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