrfs_set_extent_inline_ref_type(leaf, iref,
4733                                                  BTRFS_SHARED_BLOCK_REF_KEY);
4734                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->parent);
4735         } else {
4736                 btrfs_set_extent_inline_ref_type(leaf, iref,
4737                                                  BTRFS_TREE_BLOCK_REF_KEY);
4738                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->root);
4739         }
4740
4741         btrfs_mark_buffer_dirty(trans, leaf);
4742         btrfs_free_path(path);
4743
4744         return alloc_reserved_extent(trans, node->bytenr, fs_info->nodesize);
4745 }
4746
4747 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
4748                                      struct btrfs_root *root, u64 owner,
4749                                      u64 offset, u64 ram_bytes,
4750                                      struct btrfs_key *ins)
4751 {
4752         struct btrfs_ref generic_ref = { 0 };
4753
4754         BUG_ON(root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
4755
4756         btrfs_init_generic_ref(&generic_ref, BTRFS_ADD_DELAYED_EXTENT,
4757                                ins->objectid, ins->offset, 0);
4758         btrfs_init_data_ref(&generic_ref, root->root_key.objectid, owner,
4759                             offset, 0, false);
4760         btrfs_ref_tree_mod(root->fs_info, &generic_ref);
4761
4762         return btrfs_add_delayed_data_ref(trans, &generic_ref, ram_bytes);
4763 }
4764
4765 /*
4766  * this is used by the tree logging recovery code.  It records that
4767  * an extent has been allocated and makes sure to clear the free
4768  * space cache bits as well
4769  */
4770 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
4771                                    u64 root_objectid, u64 owner, u64 offset,
4772                                    struct btrfs_key *ins)
4773 {
4774         struct btrfs_fs_info *fs_info = trans->fs_info;
4775         int ret;
4776         struct btrfs_block_group *block_group;
4777         struct btrfs_space_info *space_info;
4778
4779         /*
4780          * Mixed block groups will exclude before processing the log so we only
4781          * need to do the exclude dance if this fs isn't mixed.
4782          */
4783         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS)) {
4784                 ret = __exclude_logged_extent(fs_info, ins->objectid,
4785                                               ins->offset);
4786                 if (ret)
4787                         return ret;
4788         }
4789
4790         block_group = btrfs_lookup_block_group(fs_info, ins->objectid);
4791         if (!block_group)
4792                 return -EINVAL;
4793
4794         space_info = block_group->space_info;
4795         spin_lock(&space_info->lock);
4796         spin_lock(&block_group->lock);
4797         space_info->bytes_reserved += ins->offset;
4798         block_group->reserved += ins->offset;
4799         spin_unlock(&block_group->lock);
4800         spin_unlock(&space_info->lock);
4801
4802         ret = alloc_reserved_file_extent(trans, 0, root_objectid, 0, owner,
4803                                          offset, ins, 1);
4804         if (ret)
4805                 btrfs_pin_extent(trans, ins->objectid, ins->offset, 1);
4806         btrfs_put_block_group(block_group);
4807         return ret;
4808 }
4809
4810 static struct extent_buffer *
4811 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
4812                       u64 bytenr, int level, u64 owner,
4813                       enum btrfs_lock_nesting nest)
4814 {
4815         struct btrfs_fs_info *fs_info = root->fs_info;
4816         struct extent_buffer *buf;
4817         u64 lockdep_owner = owner;
4818
4819         buf = btrfs_find_create_tree_block(fs_info, bytenr, owner, level);
4820         if (IS_ERR(buf))
4821                 return buf;
4822
4823         /*
4824          * Extra safety check in case the extent tree is corrupted and extent
4825          * allocator chooses to use a tree block which is already used and
4826          * locked.
4827          */
4828         if (buf->lock_owner == current->pid) {
4829                 btrfs_err_rl(fs_info,
4830 "tree block %llu owner %llu already locked by pid=%d, extent tree corruption detected",
4831                         buf->start, btrfs_header_owner(buf), current->pid);
4832                 free_extent_buffer(buf);
4833                 return ERR_PTR(-EUCLEAN);
4834         }
4835
4836         /*
4837          * The reloc trees are just snapshots, so we need them to appear to be
4838          * just like any other fs tree WRT lockdep.
4839          *
4840          * The exception however is in replace_path() in relocation, where we
4841          * hold the lock on the original fs root and then search for the reloc
4842          * root.  At that point we need to make sure any reloc root buffers are
4843          * set to the BTRFS_TREE_RELOC_OBJECTID lockdep class in order to make
4844          * lockdep happy.
4845          */
4846         if (lockdep_owner == BTRFS_TREE_RELOC_OBJECTID &&
4847             !test_bit(BTRFS_ROOT_RESET_LOCKDEP_CLASS, &root->state))
4848                 lockdep_owner = BTRFS_FS_TREE_OBJECTID;
4849
4850         /* btrfs_clear_buffer_dirty() accesses generation field. */
4851         btrfs_set_header_generation(buf, trans->transid);
4852
4853         /*
4854          * This needs to stay, because we could allocate a freed block from an
4855          * old tree into a new tree, so we need to make sure this new block is
4856          * set to the appropriate level and owner.
4857          */
4858         btrfs_set_buffer_lockdep_class(lockdep_owner, buf, level);
4859
4860         __btrfs_tree_lock(buf, nest);
4861         btrfs_clear_buffer_dirty(trans, buf);
4862         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
4863         clear_bit(EXTENT_BUFFER_NO_CHECK, &buf->bflags);
4864
4865         set_extent_buffer_uptodate(buf);
4866
4867         memzero_extent_buffer(buf, 0, sizeof(struct btrfs_header));
4868         btrfs_set_header_level(buf, level);
4869         btrfs_set_header_bytenr(buf, buf->start);
4870         btrfs_set_header_generation(buf, trans->transid);
4871         btrfs_set_header_backref_rev(buf, BTRFS_MIXED_BACKREF_REV);
4872         btrfs_set_header_owner(buf, owner);
4873         write_extent_buffer_fsid(buf, fs_info->fs_devices->metadata_uuid);
4874         write_extent_buffer_chunk_tree_uuid(buf, fs_info->chunk_tree_uuid);
4875         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
4876                 buf->log_index = root->log_transid % 2;
4877                 /*
4878                  * we allow two log transactions at a time, use different
4879                  * EXTENT bit to differentiate dirty pages.
4880                  */
4881                 if (buf->log_index == 0)
4882                         set_extent_bit(&root->dirty_log_pages, buf->start,
4883                                        buf->start + buf->len - 1,
4884                                        EXTENT_DIRTY, NULL);
4885                 else
4886                         set_extent_bit(&root->dirty_log_pages, buf->start,
4887                                        buf->start + buf->len - 1,
4888                                        EXTENT_NEW, NULL);
4889         } else {
4890                 buf->log_index = -1;
4891                 set_extent_bit(&trans->transaction->dirty_pages, buf->start,
4892                                buf->start + buf->len - 1, EXTENT_DIRTY, NULL);
4893         }
4894         /* this returns a buffer locked for blocking */
4895         return buf;
4896 }
4897
4898 /*
4899  * finds a free extent and does all the dirty work required for allocation
4900  * returns the tree buffer or an ERR_PTR on error.
4901  */
4902 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
4903                                              struct btrfs_root *root,
4904                                              u64 parent, u64 root_objectid,
4905                                              const struct btrfs_disk_key *key,
4906                                              int level, u64 hint,
4907                                              u64 empty_size,
4908                                              enum btrfs_lock_nesting nest)
4909 {
4910         struct btrfs_fs_info *fs_info = root->fs_info;
4911         struct btrfs_key ins;
4912         struct btrfs_block_rsv *block_rsv;
4913         struct extent_buffer *buf;
4914         struct btrfs_delayed_extent_op *extent_op;
4915         struct btrfs_ref generic_ref = { 0 };
4916         u64 flags = 0;
4917         int ret;
4918         u32 blocksize = fs_info->nodesize;
4919         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
4920
4921 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4922         if (btrfs_is_testing(fs_info)) {
4923                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
4924                                             level, root_objectid, nest);
4925                 if (!IS_ERR(buf))
4926                         root->alloc_bytenr += blocksize;
4927                 return buf;
4928         }
4929 #endif
4930
4931         block_rsv = btrfs_use_block_rsv(trans, root, blocksize);
4932         if (IS_ERR(block_rsv))
4933                 return ERR_CAST(block_rsv);
4934
4935         ret = btrfs_reserve_extent(root, blocksize, blocksize, blocksize,
4936                                    empty_size, hint, &ins, 0, 0);
4937         if (ret)
4938                 goto out_unuse;
4939
4940         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level,
4941                                     root_objectid, nest);
4942         if (IS_ERR(buf)) {
4943                 ret = PTR_ERR(buf);
4944                 goto out_free_reserved;
4945         }
4946
4947         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
4948                 if (parent == 0)
4949                         parent = ins.objectid;
4950                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
4951         } else
4952                 BUG_ON(parent > 0);
4953
4954         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
4955                 extent_op = btrfs_alloc_delayed_extent_op();
4956                 if (!extent_op) {
4957                         ret = -ENOMEM;
4958                         goto out_free_buf;
4959                 }
4960                 if (key)
4961                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
4962                 else
4963                         memset(&extent_op->key, 0, sizeof(extent_op->key));
4964                 extent_op->flags_to_set = flags;
4965                 extent_op->update_key = skinny_metadata ? false : true;
4966                 extent_op->update_flags = true;
4967                 extent_op->level = level;
4968
4969                 btrfs_init_generic_ref(&generic_ref, BTRFS_ADD_DELAYED_EXTENT,
4970                                        ins.objectid, ins.offset, parent);
4971                 btrfs_init_tree_ref(&generic_ref, level, root_objectid,
4972                                     root->root_key.objectid, false);
4973                 btrfs_ref_tree_mod(fs_info, &generic_ref);
4974                 ret = btrfs_add_delayed_tree_ref(trans, &generic_ref, extent_op);
4975                 if (ret)
4976                         goto out_free_delayed;
4977         }
4978         return buf;
4979
4980 out_free_delayed:
4981         btrfs_free_delayed_extent_op(extent_op);
4982 out_free_buf:
4983         btrfs_tree_unlock(buf);
4984         free_extent_buffer(buf);
4985 out_free_reserved:
4986         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 0);
4987 out_unuse:
4988         btrfs_unuse_block_rsv(fs_info, block_rsv, blocksize);
4989         return ERR_PTR(ret);
4990 }
4991
4992 struct walk_control {
4993         u64 refs[BTRFS_MAX_LEVEL];
4994         u64 flags[BTRFS_MAX_LEVEL];
4995         struct btrfs_key update_progress;
4996         struct btrfs_key drop_progress;
4997         int drop_level;
4998         int stage;
4999         int level;
5000         int shared_level;
5001         int update_ref;
5002         int keep_locks;
5003         int reada_slot;
5004         int reada_count;
5005         int restarted;
5006 };
5007
5008 #define DROP_REFERENCE  1
5009 #define UPDATE_BACKREF  2
5010
5011 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
5012                                      struct btrfs_root *root,
5013                                      struct walk_control *wc,
5014                                      struct btrfs_path *path)
5015 {
5016         struct btrfs_fs_info *fs_info = root->fs_info;
5017         u64 bytenr;
5018         u64 generation;
5019         u64 refs;
5020         u64 flags;
5021         u32 nritems;
5022         struct btrfs_key key;
5023         struct extent_buffer *eb;
5024         int ret;
5025         int slot;
5026         int nread = 0;
5027
5028         if (path->slots[wc->level] < wc->reada_slot) {
5029                 wc->reada_count = wc->reada_count * 2 / 3;
5030                 wc->reada_count = max(wc->reada_count, 2);
5031         } else {
5032                 wc->reada_count = wc->reada_count * 3 / 2;
5033                 wc->reada_count = min_t(int, wc->reada_count,
5034                                         BTRFS_NODEPTRS_PER_BLOCK(fs_info));
5035         }
5036
5037         eb = path->nodes[wc->level];
5038         nritems = btrfs_header_nritems(eb);
5039
5040         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
5041                 if (nread >= wc->reada_count)
5042                         break;
5043
5044                 cond_resched();
5045                 bytenr = btrfs_node_blockptr(eb, slot);
5046                 generation = btrfs_node_ptr_generation(eb, slot);
5047
5048                 if (slot == path->slots[wc->level])
5049                         goto reada;
5050
5051                 if (wc->stage == UPDATE_BACKREF &&
5052                     generation <= root->root_key.offset)
5053                         continue;
5054
5055                 /* We don't lock the tree block, it's OK to be racy here */
5056                 ret = btrfs_lookup_extent_info(trans, fs_info, bytenr,
5057                                                wc->level - 1, 1, &refs,
5058                                                &flags);
5059                 /* We don't care about errors in readahead. */
5060                 if (ret < 0)
5061                         continue;
5062                 BUG_ON(refs == 0);
5063
5064                 if (wc->stage == DROP_REFERENCE) {
5065                         if (refs == 1)
5066                                 goto reada;
5067
5068                         if (wc->level == 1 &&
5069                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5070                                 continue;
5071                         if (!wc->update_ref ||
5072                             generation <= root->root_key.offset)
5073                                 continue;
5074                         btrfs_node_key_to_cpu(eb, &key, slot);
5075                         ret = btrfs_comp_cpu_keys(&key,
5076                                                   &wc->update_progress);
5077                         if (ret < 0)
5078                                 continue;
5079                 } else {
5080                         if (wc->level == 1 &&
5081                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5082                                 continue;
5083                 }
5084 reada:
5085                 btrfs_readahead_node_child(eb, slot);
5086                 nread++;
5087         }
5088         wc->reada_slot = slot;
5089 }
5090
5091 /*
5092  * helper to process tree block while walking down the tree.
5093  *
5094  * when wc->stage == UPDATE_BACKREF, this function updates
5095  * back refs for pointers in the block.
5096  *
5097  * NOTE: return value 1 means we should stop walking down.
5098  */
5099 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
5100                                    struct btrfs_root *root,
5101                                    struct btrfs_path *path,
5102                                    struct walk_control *wc, int lookup_info)
5103 {
5104         struct btrfs_fs_info *fs_info = root->fs_info;
5105         int level = wc->level;
5106         struct extent_buffer *eb = path->nodes[level];
5107         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
5108         int ret;
5109
5110         if (wc->stage == UPDATE_BACKREF &&
5111             btrfs_header_owner(eb) != root->root_key.objectid)
5112                 return 1;
5113
5114         /*
5115          * when reference count of tree block is 1, it won't increase
5116          * again. once full backref flag is set, we never clear it.
5117          */
5118         if (lookup_info &&
5119             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
5120              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
5121                 BUG_ON(!path->locks[level]);
5122                 ret = btrfs_lookup_extent_info(trans, fs_info,
5123                                                eb->start, level, 1,
5124                                                &wc->refs[level],
5125                                                &wc->flags[level]);
5126                 BUG_ON(ret == -ENOMEM);
5127                 if (ret)
5128                         return ret;
5129                 BUG_ON(wc->refs[level] == 0);
5130         }
5131
5132         if (wc->stage == DROP_REFERENCE) {
5133                 if (wc->refs[level] > 1)
5134                         return 1;
5135
5136                 if (path->locks[level] && !wc->keep_locks) {
5137                         btrfs_tree_unlock_rw(eb, path->locks[level]);
5138                         path->locks[level] = 0;
5139                 }
5140                 return 0;
5141         }
5142
5143         /* wc->stage == UPDATE_BACKREF */
5144         if (!(wc->flags[level] & flag)) {
5145                 BUG_ON(!path->locks[level]);
5146                 ret = btrfs_inc_ref(trans, root, eb, 1);
5147                 BUG_ON(ret); /* -ENOMEM */
5148                 ret = btrfs_dec_ref(trans, root, eb, 0);
5149                 BUG_ON(ret); /* -ENOMEM */
5150                 ret = btrfs_set_disk_extent_flags(trans, eb, flag);
5151                 BUG_ON(ret); /* -ENOMEM */
5152                 wc->flags[level] |= flag;
5153         }
5154
5155         /*
5156          * the block is shared by multiple trees, so it's not good to
5157          * keep the tree lock
5158          */
5159         if (path->locks[level] && level > 0) {
5160                 btrfs_tree_unlock_rw(eb, path->locks[level]);
5161                 path->locks[level] = 0;
5162         }
5163         return 0;
5164 }
5165
5166 /*
5167  * This is used to verify a ref exists for this root to deal with a bug where we
5168  * would have a drop_progress key that hadn't been updated properly.
5169  */
5170 static int check_ref_exists(struct btrfs_trans_handle *trans,
5171                             struct btrfs_root *root, u64 bytenr, u64 parent,
5172                             int level)
5173 {
5174         struct btrfs_path *path;
5175         struct btrfs_extent_inline_ref *iref;
5176         int ret;
5177
5178         path = btrfs_alloc_path();
5179         if (!path)
5180                 return -ENOMEM;
5181
5182         ret = lookup_extent_backref(trans, path, &iref, bytenr,
5183                                     root->fs_info->nodesize, parent,
5184                                     root->root_key.objectid, level, 0);
5185         btrfs_free_path(path);
5186         if (ret == -ENOENT)
5187                 return 0;
5188         if (ret < 0)
5189                 return ret;
5190         return 1;
5191 }
5192
5193 /*
5194  * helper to process tree block pointer.
5195  *
5196  * when wc->stage == DROP_REFERENCE, this function checks
5197  * reference count of the block pointed to. if the block
5198  * is shared and we need update back refs for the subtree
5199  * rooted at the block, this function changes wc->stage to
5200  * UPDATE_BACKREF. if the block is shared and there is no
5201  * need to update back, this function drops the reference
5202  * to the block.
5203  *
5204  * NOTE: return value 1 means we should stop walking down.
5205  */
5206 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
5207                                  struct btrfs_root *root,
5208                                  struct btrfs_path *path,
5209                                  struct walk_control *wc, int *lookup_info)
5210 {
5211         struct btrfs_fs_info *fs_info = root->fs_info;
5212         u64 bytenr;
5213         u64 generation;
5214         u64 parent;
5215         struct btrfs_tree_parent_check check = { 0 };
5216         struct btrfs_key key;
5217         struct btrfs_ref ref = { 0 };
5218         struct extent_buffer *next;
5219         int level = wc->level;
5220         int reada = 0;
5221         int ret = 0;
5222         bool need_account = false;
5223
5224         generation = btrfs_node_ptr_generation(path->nodes[level],
5225                                                path->slots[level]);
5226         /*
5227          * if the lower level block was created before the snapshot
5228          * was created, we know there is no need to update back refs
5229          * for the subtree
5230          */
5231         if (wc->stage == UPDATE_BACKREF &&
5232             generation <= root->root_key.offset) {
5233                 *lookup_info = 1;
5234                 return 1;
5235         }
5236
5237         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
5238
5239         check.level = level - 1;
5240         check.transid = generation;
5241         check.owner_root = root->root_key.objectid;
5242         check.has_first_key = true;
5243         btrfs_node_key_to_cpu(path->nodes[level], &check.first_key,
5244                               path->slots[level]);
5245
5246         next = find_extent_buffer(fs_info, bytenr);
5247         if (!next) {
5248                 next = btrfs_find_create_tree_block(fs_info, bytenr,
5249                                 root->root_key.objectid, level - 1);
5250                 if (IS_ERR(next))
5251                         return PTR_ERR(next);
5252                 reada = 1;
5253         }
5254         btrfs_tree_lock(next);
5255
5256         ret = btrfs_lookup_extent_info(trans, fs_info, bytenr, level - 1, 1,
5257                                        &wc->refs[level - 1],
5258                                        &wc->flags[level - 1]);
5259         if (ret < 0)
5260                 goto out_unlock;
5261
5262         if (unlikely(wc->refs[level - 1] == 0)) {
5263                 btrfs_err(fs_info, "Missing references.");
5264                 ret = -EIO;
5265                 goto out_unlock;
5266         }
5267         *lookup_info = 0;
5268
5269         if (wc->stage == DROP_REFERENCE) {
5270                 if (wc->refs[level - 1] > 1) {
5271                         need_account = true;
5272                         if (level == 1 &&
5273                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5274                                 goto skip;
5275
5276                         if (!wc->update_ref ||
5277                             generation <= root->root_key.offset)
5278                                 goto skip;
5279
5280                         btrfs_node_key_to_cpu(path->nodes[level], &key,
5281                                               path->slots[level]);
5282                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
5283                         if (ret < 0)
5284                                 goto skip;
5285
5286                         wc->stage = UPDATE_BACKREF;
5287                         wc->shared_level = level - 1;
5288                 }
5289         } else {
5290                 if (level == 1 &&
5291                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5292                         goto skip;
5293         }
5294
5295         if (!btrfs_buffer_uptodate(next, generation, 0)) {
5296                 btrfs_tree_unlock(next);
5297                 free_extent_buffer(next);
5298                 next = NULL;
5299                 *lookup_info = 1;
5300         }
5301
5302         if (!next) {
5303                 if (reada && level == 1)
5304                         reada_walk_down(trans, root, wc, path);
5305                 next = read_tree_block(fs_info, bytenr, &check);
5306                 if (IS_ERR(next)) {
5307                         return PTR_ERR(next);
5308                 } else if (!extent_buffer_uptodate(next)) {
5309                         free_extent_buffer(next);
5310                         return -EIO;
5311                 }
5312                 btrfs_tree_lock(next);
5313         }
5314
5315         level--;
5316         ASSERT(level == btrfs_header_level(next));
5317         if (level != btrfs_header_level(next)) {
5318                 btrfs_err(root->fs_info, "mismatched level");
5319                 ret = -EIO;
5320                 goto out_unlock;
5321         }
5322         path->nodes[level] = next;
5323         path->slots[level] = 0;
5324         path->locks[level] = BTRFS_WRITE_LOCK;
5325         wc->level = level;
5326         if (wc->level == 1)
5327                 wc->reada_slot = 0;
5328         return 0;
5329 skip:
5330         wc->refs[level - 1] = 0;
5331         wc->flags[level - 1] = 0;
5332         if (wc->stage == DROP_REFERENCE) {
5333                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
5334                         parent = path->nodes[level]->start;
5335                 } else {
5336                         ASSERT(root->root_key.objectid ==
5337                                btrfs_header_owner(path->nodes[level]));
5338                         if (root->root_key.objectid !=
5339                             btrfs_header_owner(path->nodes[level])) {
5340                                 btrfs_err(root->fs_info,
5341                                                 "mismatched block owner");
5342                                 ret = -EIO;
5343                                 goto out_unlock;
5344                         }
5345                         parent = 0;
5346                 }
5347
5348                 /*
5349                  * If we had a drop_progress we need to verify the refs are set
5350                  * as expected.  If we find our ref then we know that from here
5351                  * on out everything should be correct, and we can clear the
5352                  * ->restarted flag.
5353                  */
5354                 if (wc->restarted) {
5355                         ret = check_ref_exists(trans, root, bytenr, parent,
5356                                                level - 1);
5357                         if (ret < 0)
5358                                 goto out_unlock;
5359                         if (ret == 0)
5360                                 goto no_delete;
5361                         ret = 0;
5362                         wc->restarted = 0;
5363                 }
5364
5365                 /*
5366                  * Reloc tree doesn't contribute to qgroup numbers, and we have
5367                  * already accounted them at merge time (replace_path),
5368                  * thus we could skip expensive subtree trace here.
5369                  */
5370                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
5371                     need_account) {
5372                         ret = btrfs_qgroup_trace_subtree(trans, next,
5373                                                          generation, level - 1);
5374                         if (ret) {
5375                                 btrfs_err_rl(fs_info,
5376                                              "Error %d accounting shared subtree. Quota is out of sync, rescan required.",
5377                                              ret);
5378                         }
5379                 }
5380
5381                 /*
5382                  * We need to update the next key in our walk control so we can
5383                  * update the drop_progress key accordingly.  We don't care if
5384                  * find_next_key doesn't find a key because that means we're at
5385                  * the end and are going to clean up now.
5386                  */
5387                 wc->drop_level = level;
5388                 find_next_key(path, level, &wc->drop_progress);
5389
5390                 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
5391                                        fs_info->nodesize, parent);
5392                 btrfs_init_tree_ref(&ref, level - 1, root->root_key.objectid,
5393                                     0, false);
5394                 ret = btrfs_free_extent(trans, &ref);
5395                 if (ret)
5396                         goto out_unlock;
5397         }
5398 no_delete:
5399         *lookup_info = 1;
5400         ret = 1;
5401
5402 out_unlock:
5403         btrfs_tree_unlock(next);
5404         free_extent_buffer(next);
5405
5406         return ret;
5407 }
5408
5409 /*
5410  * helper to process tree block while walking up the tree.
5411  *
5412  * when wc->stage == DROP_REFERENCE, this function drops
5413  * reference count on the block.
5414  *
5415  * when wc->stage == UPDATE_BACKREF, this function changes
5416  * wc->stage back to DROP_REFERENCE if we changed wc->stage
5417  * to UPDATE_BACKREF previously while processing the block.
5418  *
5419  * NOTE: return value 1 means we should stop walking up.
5420  */
5421 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
5422                                  struct btrfs_root *root,
5423                                  struct btrfs_path *path,
5424                                  struct walk_control *wc)
5425 {
5426         struct btrfs_fs_info *fs_info = root->fs_info;
5427         int ret;
5428         int level = wc->level;
5429         struct extent_buffer *eb = path->nodes[level];
5430         u64 parent = 0;
5431
5432         if (wc->stage == UPDATE_BACKREF) {
5433                 BUG_ON(wc->shared_level < level);
5434                 if (level < wc->shared_level)
5435                         goto out;
5436
5437                 ret = find_next_key(path, level + 1, &wc->update_progress);
5438                 if (ret > 0)
5439                         wc->update_ref = 0;
5440
5441                 wc->stage = DROP_REFERENCE;
5442                 wc->shared_level = -1;
5443                 path->slots[level] = 0;
5444
5445                 /*
5446                  * check reference count again if the block isn't locked.
5447                  * we should start walking down the tree again if reference
5448                  * count is one.
5449                  */
5450                 if (!path->locks[level]) {
5451                         BUG_ON(level == 0);
5452                         btrfs_tree_lock(eb);
5453                         path->locks[level] = BTRFS_WRITE_LOCK;
5454
5455                         ret = btrfs_lookup_extent_info(trans, fs_info,
5456                                                        eb->start, level, 1,
5457                                                        &wc->refs[level],
5458                                                        &wc->flags[level]);
5459                         if (ret < 0) {
5460                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
5461                                 path->locks[level] = 0;
5462                                 return ret;
5463                         }
5464                         BUG_ON(wc->refs[level] == 0);
5465                         if (wc->refs[level] == 1) {
5466                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
5467                                 path->locks[level] = 0;
5468                                 return 1;
5469                         }
5470                 }
5471         }
5472
5473         /* wc->stage == DROP_REFERENCE */
5474         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
5475
5476         if (wc->refs[level] == 1) {
5477                 if (level == 0) {
5478                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
5479                                 ret = btrfs_dec_ref(trans, root, eb, 1);
5480                         else
5481                                 ret = btrfs_dec_ref(trans, root, eb, 0);
5482                         BUG_ON(ret); /* -ENOMEM */
5483                         if (is_fstree(root->root_key.objectid)) {
5484                                 ret = btrfs_qgroup_trace_leaf_items(trans, eb);
5485                                 if (ret) {
5486                                         btrfs_err_rl(fs_info,
5487         "error %d accounting leaf items, quota is out of sync, rescan required",
5488                                              ret);
5489                                 }
5490                         }
5491                 }
5492                 /* Make block locked assertion in btrfs_clear_buffer_dirty happy. */
5493                 if (!path->locks[level]) {
5494                         btrfs_tree_lock(eb);
5495                         path->locks[level] = BTRFS_WRITE_LOCK;
5496                 }
5497                 btrfs_clear_buffer_dirty(trans, eb);
5498         }
5499
5500         if (eb == root->node) {
5501                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
5502                         parent = eb->start;
5503                 else if (root->root_key.objectid != btrfs_header_owner(eb))
5504                         goto owner_mismatch;
5505         } else {
5506                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
5507                         parent = path->nodes[level + 1]->start;
5508                 else if (root->root_key.objectid !=
5509                          btrfs_header_owner(path->nodes[level + 1]))
5510                         goto owner_mismatch;
5511         }
5512
5513         btrfs_free_tree_block(trans, btrfs_root_id(root), eb, parent,
5514                               wc->refs[level] == 1);
5515 out:
5516         wc->refs[level] = 0;
5517         wc->flags[level] = 0;
5518         return 0;
5519
5520 owner_mismatch:
5521         btrfs_err_rl(fs_info, "unexpected tree owner, have %llu expect %llu",
5522                      btrfs_header_owner(eb), root->root_key.objectid);
5523         return -EUCLEAN;
5524 }
5525
5526 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
5527                                    struct btrfs_root *root,
5528                                    struct btrfs_path *path,
5529                                    struct walk_control *wc)
5530 {
5531         int level = wc->level;
5532         int lookup_info = 1;
5533         int ret = 0;
5534
5535         while (level >= 0) {
5536                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
5537                 if (ret)
5538                         break;
5539
5540                 if (level == 0)
5541                         break;
5542
5543                 if (path->slots[level] >=
5544                     btrfs_header_nritems(path->nodes[level]))
5545                         break;
5546
5547                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
5548                 if (ret > 0) {
5549                         path->slots[level]++;
5550                         continue;
5551                 } else if (ret < 0)
5552                         break;
5553                 level = wc->level;
5554         }
5555         return (ret == 1) ? 0 : ret;
5556 }
5557
5558 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
5559                                  struct btrfs_root *root,
5560                                  struct btrfs_path *path,
5561                                  struct walk_control *wc, int max_level)
5562 {
5563         int level = wc->level;
5564         int ret;
5565
5566         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
5567         while (level < max_level && path->nodes[level]) {
5568                 wc->level = level;
5569                 if (path->slots[level] + 1 <
5570                     btrfs_header_nritems(path->nodes[level])) {
5571                         path->slots[level]++;
5572                         return 0;
5573                 } else {
5574                         ret = walk_up_proc(trans, root, path, wc);
5575                         if (ret > 0)
5576                                 return 0;
5577                         if (ret < 0)
5578                                 return ret;
5579
5580                         if (path->locks[level]) {
5581                                 btrfs_tree_unlock_rw(path->nodes[level],
5582                                                      path->locks[level]);
5583                                 path->locks[level] = 0;
5584                         }
5585                         free_extent_buffer(path->nodes[level]);
5586                         path->nodes[level] = NULL;
5587                         level++;
5588                 }
5589         }
5590         return 1;
5591 }
5592
5593 /*
5594  * drop a subvolume tree.
5595  *
5596  * this function traverses the tree freeing any blocks that only
5597  * referenced by the tree.
5598  *
5599  * when a shared tree block is found. this function decreases its
5600  * reference count by one. if update_ref is true, this function
5601  * also make sure backrefs for the shared block and all lower level
5602  * blocks are properly updated.
5603  *
5604  * If called with for_reloc == 0, may exit early with -EAGAIN
5605  */
5606 int btrfs_drop_snapshot(struct btrfs_root *root, int update_ref, int for_reloc)
5607 {
5608         const bool is_reloc_root = (root->root_key.objectid ==
5609                                     BTRFS_TREE_RELOC_OBJECTID);
5610         struct btrfs_fs_info *fs_info = root->fs_info;
5611         struct btrfs_path *path;
5612         struct btrfs_trans_handle *trans;
5613         struct btrfs_root *tree_root = fs_info->tree_root;
5614         struct btrfs_root_item *root_item = &root->root_item;
5615         struct walk_control *wc;
5616         struct btrfs_key key;
5617         int err = 0;
5618         int ret;
5619         int level;
5620         bool root_dropped = false;
5621         bool unfinished_drop = false;
5622
5623         btrfs_debug(fs_info, "Drop subvolume %llu", root->root_key.objectid);
5624
5625         path = btrfs_alloc_path();
5626         if (!path) {
5627                 err = -ENOMEM;
5628                 goto out;
5629         }
5630
5631         wc = kzalloc(sizeof(*wc), GFP_NOFS);
5632         if (!wc) {
5633                 btrfs_free_path(path);
5634                 err = -ENOMEM;
5635                 goto out;
5636         }
5637
5638         /*
5639          * Use join to avoid potential EINTR from transaction start. See
5640          * wait_reserve_ticket and the whole reservation callchain.
5641          */
5642         if (for_reloc)
5643                 trans = btrfs_join_transaction(tree_root);
5644         else
5645                 trans = btrfs_start_transaction(tree_root, 0);
5646         if (IS_ERR(trans)) {
5647                 err = PTR_ERR(trans);
5648                 goto out_free;
5649         }
5650
5651         err = btrfs_run_delayed_items(trans);
5652         if (err)
5653                 goto out_end_trans;
5654
5655         /*
5656          * This will help us catch people modifying the fs tree while we're
5657          * dropping it.  It is unsafe to mess with the fs tree while it's being
5658          * dropped as we unlock the root node and parent nodes as we walk down
5659          * the tree, assuming nothing will change.  If something does change
5660          * then we'll have stale information and drop references to blocks we've
5661          * already dropped.
5662          */
5663         set_bit(BTRFS_ROOT_DELETING, &root->state);
5664         unfinished_drop = test_bit(BTRFS_ROOT_UNFINISHED_DROP, &root->state);
5665
5666         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
5667                 level = btrfs_header_level(root->node);
5668                 path->nodes[level] = btrfs_lock_root_node(root);
5669                 path->slots[level] = 0;
5670                 path->locks[level] = BTRFS_WRITE_LOCK;
5671                 memset(&wc->update_progress, 0,
5672                        sizeof(wc->update_progress));
5673         } else {
5674                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
5675                 memcpy(&wc->update_progress, &key,
5676                        sizeof(wc->update_progress));
5677
5678                 level = btrfs_root_drop_level(root_item);
5679                 BUG_ON(level == 0);
5680                 path->lowest_level = level;
5681                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5682                 path->lowest_level = 0;
5683                 if (ret < 0) {
5684                         err = ret;
5685                         goto out_end_trans;
5686                 }
5687                 WARN_ON(ret > 0);
5688
5689                 /*
5690                  * unlock our path, this is safe because only this
5691                  * function is allowed to delete this snapshot
5692                  */
5693                 btrfs_unlock_up_safe(path, 0);
5694
5695                 level = btrfs_header_level(root->node);
5696                 while (1) {
5697                         btrfs_tree_lock(path->nodes[level]);
5698                         path->locks[level] = BTRFS_WRITE_LOCK;
5699
5700                         ret = btrfs_lookup_extent_info(trans, fs_info,
5701                                                 path->nodes[level]->start,
5702                                                 level, 1, &wc->refs[level],
5703                                                 &wc->flags[level]);
5704                         if (ret < 0) {
5705                                 err = ret;
5706                                 goto out_end_trans;
5707                         }
5708                         BUG_ON(wc->refs[level] == 0);
5709
5710                         if (level == btrfs_root_drop_level(root_item))
5711                                 break;
5712
5713                         btrfs_tree_unlock(path->nodes[level]);
5714                         path->locks[level] = 0;
5715                         WARN_ON(wc->refs[level] != 1);
5716                         level--;
5717                 }
5718         }
5719
5720         wc->restarted = test_bit(BTRFS_ROOT_DEAD_TREE, &root->state);
5721         wc->level = level;
5722         wc->shared_level = -1;
5723         wc->stage = DROP_REFERENCE;
5724         wc->update_ref = update_ref;
5725         wc->keep_locks = 0;
5726         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
5727
5728         while (1) {
5729
5730                 ret = walk_down_tree(trans, root, path, wc);
5731                 if (ret < 0) {
5732                         btrfs_abort_transaction(trans, ret);
5733                         err = ret;
5734                         break;
5735                 }
5736
5737                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
5738                 if (ret < 0) {
5739                         btrfs_abort_transaction(trans, ret);
5740                         err = ret;
5741                         break;
5742                 }
5743
5744                 if (ret > 0) {
5745                         BUG_ON(wc->stage != DROP_REFERENCE);
5746                         break;
5747                 }
5748
5749                 if (wc->stage == DROP_REFERENCE) {
5750                         wc->drop_level = wc->level;
5751                         btrfs_node_key_to_cpu(path->nodes[wc->drop_level],
5752                                               &wc->drop_progress,
5753                                               path->slots[wc->drop_level]);
5754                 }
5755                 btrfs_cpu_key_to_disk(&root_item->drop_progress,
5756                                       &wc->drop_progress);
5757                 btrfs_set_root_drop_level(root_item, wc->drop_level);
5758
5759                 BUG_ON(wc->level == 0);
5760                 if (btrfs_should_end_transaction(trans) ||
5761                     (!for_reloc && btrfs_need_cleaner_sleep(fs_info))) {
5762                         ret = btrfs_update_root(trans, tree_root,
5763                                                 &root->root_key,
5764                                                 root_item);
5765                         if (ret) {
5766                                 btrfs_abort_transaction(trans, ret);
5767                                 err = ret;
5768                                 goto out_end_trans;
5769                         }
5770
5771                         if (!is_reloc_root)
5772                                 btrfs_set_last_root_drop_gen(fs_info, trans->transid);
5773
5774                         btrfs_end_transaction_throttle(trans);
5775                         if (!for_reloc && btrfs_need_cleaner_sleep(fs_info)) {
5776                                 btrfs_debug(fs_info,
5777                                             "drop snapshot early exit");
5778                                 err = -EAGAIN;
5779                                 goto out_free;
5780                         }
5781
5782                        /*
5783                         * Use join to avoid potential EINTR from transaction
5784                         * start. See wait_reserve_ticket and the whole
5785                         * reservation callchain.
5786                         */
5787                         if (for_reloc)
5788                                 trans = btrfs_join_transaction(tree_root);
5789                         else
5790                                 trans = btrfs_start_transaction(tree_root, 0);
5791                         if (IS_ERR(trans)) {
5792                                 err = PTR_ERR(trans);
5793                                 goto out_free;
5794                         }
5795                 }
5796         }
5797         btrfs_release_path(path);
5798         if (err)
5799                 goto out_end_trans;
5800
5801         ret = btrfs_del_root(trans, &root->root_key);
5802         if (ret) {
5803                 btrfs_abort_transaction(trans, ret);
5804                 err = ret;
5805                 goto out_end_trans;
5806         }
5807
5808         if (!is_reloc_root) {
5809                 ret = btrfs_find_root(tree_root, &root->root_key, path,
5810                                       NULL, NULL);
5811                 if (ret < 0) {
5812                         btrfs_abort_transaction(trans, ret);
5813                         err = ret;
5814                         goto out_end_trans;
5815                 } else if (ret > 0) {
5816                         /* if we fail to delete the orphan item this time
5817                          * around, it'll get picked up the next time.
5818                          *
5819                          * The most common failure here is just -ENOENT.
5820                          */
5821                         btrfs_del_orphan_item(trans, tree_root,
5822                                               root->root_key.objectid);
5823                 }
5824         }
5825
5826         /*
5827          * This subvolume is going to be completely dropped, and won't be
5828          * recorded as dirty roots, thus pertrans meta rsv will not be freed at
5829          * commit transaction time.  So free it here manually.
5830          */
5831         btrfs_qgroup_convert_reserved_meta(root, INT_MAX);
5832         btrfs_qgroup_free_meta_all_pertrans(root);
5833
5834         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state))
5835                 btrfs_add_dropped_root(trans, root);
5836         else
5837                 btrfs_put_root(root);
5838         root_dropped = true;
5839 out_end_trans:
5840         if (!is_reloc_root)
5841                 btrfs_set_last_root_drop_gen(fs_info, trans->transid);
5842
5843         btrfs_end_transaction_throttle(trans);
5844 out_free:
5845         kfree(wc);
5846         btrfs_free_path(path);
5847 out:
5848         /*
5849          * We were an unfinished drop root, check to see if there are any
5850          * pending, and if not clear and wake up any waiters.
5851          */
5852         if (!err && unfinished_drop)
5853                 btrfs_maybe_wake_unfinished_drop(fs_info);
5854
5855         /*
5856          * So if we need to stop dropping the snapshot for whatever reason we
5857          * need to make sure to add it back to the dead root list so that we
5858          * keep trying to do the work later.  This also cleans up roots if we
5859          * don't have it in the radix (like when we recover after a power fail
5860          * or unmount) so we don't leak memory.
5861          */
5862         if (!for_reloc && !root_dropped)
5863                 btrfs_add_dead_root(root);
5864         return err;
5865 }
5866
5867 /*
5868  * drop subtree rooted at tree block 'node'.
5869  *
5870  * NOTE: this function will unlock and release tree block 'node'
5871  * only used by relocation code
5872  */
5873 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
5874                         struct btrfs_root *root,
5875                         struct extent_buffer *node,
5876                         struct extent_buffer *parent)
5877 {
5878         struct btrfs_fs_info *fs_info = root->fs_info;
5879         struct btrfs_path *path;
5880         struct walk_control *wc;
5881         int level;
5882         int parent_level;
5883         int ret = 0;
5884         int wret;
5885
5886         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
5887
5888         path = btrfs_alloc_path();
5889         if (!path)
5890                 return -ENOMEM;
5891
5892         wc = kzalloc(sizeof(*wc), GFP_NOFS);
5893         if (!wc) {
5894                 btrfs_free_path(path);
5895                 return -ENOMEM;
5896         }
5897
5898         btrfs_assert_tree_write_locked(parent);
5899         parent_level = btrfs_header_level(parent);
5900         atomic_inc(&parent->refs);
5901         path->nodes[parent_level] = parent;
5902         path->slots[parent_level] = btrfs_header_nritems(parent);
5903
5904         btrfs_assert_tree_write_locked(node);
5905         level = btrfs_header_level(node);
5906         path->nodes[level] = node;
5907         path->slots[level] = 0;
5908         path->locks[level] = BTRFS_WRITE_LOCK;
5909
5910         wc->refs[parent_level] = 1;
5911         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
5912         wc->level = level;
5913         wc->shared_level = -1;
5914         wc->stage = DROP_REFERENCE;
5915         wc->update_ref = 0;
5916         wc->keep_locks = 1;
5917         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
5918
5919         while (1) {
5920                 wret = walk_down_tree(trans, root, path, wc);
5921                 if (wret < 0) {
5922                         ret = wret;
5923                         break;
5924                 }
5925
5926                 wret = walk_up_tree(trans, root, path, wc, parent_level);
5927                 if (wret < 0)
5928                         ret = wret;
5929                 if (wret != 0)
5930                         break;
5931         }
5932
5933         kfree(wc);
5934         btrfs_free_path(path);
5935         return ret;
5936 }
5937
5938 int btrfs_error_unpin_extent_range(struct btrfs_fs_info *fs_info,
5939                                    u64 start, u64 end)
5940 {
5941         return unpin_extent_range(fs_info, start, end, false);
5942 }
5943
5944 /*
5945  * It used to be that old block groups would be left around forever.
5946  * Iterating over them would be enough to trim unused space.  Since we
5947  * now automatically remove them, we also need to iterate over unallocated
5948  * space.
5949  *
5950  * We don't want a transaction for this since the discard may take a
5951  * substantial amount of time.  We don't require that a transaction be
5952  * running, but we do need to take a running transaction into account
5953  * to ensure that we're not discarding chunks that were released or
5954  * allocated in the current transaction.
5955  *
5956  * Holding the chunks lock will prevent other threads from allocating
5957  * or releasing chunks, but it won't prevent a running transaction
5958  * from committing and releasing the memory that the pending chunks
5959  * list head uses.  For that, we need to take a reference to the
5960  * transaction and hold the commit root sem.  We only need to hold
5961  * it while performing the free space search since we have already
5962  * held back allocations.
5963  */
5964 static int btrfs_trim_free_extents(struct btrfs_device *device, u64 *trimmed)
5965 {
5966         u64 start = BTRFS_DEVICE_RANGE_RESERVED, len = 0, end = 0;
5967         int ret;
5968
5969         *trimmed = 0;
5970
5971         /* Discard not supported = nothing to do. */
5972         if (!bdev_max_discard_sectors(device->bdev))
5973                 return 0;
5974
5975         /* Not writable = nothing to do. */
5976         if (!test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state))
5977                 return 0;
5978
5979         /* No free space = nothing to do. */
5980         if (device->total_bytes <= device->bytes_used)
5981                 return 0;
5982
5983         ret = 0;
5984
5985         while (1) {
5986                 struct btrfs_fs_info *fs_info = device->fs_info;
5987                 u64 bytes;
5988
5989                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
5990                 if (ret)
5991                         break;
5992
5993                 find_first_clear_extent_bit(&device->alloc_state, start,
5994                                             &start, &end,
5995                                             CHUNK_TRIMMED | CHUNK_ALLOCATED);
5996
5997                 /* Check if there are any CHUNK_* bits left */
5998                 if (start > device->total_bytes) {
5999                         WARN_ON(IS_ENABLED(CONFIG_BTRFS_DEBUG));
6000                         btrfs_warn_in_rcu(fs_info,
6001 "ignoring attempt to trim beyond device size: offset %llu length %llu device %s device size %llu",
6002                                           start, end - start + 1,
6003                                           btrfs_dev_name(device),
6004                                           device->total_bytes);
6005                         mutex_unlock(&fs_info->chunk_mutex);
6006                         ret = 0;
6007                         break;
6008                 }
6009
6010                 /* Ensure we skip the reserved space on each device. */
6011                 start = max_t(u64, start, BTRFS_DEVICE_RANGE_RESERVED);
6012
6013                 /*
6014                  * If find_first_clear_extent_bit find a range that spans the
6015                  * end of the device it will set end to -1, in this case it's up
6016                  * to the caller to trim the value to the size of the device.
6017                  */
6018                 end = min(end, device->total_bytes - 1);
6019
6020                 len = end - start + 1;
6021
6022                 /* We didn't find any extents */
6023                 if (!len) {
6024                         mutex_unlock(&fs_info->chunk_mutex);
6025                         ret = 0;
6026                         break;
6027                 }
6028
6029                 ret = btrfs_issue_discard(device->bdev, start, len,
6030                                           &bytes);
6031                 if (!ret)
6032                         set_extent_bit(&device->alloc_state, start,
6033                                        start + bytes - 1, CHUNK_TRIMMED, NULL);
6034                 mutex_unlock(&fs_info->chunk_mutex);
6035
6036                 if (ret)
6037                         break;
6038
6039                 start += len;
6040                 *trimmed += bytes;
6041
6042                 if (fatal_signal_pending(current)) {
6043                         ret = -ERESTARTSYS;
6044                         break;
6045                 }
6046
6047                 cond_resched();
6048         }
6049
6050         return ret;
6051 }
6052
6053 /*
6054  * Trim the whole filesystem by:
6055  * 1) trimming the free space in each block group
6056  * 2) trimming the unallocated space on each device
6057  *
6058  * This will also continue trimming even if a block group or device encounters
6059  * an error.  The return value will be the last error, or 0 if nothing bad
6060  * happens.
6061  */
6062 int btrfs_trim_fs(struct btrfs_fs_info *fs_info, struct fstrim_range *range)
6063 {
6064         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
6065         struct btrfs_block_group *cache = NULL;
6066         struct btrfs_device *device;
6067         u64 group_trimmed;
6068         u64 range_end = U64_MAX;
6069         u64 start;
6070         u64 end;
6071         u64 trimmed = 0;
6072         u64 bg_failed = 0;
6073         u64 dev_failed = 0;
6074         int bg_ret = 0;
6075         int dev_ret = 0;
6076         int ret = 0;
6077
6078         if (range->start == U64_MAX)
6079                 return -EINVAL;
6080
6081         /*
6082          * Check range overflow if range->len is set.
6083          * The default range->len is U64_MAX.
6084          */
6085         if (range->len != U64_MAX &&
6086             check_add_overflow(range->start, range->len, &range_end))
6087                 return -EINVAL;
6088
6089         cache = btrfs_lookup_first_block_group(fs_info, range->start);
6090         for (; cache; cache = btrfs_next_block_group(cache)) {
6091                 if (cache->start >= range_end) {
6092                         btrfs_put_block_group(cache);
6093                         break;
6094                 }
6095
6096                 start = max(range->start, cache->start);
6097                 end = min(range_end, cache->start + cache->length);
6098
6099                 if (end - start >= range->minlen) {
6100                         if (!btrfs_block_group_done(cache)) {
6101                                 ret = btrfs_cache_block_group(cache, true);
6102                                 if (ret) {
6103                                         bg_failed++;
6104                                         bg_ret = ret;
6105                                         continue;
6106                                 }
6107                         }
6108                         ret = btrfs_trim_block_group(cache,
6109                                                      &group_trimmed,
6110                                                      start,
6111                                                      end,
6112                                                      range->minlen);
6113
6114                         trimmed += group_trimmed;
6115                         if (ret) {
6116                                 bg_failed++;
6117                                 bg_ret = ret;
6118                                 continue;
6119                         }
6120                 }
6121         }
6122
6123         if (bg_failed)
6124                 btrfs_warn(fs_info,
6125                         "failed to trim %llu block group(s), last error %d",
6126                         bg_failed, bg_ret);
6127
6128         mutex_lock(&fs_devices->device_list_mutex);
6129         list_for_each_entry(device, &fs_devices->devices, dev_list) {
6130                 if (test_bit(BTRFS_DEV_STATE_MISSING, &device->dev_state))
6131                         continue;
6132
6133                 ret = btrfs_trim_free_extents(device, &group_trimmed);
6134                 if (ret) {
6135                         dev_failed++;
6136                         dev_ret = ret;
6137                         break;
6138                 }
6139
6140                 trimmed += group_trimmed;
6141         }
6142         mutex_unlock(&fs_devices->device_list_mutex);
6143
6144         if (dev_failed)
6145                 btrfs_warn(fs_info,
6146                         "failed to trim %llu device(s), last error %d",
6147                         dev_failed, dev_ret);
6148         range->len = trimmed;
6149         if (bg_ret)
6150                 return bg_ret;
6151         return dev_ret;
6152 }