fs/btrfs/extent-tree.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2007 Oracle.  All rights reserved.
   4  */
   5
   6 #include <linux/sched.h>
   7 #include <linux/sched/signal.h>
   8 #include <linux/pagemap.h>
   9 #include <linux/writeback.h>
  10 #include <linux/blkdev.h>
  11 #include <linux/sort.h>
  12 #include <linux/rcupdate.h>
  13 #include <linux/kthread.h>
  14 #include <linux/slab.h>
  15 #include <linux/ratelimit.h>
  16 #include <linux/percpu_counter.h>
  17 #include <linux/lockdep.h>
  18 #include <linux/crc32c.h>
  19 #include "ctree.h"
  20 #include "extent-tree.h"
  21 #include "transaction.h"
  22 #include "disk-io.h"
  23 #include "print-tree.h"
  24 #include "volumes.h"
  25 #include "raid56.h"
  26 #include "locking.h"
  27 #include "free-space-cache.h"
  28 #include "free-space-tree.h"
  29 #include "qgroup.h"
  30 #include "ref-verify.h"
  31 #include "space-info.h"
  32 #include "block-rsv.h"
  33 #include "discard.h"
  34 #include "zoned.h"
  35 #include "dev-replace.h"
  36 #include "fs.h"
  37 #include "accessors.h"
  38 #include "root-tree.h"
  39 #include "file-item.h"
  40 #include "orphan.h"
  41 #include "tree-checker.h"
  42 #include "raid-stripe-tree.h"
  43
  44 #undef SCRAMBLE_DELAYED_REFS
  45
  46
  47 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  48                                struct btrfs_delayed_ref_head *href,
  49                                struct btrfs_delayed_ref_node *node,
  50                                struct btrfs_delayed_extent_op *extra_op);
  51 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  52                                     struct extent_buffer *leaf,
  53                                     struct btrfs_extent_item *ei);
  54 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  55                                       u64 parent, u64 root_objectid,
  56                                       u64 flags, u64 owner, u64 offset,
  57                                       struct btrfs_key *ins, int ref_mod, u64 oref_root);
  58 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  59                                      struct btrfs_delayed_ref_node *node,
  60                                      struct btrfs_delayed_extent_op *extent_op);
  61 static int find_next_key(struct btrfs_path *path, int level,
  62                          struct btrfs_key *key);
  63
  64 static int block_group_bits(struct btrfs_block_group *cache, u64 bits)
  65 {
  66         return (cache->flags & bits) == bits;
  67 }
  68
  69 /* simple helper to search for an existing data extent at a given offset */
  70 int btrfs_lookup_data_extent(struct btrfs_fs_info *fs_info, u64 start, u64 len)
  71 {
  72         struct btrfs_root *root = btrfs_extent_root(fs_info, start);
  73         int ret;
  74         struct btrfs_key key;
  75         struct btrfs_path *path;
  76
  77         path = btrfs_alloc_path();
  78         if (!path)
  79                 return -ENOMEM;
  80
  81         key.objectid = start;
  82         key.offset = len;
  83         key.type = BTRFS_EXTENT_ITEM_KEY;
  84         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
  85         btrfs_free_path(path);
  86         return ret;
  87 }
  88
  89 /*
  90  * helper function to lookup reference count and flags of a tree block.
  91  *
  92  * the head node for delayed ref is used to store the sum of all the
  93  * reference count modifications queued up in the rbtree. the head
  94  * node may also store the extent flags to set. This way you can check
  95  * to see what the reference count and extent flags would be if all of
  96  * the delayed refs are not processed.
  97  */
  98 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
  99                              struct btrfs_fs_info *fs_info, u64 bytenr,
 100                              u64 offset, int metadata, u64 *refs, u64 *flags,
 101                              u64 *owning_root)
 102 {
 103         struct btrfs_root *extent_root;
 104         struct btrfs_delayed_ref_head *head;
 105         struct btrfs_delayed_ref_root *delayed_refs;
 106         struct btrfs_path *path;
 107         struct btrfs_key key;
 108         u64 num_refs;
 109         u64 extent_flags;
 110         u64 owner = 0;
 111         int ret;
 112
 113         /*
 114          * If we don't have skinny metadata, don't bother doing anything
 115          * different
 116          */
 117         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA)) {
 118                 offset = fs_info->nodesize;
 119                 metadata = 0;
 120         }
 121
 122         path = btrfs_alloc_path();
 123         if (!path)
 124                 return -ENOMEM;
 125
 126 search_again:
 127         key.objectid = bytenr;
 128         key.offset = offset;
 129         if (metadata)
 130                 key.type = BTRFS_METADATA_ITEM_KEY;
 131         else
 132                 key.type = BTRFS_EXTENT_ITEM_KEY;
 133
 134         extent_root = btrfs_extent_root(fs_info, bytenr);
 135         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 136         if (ret < 0)
 137                 goto out_free;
 138
 139         if (ret > 0 && key.type == BTRFS_METADATA_ITEM_KEY) {
 140                 if (path->slots[0]) {
 141                         path->slots[0]--;
 142                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 143                                               path->slots[0]);
 144                         if (key.objectid == bytenr &&
 145                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 146                             key.offset == fs_info->nodesize)
 147                                 ret = 0;
 148                 }
 149         }
 150
 151         if (ret == 0) {
 152                 struct extent_buffer *leaf = path->nodes[0];
 153                 struct btrfs_extent_item *ei;
 154                 const u32 item_size = btrfs_item_size(leaf, path->slots[0]);
 155
 156                 if (unlikely(item_size < sizeof(*ei))) {
 157                         ret = -EUCLEAN;
 158                         btrfs_err(fs_info,
 159                         "unexpected extent item size, has %u expect >= %zu",
 160                                   item_size, sizeof(*ei));
 161                         btrfs_abort_transaction(trans, ret);
 162                         goto out_free;
 163                 }
 164
 165                 ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
 166                 num_refs = btrfs_extent_refs(leaf, ei);
 167                 if (unlikely(num_refs == 0)) {
 168                         ret = -EUCLEAN;
 169                         btrfs_err(fs_info,
 170                 "unexpected zero reference count for extent item (%llu %u %llu)",
 171                                   key.objectid, key.type, key.offset);
 172                         btrfs_abort_transaction(trans, ret);
 173                         goto out_free;
 174                 }
 175                 extent_flags = btrfs_extent_flags(leaf, ei);
 176                 owner = btrfs_get_extent_owner_root(fs_info, leaf, path->slots[0]);
 177         } else {
 178                 num_refs = 0;
 179                 extent_flags = 0;
 180                 ret = 0;
 181         }
 182
 183         delayed_refs = &trans->transaction->delayed_refs;
 184         spin_lock(&delayed_refs->lock);
 185         head = btrfs_find_delayed_ref_head(fs_info, delayed_refs, bytenr);
 186         if (head) {
 187                 if (!mutex_trylock(&head->mutex)) {
 188                         refcount_inc(&head->refs);
 189                         spin_unlock(&delayed_refs->lock);
 190
 191                         btrfs_release_path(path);
 192
 193                         /*
 194                          * Mutex was contended, block until it's released and try
 195                          * again
 196                          */
 197                         mutex_lock(&head->mutex);
 198                         mutex_unlock(&head->mutex);
 199                         btrfs_put_delayed_ref_head(head);
 200                         goto search_again;
 201                 }
 202                 spin_lock(&head->lock);
 203                 if (head->extent_op && head->extent_op->update_flags)
 204                         extent_flags |= head->extent_op->flags_to_set;
 205
 206                 num_refs += head->ref_mod;
 207                 spin_unlock(&head->lock);
 208                 mutex_unlock(&head->mutex);
 209         }
 210         spin_unlock(&delayed_refs->lock);
 211
 212         WARN_ON(num_refs == 0);
 213         if (refs)
 214                 *refs = num_refs;
 215         if (flags)
 216                 *flags = extent_flags;
 217         if (owning_root)
 218                 *owning_root = owner;
 219 out_free:
 220         btrfs_free_path(path);
 221         return ret;
 222 }
 223
 224 /*
 225  * Back reference rules.  Back refs have three main goals:
 226  *
 227  * 1) differentiate between all holders of references to an extent so that
 228  *    when a reference is dropped we can make sure it was a valid reference
 229  *    before freeing the extent.
 230  *
 231  * 2) Provide enough information to quickly find the holders of an extent
 232  *    if we notice a given block is corrupted or bad.
 233  *
 234  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 235  *    maintenance.  This is actually the same as #2, but with a slightly
 236  *    different use case.
 237  *
 238  * There are two kinds of back refs. The implicit back refs is optimized
 239  * for pointers in non-shared tree blocks. For a given pointer in a block,
 240  * back refs of this kind provide information about the block's owner tree
 241  * and the pointer's key. These information allow us to find the block by
 242  * b-tree searching. The full back refs is for pointers in tree blocks not
 243  * referenced by their owner trees. The location of tree block is recorded
 244  * in the back refs. Actually the full back refs is generic, and can be
 245  * used in all cases the implicit back refs is used. The major shortcoming
 246  * of the full back refs is its overhead. Every time a tree block gets
 247  * COWed, we have to update back refs entry for all pointers in it.
 248  *
 249  * For a newly allocated tree block, we use implicit back refs for
 250  * pointers in it. This means most tree related operations only involve
 251  * implicit back refs. For a tree block created in old transaction, the
 252  * only way to drop a reference to it is COW it. So we can detect the
 253  * event that tree block loses its owner tree's reference and do the
 254  * back refs conversion.
 255  *
 256  * When a tree block is COWed through a tree, there are four cases:
 257  *
 258  * The reference count of the block is one and the tree is the block's
 259  * owner tree. Nothing to do in this case.
 260  *
 261  * The reference count of the block is one and the tree is not the
 262  * block's owner tree. In this case, full back refs is used for pointers
 263  * in the block. Remove these full back refs, add implicit back refs for
 264  * every pointers in the new block.
 265  *
 266  * The reference count of the block is greater than one and the tree is
 267  * the block's owner tree. In this case, implicit back refs is used for
 268  * pointers in the block. Add full back refs for every pointers in the
 269  * block, increase lower level extents' reference counts. The original
 270  * implicit back refs are entailed to the new block.
 271  *
 272  * The reference count of the block is greater than one and the tree is
 273  * not the block's owner tree. Add implicit back refs for every pointer in
 274  * the new block, increase lower level extents' reference count.
 275  *
 276  * Back Reference Key composing:
 277  *
 278  * The key objectid corresponds to the first byte in the extent,
 279  * The key type is used to differentiate between types of back refs.
 280  * There are different meanings of the key offset for different types
 281  * of back refs.
 282  *
 283  * File extents can be referenced by:
 284  *
 285  * - multiple snapshots, subvolumes, or different generations in one subvol
 286  * - different files inside a single subvolume
 287  * - different offsets inside a file (bookend extents in file.c)
 288  *
 289  * The extent ref structure for the implicit back refs has fields for:
 290  *
 291  * - Objectid of the subvolume root
 292  * - objectid of the file holding the reference
 293  * - original offset in the file
 294  * - how many bookend extents
 295  *
 296  * The key offset for the implicit back refs is hash of the first
 297  * three fields.
 298  *
 299  * The extent ref structure for the full back refs has field for:
 300  *
 301  * - number of pointers in the tree leaf
 302  *
 303  * The key offset for the implicit back refs is the first byte of
 304  * the tree leaf
 305  *
 306  * When a file extent is allocated, The implicit back refs is used.
 307  * the fields are filled in:
 308  *
 309  *     (root_key.objectid, inode objectid, offset in file, 1)
 310  *
 311  * When a file extent is removed file truncation, we find the
 312  * corresponding implicit back refs and check the following fields:
 313  *
 314  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
 315  *
 316  * Btree extents can be referenced by:
 317  *
 318  * - Different subvolumes
 319  *
 320  * Both the implicit back refs and the full back refs for tree blocks
 321  * only consist of key. The key offset for the implicit back refs is
 322  * objectid of block's owner tree. The key offset for the full back refs
 323  * is the first byte of parent block.
 324  *
 325  * When implicit back refs is used, information about the lowest key and
 326  * level of the tree block are required. These information are stored in
 327  * tree block info structure.
 328  */
 329
 330 /*
 331  * is_data == BTRFS_REF_TYPE_BLOCK, tree block type is required,
 332  * is_data == BTRFS_REF_TYPE_DATA, data type is requiried,
 333  * is_data == BTRFS_REF_TYPE_ANY, either type is OK.
 334  */
 335 int btrfs_get_extent_inline_ref_type(const struct extent_buffer *eb,
 336                                      struct btrfs_extent_inline_ref *iref,
 337                                      enum btrfs_inline_ref_type is_data)
 338 {
 339         struct btrfs_fs_info *fs_info = eb->fs_info;
 340         int type = btrfs_extent_inline_ref_type(eb, iref);
 341         u64 offset = btrfs_extent_inline_ref_offset(eb, iref);
 342
 343         if (type == BTRFS_EXTENT_OWNER_REF_KEY) {
 344                 ASSERT(btrfs_fs_incompat(fs_info, SIMPLE_QUOTA));
 345                 return type;
 346         }
 347
 348         if (type == BTRFS_TREE_BLOCK_REF_KEY ||
 349             type == BTRFS_SHARED_BLOCK_REF_KEY ||
 350             type == BTRFS_SHARED_DATA_REF_KEY ||
 351             type == BTRFS_EXTENT_DATA_REF_KEY) {
 352                 if (is_data == BTRFS_REF_TYPE_BLOCK) {
 353                         if (type == BTRFS_TREE_BLOCK_REF_KEY)
 354                                 return type;
 355                         if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
 356                                 ASSERT(fs_info);
 357                                 /*
 358                                  * Every shared one has parent tree block,
 359                                  * which must be aligned to sector size.
 360                                  */
 361                                 if (offset && IS_ALIGNED(offset, fs_info->sectorsize))
 362                                         return type;
 363                         }
 364                 } else if (is_data == BTRFS_REF_TYPE_DATA) {
 365                         if (type == BTRFS_EXTENT_DATA_REF_KEY)
 366                                 return type;
 367                         if (type == BTRFS_SHARED_DATA_REF_KEY) {
 368                                 ASSERT(fs_info);
 369                                 /*
 370                                  * Every shared one has parent tree block,
 371                                  * which must be aligned to sector size.
 372                                  */
 373                                 if (offset &&
 374                                     IS_ALIGNED(offset, fs_info->sectorsize))
 375                                         return type;
 376                         }
 377                 } else {
 378                         ASSERT(is_data == BTRFS_REF_TYPE_ANY);
 379                         return type;
 380                 }
 381         }
 382
 383         WARN_ON(1);
 384         btrfs_print_leaf(eb);
 385         btrfs_err(fs_info,
 386                   "eb %llu iref 0x%lx invalid extent inline ref type %d",
 387                   eb->start, (unsigned long)iref, type);
 388
 389         return BTRFS_REF_TYPE_INVALID;
 390 }
 391
 392 u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
 393 {
 394         u32 high_crc = ~(u32)0;
 395         u32 low_crc = ~(u32)0;
 396         __le64 lenum;
 397
 398         lenum = cpu_to_le64(root_objectid);
 399         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
 400         lenum = cpu_to_le64(owner);
 401         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
 402         lenum = cpu_to_le64(offset);
 403         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
 404
 405         return ((u64)high_crc << 31) ^ (u64)low_crc;
 406 }
 407
 408 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
 409                                      struct btrfs_extent_data_ref *ref)
 410 {
 411         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
 412                                     btrfs_extent_data_ref_objectid(leaf, ref),
 413                                     btrfs_extent_data_ref_offset(leaf, ref));
 414 }
 415
 416 static int match_extent_data_ref(struct extent_buffer *leaf,
 417                                  struct btrfs_extent_data_ref *ref,
 418                                  u64 root_objectid, u64 owner, u64 offset)
 419 {
 420         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
 421             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
 422             btrfs_extent_data_ref_offset(leaf, ref) != offset)
 423                 return 0;
 424         return 1;
 425 }
 426
 427 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
 428                                            struct btrfs_path *path,
 429                                            u64 bytenr, u64 parent,
 430                                            u64 root_objectid,
 431                                            u64 owner, u64 offset)
 432 {
 433         struct btrfs_root *root = btrfs_extent_root(trans->fs_info, bytenr);
 434         struct btrfs_key key;
 435         struct btrfs_extent_data_ref *ref;
 436         struct extent_buffer *leaf;
 437         u32 nritems;
 438         int recow;
 439         int ret;
 440
 441         key.objectid = bytenr;
 442         if (parent) {
 443                 key.type = BTRFS_SHARED_DATA_REF_KEY;
 444                 key.offset = parent;
 445         } else {
 446                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
 447                 key.offset = hash_extent_data_ref(root_objectid,
 448                                                   owner, offset);
 449         }
 450 again:
 451         recow = 0;
 452         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
 453         if (ret < 0)
 454                 return ret;
 455
 456         if (parent) {
 457                 if (ret)
 458                         return -ENOENT;
 459                 return 0;
 460         }
 461
 462         ret = -ENOENT;
 463         leaf = path->nodes[0];
 464         nritems = btrfs_header_nritems(leaf);
 465         while (1) {
 466                 if (path->slots[0] >= nritems) {
 467                         ret = btrfs_next_leaf(root, path);
 468                         if (ret) {
 469                                 if (ret > 0)
 470                                         return -ENOENT;
 471                                 return ret;
 472                         }
 473
 474                         leaf = path->nodes[0];
 475                         nritems = btrfs_header_nritems(leaf);
 476                         recow = 1;
 477                 }
 478
 479                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 480                 if (key.objectid != bytenr ||
 481                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
 482                         goto fail;
 483
 484                 ref = btrfs_item_ptr(leaf, path->slots[0],
 485                                      struct btrfs_extent_data_ref);
 486
 487                 if (match_extent_data_ref(leaf, ref, root_objectid,
 488                                           owner, offset)) {
 489                         if (recow) {
 490                                 btrfs_release_path(path);
 491                                 goto again;
 492                         }
 493                         ret = 0;
 494                         break;
 495                 }
 496                 path->slots[0]++;
 497         }
 498 fail:
 499         return ret;
 500 }
 501
 502 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
 503                                            struct btrfs_path *path,
 504                                            struct btrfs_delayed_ref_node *node,
 505                                            u64 bytenr)
 506 {
 507         struct btrfs_root *root = btrfs_extent_root(trans->fs_info, bytenr);
 508         struct btrfs_key key;
 509         struct extent_buffer *leaf;
 510         u64 owner = btrfs_delayed_ref_owner(node);
 511         u64 offset = btrfs_delayed_ref_offset(node);
 512         u32 size;
 513         u32 num_refs;
 514         int ret;
 515
 516         key.objectid = bytenr;
 517         if (node->parent) {
 518                 key.type = BTRFS_SHARED_DATA_REF_KEY;
 519                 key.offset = node->parent;
 520                 size = sizeof(struct btrfs_shared_data_ref);
 521         } else {
 522                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
 523                 key.offset = hash_extent_data_ref(node->ref_root, owner, offset);
 524                 size = sizeof(struct btrfs_extent_data_ref);
 525         }
 526
 527         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
 528         if (ret && ret != -EEXIST)
 529                 goto fail;
 530
 531         leaf = path->nodes[0];
 532         if (node->parent) {
 533                 struct btrfs_shared_data_ref *ref;
 534                 ref = btrfs_item_ptr(leaf, path->slots[0],
 535                                      struct btrfs_shared_data_ref);
 536                 if (ret == 0) {
 537                         btrfs_set_shared_data_ref_count(leaf, ref, node->ref_mod);
 538                 } else {
 539                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
 540                         num_refs += node->ref_mod;
 541                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
 542                 }
 543         } else {
 544                 struct btrfs_extent_data_ref *ref;
 545                 while (ret == -EEXIST) {
 546                         ref = btrfs_item_ptr(leaf, path->slots[0],
 547                                              struct btrfs_extent_data_ref);
 548                         if (match_extent_data_ref(leaf, ref, node->ref_root,
 549                                                   owner, offset))
 550                                 break;
 551                         btrfs_release_path(path);
 552                         key.offset++;
 553                         ret = btrfs_insert_empty_item(trans, root, path, &key,
 554                                                       size);
 555                         if (ret && ret != -EEXIST)
 556                                 goto fail;
 557
 558                         leaf = path->nodes[0];
 559                 }
 560                 ref = btrfs_item_ptr(leaf, path->slots[0],
 561                                      struct btrfs_extent_data_ref);
 562                 if (ret == 0) {
 563                         btrfs_set_extent_data_ref_root(leaf, ref, node->ref_root);
 564                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
 565                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
 566                         btrfs_set_extent_data_ref_count(leaf, ref, node->ref_mod);
 567                 } else {
 568                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
 569                         num_refs += node->ref_mod;
 570                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
 571                 }
 572         }
 573         btrfs_mark_buffer_dirty(trans, leaf);
 574         ret = 0;
 575 fail:
 576         btrfs_release_path(path);
 577         return ret;
 578 }
 579
 580 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
 581                                            struct btrfs_root *root,
 582                                            struct btrfs_path *path,
 583                                            int refs_to_drop)
 584 {
 585         struct btrfs_key key;
 586         struct btrfs_extent_data_ref *ref1 = NULL;
 587         struct btrfs_shared_data_ref *ref2 = NULL;
 588         struct extent_buffer *leaf;
 589         u32 num_refs = 0;
 590         int ret = 0;
 591
 592         leaf = path->nodes[0];
 593         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 594
 595         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
 596                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
 597                                       struct btrfs_extent_data_ref);
 598                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
 599         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
 600                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
 601                                       struct btrfs_shared_data_ref);
 602                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
 603         } else {
 604                 btrfs_err(trans->fs_info,
 605                           "unrecognized backref key (%llu %u %llu)",
 606                           key.objectid, key.type, key.offset);
 607                 btrfs_abort_transaction(trans, -EUCLEAN);
 608                 return -EUCLEAN;
 609         }
 610
 611         BUG_ON(num_refs < refs_to_drop);
 612         num_refs -= refs_to_drop;
 613
 614         if (num_refs == 0) {
 615                 ret = btrfs_del_item(trans, root, path);
 616         } else {
 617                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
 618                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
 619                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
 620                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
 621                 btrfs_mark_buffer_dirty(trans, leaf);
 622         }
 623         return ret;
 624 }
 625
 626 static noinline u32 extent_data_ref_count(struct btrfs_path *path,
 627                                           struct btrfs_extent_inline_ref *iref)
 628 {
 629         struct btrfs_key key;
 630         struct extent_buffer *leaf;
 631         struct btrfs_extent_data_ref *ref1;
 632         struct btrfs_shared_data_ref *ref2;
 633         u32 num_refs = 0;
 634         int type;
 635
 636         leaf = path->nodes[0];
 637         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 638
 639         if (iref) {
 640                 /*
 641                  * If type is invalid, we should have bailed out earlier than
 642                  * this call.
 643                  */
 644                 type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
 645                 ASSERT(type != BTRFS_REF_TYPE_INVALID);
 646                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
 647                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
 648                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
 649                 } else {
 650                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
 651                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
 652                 }
 653         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
 654                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
 655                                       struct btrfs_extent_data_ref);
 656                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
 657         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
 658                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
 659                                       struct btrfs_shared_data_ref);
 660                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
 661         } else {
 662                 WARN_ON(1);
 663         }
 664         return num_refs;
 665 }
 666
 667 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
 668                                           struct btrfs_path *path,
 669                                           u64 bytenr, u64 parent,
 670                                           u64 root_objectid)
 671 {
 672         struct btrfs_root *root = btrfs_extent_root(trans->fs_info, bytenr);
 673         struct btrfs_key key;
 674         int ret;
 675
 676         key.objectid = bytenr;
 677         if (parent) {
 678                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
 679                 key.offset = parent;
 680         } else {
 681                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
 682                 key.offset = root_objectid;
 683         }
 684
 685         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
 686         if (ret > 0)
 687                 ret = -ENOENT;
 688         return ret;
 689 }
 690
 691 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
 692                                           struct btrfs_path *path,
 693                                           struct btrfs_delayed_ref_node *node,
 694                                           u64 bytenr)
 695 {
 696         struct btrfs_root *root = btrfs_extent_root(trans->fs_info, bytenr);
 697         struct btrfs_key key;
 698         int ret;
 699
 700         key.objectid = bytenr;
 701         if (node->parent) {
 702                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
 703                 key.offset = node->parent;
 704         } else {
 705                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
 706                 key.offset = node->ref_root;
 707         }
 708
 709         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
 710         btrfs_release_path(path);
 711         return ret;
 712 }
 713
 714 static inline int extent_ref_type(u64 parent, u64 owner)
 715 {
 716         int type;
 717         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
 718                 if (parent > 0)
 719                         type = BTRFS_SHARED_BLOCK_REF_KEY;
 720                 else
 721                         type = BTRFS_TREE_BLOCK_REF_KEY;
 722         } else {
 723                 if (parent > 0)
 724                         type = BTRFS_SHARED_DATA_REF_KEY;
 725                 else
 726                         type = BTRFS_EXTENT_DATA_REF_KEY;
 727         }
 728         return type;
 729 }
 730
 731 static int find_next_key(struct btrfs_path *path, int level,
 732                          struct btrfs_key *key)
 733
 734 {
 735         for (; level < BTRFS_MAX_LEVEL; level++) {
 736                 if (!path->nodes[level])
 737                         break;
 738                 if (path->slots[level] + 1 >=
 739                     btrfs_header_nritems(path->nodes[level]))
 740                         continue;
 741                 if (level == 0)
 742                         btrfs_item_key_to_cpu(path->nodes[level], key,
 743                                               path->slots[level] + 1);
 744                 else
 745                         btrfs_node_key_to_cpu(path->nodes[level], key,
 746                                               path->slots[level] + 1);
 747                 return 0;
 748         }
 749         return 1;
 750 }
 751
 752 /*
 753  * look for inline back ref. if back ref is found, *ref_ret is set
 754  * to the address of inline back ref, and 0 is returned.
 755  *
 756  * if back ref isn't found, *ref_ret is set to the address where it
 757  * should be inserted, and -ENOENT is returned.
 758  *
 759  * if insert is true and there are too many inline back refs, the path
 760  * points to the extent item, and -EAGAIN is returned.
 761  *
 762  * NOTE: inline back refs are ordered in the same way that back ref
 763  *       items in the tree are ordered.
 764  */
 765 static noinline_for_stack
 766 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
 767                                  struct btrfs_path *path,
 768                                  struct btrfs_extent_inline_ref **ref_ret,
 769                                  u64 bytenr, u64 num_bytes,
 770                                  u64 parent, u64 root_objectid,
 771                                  u64 owner, u64 offset, int insert)
 772 {
 773         struct btrfs_fs_info *fs_info = trans->fs_info;
 774         struct btrfs_root *root = btrfs_extent_root(fs_info, bytenr);
 775         struct btrfs_key key;
 776         struct extent_buffer *leaf;
 777         struct btrfs_extent_item *ei;
 778         struct btrfs_extent_inline_ref *iref;
 779         u64 flags;
 780         u64 item_size;
 781         unsigned long ptr;
 782         unsigned long end;
 783         int extra_size;
 784         int type;
 785         int want;
 786         int ret;
 787         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
 788         int needed;
 789
 790         key.objectid = bytenr;
 791         key.type = BTRFS_EXTENT_ITEM_KEY;
 792         key.offset = num_bytes;
 793
 794         want = extent_ref_type(parent, owner);
 795         if (insert) {
 796                 extra_size = btrfs_extent_inline_ref_size(want);
 797                 path->search_for_extension = 1;
 798         } else
 799                 extra_size = -1;
 800
 801         /*
 802          * Owner is our level, so we can just add one to get the level for the
 803          * block we are interested in.
 804          */
 805         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
 806                 key.type = BTRFS_METADATA_ITEM_KEY;
 807                 key.offset = owner;
 808         }
 809
 810 again:
 811         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
 812         if (ret < 0)
 813                 goto out;
 814
 815         /*
 816          * We may be a newly converted file system which still has the old fat
 817          * extent entries for metadata, so try and see if we have one of those.
 818          */
 819         if (ret > 0 && skinny_metadata) {
 820                 skinny_metadata = false;
 821                 if (path->slots[0]) {
 822                         path->slots[0]--;
 823                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 824                                               path->slots[0]);
 825                         if (key.objectid == bytenr &&
 826                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 827                             key.offset == num_bytes)
 828                                 ret = 0;
 829                 }
 830                 if (ret) {
 831                         key.objectid = bytenr;
 832                         key.type = BTRFS_EXTENT_ITEM_KEY;
 833                         key.offset = num_bytes;
 834                         btrfs_release_path(path);
 835                         goto again;
 836                 }
 837         }
 838
 839         if (ret && !insert) {
 840                 ret = -ENOENT;
 841                 goto out;
 842         } else if (WARN_ON(ret)) {
 843                 btrfs_print_leaf(path->nodes[0]);
 844                 btrfs_err(fs_info,
 845 "extent item not found for insert, bytenr %llu num_bytes %llu parent %llu root_objectid %llu owner %llu offset %llu",
 846                           bytenr, num_bytes, parent, root_objectid, owner,
 847                           offset);
 848                 ret = -EUCLEAN;
 849                 goto out;
 850         }
 851
 852         leaf = path->nodes[0];
 853         item_size = btrfs_item_size(leaf, path->slots[0]);
 854         if (unlikely(item_size < sizeof(*ei))) {
 855                 ret = -EUCLEAN;
 856                 btrfs_err(fs_info,
 857                           "unexpected extent item size, has %llu expect >= %zu",
 858                           item_size, sizeof(*ei));
 859                 btrfs_abort_transaction(trans, ret);
 860                 goto out;
 861         }
 862
 863         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
 864         flags = btrfs_extent_flags(leaf, ei);
 865
 866         ptr = (unsigned long)(ei + 1);
 867         end = (unsigned long)ei + item_size;
 868
 869         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
 870                 ptr += sizeof(struct btrfs_tree_block_info);
 871                 BUG_ON(ptr > end);
 872         }
 873
 874         if (owner >= BTRFS_FIRST_FREE_OBJECTID)
 875                 needed = BTRFS_REF_TYPE_DATA;
 876         else
 877                 needed = BTRFS_REF_TYPE_BLOCK;
 878
 879         ret = -ENOENT;
 880         while (ptr < end) {
 881                 iref = (struct btrfs_extent_inline_ref *)ptr;
 882                 type = btrfs_get_extent_inline_ref_type(leaf, iref, needed);
 883                 if (type == BTRFS_EXTENT_OWNER_REF_KEY) {
 884                         ASSERT(btrfs_fs_incompat(fs_info, SIMPLE_QUOTA));
 885                         ptr += btrfs_extent_inline_ref_size(type);
 886                         continue;
 887                 }
 888                 if (type == BTRFS_REF_TYPE_INVALID) {
 889                         ret = -EUCLEAN;
 890                         goto out;
 891                 }
 892
 893                 if (want < type)
 894                         break;
 895                 if (want > type) {
 896                         ptr += btrfs_extent_inline_ref_size(type);
 897                         continue;
 898                 }
 899
 900                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
 901                         struct btrfs_extent_data_ref *dref;
 902                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
 903                         if (match_extent_data_ref(leaf, dref, root_objectid,
 904                                                   owner, offset)) {
 905                                 ret = 0;
 906                                 break;
 907                         }
 908                         if (hash_extent_data_ref_item(leaf, dref) <
 909                             hash_extent_data_ref(root_objectid, owner, offset))
 910                                 break;
 911                 } else {
 912                         u64 ref_offset;
 913                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
 914                         if (parent > 0) {
 915                                 if (parent == ref_offset) {
 916                                         ret = 0;
 917                                         break;
 918                                 }
 919                                 if (ref_offset < parent)
 920                                         break;
 921                         } else {
 922                                 if (root_objectid == ref_offset) {
 923                                         ret = 0;
 924                                         break;
 925                                 }
 926                                 if (ref_offset < root_objectid)
 927                                         break;
 928                         }
 929                 }
 930                 ptr += btrfs_extent_inline_ref_size(type);
 931         }
 932
 933         if (unlikely(ptr > end)) {
 934                 ret = -EUCLEAN;
 935                 btrfs_print_leaf(path->nodes[0]);
 936                 btrfs_crit(fs_info,
 937 "overrun extent record at slot %d while looking for inline extent for root %llu owner %llu offset %llu parent %llu",
 938                            path->slots[0], root_objectid, owner, offset, parent);
 939                 goto out;
 940         }
 941
 942         if (ret == -ENOENT && insert) {
 943                 if (item_size + extra_size >=
 944                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
 945                         ret = -EAGAIN;
 946                         goto out;
 947                 }
 948
 949                 if (path->slots[0] + 1 < btrfs_header_nritems(path->nodes[0])) {
 950                         struct btrfs_key tmp_key;
 951
 952                         btrfs_item_key_to_cpu(path->nodes[0], &tmp_key, path->slots[0] + 1);
 953                         if (tmp_key.objectid == bytenr &&
 954                             tmp_key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
 955                                 ret = -EAGAIN;
 956                                 goto out;
 957                         }
 958                         goto out_no_entry;
 959                 }
 960
 961                 if (!path->keep_locks) {
 962                         btrfs_release_path(path);
 963                         path->keep_locks = 1;
 964                         goto again;
 965                 }
 966
 967                 /*
 968                  * To add new inline back ref, we have to make sure
 969                  * there is no corresponding back ref item.
 970                  * For simplicity, we just do not add new inline back
 971                  * ref if there is any kind of item for this block
 972                  */
 973                 if (find_next_key(path, 0, &key) == 0 &&
 974                     key.objectid == bytenr &&
 975                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
 976                         ret = -EAGAIN;
 977                         goto out;
 978                 }
 979         }
 980 out_no_entry:
 981         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
 982 out:
 983         if (path->keep_locks) {
 984                 path->keep_locks = 0;
 985                 btrfs_unlock_up_safe(path, 1);
 986         }
 987         if (insert)
 988                 path->search_for_extension = 0;
 989         return ret;
 990 }
 991
 992 /*
 993  * helper to add new inline back ref
 994  */
 995 static noinline_for_stack
 996 void setup_inline_extent_backref(struct btrfs_trans_handle *trans,
 997                                  struct btrfs_path *path,
 998                                  struct btrfs_extent_inline_ref *iref,
 999                                  u64 parent, u64 root_objectid,
1000                                  u64 owner, u64 offset, int refs_to_add,
1001                                  struct btrfs_delayed_extent_op *extent_op)
1002 {
1003         struct extent_buffer *leaf;
1004         struct btrfs_extent_item *ei;
1005         unsigned long ptr;
1006         unsigned long end;
1007         unsigned long item_offset;
1008         u64 refs;
1009         int size;
1010         int type;
1011
1012         leaf = path->nodes[0];
1013         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1014         item_offset = (unsigned long)iref - (unsigned long)ei;
1015
1016         type = extent_ref_type(parent, owner);
1017         size = btrfs_extent_inline_ref_size(type);
1018
1019         btrfs_extend_item(trans, path, size);
1020
1021         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1022         refs = btrfs_extent_refs(leaf, ei);
1023         refs += refs_to_add;
1024         btrfs_set_extent_refs(leaf, ei, refs);
1025         if (extent_op)
1026                 __run_delayed_extent_op(extent_op, leaf, ei);
1027
1028         ptr = (unsigned long)ei + item_offset;
1029         end = (unsigned long)ei + btrfs_item_size(leaf, path->slots[0]);
1030         if (ptr < end - size)
1031                 memmove_extent_buffer(leaf, ptr + size, ptr,
1032                                       end - size - ptr);
1033
1034         iref = (struct btrfs_extent_inline_ref *)ptr;
1035         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1036         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1037                 struct btrfs_extent_data_ref *dref;
1038                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1039                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1040                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1041                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1042                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1043         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1044                 struct btrfs_shared_data_ref *sref;
1045                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1046                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1047                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1048         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1049                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1050         } else {
1051                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1052         }
1053         btrfs_mark_buffer_dirty(trans, leaf);
1054 }
1055
1056 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1057                                  struct btrfs_path *path,
1058                                  struct btrfs_extent_inline_ref **ref_ret,
1059                                  u64 bytenr, u64 num_bytes, u64 parent,
1060                                  u64 root_objectid, u64 owner, u64 offset)
1061 {
1062         int ret;
1063
1064         ret = lookup_inline_extent_backref(trans, path, ref_ret, bytenr,
1065                                            num_bytes, parent, root_objectid,
1066                                            owner, offset, 0);
1067         if (ret != -ENOENT)
1068                 return ret;
1069
1070         btrfs_release_path(path);
1071         *ref_ret = NULL;
1072
1073         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1074                 ret = lookup_tree_block_ref(trans, path, bytenr, parent,
1075                                             root_objectid);
1076         } else {
1077                 ret = lookup_extent_data_ref(trans, path, bytenr, parent,
1078                                              root_objectid, owner, offset);
1079         }
1080         return ret;
1081 }
1082
1083 /*
1084  * helper to update/remove inline back ref
1085  */
1086 static noinline_for_stack int update_inline_extent_backref(
1087                                   struct btrfs_trans_handle *trans,
1088                                   struct btrfs_path *path,
1089                                   struct btrfs_extent_inline_ref *iref,
1090                                   int refs_to_mod,
1091                                   struct btrfs_delayed_extent_op *extent_op)
1092 {
1093         struct extent_buffer *leaf = path->nodes[0];
1094         struct btrfs_fs_info *fs_info = leaf->fs_info;
1095         struct btrfs_extent_item *ei;
1096         struct btrfs_extent_data_ref *dref = NULL;
1097         struct btrfs_shared_data_ref *sref = NULL;
1098         unsigned long ptr;
1099         unsigned long end;
1100         u32 item_size;
1101         int size;
1102         int type;
1103         u64 refs;
1104
1105         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1106         refs = btrfs_extent_refs(leaf, ei);
1107         if (unlikely(refs_to_mod < 0 && refs + refs_to_mod <= 0)) {
1108                 struct btrfs_key key;
1109                 u32 extent_size;
1110
1111                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1112                 if (key.type == BTRFS_METADATA_ITEM_KEY)
1113                         extent_size = fs_info->nodesize;
1114                 else
1115                         extent_size = key.offset;
1116                 btrfs_print_leaf(leaf);
1117                 btrfs_err(fs_info,
1118         "invalid refs_to_mod for extent %llu num_bytes %u, has %d expect >= -%llu",
1119                           key.objectid, extent_size, refs_to_mod, refs);
1120                 return -EUCLEAN;
1121         }
1122         refs += refs_to_mod;
1123         btrfs_set_extent_refs(leaf, ei, refs);
1124         if (extent_op)
1125                 __run_delayed_extent_op(extent_op, leaf, ei);
1126
1127         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_ANY);
1128         /*
1129          * Function btrfs_get_extent_inline_ref_type() has already printed
1130          * error messages.
1131          */
1132         if (unlikely(type == BTRFS_REF_TYPE_INVALID))
1133                 return -EUCLEAN;
1134
1135         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1136                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1137                 refs = btrfs_extent_data_ref_count(leaf, dref);
1138         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1139                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1140                 refs = btrfs_shared_data_ref_count(leaf, sref);
1141         } else {
1142                 refs = 1;
1143                 /*
1144                  * For tree blocks we can only drop one ref for it, and tree
1145                  * blocks should not have refs > 1.
1146                  *
1147                  * Furthermore if we're inserting a new inline backref, we
1148                  * won't reach this path either. That would be
1149                  * setup_inline_extent_backref().
1150                  */
1151                 if (unlikely(refs_to_mod != -1)) {
1152                         struct btrfs_key key;
1153
1154                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1155
1156                         btrfs_print_leaf(leaf);
1157                         btrfs_err(fs_info,
1158                         "invalid refs_to_mod for tree block %llu, has %d expect -1",
1159                                   key.objectid, refs_to_mod);
1160                         return -EUCLEAN;
1161                 }
1162         }
1163
1164         if (unlikely(refs_to_mod < 0 && refs < -refs_to_mod)) {
1165                 struct btrfs_key key;
1166                 u32 extent_size;
1167
1168                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1169                 if (key.type == BTRFS_METADATA_ITEM_KEY)
1170                         extent_size = fs_info->nodesize;
1171                 else
1172                         extent_size = key.offset;
1173                 btrfs_print_leaf(leaf);
1174                 btrfs_err(fs_info,
1175 "invalid refs_to_mod for backref entry, iref %lu extent %llu num_bytes %u, has %d expect >= -%llu",
1176                           (unsigned long)iref, key.objectid, extent_size,
1177                           refs_to_mod, refs);
1178                 return -EUCLEAN;
1179         }
1180         refs += refs_to_mod;
1181
1182         if (refs > 0) {
1183                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1184                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1185                 else
1186                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1187         } else {
1188                 size =  btrfs_extent_inline_ref_size(type);
1189                 item_size = btrfs_item_size(leaf, path->slots[0]);
1190                 ptr = (unsigned long)iref;
1191                 end = (unsigned long)ei + item_size;
1192                 if (ptr + size < end)
1193                         memmove_extent_buffer(leaf, ptr, ptr + size,
1194                                               end - ptr - size);
1195                 item_size -= size;
1196                 btrfs_truncate_item(trans, path, item_size, 1);
1197         }
1198         btrfs_mark_buffer_dirty(trans, leaf);
1199         return 0;
1200 }
1201
1202 static noinline_for_stack
1203 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1204                                  struct btrfs_path *path,
1205                                  u64 bytenr, u64 num_bytes, u64 parent,
1206                                  u64 root_objectid, u64 owner,
1207                                  u64 offset, int refs_to_add,
1208                                  struct btrfs_delayed_extent_op *extent_op)
1209 {
1210         struct btrfs_extent_inline_ref *iref;
1211         int ret;
1212
1213         ret = lookup_inline_extent_backref(trans, path, &iref, bytenr,
1214                                            num_bytes, parent, root_objectid,
1215                                            owner, offset, 1);
1216         if (ret == 0) {
1217                 /*
1218                  * We're adding refs to a tree block we already own, this
1219                  * should not happen at all.
1220                  */
1221                 if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1222                         btrfs_print_leaf(path->nodes[0]);
1223                         btrfs_crit(trans->fs_info,
1224 "adding refs to an existing tree ref, bytenr %llu num_bytes %llu root_objectid %llu slot %u",
1225                                    bytenr, num_bytes, root_objectid, path->slots[0]);
1226                         return -EUCLEAN;
1227                 }
1228                 ret = update_inline_extent_backref(trans, path, iref,
1229                                                    refs_to_add, extent_op);
1230         } else if (ret == -ENOENT) {
1231                 setup_inline_extent_backref(trans, path, iref, parent,
1232                                             root_objectid, owner, offset,
1233                                             refs_to_add, extent_op);
1234                 ret = 0;
1235         }
1236         return ret;
1237 }
1238
1239 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1240                                  struct btrfs_root *root,
1241                                  struct btrfs_path *path,
1242                                  struct btrfs_extent_inline_ref *iref,
1243                                  int refs_to_drop, int is_data)
1244 {
1245         int ret = 0;
1246
1247         BUG_ON(!is_data && refs_to_drop != 1);
1248         if (iref)
1249                 ret = update_inline_extent_backref(trans, path, iref,
1250                                                    -refs_to_drop, NULL);
1251         else if (is_data)
1252                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop);
1253         else
1254                 ret = btrfs_del_item(trans, root, path);
1255         return ret;
1256 }
1257
1258 static int btrfs_issue_discard(struct block_device *bdev, u64 start, u64 len,
1259                                u64 *discarded_bytes)
1260 {
1261         int j, ret = 0;
1262         u64 bytes_left, end;
1263         u64 aligned_start = ALIGN(start, 1 << SECTOR_SHIFT);
1264
1265         /* Adjust the range to be aligned to 512B sectors if necessary. */
1266         if (start != aligned_start) {
1267                 len -= aligned_start - start;
1268                 len = round_down(len, 1 << SECTOR_SHIFT);
1269                 start = aligned_start;
1270         }
1271
1272         *discarded_bytes = 0;
1273
1274         if (!len)
1275                 return 0;
1276
1277         end = start + len;
1278         bytes_left = len;
1279
1280         /* Skip any superblocks on this device. */
1281         for (j = 0; j < BTRFS_SUPER_MIRROR_MAX; j++) {
1282                 u64 sb_start = btrfs_sb_offset(j);
1283                 u64 sb_end = sb_start + BTRFS_SUPER_INFO_SIZE;
1284                 u64 size = sb_start - start;
1285
1286                 if (!in_range(sb_start, start, bytes_left) &&
1287                     !in_range(sb_end, start, bytes_left) &&
1288                     !in_range(start, sb_start, BTRFS_SUPER_INFO_SIZE))
1289                         continue;
1290
1291                 /*
1292                  * Superblock spans beginning of range.  Adjust start and
1293                  * try again.
1294                  */
1295                 if (sb_start <= start) {
1296                         start += sb_end - start;
1297                         if (start > end) {
1298                                 bytes_left = 0;
1299                                 break;
1300                         }
1301                         bytes_left = end - start;
1302                         continue;
1303                 }
1304
1305                 if (size) {
1306                         ret = blkdev_issue_discard(bdev, start >> SECTOR_SHIFT,
1307                                                    size >> SECTOR_SHIFT,
1308                                                    GFP_NOFS);
1309                         if (!ret)
1310                                 *discarded_bytes += size;
1311                         else if (ret != -EOPNOTSUPP)
1312                                 return ret;
1313                 }
1314
1315                 start = sb_end;
1316                 if (start > end) {
1317                         bytes_left = 0;
1318                         break;
1319                 }
1320                 bytes_left = end - start;
1321         }
1322
1323         while (bytes_left) {
1324                 u64 bytes_to_discard = min(BTRFS_MAX_DISCARD_CHUNK_SIZE, bytes_left);
1325
1326                 ret = blkdev_issue_discard(bdev, start >> SECTOR_SHIFT,
1327                                            bytes_to_discard >> SECTOR_SHIFT,
1328                                            GFP_NOFS);
1329
1330                 if (ret) {
1331                         if (ret != -EOPNOTSUPP)
1332                                 break;
1333                         continue;
1334                 }
1335
1336                 start += bytes_to_discard;
1337                 bytes_left -= bytes_to_discard;
1338                 *discarded_bytes += bytes_to_discard;
1339
1340                 if (btrfs_trim_interrupted()) {
1341                         ret = -ERESTARTSYS;
1342                         break;
1343                 }
1344         }
1345
1346         return ret;
1347 }
1348
1349 static int do_discard_extent(struct btrfs_discard_stripe *stripe, u64 *bytes)
1350 {
1351         struct btrfs_device *dev = stripe->dev;
1352         struct btrfs_fs_info *fs_info = dev->fs_info;
1353         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1354         u64 phys = stripe->physical;
1355         u64 len = stripe->length;
1356         u64 discarded = 0;
1357         int ret = 0;
1358
1359         /* Zone reset on a zoned filesystem */
1360         if (btrfs_can_zone_reset(dev, phys, len)) {
1361                 u64 src_disc;
1362
1363                 ret = btrfs_reset_device_zone(dev, phys, len, &discarded);
1364                 if (ret)
1365                         goto out;
1366
1367                 if (!btrfs_dev_replace_is_ongoing(dev_replace) ||
1368                     dev != dev_replace->srcdev)
1369                         goto out;
1370
1371                 src_disc = discarded;
1372
1373                 /* Send to replace target as well */
1374                 ret = btrfs_reset_device_zone(dev_replace->tgtdev, phys, len,
1375                                               &discarded);
1376                 discarded += src_disc;
1377         } else if (bdev_max_discard_sectors(stripe->dev->bdev)) {
1378                 ret = btrfs_issue_discard(dev->bdev, phys, len, &discarded);
1379         } else {
1380                 ret = 0;
1381                 *bytes = 0;
1382         }
1383
1384 out:
1385         *bytes = discarded;
1386         return ret;
1387 }
1388
1389 int btrfs_discard_extent(struct btrfs_fs_info *fs_info, u64 bytenr,
1390                          u64 num_bytes, u64 *actual_bytes)
1391 {
1392         int ret = 0;
1393         u64 discarded_bytes = 0;
1394         u64 end = bytenr + num_bytes;
1395         u64 cur = bytenr;
1396
1397         /*
1398          * Avoid races with device replace and make sure the devices in the
1399          * stripes don't go away while we are discarding.
1400          */
1401         btrfs_bio_counter_inc_blocked(fs_info);
1402         while (cur < end) {
1403                 struct btrfs_discard_stripe *stripes;
1404                 unsigned int num_stripes;
1405                 int i;
1406
1407                 num_bytes = end - cur;
1408                 stripes = btrfs_map_discard(fs_info, cur, &num_bytes, &num_stripes);
1409                 if (IS_ERR(stripes)) {
1410                         ret = PTR_ERR(stripes);
1411                         if (ret == -EOPNOTSUPP)
1412                                 ret = 0;
1413                         break;
1414                 }
1415
1416                 for (i = 0; i < num_stripes; i++) {
1417                         struct btrfs_discard_stripe *stripe = stripes + i;
1418                         u64 bytes;
1419
1420                         if (!stripe->dev->bdev) {
1421                                 ASSERT(btrfs_test_opt(fs_info, DEGRADED));
1422                                 continue;
1423                         }
1424
1425                         if (!test_bit(BTRFS_DEV_STATE_WRITEABLE,
1426                                         &stripe->dev->dev_state))
1427                                 continue;
1428
1429                         ret = do_discard_extent(stripe, &bytes);
1430                         if (ret) {
1431                                 /*
1432                                  * Keep going if discard is not supported by the
1433                                  * device.
1434                                  */
1435                                 if (ret != -EOPNOTSUPP)
1436                                         break;
1437                                 ret = 0;
1438                         } else {
1439                                 discarded_bytes += bytes;
1440                         }
1441                 }
1442                 kfree(stripes);
1443                 if (ret)
1444                         break;
1445                 cur += num_bytes;
1446         }
1447         btrfs_bio_counter_dec(fs_info);
1448         if (actual_bytes)
1449                 *actual_bytes = discarded_bytes;
1450         return ret;
1451 }
1452
1453 /* Can return -ENOMEM */
1454 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1455                          struct btrfs_ref *generic_ref)
1456 {
1457         struct btrfs_fs_info *fs_info = trans->fs_info;
1458         int ret;
1459
1460         ASSERT(generic_ref->type != BTRFS_REF_NOT_SET &&
1461                generic_ref->action);
1462         BUG_ON(generic_ref->type == BTRFS_REF_METADATA &&
1463                generic_ref->ref_root == BTRFS_TREE_LOG_OBJECTID);
1464
1465         if (generic_ref->type == BTRFS_REF_METADATA)
1466                 ret = btrfs_add_delayed_tree_ref(trans, generic_ref, NULL);
1467         else
1468                 ret = btrfs_add_delayed_data_ref(trans, generic_ref, 0);
1469
1470         btrfs_ref_tree_mod(fs_info, generic_ref);
1471
1472         return ret;
1473 }
1474
1475 /*
1476  * Insert backreference for a given extent.
1477  *
1478  * The counterpart is in __btrfs_free_extent(), with examples and more details
1479  * how it works.
1480  *
1481  * @trans:          Handle of transaction
1482  *
1483  * @node:           The delayed ref node used to get the bytenr/length for
1484  *                  extent whose references are incremented.
1485  *
1486  * @extent_op       Pointer to a structure, holding information necessary when
1487  *                  updating a tree block's flags
1488  *
1489  */
1490 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1491                                   struct btrfs_delayed_ref_node *node,
1492                                   struct btrfs_delayed_extent_op *extent_op)
1493 {
1494         struct btrfs_path *path;
1495         struct extent_buffer *leaf;
1496         struct btrfs_extent_item *item;
1497         struct btrfs_key key;
1498         u64 bytenr = node->bytenr;
1499         u64 num_bytes = node->num_bytes;
1500         u64 owner = btrfs_delayed_ref_owner(node);
1501         u64 offset = btrfs_delayed_ref_offset(node);
1502         u64 refs;
1503         int refs_to_add = node->ref_mod;
1504         int ret;
1505
1506         path = btrfs_alloc_path();
1507         if (!path)
1508                 return -ENOMEM;
1509
1510         /* this will setup the path even if it fails to insert the back ref */
1511         ret = insert_inline_extent_backref(trans, path, bytenr, num_bytes,
1512                                            node->parent, node->ref_root, owner,
1513                                            offset, refs_to_add, extent_op);
1514         if ((ret < 0 && ret != -EAGAIN) || !ret)
1515                 goto out;
1516
1517         /*
1518          * Ok we had -EAGAIN which means we didn't have space to insert and
1519          * inline extent ref, so just update the reference count and add a
1520          * normal backref.
1521          */
1522         leaf = path->nodes[0];
1523         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1524         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1525         refs = btrfs_extent_refs(leaf, item);
1526         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
1527         if (extent_op)
1528                 __run_delayed_extent_op(extent_op, leaf, item);
1529
1530         btrfs_mark_buffer_dirty(trans, leaf);
1531         btrfs_release_path(path);
1532
1533         /* now insert the actual backref */
1534         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1535                 ret = insert_tree_block_ref(trans, path, node, bytenr);
1536         else
1537                 ret = insert_extent_data_ref(trans, path, node, bytenr);
1538
1539         if (ret)
1540                 btrfs_abort_transaction(trans, ret);
1541 out:
1542         btrfs_free_path(path);
1543         return ret;
1544 }
1545
1546 static void free_head_ref_squota_rsv(struct btrfs_fs_info *fs_info,
1547                                      struct btrfs_delayed_ref_head *href)
1548 {
1549         u64 root = href->owning_root;
1550
1551         /*
1552          * Don't check must_insert_reserved, as this is called from contexts
1553          * where it has already been unset.
1554          */
1555         if (btrfs_qgroup_mode(fs_info) != BTRFS_QGROUP_MODE_SIMPLE ||
1556             !href->is_data || !is_fstree(root))
1557                 return;
1558
1559         btrfs_qgroup_free_refroot(fs_info, root, href->reserved_bytes,
1560                                   BTRFS_QGROUP_RSV_DATA);
1561 }
1562
1563 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
1564                                 struct btrfs_delayed_ref_head *href,
1565                                 struct btrfs_delayed_ref_node *node,
1566                                 struct btrfs_delayed_extent_op *extent_op,
1567                                 bool insert_reserved)
1568 {
1569         int ret = 0;
1570         u64 parent = 0;
1571         u64 flags = 0;
1572
1573         trace_run_delayed_data_ref(trans->fs_info, node);
1574
1575         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
1576                 parent = node->parent;
1577
1578         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1579                 struct btrfs_key key;
1580                 struct btrfs_squota_delta delta = {
1581                         .root = href->owning_root,
1582                         .num_bytes = node->num_bytes,
1583                         .is_data = true,
1584                         .is_inc = true,
1585                         .generation = trans->transid,
1586                 };
1587                 u64 owner = btrfs_delayed_ref_owner(node);
1588                 u64 offset = btrfs_delayed_ref_offset(node);
1589
1590                 if (extent_op)
1591                         flags |= extent_op->flags_to_set;
1592
1593                 key.objectid = node->bytenr;
1594                 key.type = BTRFS_EXTENT_ITEM_KEY;
1595                 key.offset = node->num_bytes;
1596
1597                 ret = alloc_reserved_file_extent(trans, parent, node->ref_root,
1598                                                  flags, owner, offset, &key,
1599                                                  node->ref_mod,
1600                                                  href->owning_root);
1601                 free_head_ref_squota_rsv(trans->fs_info, href);
1602                 if (!ret)
1603                         ret = btrfs_record_squota_delta(trans->fs_info, &delta);
1604         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1605                 ret = __btrfs_inc_extent_ref(trans, node, extent_op);
1606         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1607                 ret = __btrfs_free_extent(trans, href, node, extent_op);
1608         } else {
1609                 BUG();
1610         }
1611         return ret;
1612 }
1613
1614 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
1615                                     struct extent_buffer *leaf,
1616                                     struct btrfs_extent_item *ei)
1617 {
1618         u64 flags = btrfs_extent_flags(leaf, ei);
1619         if (extent_op->update_flags) {
1620                 flags |= extent_op->flags_to_set;
1621                 btrfs_set_extent_flags(leaf, ei, flags);
1622         }
1623
1624         if (extent_op->update_key) {
1625                 struct btrfs_tree_block_info *bi;
1626                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
1627                 bi = (struct btrfs_tree_block_info *)(ei + 1);
1628                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
1629         }
1630 }
1631
1632 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
1633                                  struct btrfs_delayed_ref_head *head,
1634                                  struct btrfs_delayed_extent_op *extent_op)
1635 {
1636         struct btrfs_fs_info *fs_info = trans->fs_info;
1637         struct btrfs_root *root;
1638         struct btrfs_key key;
1639         struct btrfs_path *path;
1640         struct btrfs_extent_item *ei;
1641         struct extent_buffer *leaf;
1642         u32 item_size;
1643         int ret;
1644         int metadata = 1;
1645
1646         if (TRANS_ABORTED(trans))
1647                 return 0;
1648
1649         if (!btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1650                 metadata = 0;
1651
1652         path = btrfs_alloc_path();
1653         if (!path)
1654                 return -ENOMEM;
1655
1656         key.objectid = head->bytenr;
1657
1658         if (metadata) {
1659                 key.type = BTRFS_METADATA_ITEM_KEY;
1660                 key.offset = head->level;
1661         } else {
1662                 key.type = BTRFS_EXTENT_ITEM_KEY;
1663                 key.offset = head->num_bytes;
1664         }
1665
1666         root = btrfs_extent_root(fs_info, key.objectid);
1667 again:
1668         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
1669         if (ret < 0) {
1670                 goto out;
1671         } else if (ret > 0) {
1672                 if (metadata) {
1673                         if (path->slots[0] > 0) {
1674                                 path->slots[0]--;
1675                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
1676                                                       path->slots[0]);
1677                                 if (key.objectid == head->bytenr &&
1678                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
1679                                     key.offset == head->num_bytes)
1680                                         ret = 0;
1681                         }
1682                         if (ret > 0) {
1683                                 btrfs_release_path(path);
1684                                 metadata = 0;
1685
1686                                 key.objectid = head->bytenr;
1687                                 key.offset = head->num_bytes;
1688                                 key.type = BTRFS_EXTENT_ITEM_KEY;
1689                                 goto again;
1690                         }
1691                 } else {
1692                         ret = -EUCLEAN;
1693                         btrfs_err(fs_info,
1694                   "missing extent item for extent %llu num_bytes %llu level %d",
1695                                   head->bytenr, head->num_bytes, head->level);
1696                         goto out;
1697                 }
1698         }
1699
1700         leaf = path->nodes[0];
1701         item_size = btrfs_item_size(leaf, path->slots[0]);
1702
1703         if (unlikely(item_size < sizeof(*ei))) {
1704                 ret = -EUCLEAN;
1705                 btrfs_err(fs_info,
1706                           "unexpected extent item size, has %u expect >= %zu",
1707                           item_size, sizeof(*ei));
1708                 btrfs_abort_transaction(trans, ret);
1709                 goto out;
1710         }
1711
1712         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1713         __run_delayed_extent_op(extent_op, leaf, ei);
1714
1715         btrfs_mark_buffer_dirty(trans, leaf);
1716 out:
1717         btrfs_free_path(path);
1718         return ret;
1719 }
1720
1721 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
1722                                 struct btrfs_delayed_ref_head *href,
1723                                 struct btrfs_delayed_ref_node *node,
1724                                 struct btrfs_delayed_extent_op *extent_op,
1725                                 bool insert_reserved)
1726 {
1727         int ret = 0;
1728         struct btrfs_fs_info *fs_info = trans->fs_info;
1729         u64 parent = 0;
1730         u64 ref_root = 0;
1731
1732         trace_run_delayed_tree_ref(trans->fs_info, node);
1733
1734         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
1735                 parent = node->parent;
1736         ref_root = node->ref_root;
1737
1738         if (unlikely(node->ref_mod != 1)) {
1739                 btrfs_err(trans->fs_info,
1740         "btree block %llu has %d references rather than 1: action %d ref_root %llu parent %llu",
1741                           node->bytenr, node->ref_mod, node->action, ref_root,
1742                           parent);
1743                 return -EUCLEAN;
1744         }
1745         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1746                 struct btrfs_squota_delta delta = {
1747                         .root = href->owning_root,
1748                         .num_bytes = fs_info->nodesize,
1749                         .is_data = false,
1750                         .is_inc = true,
1751                         .generation = trans->transid,
1752                 };
1753
1754                 ret = alloc_reserved_tree_block(trans, node, extent_op);
1755                 if (!ret)
1756                         btrfs_record_squota_delta(fs_info, &delta);
1757         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1758                 ret = __btrfs_inc_extent_ref(trans, node, extent_op);
1759         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1760                 ret = __btrfs_free_extent(trans, href, node, extent_op);
1761         } else {
1762                 BUG();
1763         }
1764         return ret;
1765 }
1766
1767 /* helper function to actually process a single delayed ref entry */
1768 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
1769                                struct btrfs_delayed_ref_head *href,
1770                                struct btrfs_delayed_ref_node *node,
1771                                struct btrfs_delayed_extent_op *extent_op,
1772                                bool insert_reserved)
1773 {
1774         int ret = 0;
1775
1776         if (TRANS_ABORTED(trans)) {
1777                 if (insert_reserved) {
1778                         btrfs_pin_extent(trans, node->bytenr, node->num_bytes, 1);
1779                         free_head_ref_squota_rsv(trans->fs_info, href);
1780                 }
1781                 return 0;
1782         }
1783
1784         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
1785             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
1786                 ret = run_delayed_tree_ref(trans, href, node, extent_op,
1787                                            insert_reserved);
1788         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
1789                  node->type == BTRFS_SHARED_DATA_REF_KEY)
1790                 ret = run_delayed_data_ref(trans, href, node, extent_op,
1791                                            insert_reserved);
1792         else if (node->type == BTRFS_EXTENT_OWNER_REF_KEY)
1793                 ret = 0;
1794         else
1795                 BUG();
1796         if (ret && insert_reserved)
1797                 btrfs_pin_extent(trans, node->bytenr, node->num_bytes, 1);
1798         if (ret < 0)
1799                 btrfs_err(trans->fs_info,
1800 "failed to run delayed ref for logical %llu num_bytes %llu type %u action %u ref_mod %d: %d",
1801                           node->bytenr, node->num_bytes, node->type,
1802                           node->action, node->ref_mod, ret);
1803         return ret;
1804 }
1805
1806 static inline struct btrfs_delayed_ref_node *
1807 select_delayed_ref(struct btrfs_delayed_ref_head *head)
1808 {
1809         struct btrfs_delayed_ref_node *ref;
1810
1811         if (RB_EMPTY_ROOT(&head->ref_tree.rb_root))
1812                 return NULL;
1813
1814         /*
1815          * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
1816          * This is to prevent a ref count from going down to zero, which deletes
1817          * the extent item from the extent tree, when there still are references
1818          * to add, which would fail because they would not find the extent item.
1819          */
1820         if (!list_empty(&head->ref_add_list))
1821                 return list_first_entry(&head->ref_add_list,
1822                                 struct btrfs_delayed_ref_node, add_list);
1823
1824         ref = rb_entry(rb_first_cached(&head->ref_tree),
1825                        struct btrfs_delayed_ref_node, ref_node);
1826         ASSERT(list_empty(&ref->add_list));
1827         return ref;
1828 }
1829
1830 static struct btrfs_delayed_extent_op *cleanup_extent_op(
1831                                 struct btrfs_delayed_ref_head *head)
1832 {
1833         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
1834
1835         if (!extent_op)
1836                 return NULL;
1837
1838         if (head->must_insert_reserved) {
1839                 head->extent_op = NULL;
1840                 btrfs_free_delayed_extent_op(extent_op);
1841                 return NULL;
1842         }
1843         return extent_op;
1844 }
1845
1846 static int run_and_cleanup_extent_op(struct btrfs_trans_handle *trans,
1847                                      struct btrfs_delayed_ref_head *head)
1848 {
1849         struct btrfs_delayed_extent_op *extent_op;
1850         int ret;
1851
1852         extent_op = cleanup_extent_op(head);
1853         if (!extent_op)
1854                 return 0;
1855         head->extent_op = NULL;
1856         spin_unlock(&head->lock);
1857         ret = run_delayed_extent_op(trans, head, extent_op);
1858         btrfs_free_delayed_extent_op(extent_op);
1859         return ret ? ret : 1;
1860 }
1861
1862 u64 btrfs_cleanup_ref_head_accounting(struct btrfs_fs_info *fs_info,
1863                                   struct btrfs_delayed_ref_root *delayed_refs,
1864                                   struct btrfs_delayed_ref_head *head)
1865 {
1866         u64 ret = 0;
1867
1868         /*
1869          * We had csum deletions accounted for in our delayed refs rsv, we need
1870          * to drop the csum leaves for this update from our delayed_refs_rsv.
1871          */
1872         if (head->total_ref_mod < 0 && head->is_data) {
1873                 int nr_csums;
1874
1875                 spin_lock(&delayed_refs->lock);
1876                 delayed_refs->pending_csums -= head->num_bytes;
1877                 spin_unlock(&delayed_refs->lock);
1878                 nr_csums = btrfs_csum_bytes_to_leaves(fs_info, head->num_bytes);
1879
1880                 btrfs_delayed_refs_rsv_release(fs_info, 0, nr_csums);
1881
1882                 ret = btrfs_calc_delayed_ref_csum_bytes(fs_info, nr_csums);
1883         }
1884         /* must_insert_reserved can be set only if we didn't run the head ref. */
1885         if (head->must_insert_reserved)
1886                 free_head_ref_squota_rsv(fs_info, head);
1887
1888         return ret;
1889 }
1890
1891 static int cleanup_ref_head(struct btrfs_trans_handle *trans,
1892                             struct btrfs_delayed_ref_head *head,
1893                             u64 *bytes_released)
1894 {
1895
1896         struct btrfs_fs_info *fs_info = trans->fs_info;
1897         struct btrfs_delayed_ref_root *delayed_refs;
1898         int ret;
1899
1900         delayed_refs = &trans->transaction->delayed_refs;
1901
1902         ret = run_and_cleanup_extent_op(trans, head);
1903         if (ret < 0) {
1904                 btrfs_unselect_ref_head(delayed_refs, head);
1905                 btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
1906                 return ret;
1907         } else if (ret) {
1908                 return ret;
1909         }
1910
1911         /*
1912          * Need to drop our head ref lock and re-acquire the delayed ref lock
1913          * and then re-check to make sure nobody got added.
1914          */
1915         spin_unlock(&head->lock);
1916         spin_lock(&delayed_refs->lock);
1917         spin_lock(&head->lock);
1918         if (!RB_EMPTY_ROOT(&head->ref_tree.rb_root) || head->extent_op) {
1919                 spin_unlock(&head->lock);
1920                 spin_unlock(&delayed_refs->lock);
1921                 return 1;
1922         }
1923         btrfs_delete_ref_head(fs_info, delayed_refs, head);
1924         spin_unlock(&head->lock);
1925         spin_unlock(&delayed_refs->lock);
1926
1927         if (head->must_insert_reserved) {
1928                 btrfs_pin_extent(trans, head->bytenr, head->num_bytes, 1);
1929                 if (head->is_data) {
1930                         struct btrfs_root *csum_root;
1931
1932                         csum_root = btrfs_csum_root(fs_info, head->bytenr);
1933                         ret = btrfs_del_csums(trans, csum_root, head->bytenr,
1934                                               head->num_bytes);
1935                 }
1936         }
1937
1938         *bytes_released += btrfs_cleanup_ref_head_accounting(fs_info, delayed_refs, head);
1939
1940         trace_run_delayed_ref_head(fs_info, head, 0);
1941         btrfs_delayed_ref_unlock(head);
1942         btrfs_put_delayed_ref_head(head);
1943         return ret;
1944 }
1945
1946 static int btrfs_run_delayed_refs_for_head(struct btrfs_trans_handle *trans,
1947                                            struct btrfs_delayed_ref_head *locked_ref,
1948                                            u64 *bytes_released)
1949 {
1950         struct btrfs_fs_info *fs_info = trans->fs_info;
1951         struct btrfs_delayed_ref_root *delayed_refs;
1952         struct btrfs_delayed_extent_op *extent_op;
1953         struct btrfs_delayed_ref_node *ref;
1954         bool must_insert_reserved;
1955         int ret;
1956
1957         delayed_refs = &trans->transaction->delayed_refs;
1958
1959         lockdep_assert_held(&locked_ref->mutex);
1960         lockdep_assert_held(&locked_ref->lock);
1961
1962         while ((ref = select_delayed_ref(locked_ref))) {
1963                 if (ref->seq &&
1964                     btrfs_check_delayed_seq(fs_info, ref->seq)) {
1965                         spin_unlock(&locked_ref->lock);
1966                         btrfs_unselect_ref_head(delayed_refs, locked_ref);
1967                         return -EAGAIN;
1968                 }
1969
1970                 rb_erase_cached(&ref->ref_node, &locked_ref->ref_tree);
1971                 RB_CLEAR_NODE(&ref->ref_node);
1972                 if (!list_empty(&ref->add_list))
1973                         list_del(&ref->add_list);
1974                 /*
1975                  * When we play the delayed ref, also correct the ref_mod on
1976                  * head
1977                  */
1978                 switch (ref->action) {
1979                 case BTRFS_ADD_DELAYED_REF:
1980                 case BTRFS_ADD_DELAYED_EXTENT:
1981                         locked_ref->ref_mod -= ref->ref_mod;
1982                         break;
1983                 case BTRFS_DROP_DELAYED_REF:
1984                         locked_ref->ref_mod += ref->ref_mod;
1985                         break;
1986                 default:
1987                         WARN_ON(1);
1988                 }
1989
1990                 /*
1991                  * Record the must_insert_reserved flag before we drop the
1992                  * spin lock.
1993                  */
1994                 must_insert_reserved = locked_ref->must_insert_reserved;
1995                 /*
1996                  * Unsetting this on the head ref relinquishes ownership of
1997                  * the rsv_bytes, so it is critical that every possible code
1998                  * path from here forward frees all reserves including qgroup
1999                  * reserve.
2000                  */
2001                 locked_ref->must_insert_reserved = false;
2002
2003                 extent_op = locked_ref->extent_op;
2004                 locked_ref->extent_op = NULL;
2005                 spin_unlock(&locked_ref->lock);
2006
2007                 ret = run_one_delayed_ref(trans, locked_ref, ref, extent_op,
2008                                           must_insert_reserved);
2009                 btrfs_delayed_refs_rsv_release(fs_info, 1, 0);
2010                 *bytes_released += btrfs_calc_delayed_ref_bytes(fs_info, 1);
2011
2012                 btrfs_free_delayed_extent_op(extent_op);
2013                 if (ret) {
2014                         btrfs_unselect_ref_head(delayed_refs, locked_ref);
2015                         btrfs_put_delayed_ref(ref);
2016                         return ret;
2017                 }
2018
2019                 btrfs_put_delayed_ref(ref);
2020                 cond_resched();
2021
2022                 spin_lock(&locked_ref->lock);
2023                 btrfs_merge_delayed_refs(fs_info, delayed_refs, locked_ref);
2024         }
2025
2026         return 0;
2027 }
2028
2029 /*
2030  * Returns 0 on success or if called with an already aborted transaction.
2031  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2032  */
2033 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2034                                              u64 min_bytes)
2035 {
2036         struct btrfs_fs_info *fs_info = trans->fs_info;
2037         struct btrfs_delayed_ref_root *delayed_refs;
2038         struct btrfs_delayed_ref_head *locked_ref = NULL;
2039         int ret;
2040         unsigned long count = 0;
2041         unsigned long max_count = 0;
2042         u64 bytes_processed = 0;
2043
2044         delayed_refs = &trans->transaction->delayed_refs;
2045         if (min_bytes == 0) {
2046                 max_count = delayed_refs->num_heads_ready;
2047                 min_bytes = U64_MAX;
2048         }
2049
2050         do {
2051                 if (!locked_ref) {
2052                         locked_ref = btrfs_select_ref_head(fs_info, delayed_refs);
2053                         if (IS_ERR_OR_NULL(locked_ref)) {
2054                                 if (PTR_ERR(locked_ref) == -EAGAIN) {
2055                                         continue;
2056                                 } else {
2057                                         break;
2058                                 }
2059                         }
2060                         count++;
2061                 }
2062                 /*
2063                  * We need to try and merge add/drops of the same ref since we
2064                  * can run into issues with relocate dropping the implicit ref
2065                  * and then it being added back again before the drop can
2066                  * finish.  If we merged anything we need to re-loop so we can
2067                  * get a good ref.
2068                  * Or we can get node references of the same type that weren't
2069                  * merged when created due to bumps in the tree mod seq, and
2070                  * we need to merge them to prevent adding an inline extent
2071                  * backref before dropping it (triggering a BUG_ON at
2072                  * insert_inline_extent_backref()).
2073                  */
2074                 spin_lock(&locked_ref->lock);
2075                 btrfs_merge_delayed_refs(fs_info, delayed_refs, locked_ref);
2076
2077                 ret = btrfs_run_delayed_refs_for_head(trans, locked_ref, &bytes_processed);
2078                 if (ret < 0 && ret != -EAGAIN) {
2079                         /*
2080                          * Error, btrfs_run_delayed_refs_for_head already
2081                          * unlocked everything so just bail out
2082                          */
2083                         return ret;
2084                 } else if (!ret) {
2085                         /*
2086                          * Success, perform the usual cleanup of a processed
2087                          * head
2088                          */
2089                         ret = cleanup_ref_head(trans, locked_ref, &bytes_processed);
2090                         if (ret > 0 ) {
2091                                 /* We dropped our lock, we need to loop. */
2092                                 ret = 0;
2093                                 continue;
2094                         } else if (ret) {
2095                                 return ret;
2096                         }
2097                 }
2098
2099                 /*
2100                  * Either success case or btrfs_run_delayed_refs_for_head
2101                  * returned -EAGAIN, meaning we need to select another head
2102                  */
2103
2104                 locked_ref = NULL;
2105                 cond_resched();
2106         } while ((min_bytes != U64_MAX && bytes_processed < min_bytes) ||
2107                  (max_count > 0 && count < max_count) ||
2108                  locked_ref);
2109
2110         return 0;
2111 }
2112
2113 #ifdef SCRAMBLE_DELAYED_REFS
2114 /*
2115  * Normally delayed refs get processed in ascending bytenr order. This
2116  * correlates in most cases to the order added. To expose dependencies on this
2117  * order, we start to process the tree in the middle instead of the beginning
2118  */
2119 static u64 find_middle(struct rb_root *root)
2120 {
2121         struct rb_node *n = root->rb_node;
2122         struct btrfs_delayed_ref_node *entry;
2123         int alt = 1;
2124         u64 middle;
2125         u64 first = 0, last = 0;
2126
2127         n = rb_first(root);
2128         if (n) {
2129                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2130                 first = entry->bytenr;
2131         }
2132         n = rb_last(root);
2133         if (n) {
2134                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2135                 last = entry->bytenr;
2136         }
2137         n = root->rb_node;
2138
2139         while (n) {
2140                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2141                 WARN_ON(!entry->in_tree);
2142
2143                 middle = entry->bytenr;
2144
2145                 if (alt)
2146                         n = n->rb_left;
2147                 else
2148                         n = n->rb_right;
2149
2150                 alt = 1 - alt;
2151         }
2152         return middle;
2153 }
2154 #endif
2155
2156 /*
2157  * Start processing the delayed reference count updates and extent insertions
2158  * we have queued up so far.
2159  *
2160  * @trans:      Transaction handle.
2161  * @min_bytes:  How many bytes of delayed references to process. After this
2162  *              many bytes we stop processing delayed references if there are
2163  *              any more. If 0 it means to run all existing delayed references,
2164  *              but not new ones added after running all existing ones.
2165  *              Use (u64)-1 (U64_MAX) to run all existing delayed references
2166  *              plus any new ones that are added.
2167  *
2168  * Returns 0 on success or if called with an aborted transaction
2169  * Returns <0 on error and aborts the transaction
2170  */
2171 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans, u64 min_bytes)
2172 {
2173         struct btrfs_fs_info *fs_info = trans->fs_info;
2174         struct btrfs_delayed_ref_root *delayed_refs;
2175         int ret;
2176
2177         /* We'll clean this up in btrfs_cleanup_transaction */
2178         if (TRANS_ABORTED(trans))
2179                 return 0;
2180
2181         if (test_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags))
2182                 return 0;
2183
2184         delayed_refs = &trans->transaction->delayed_refs;
2185 again:
2186 #ifdef SCRAMBLE_DELAYED_REFS
2187         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2188 #endif
2189         ret = __btrfs_run_delayed_refs(trans, min_bytes);
2190         if (ret < 0) {
2191                 btrfs_abort_transaction(trans, ret);
2192                 return ret;
2193         }
2194
2195         if (min_bytes == U64_MAX) {
2196                 btrfs_create_pending_block_groups(trans);
2197
2198                 spin_lock(&delayed_refs->lock);
2199                 if (xa_empty(&delayed_refs->head_refs)) {
2200                         spin_unlock(&delayed_refs->lock);
2201                         return 0;
2202                 }
2203                 spin_unlock(&delayed_refs->lock);
2204
2205                 cond_resched();
2206                 goto again;
2207         }
2208
2209         return 0;
2210 }
2211
2212 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2213                                 struct extent_buffer *eb, u64 flags)
2214 {
2215         struct btrfs_delayed_extent_op *extent_op;
2216         int ret;
2217
2218         extent_op = btrfs_alloc_delayed_extent_op();
2219         if (!extent_op)
2220                 return -ENOMEM;
2221
2222         extent_op->flags_to_set = flags;
2223         extent_op->update_flags = true;
2224         extent_op->update_key = false;
2225
2226         ret = btrfs_add_delayed_extent_op(trans, eb->start, eb->len,
2227                                           btrfs_header_level(eb), extent_op);
2228         if (ret)
2229                 btrfs_free_delayed_extent_op(extent_op);
2230         return ret;
2231 }
2232
2233 static noinline int check_delayed_ref(struct btrfs_root *root,
2234                                       struct btrfs_path *path,
2235                                       u64 objectid, u64 offset, u64 bytenr)
2236 {
2237         struct btrfs_delayed_ref_head *head;
2238         struct btrfs_delayed_ref_node *ref;
2239         struct btrfs_delayed_ref_root *delayed_refs;
2240         struct btrfs_transaction *cur_trans;
2241         struct rb_node *node;
2242         int ret = 0;
2243
2244         spin_lock(&root->fs_info->trans_lock);
2245         cur_trans = root->fs_info->running_transaction;
2246         if (cur_trans)
2247                 refcount_inc(&cur_trans->use_count);
2248         spin_unlock(&root->fs_info->trans_lock);
2249         if (!cur_trans)
2250                 return 0;
2251
2252         delayed_refs = &cur_trans->delayed_refs;
2253         spin_lock(&delayed_refs->lock);
2254         head = btrfs_find_delayed_ref_head(root->fs_info, delayed_refs, bytenr);
2255         if (!head) {
2256                 spin_unlock(&delayed_refs->lock);
2257                 btrfs_put_transaction(cur_trans);
2258                 return 0;
2259         }
2260
2261         if (!mutex_trylock(&head->mutex)) {
2262                 if (path->nowait) {
2263                         spin_unlock(&delayed_refs->lock);
2264                         btrfs_put_transaction(cur_trans);
2265                         return -EAGAIN;
2266                 }
2267
2268                 refcount_inc(&head->refs);
2269                 spin_unlock(&delayed_refs->lock);
2270
2271                 btrfs_release_path(path);
2272
2273                 /*
2274                  * Mutex was contended, block until it's released and let
2275                  * caller try again
2276                  */
2277                 mutex_lock(&head->mutex);
2278                 mutex_unlock(&head->mutex);
2279                 btrfs_put_delayed_ref_head(head);
2280                 btrfs_put_transaction(cur_trans);
2281                 return -EAGAIN;
2282         }
2283         spin_unlock(&delayed_refs->lock);
2284
2285         spin_lock(&head->lock);
2286         /*
2287          * XXX: We should replace this with a proper search function in the
2288          * future.
2289          */
2290         for (node = rb_first_cached(&head->ref_tree); node;
2291              node = rb_next(node)) {
2292                 u64 ref_owner;
2293                 u64 ref_offset;
2294
2295                 ref = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
2296                 /* If it's a shared ref we know a cross reference exists */
2297                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
2298                         ret = 1;
2299                         break;
2300                 }
2301
2302                 ref_owner = btrfs_delayed_ref_owner(ref);
2303                 ref_offset = btrfs_delayed_ref_offset(ref);
2304
2305                 /*
2306                  * If our ref doesn't match the one we're currently looking at
2307                  * then we have a cross reference.
2308                  */
2309                 if (ref->ref_root != btrfs_root_id(root) ||
2310                     ref_owner != objectid || ref_offset != offset) {
2311                         ret = 1;
2312                         break;
2313                 }
2314         }
2315         spin_unlock(&head->lock);
2316         mutex_unlock(&head->mutex);
2317         btrfs_put_transaction(cur_trans);
2318         return ret;
2319 }
2320
2321 static noinline int check_committed_ref(struct btrfs_root *root,
2322                                         struct btrfs_path *path,
2323                                         u64 objectid, u64 offset, u64 bytenr,
2324                                         bool strict)
2325 {
2326         struct btrfs_fs_info *fs_info = root->fs_info;
2327         struct btrfs_root *extent_root = btrfs_extent_root(fs_info, bytenr);
2328         struct extent_buffer *leaf;
2329         struct btrfs_extent_data_ref *ref;
2330         struct btrfs_extent_inline_ref *iref;
2331         struct btrfs_extent_item *ei;
2332         struct btrfs_key key;
2333         u32 item_size;
2334         u32 expected_size;
2335         int type;
2336         int ret;
2337
2338         key.objectid = bytenr;
2339         key.offset = (u64)-1;
2340         key.type = BTRFS_EXTENT_ITEM_KEY;
2341
2342         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2343         if (ret < 0)
2344                 goto out;
2345         if (ret == 0) {
2346                 /*
2347                  * Key with offset -1 found, there would have to exist an extent
2348                  * item with such offset, but this is out of the valid range.
2349                  */
2350                 ret = -EUCLEAN;
2351                 goto out;
2352         }
2353
2354         ret = -ENOENT;
2355         if (path->slots[0] == 0)
2356                 goto out;
2357
2358         path->slots[0]--;
2359         leaf = path->nodes[0];
2360         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2361
2362         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2363                 goto out;
2364
2365         ret = 1;
2366         item_size = btrfs_item_size(leaf, path->slots[0]);
2367         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2368         expected_size = sizeof(*ei) + btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY);
2369
2370         /* No inline refs; we need to bail before checking for owner ref. */
2371         if (item_size == sizeof(*ei))
2372                 goto out;
2373
2374         /* Check for an owner ref; skip over it to the real inline refs. */
2375         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2376         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
2377         if (btrfs_fs_incompat(fs_info, SIMPLE_QUOTA) && type == BTRFS_EXTENT_OWNER_REF_KEY) {
2378                 expected_size += btrfs_extent_inline_ref_size(BTRFS_EXTENT_OWNER_REF_KEY);
2379                 iref = (struct btrfs_extent_inline_ref *)(iref + 1);
2380         }
2381
2382         /* If extent item has more than 1 inline ref then it's shared */
2383         if (item_size != expected_size)
2384                 goto out;
2385
2386         /*
2387          * If extent created before last snapshot => it's shared unless the
2388          * snapshot has been deleted. Use the heuristic if strict is false.
2389          */
2390         if (!strict &&
2391             (btrfs_extent_generation(leaf, ei) <=
2392              btrfs_root_last_snapshot(&root->root_item)))
2393                 goto out;
2394
2395         /* If this extent has SHARED_DATA_REF then it's shared */
2396         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
2397         if (type != BTRFS_EXTENT_DATA_REF_KEY)
2398                 goto out;
2399
2400         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2401         if (btrfs_extent_refs(leaf, ei) !=
2402             btrfs_extent_data_ref_count(leaf, ref) ||
2403             btrfs_extent_data_ref_root(leaf, ref) != btrfs_root_id(root) ||
2404             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2405             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2406                 goto out;
2407
2408         ret = 0;
2409 out:
2410         return ret;
2411 }
2412
2413 int btrfs_cross_ref_exist(struct btrfs_root *root, u64 objectid, u64 offset,
2414                           u64 bytenr, bool strict, struct btrfs_path *path)
2415 {
2416         int ret;
2417
2418         do {
2419                 ret = check_committed_ref(root, path, objectid,
2420                                           offset, bytenr, strict);
2421                 if (ret && ret != -ENOENT)
2422                         goto out;
2423
2424                 ret = check_delayed_ref(root, path, objectid, offset, bytenr);
2425         } while (ret == -EAGAIN && !path->nowait);
2426
2427 out:
2428         btrfs_release_path(path);
2429         if (btrfs_is_data_reloc_root(root))
2430                 WARN_ON(ret > 0);
2431         return ret;
2432 }
2433
2434 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2435                            struct btrfs_root *root,
2436                            struct extent_buffer *buf,
2437                            int full_backref, int inc)
2438 {
2439         struct btrfs_fs_info *fs_info = root->fs_info;
2440         u64 parent;
2441         u64 ref_root;
2442         u32 nritems;
2443         struct btrfs_key key;
2444         struct btrfs_file_extent_item *fi;
2445         bool for_reloc = btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC);
2446         int i;
2447         int action;
2448         int level;
2449         int ret = 0;
2450
2451         if (btrfs_is_testing(fs_info))
2452                 return 0;
2453
2454         ref_root = btrfs_header_owner(buf);
2455         nritems = btrfs_header_nritems(buf);
2456         level = btrfs_header_level(buf);
2457
2458         if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state) && level == 0)
2459                 return 0;
2460
2461         if (full_backref)
2462                 parent = buf->start;
2463         else
2464                 parent = 0;
2465         if (inc)
2466                 action = BTRFS_ADD_DELAYED_REF;
2467         else
2468                 action = BTRFS_DROP_DELAYED_REF;
2469
2470         for (i = 0; i < nritems; i++) {
2471                 struct btrfs_ref ref = {
2472                         .action = action,
2473                         .parent = parent,
2474                         .ref_root = ref_root,
2475                 };
2476
2477                 if (level == 0) {
2478                         btrfs_item_key_to_cpu(buf, &key, i);
2479                         if (key.type != BTRFS_EXTENT_DATA_KEY)
2480                                 continue;
2481                         fi = btrfs_item_ptr(buf, i,
2482                                             struct btrfs_file_extent_item);
2483                         if (btrfs_file_extent_type(buf, fi) ==
2484                             BTRFS_FILE_EXTENT_INLINE)
2485                                 continue;
2486                         ref.bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2487                         if (ref.bytenr == 0)
2488                                 continue;
2489
2490                         ref.num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
2491                         ref.owning_root = ref_root;
2492
2493                         key.offset -= btrfs_file_extent_offset(buf, fi);
2494                         btrfs_init_data_ref(&ref, key.objectid, key.offset,
2495                                             btrfs_root_id(root), for_reloc);
2496                         if (inc)
2497                                 ret = btrfs_inc_extent_ref(trans, &ref);
2498                         else
2499                                 ret = btrfs_free_extent(trans, &ref);
2500                         if (ret)
2501                                 goto fail;
2502                 } else {
2503                         /* We don't know the owning_root, leave as 0. */
2504                         ref.bytenr = btrfs_node_blockptr(buf, i);
2505                         ref.num_bytes = fs_info->nodesize;
2506
2507                         btrfs_init_tree_ref(&ref, level - 1,
2508                                             btrfs_root_id(root), for_reloc);
2509                         if (inc)
2510                                 ret = btrfs_inc_extent_ref(trans, &ref);
2511                         else
2512                                 ret = btrfs_free_extent(trans, &ref);
2513                         if (ret)
2514                                 goto fail;
2515                 }
2516         }
2517         return 0;
2518 fail:
2519         return ret;
2520 }
2521
2522 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2523                   struct extent_buffer *buf, int full_backref)
2524 {
2525         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
2526 }
2527
2528 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2529                   struct extent_buffer *buf, int full_backref)
2530 {
2531         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
2532 }
2533
2534 static u64 get_alloc_profile_by_root(struct btrfs_root *root, int data)
2535 {
2536         struct btrfs_fs_info *fs_info = root->fs_info;
2537         u64 flags;
2538         u64 ret;
2539
2540         if (data)
2541                 flags = BTRFS_BLOCK_GROUP_DATA;
2542         else if (root == fs_info->chunk_root)
2543                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
2544         else
2545                 flags = BTRFS_BLOCK_GROUP_METADATA;
2546
2547         ret = btrfs_get_alloc_profile(fs_info, flags);
2548         return ret;
2549 }
2550
2551 static u64 first_logical_byte(struct btrfs_fs_info *fs_info)
2552 {
2553         struct rb_node *leftmost;
2554         u64 bytenr = 0;
2555
2556         read_lock(&fs_info->block_group_cache_lock);
2557         /* Get the block group with the lowest logical start address. */
2558         leftmost = rb_first_cached(&fs_info->block_group_cache_tree);
2559         if (leftmost) {
2560                 struct btrfs_block_group *bg;
2561
2562                 bg = rb_entry(leftmost, struct btrfs_block_group, cache_node);
2563                 bytenr = bg->start;
2564         }
2565         read_unlock(&fs_info->block_group_cache_lock);
2566
2567         return bytenr;
2568 }
2569
2570 static int pin_down_extent(struct btrfs_trans_handle *trans,
2571                            struct btrfs_block_group *cache,
2572                            u64 bytenr, u64 num_bytes, int reserved)
2573 {
2574         struct btrfs_fs_info *fs_info = cache->fs_info;
2575
2576         spin_lock(&cache->space_info->lock);
2577         spin_lock(&cache->lock);
2578         cache->pinned += num_bytes;
2579         btrfs_space_info_update_bytes_pinned(fs_info, cache->space_info,
2580                                              num_bytes);
2581         if (reserved) {
2582                 cache->reserved -= num_bytes;
2583                 cache->space_info->bytes_reserved -= num_bytes;
2584         }
2585         spin_unlock(&cache->lock);
2586         spin_unlock(&cache->space_info->lock);
2587
2588         set_extent_bit(&trans->transaction->pinned_extents, bytenr,
2589                        bytenr + num_bytes - 1, EXTENT_DIRTY, NULL);
2590         return 0;
2591 }
2592
2593 int btrfs_pin_extent(struct btrfs_trans_handle *trans,
2594                      u64 bytenr, u64 num_bytes, int reserved)
2595 {
2596         struct btrfs_block_group *cache;
2597
2598         cache = btrfs_lookup_block_group(trans->fs_info, bytenr);
2599         BUG_ON(!cache); /* Logic error */
2600
2601         pin_down_extent(trans, cache, bytenr, num_bytes, reserved);
2602
2603         btrfs_put_block_group(cache);
2604         return 0;
2605 }
2606
2607 int btrfs_pin_extent_for_log_replay(struct btrfs_trans_handle *trans,
2608                                     const struct extent_buffer *eb)
2609 {
2610         struct btrfs_block_group *cache;
2611         int ret;
2612
2613         cache = btrfs_lookup_block_group(trans->fs_info, eb->start);
2614         if (!cache)
2615                 return -EINVAL;
2616
2617         /*
2618          * Fully cache the free space first so that our pin removes the free space
2619          * from the cache.
2620          */
2621         ret = btrfs_cache_block_group(cache, true);
2622         if (ret)
2623                 goto out;
2624
2625         pin_down_extent(trans, cache, eb->start, eb->len, 0);
2626
2627         /* remove us from the free space cache (if we're there at all) */
2628         ret = btrfs_remove_free_space(cache, eb->start, eb->len);
2629 out:
2630         btrfs_put_block_group(cache);
2631         return ret;
2632 }
2633
2634 static int __exclude_logged_extent(struct btrfs_fs_info *fs_info,
2635                                    u64 start, u64 num_bytes)
2636 {
2637         int ret;
2638         struct btrfs_block_group *block_group;
2639
2640         block_group = btrfs_lookup_block_group(fs_info, start);
2641         if (!block_group)
2642                 return -EINVAL;
2643
2644         ret = btrfs_cache_block_group(block_group, true);
2645         if (ret)
2646                 goto out;
2647
2648         ret = btrfs_remove_free_space(block_group, start, num_bytes);
2649 out:
2650         btrfs_put_block_group(block_group);
2651         return ret;
2652 }
2653
2654 int btrfs_exclude_logged_extents(struct extent_buffer *eb)
2655 {
2656         struct btrfs_fs_info *fs_info = eb->fs_info;
2657         struct btrfs_file_extent_item *item;
2658         struct btrfs_key key;
2659         int found_type;
2660         int i;
2661         int ret = 0;
2662
2663         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS))
2664                 return 0;
2665
2666         for (i = 0; i < btrfs_header_nritems(eb); i++) {
2667                 btrfs_item_key_to_cpu(eb, &key, i);
2668                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2669                         continue;
2670                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
2671                 found_type = btrfs_file_extent_type(eb, item);
2672                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
2673                         continue;
2674                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
2675                         continue;
2676                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
2677                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
2678                 ret = __exclude_logged_extent(fs_info, key.objectid, key.offset);
2679                 if (ret)
2680                         break;
2681         }
2682
2683         return ret;
2684 }
2685
2686 static void
2687 btrfs_inc_block_group_reservations(struct btrfs_block_group *bg)
2688 {
2689         atomic_inc(&bg->reservations);
2690 }
2691
2692 /*
2693  * Returns the free cluster for the given space info and sets empty_cluster to
2694  * what it should be based on the mount options.
2695  */
2696 static struct btrfs_free_cluster *
2697 fetch_cluster_info(struct btrfs_fs_info *fs_info,
2698                    struct btrfs_space_info *space_info, u64 *empty_cluster)
2699 {
2700         struct btrfs_free_cluster *ret = NULL;
2701
2702         *empty_cluster = 0;
2703         if (btrfs_mixed_space_info(space_info))
2704                 return ret;
2705
2706         if (space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
2707                 ret = &fs_info->meta_alloc_cluster;
2708                 if (btrfs_test_opt(fs_info, SSD))
2709                         *empty_cluster = SZ_2M;
2710                 else
2711                         *empty_cluster = SZ_64K;
2712         } else if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) &&
2713                    btrfs_test_opt(fs_info, SSD_SPREAD)) {
2714                 *empty_cluster = SZ_2M;
2715                 ret = &fs_info->data_alloc_cluster;
2716         }
2717
2718         return ret;
2719 }
2720
2721 static int unpin_extent_range(struct btrfs_fs_info *fs_info,
2722                               u64 start, u64 end,
2723                               const bool return_free_space)
2724 {
2725         struct btrfs_block_group *cache = NULL;
2726         struct btrfs_space_info *space_info;
2727         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
2728         struct btrfs_free_cluster *cluster = NULL;
2729         u64 len;
2730         u64 total_unpinned = 0;
2731         u64 empty_cluster = 0;
2732         bool readonly;
2733         int ret = 0;
2734
2735         while (start <= end) {
2736                 readonly = false;
2737                 if (!cache ||
2738                     start >= cache->start + cache->length) {
2739                         if (cache)
2740                                 btrfs_put_block_group(cache);
2741                         total_unpinned = 0;
2742                         cache = btrfs_lookup_block_group(fs_info, start);
2743                         if (cache == NULL) {
2744                                 /* Logic error, something removed the block group. */
2745                                 ret = -EUCLEAN;
2746                                 goto out;
2747                         }
2748
2749                         cluster = fetch_cluster_info(fs_info,
2750                                                      cache->space_info,
2751                                                      &empty_cluster);
2752                         empty_cluster <<= 1;
2753                 }
2754
2755                 len = cache->start + cache->length - start;
2756                 len = min(len, end + 1 - start);
2757
2758                 if (return_free_space)
2759                         btrfs_add_free_space(cache, start, len);
2760
2761                 start += len;
2762                 total_unpinned += len;
2763                 space_info = cache->space_info;
2764
2765                 /*
2766                  * If this space cluster has been marked as fragmented and we've
2767                  * unpinned enough in this block group to potentially allow a
2768                  * cluster to be created inside of it go ahead and clear the
2769                  * fragmented check.
2770                  */
2771                 if (cluster && cluster->fragmented &&
2772                     total_unpinned > empty_cluster) {
2773                         spin_lock(&cluster->lock);
2774                         cluster->fragmented = 0;
2775                         spin_unlock(&cluster->lock);
2776                 }
2777
2778                 spin_lock(&space_info->lock);
2779                 spin_lock(&cache->lock);
2780                 cache->pinned -= len;
2781                 btrfs_space_info_update_bytes_pinned(fs_info, space_info, -len);
2782                 space_info->max_extent_size = 0;
2783                 if (cache->ro) {
2784                         space_info->bytes_readonly += len;
2785                         readonly = true;
2786                 } else if (btrfs_is_zoned(fs_info)) {
2787                         /* Need reset before reusing in a zoned block group */
2788                         btrfs_space_info_update_bytes_zone_unusable(fs_info, space_info,
2789                                                                     len);
2790                         readonly = true;
2791                 }
2792                 spin_unlock(&cache->lock);
2793                 if (!readonly && return_free_space &&
2794                     global_rsv->space_info == space_info) {
2795                         spin_lock(&global_rsv->lock);
2796                         if (!global_rsv->full) {
2797                                 u64 to_add = min(len, global_rsv->size -
2798                                                       global_rsv->reserved);
2799
2800                                 global_rsv->reserved += to_add;
2801                                 btrfs_space_info_update_bytes_may_use(fs_info,
2802                                                 space_info, to_add);
2803                                 if (global_rsv->reserved >= global_rsv->size)
2804                                         global_rsv->full = 1;
2805                                 len -= to_add;
2806                         }
2807                         spin_unlock(&global_rsv->lock);
2808                 }
2809                 /* Add to any tickets we may have */
2810                 if (!readonly && return_free_space && len)
2811                         btrfs_try_granting_tickets(fs_info, space_info);
2812                 spin_unlock(&space_info->lock);
2813         }
2814
2815         if (cache)
2816                 btrfs_put_block_group(cache);
2817 out:
2818         return ret;
2819 }
2820
2821 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans)
2822 {
2823         struct btrfs_fs_info *fs_info = trans->fs_info;
2824         struct btrfs_block_group *block_group, *tmp;
2825         struct list_head *deleted_bgs;
2826         struct extent_io_tree *unpin;
2827         u64 start;
2828         u64 end;
2829         int ret;
2830
2831         unpin = &trans->transaction->pinned_extents;
2832
2833         while (!TRANS_ABORTED(trans)) {
2834                 struct extent_state *cached_state = NULL;
2835
2836                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
2837                 if (!find_first_extent_bit(unpin, 0, &start, &end,
2838                                            EXTENT_DIRTY, &cached_state)) {
2839                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
2840                         break;
2841                 }
2842
2843                 if (btrfs_test_opt(fs_info, DISCARD_SYNC))
2844                         ret = btrfs_discard_extent(fs_info, start,
2845                                                    end + 1 - start, NULL);
2846
2847                 clear_extent_dirty(unpin, start, end, &cached_state);
2848                 ret = unpin_extent_range(fs_info, start, end, true);
2849                 BUG_ON(ret);
2850                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
2851                 free_extent_state(cached_state);
2852                 cond_resched();
2853         }
2854
2855         if (btrfs_test_opt(fs_info, DISCARD_ASYNC)) {
2856                 btrfs_discard_calc_delay(&fs_info->discard_ctl);
2857                 btrfs_discard_schedule_work(&fs_info->discard_ctl, true);
2858         }
2859
2860         /*
2861          * Transaction is finished.  We don't need the lock anymore.  We
2862          * do need to clean up the block groups in case of a transaction
2863          * abort.
2864          */
2865         deleted_bgs = &trans->transaction->deleted_bgs;
2866         list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
2867                 u64 trimmed = 0;
2868
2869                 ret = -EROFS;
2870                 if (!TRANS_ABORTED(trans))
2871                         ret = btrfs_discard_extent(fs_info,
2872                                                    block_group->start,
2873                                                    block_group->length,
2874                                                    &trimmed);
2875
2876                 list_del_init(&block_group->bg_list);
2877                 btrfs_unfreeze_block_group(block_group);
2878                 btrfs_put_block_group(block_group);
2879
2880                 if (ret) {
2881                         const char *errstr = btrfs_decode_error(ret);
2882                         btrfs_warn(fs_info,
2883                            "discard failed while removing blockgroup: errno=%d %s",
2884                                    ret, errstr);
2885                 }
2886         }
2887
2888         return 0;
2889 }
2890
2891 /*
2892  * Parse an extent item's inline extents looking for a simple quotas owner ref.
2893  *
2894  * @fs_info:    the btrfs_fs_info for this mount
2895  * @leaf:       a leaf in the extent tree containing the extent item
2896  * @slot:       the slot in the leaf where the extent item is found
2897  *
2898  * Returns the objectid of the root that originally allocated the extent item
2899  * if the inline owner ref is expected and present, otherwise 0.
2900  *
2901  * If an extent item has an owner ref item, it will be the first inline ref
2902  * item. Therefore the logic is to check whether there are any inline ref
2903  * items, then check the type of the first one.
2904  */
2905 u64 btrfs_get_extent_owner_root(struct btrfs_fs_info *fs_info,
2906                                 struct extent_buffer *leaf, int slot)
2907 {
2908         struct btrfs_extent_item *ei;
2909         struct btrfs_extent_inline_ref *iref;
2910         struct btrfs_extent_owner_ref *oref;
2911         unsigned long ptr;
2912         unsigned long end;
2913         int type;
2914
2915         if (!btrfs_fs_incompat(fs_info, SIMPLE_QUOTA))
2916                 return 0;
2917
2918         ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
2919         ptr = (unsigned long)(ei + 1);
2920         end = (unsigned long)ei + btrfs_item_size(leaf, slot);
2921
2922         /* No inline ref items of any kind, can't check type. */
2923         if (ptr == end)
2924                 return 0;
2925
2926         iref = (struct btrfs_extent_inline_ref *)ptr;
2927         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_ANY);
2928
2929         /* We found an owner ref, get the root out of it. */
2930         if (type == BTRFS_EXTENT_OWNER_REF_KEY) {
2931                 oref = (struct btrfs_extent_owner_ref *)(&iref->offset);
2932                 return btrfs_extent_owner_ref_root_id(leaf, oref);
2933         }
2934
2935         /* We have inline refs, but not an owner ref. */
2936         return 0;
2937 }
2938
2939 static int do_free_extent_accounting(struct btrfs_trans_handle *trans,
2940                                      u64 bytenr, struct btrfs_squota_delta *delta)
2941 {
2942         int ret;
2943         u64 num_bytes = delta->num_bytes;
2944
2945         if (delta->is_data) {
2946                 struct btrfs_root *csum_root;
2947
2948                 csum_root = btrfs_csum_root(trans->fs_info, bytenr);
2949                 ret = btrfs_del_csums(trans, csum_root, bytenr, num_bytes);
2950                 if (ret) {
2951                         btrfs_abort_transaction(trans, ret);
2952                         return ret;
2953                 }
2954
2955                 ret = btrfs_delete_raid_extent(trans, bytenr, num_bytes);
2956                 if (ret) {
2957                         btrfs_abort_transaction(trans, ret);
2958                         return ret;
2959                 }
2960         }
2961
2962         ret = btrfs_record_squota_delta(trans->fs_info, delta);
2963         if (ret) {
2964                 btrfs_abort_transaction(trans, ret);
2965                 return ret;
2966         }
2967
2968         ret = add_to_free_space_tree(trans, bytenr, num_bytes);
2969         if (ret) {
2970                 btrfs_abort_transaction(trans, ret);
2971                 return ret;
2972         }
2973
2974         ret = btrfs_update_block_group(trans, bytenr, num_bytes, false);
2975         if (ret)
2976                 btrfs_abort_transaction(trans, ret);
2977
2978         return ret;
2979 }
2980
2981 #define abort_and_dump(trans, path, fmt, args...)       \
2982 ({                                                      \
2983         btrfs_abort_transaction(trans, -EUCLEAN);       \
2984         btrfs_print_leaf(path->nodes[0]);               \
2985         btrfs_crit(trans->fs_info, fmt, ##args);        \
2986 })
2987
2988 /*
2989  * Drop one or more refs of @node.
2990  *
2991  * 1. Locate the extent refs.
2992  *    It's either inline in EXTENT/METADATA_ITEM or in keyed SHARED_* item.
2993  *    Locate it, then reduce the refs number or remove the ref line completely.
2994  *
2995  * 2. Update the refs count in EXTENT/METADATA_ITEM
2996  *
2997  * Inline backref case:
2998  *
2999  * in extent tree we have:
3000  *
3001  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 16201 itemsize 82
3002  *              refs 2 gen 6 flags DATA
3003  *              extent data backref root FS_TREE objectid 258 offset 0 count 1
3004  *              extent data backref root FS_TREE objectid 257 offset 0 count 1
3005  *
3006  * This function gets called with:
3007  *
3008  *    node->bytenr = 13631488
3009  *    node->num_bytes = 1048576
3010  *    root_objectid = FS_TREE
3011  *    owner_objectid = 257
3012  *    owner_offset = 0
3013  *    refs_to_drop = 1
3014  *
3015  * Then we should get some like:
3016  *
3017  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 16201 itemsize 82
3018  *              refs 1 gen 6 flags DATA
3019  *              extent data backref root FS_TREE objectid 258 offset 0 count 1
3020  *
3021  * Keyed backref case:
3022  *
3023  * in extent tree we have:
3024  *
3025  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 3971 itemsize 24
3026  *              refs 754 gen 6 flags DATA
3027  *      [...]
3028  *      item 2 key (13631488 EXTENT_DATA_REF <HASH>) itemoff 3915 itemsize 28
3029  *              extent data backref root FS_TREE objectid 866 offset 0 count 1
3030  *
3031  * This function get called with:
3032  *
3033  *    node->bytenr = 13631488
3034  *    node->num_bytes = 1048576
3035  *    root_objectid = FS_TREE
3036  *    owner_objectid = 866
3037  *    owner_offset = 0
3038  *    refs_to_drop = 1
3039  *
3040  * Then we should get some like:
3041  *
3042  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 3971 itemsize 24
3043  *              refs 753 gen 6 flags DATA
3044  *
3045  * And that (13631488 EXTENT_DATA_REF <HASH>) gets removed.
3046  */
3047 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
3048                                struct btrfs_delayed_ref_head *href,
3049                                struct btrfs_delayed_ref_node *node,
3050                                struct btrfs_delayed_extent_op *extent_op)
3051 {
3052         struct btrfs_fs_info *info = trans->fs_info;
3053         struct btrfs_key key;
3054         struct btrfs_path *path;
3055         struct btrfs_root *extent_root;
3056         struct extent_buffer *leaf;
3057         struct btrfs_extent_item *ei;
3058         struct btrfs_extent_inline_ref *iref;
3059         int ret;
3060         int is_data;
3061         int extent_slot = 0;
3062         int found_extent = 0;
3063         int num_to_del = 1;
3064         int refs_to_drop = node->ref_mod;
3065         u32 item_size;
3066         u64 refs;
3067         u64 bytenr = node->bytenr;
3068         u64 num_bytes = node->num_bytes;
3069         u64 owner_objectid = btrfs_delayed_ref_owner(node);
3070         u64 owner_offset = btrfs_delayed_ref_offset(node);
3071         bool skinny_metadata = btrfs_fs_incompat(info, SKINNY_METADATA);
3072         u64 delayed_ref_root = href->owning_root;
3073
3074         extent_root = btrfs_extent_root(info, bytenr);
3075         ASSERT(extent_root);
3076
3077         path = btrfs_alloc_path();
3078         if (!path)
3079                 return -ENOMEM;
3080
3081         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
3082
3083         if (!is_data && refs_to_drop != 1) {
3084                 btrfs_crit(info,
3085 "invalid refs_to_drop, dropping more than 1 refs for tree block %llu refs_to_drop %u",
3086                            node->bytenr, refs_to_drop);
3087                 ret = -EINVAL;
3088                 btrfs_abort_transaction(trans, ret);
3089                 goto out;
3090         }
3091
3092         if (is_data)
3093                 skinny_metadata = false;
3094
3095         ret = lookup_extent_backref(trans, path, &iref, bytenr, num_bytes,
3096                                     node->parent, node->ref_root, owner_objectid,
3097                                     owner_offset);
3098         if (ret == 0) {
3099                 /*
3100                  * Either the inline backref or the SHARED_DATA_REF/
3101                  * SHARED_BLOCK_REF is found
3102                  *
3103                  * Here is a quick path to locate EXTENT/METADATA_ITEM.
3104                  * It's possible the EXTENT/METADATA_ITEM is near current slot.
3105                  */
3106                 extent_slot = path->slots[0];
3107                 while (extent_slot >= 0) {
3108                         btrfs_item_key_to_cpu(path->nodes[0], &key,
3109                                               extent_slot);
3110                         if (key.objectid != bytenr)
3111                                 break;
3112                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
3113                             key.offset == num_bytes) {
3114                                 found_extent = 1;
3115                                 break;
3116                         }
3117                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
3118                             key.offset == owner_objectid) {
3119                                 found_extent = 1;
3120                                 break;
3121                         }
3122
3123                         /* Quick path didn't find the EXTENT/METADATA_ITEM */
3124                         if (path->slots[0] - extent_slot > 5)
3125                                 break;
3126                         extent_slot--;
3127                 }
3128
3129                 if (!found_extent) {
3130                         if (iref) {
3131                                 abort_and_dump(trans, path,
3132 "invalid iref slot %u, no EXTENT/METADATA_ITEM found but has inline extent ref",
3133                                            path->slots[0]);
3134                                 ret = -EUCLEAN;
3135                                 goto out;
3136                         }
3137                         /* Must be SHARED_* item, remove the backref first */
3138                         ret = remove_extent_backref(trans, extent_root, path,
3139                                                     NULL, refs_to_drop, is_data);
3140                         if (ret) {
3141                                 btrfs_abort_transaction(trans, ret);
3142                                 goto out;
3143                         }
3144                         btrfs_release_path(path);
3145
3146                         /* Slow path to locate EXTENT/METADATA_ITEM */
3147                         key.objectid = bytenr;
3148                         key.type = BTRFS_EXTENT_ITEM_KEY;
3149                         key.offset = num_bytes;
3150
3151                         if (!is_data && skinny_metadata) {
3152                                 key.type = BTRFS_METADATA_ITEM_KEY;
3153                                 key.offset = owner_objectid;
3154                         }
3155
3156                         ret = btrfs_search_slot(trans, extent_root,
3157                                                 &key, path, -1, 1);
3158                         if (ret > 0 && skinny_metadata && path->slots[0]) {
3159                                 /*
3160                                  * Couldn't find our skinny metadata item,
3161                                  * see if we have ye olde extent item.
3162                                  */
3163                                 path->slots[0]--;
3164                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
3165                                                       path->slots[0]);
3166                                 if (key.objectid == bytenr &&
3167                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
3168                                     key.offset == num_bytes)
3169                                         ret = 0;
3170                         }
3171
3172                         if (ret > 0 && skinny_metadata) {
3173                                 skinny_metadata = false;
3174                                 key.objectid = bytenr;
3175                                 key.type = BTRFS_EXTENT_ITEM_KEY;
3176                                 key.offset = num_bytes;
3177                                 btrfs_release_path(path);
3178                                 ret = btrfs_search_slot(trans, extent_root,
3179                                                         &key, path, -1, 1);
3180                         }
3181
3182                         if (ret) {
3183                                 if (ret > 0)
3184                                         btrfs_print_leaf(path->nodes[0]);
3185                                 btrfs_err(info,
3186                         "umm, got %d back from search, was looking for %llu, slot %d",
3187                                           ret, bytenr, path->slots[0]);
3188                         }
3189                         if (ret < 0) {
3190                                 btrfs_abort_transaction(trans, ret);
3191                                 goto out;
3192                         }
3193                         extent_slot = path->slots[0];
3194                 }
3195         } else if (WARN_ON(ret == -ENOENT)) {
3196                 abort_and_dump(trans, path,
3197 "unable to find ref byte nr %llu parent %llu root %llu owner %llu offset %llu slot %d",
3198                                bytenr, node->parent, node->ref_root, owner_objectid,
3199                                owner_offset, path->slots[0]);
3200                 goto out;
3201         } else {
3202                 btrfs_abort_transaction(trans, ret);
3203                 goto out;
3204         }
3205
3206         leaf = path->nodes[0];
3207         item_size = btrfs_item_size(leaf, extent_slot);
3208         if (unlikely(item_size < sizeof(*ei))) {
3209                 ret = -EUCLEAN;
3210                 btrfs_err(trans->fs_info,
3211                           "unexpected extent item size, has %u expect >= %zu",
3212                           item_size, sizeof(*ei));
3213                 btrfs_abort_transaction(trans, ret);
3214                 goto out;
3215         }
3216         ei = btrfs_item_ptr(leaf, extent_slot,
3217                             struct btrfs_extent_item);
3218         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
3219             key.type == BTRFS_EXTENT_ITEM_KEY) {
3220                 struct btrfs_tree_block_info *bi;
3221
3222                 if (item_size < sizeof(*ei) + sizeof(*bi)) {
3223                         abort_and_dump(trans, path,
3224 "invalid extent item size for key (%llu, %u, %llu) slot %u owner %llu, has %u expect >= %zu",
3225                                        key.objectid, key.type, key.offset,
3226                                        path->slots[0], owner_objectid, item_size,
3227                                        sizeof(*ei) + sizeof(*bi));
3228                         ret = -EUCLEAN;
3229                         goto out;
3230                 }
3231                 bi = (struct btrfs_tree_block_info *)(ei + 1);
3232                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
3233         }
3234
3235         refs = btrfs_extent_refs(leaf, ei);
3236         if (refs < refs_to_drop) {
3237                 abort_and_dump(trans, path,
3238                 "trying to drop %d refs but we only have %llu for bytenr %llu slot %u",
3239                                refs_to_drop, refs, bytenr, path->slots[0]);
3240                 ret = -EUCLEAN;
3241                 goto out;
3242         }
3243         refs -= refs_to_drop;
3244
3245         if (refs > 0) {
3246                 if (extent_op)
3247                         __run_delayed_extent_op(extent_op, leaf, ei);
3248                 /*
3249                  * In the case of inline back ref, reference count will
3250                  * be updated by remove_extent_backref
3251                  */
3252                 if (iref) {
3253                         if (!found_extent) {
3254                                 abort_and_dump(trans, path,
3255 "invalid iref, got inlined extent ref but no EXTENT/METADATA_ITEM found, slot %u",
3256                                                path->slots[0]);
3257                                 ret = -EUCLEAN;
3258                                 goto out;
3259                         }
3260                 } else {
3261                         btrfs_set_extent_refs(leaf, ei, refs);
3262                         btrfs_mark_buffer_dirty(trans, leaf);
3263                 }
3264                 if (found_extent) {
3265                         ret = remove_extent_backref(trans, extent_root, path,
3266                                                     iref, refs_to_drop, is_data);
3267                         if (ret) {
3268                                 btrfs_abort_transaction(trans, ret);
3269                                 goto out;
3270                         }
3271                 }
3272         } else {
3273                 struct btrfs_squota_delta delta = {
3274                         .root = delayed_ref_root,
3275                         .num_bytes = num_bytes,
3276                         .is_data = is_data,
3277                         .is_inc = false,
3278                         .generation = btrfs_extent_generation(leaf, ei),
3279                 };
3280
3281                 /* In this branch refs == 1 */
3282                 if (found_extent) {
3283                         if (is_data && refs_to_drop !=
3284                             extent_data_ref_count(path, iref)) {
3285                                 abort_and_dump(trans, path,
3286                 "invalid refs_to_drop, current refs %u refs_to_drop %u slot %u",
3287                                                extent_data_ref_count(path, iref),
3288                                                refs_to_drop, path->slots[0]);
3289                                 ret = -EUCLEAN;
3290                                 goto out;
3291                         }
3292                         if (iref) {
3293                                 if (path->slots[0] != extent_slot) {
3294                                         abort_and_dump(trans, path,
3295 "invalid iref, extent item key (%llu %u %llu) slot %u doesn't have wanted iref",
3296                                                        key.objectid, key.type,
3297                                                        key.offset, path->slots[0]);
3298                                         ret = -EUCLEAN;
3299                                         goto out;
3300                                 }
3301                         } else {
3302                                 /*
3303                                  * No inline ref, we must be at SHARED_* item,
3304                                  * And it's single ref, it must be:
3305                                  * |    extent_slot       ||extent_slot + 1|
3306                                  * [ EXTENT/METADATA_ITEM ][ SHARED_* ITEM ]
3307                                  */
3308                                 if (path->slots[0] != extent_slot + 1) {
3309                                         abort_and_dump(trans, path,
3310         "invalid SHARED_* item slot %u, previous item is not EXTENT/METADATA_ITEM",
3311                                                        path->slots[0]);
3312                                         ret = -EUCLEAN;
3313                                         goto out;
3314                                 }
3315                                 path->slots[0] = extent_slot;
3316                                 num_to_del = 2;
3317                         }
3318                 }
3319                 /*
3320                  * We can't infer the data owner from the delayed ref, so we need
3321                  * to try to get it from the owning ref item.
3322                  *
3323                  * If it is not present, then that extent was not written under
3324                  * simple quotas mode, so we don't need to account for its deletion.
3325                  */
3326                 if (is_data)
3327                         delta.root = btrfs_get_extent_owner_root(trans->fs_info,
3328                                                                  leaf, extent_slot);
3329
3330                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
3331                                       num_to_del);
3332                 if (ret) {
3333                         btrfs_abort_transaction(trans, ret);
3334                         goto out;
3335                 }
3336                 btrfs_release_path(path);
3337
3338                 ret = do_free_extent_accounting(trans, bytenr, &delta);
3339         }
3340         btrfs_release_path(path);
3341
3342 out:
3343         btrfs_free_path(path);
3344         return ret;
3345 }
3346
3347 /*
3348  * when we free an block, it is possible (and likely) that we free the last
3349  * delayed ref for that extent as well.  This searches the delayed ref tree for
3350  * a given extent, and if there are no other delayed refs to be processed, it
3351  * removes it from the tree.
3352  */
3353 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
3354                                       u64 bytenr)
3355 {
3356         struct btrfs_fs_info *fs_info = trans->fs_info;
3357         struct btrfs_delayed_ref_head *head;
3358         struct btrfs_delayed_ref_root *delayed_refs;
3359         int ret = 0;
3360
3361         delayed_refs = &trans->transaction->delayed_refs;
3362         spin_lock(&delayed_refs->lock);
3363         head = btrfs_find_delayed_ref_head(fs_info, delayed_refs, bytenr);
3364         if (!head)
3365                 goto out_delayed_unlock;
3366
3367         spin_lock(&head->lock);
3368         if (!RB_EMPTY_ROOT(&head->ref_tree.rb_root))
3369                 goto out;
3370
3371         if (cleanup_extent_op(head) != NULL)
3372                 goto out;
3373
3374         /*
3375          * waiting for the lock here would deadlock.  If someone else has it
3376          * locked they are already in the process of dropping it anyway
3377          */
3378         if (!mutex_trylock(&head->mutex))
3379                 goto out;
3380
3381         btrfs_delete_ref_head(fs_info, delayed_refs, head);
3382         head->processing = false;
3383
3384         spin_unlock(&head->lock);
3385         spin_unlock(&delayed_refs->lock);
3386
3387         BUG_ON(head->extent_op);
3388         if (head->must_insert_reserved)
3389                 ret = 1;
3390
3391         btrfs_cleanup_ref_head_accounting(fs_info, delayed_refs, head);
3392         mutex_unlock(&head->mutex);
3393         btrfs_put_delayed_ref_head(head);
3394         return ret;
3395 out:
3396         spin_unlock(&head->lock);
3397
3398 out_delayed_unlock:
3399         spin_unlock(&delayed_refs->lock);
3400         return 0;
3401 }
3402
3403 int btrfs_free_tree_block(struct btrfs_trans_handle *trans,
3404                           u64 root_id,
3405                           struct extent_buffer *buf,
3406                           u64 parent, int last_ref)
3407 {
3408         struct btrfs_fs_info *fs_info = trans->fs_info;
3409         struct btrfs_block_group *bg;
3410         int ret;
3411
3412         if (root_id != BTRFS_TREE_LOG_OBJECTID) {
3413                 struct btrfs_ref generic_ref = {
3414                         .action = BTRFS_DROP_DELAYED_REF,
3415                         .bytenr = buf->start,
3416                         .num_bytes = buf->len,
3417                         .parent = parent,
3418                         .owning_root = btrfs_header_owner(buf),
3419                         .ref_root = root_id,
3420                 };
3421
3422                 /*
3423                  * Assert that the extent buffer is not cleared due to
3424                  * EXTENT_BUFFER_ZONED_ZEROOUT. Please refer
3425                  * btrfs_clear_buffer_dirty() and btree_csum_one_bio() for
3426                  * detail.
3427                  */
3428                 ASSERT(btrfs_header_bytenr(buf) != 0);
3429
3430                 btrfs_init_tree_ref(&generic_ref, btrfs_header_level(buf), 0, false);
3431                 btrfs_ref_tree_mod(fs_info, &generic_ref);
3432                 ret = btrfs_add_delayed_tree_ref(trans, &generic_ref, NULL);
3433                 if (ret < 0)
3434                         return ret;
3435         }
3436
3437         if (!last_ref)
3438                 return 0;
3439
3440         if (btrfs_header_generation(buf) != trans->transid)
3441                 goto out;
3442
3443         if (root_id != BTRFS_TREE_LOG_OBJECTID) {
3444                 ret = check_ref_cleanup(trans, buf->start);
3445                 if (!ret)
3446                         goto out;
3447         }
3448
3449         bg = btrfs_lookup_block_group(fs_info, buf->start);
3450
3451         if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
3452                 pin_down_extent(trans, bg, buf->start, buf->len, 1);
3453                 btrfs_put_block_group(bg);
3454                 goto out;
3455         }
3456
3457         /*
3458          * If there are tree mod log users we may have recorded mod log
3459          * operations for this node.  If we re-allocate this node we
3460          * could replay operations on this node that happened when it
3461          * existed in a completely different root.  For example if it
3462          * was part of root A, then was reallocated to root B, and we
3463          * are doing a btrfs_old_search_slot(root b), we could replay
3464          * operations that happened when the block was part of root A,
3465          * giving us an inconsistent view of the btree.
3466          *
3467          * We are safe from races here because at this point no other
3468          * node or root points to this extent buffer, so if after this
3469          * check a new tree mod log user joins we will not have an
3470          * existing log of operations on this node that we have to
3471          * contend with.
3472          */
3473
3474         if (test_bit(BTRFS_FS_TREE_MOD_LOG_USERS, &fs_info->flags)
3475                      || btrfs_is_zoned(fs_info)) {
3476                 pin_down_extent(trans, bg, buf->start, buf->len, 1);
3477                 btrfs_put_block_group(bg);
3478                 goto out;
3479         }
3480
3481         WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
3482
3483         btrfs_add_free_space(bg, buf->start, buf->len);
3484         btrfs_free_reserved_bytes(bg, buf->len, 0);
3485         btrfs_put_block_group(bg);
3486         trace_btrfs_reserved_extent_free(fs_info, buf->start, buf->len);
3487
3488 out:
3489
3490         /*
3491          * Deleting the buffer, clear the corrupt flag since it doesn't
3492          * matter anymore.
3493          */
3494         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
3495         return 0;
3496 }
3497
3498 /* Can return -ENOMEM */
3499 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_ref *ref)
3500 {
3501         struct btrfs_fs_info *fs_info = trans->fs_info;
3502         int ret;
3503
3504         if (btrfs_is_testing(fs_info))
3505                 return 0;
3506
3507         /*
3508          * tree log blocks never actually go into the extent allocation
3509          * tree, just update pinning info and exit early.
3510          */
3511         if (ref->ref_root == BTRFS_TREE_LOG_OBJECTID) {
3512                 btrfs_pin_extent(trans, ref->bytenr, ref->num_bytes, 1);
3513                 ret = 0;
3514         } else if (ref->type == BTRFS_REF_METADATA) {
3515                 ret = btrfs_add_delayed_tree_ref(trans, ref, NULL);
3516         } else {
3517                 ret = btrfs_add_delayed_data_ref(trans, ref, 0);
3518         }
3519
3520         if (ref->ref_root != BTRFS_TREE_LOG_OBJECTID)
3521                 btrfs_ref_tree_mod(fs_info, ref);
3522
3523         return ret;
3524 }
3525
3526 enum btrfs_loop_type {
3527         /*
3528          * Start caching block groups but do not wait for progress or for them
3529          * to be done.
3530          */
3531         LOOP_CACHING_NOWAIT,
3532
3533         /*
3534          * Wait for the block group free_space >= the space we're waiting for if
3535          * the block group isn't cached.
3536          */
3537         LOOP_CACHING_WAIT,
3538
3539         /*
3540          * Allow allocations to happen from block groups that do not yet have a
3541          * size classification.
3542          */
3543         LOOP_UNSET_SIZE_CLASS,
3544
3545         /*
3546          * Allocate a chunk and then retry the allocation.
3547          */
3548         LOOP_ALLOC_CHUNK,
3549
3550         /*
3551          * Ignore the size class restrictions for this allocation.
3552          */
3553         LOOP_WRONG_SIZE_CLASS,
3554
3555         /*
3556          * Ignore the empty size, only try to allocate the number of bytes
3557          * needed for this allocation.
3558          */
3559         LOOP_NO_EMPTY_SIZE,
3560 };
3561
3562 static inline void
3563 btrfs_lock_block_group(struct btrfs_block_group *cache,
3564                        int delalloc)
3565 {
3566         if (delalloc)
3567                 down_read(&cache->data_rwsem);
3568 }
3569
3570 static inline void btrfs_grab_block_group(struct btrfs_block_group *cache,
3571                        int delalloc)
3572 {
3573         btrfs_get_block_group(cache);
3574         if (delalloc)
3575                 down_read(&cache->data_rwsem);
3576 }
3577
3578 static struct btrfs_block_group *btrfs_lock_cluster(
3579                    struct btrfs_block_group *block_group,
3580                    struct btrfs_free_cluster *cluster,
3581                    int delalloc)
3582         __acquires(&cluster->refill_lock)
3583 {
3584         struct btrfs_block_group *used_bg = NULL;
3585
3586         spin_lock(&cluster->refill_lock);
3587         while (1) {
3588                 used_bg = cluster->block_group;
3589                 if (!used_bg)
3590                         return NULL;
3591
3592                 if (used_bg == block_group)
3593                         return used_bg;
3594
3595                 btrfs_get_block_group(used_bg);
3596
3597                 if (!delalloc)
3598                         return used_bg;
3599
3600                 if (down_read_trylock(&used_bg->data_rwsem))
3601                         return used_bg;
3602
3603                 spin_unlock(&cluster->refill_lock);
3604
3605                 /* We should only have one-level nested. */
3606                 down_read_nested(&used_bg->data_rwsem, SINGLE_DEPTH_NESTING);
3607
3608                 spin_lock(&cluster->refill_lock);
3609                 if (used_bg == cluster->block_group)
3610                         return used_bg;
3611
3612                 up_read(&used_bg->data_rwsem);
3613                 btrfs_put_block_group(used_bg);
3614         }
3615 }
3616
3617 static inline void
3618 btrfs_release_block_group(struct btrfs_block_group *cache,
3619                          int delalloc)
3620 {
3621         if (delalloc)
3622                 up_read(&cache->data_rwsem);
3623         btrfs_put_block_group(cache);
3624 }
3625
3626 /*
3627  * Helper function for find_free_extent().
3628  *
3629  * Return -ENOENT to inform caller that we need fallback to unclustered mode.
3630  * Return >0 to inform caller that we find nothing
3631  * Return 0 means we have found a location and set ffe_ctl->found_offset.
3632  */
3633 static int find_free_extent_clustered(struct btrfs_block_group *bg,
3634                                       struct find_free_extent_ctl *ffe_ctl,
3635                                       struct btrfs_block_group **cluster_bg_ret)
3636 {
3637         struct btrfs_block_group *cluster_bg;
3638         struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
3639         u64 aligned_cluster;
3640         u64 offset;
3641         int ret;
3642
3643         cluster_bg = btrfs_lock_cluster(bg, last_ptr, ffe_ctl->delalloc);
3644         if (!cluster_bg)
3645                 goto refill_cluster;
3646         if (cluster_bg != bg && (cluster_bg->ro ||
3647             !block_group_bits(cluster_bg, ffe_ctl->flags)))
3648                 goto release_cluster;
3649
3650         offset = btrfs_alloc_from_cluster(cluster_bg, last_ptr,
3651                         ffe_ctl->num_bytes, cluster_bg->start,
3652                         &ffe_ctl->max_extent_size);
3653         if (offset) {
3654                 /* We have a block, we're done */
3655                 spin_unlock(&last_ptr->refill_lock);
3656                 trace_btrfs_reserve_extent_cluster(cluster_bg, ffe_ctl);
3657                 *cluster_bg_ret = cluster_bg;
3658                 ffe_ctl->found_offset = offset;
3659                 return 0;
3660         }
3661         WARN_ON(last_ptr->block_group != cluster_bg);
3662
3663 release_cluster:
3664         /*
3665          * If we are on LOOP_NO_EMPTY_SIZE, we can't set up a new clusters, so
3666          * lets just skip it and let the allocator find whatever block it can
3667          * find. If we reach this point, we will have tried the cluster
3668          * allocator plenty of times and not have found anything, so we are
3669          * likely way too fragmented for the clustering stuff to find anything.
3670          *
3671          * However, if the cluster is taken from the current block group,
3672          * release the cluster first, so that we stand a better chance of
3673          * succeeding in the unclustered allocation.
3674          */
3675         if (ffe_ctl->loop >= LOOP_NO_EMPTY_SIZE && cluster_bg != bg) {
3676                 spin_unlock(&last_ptr->refill_lock);
3677                 btrfs_release_block_group(cluster_bg, ffe_ctl->delalloc);
3678                 return -ENOENT;
3679         }
3680
3681         /* This cluster didn't work out, free it and start over */
3682         btrfs_return_cluster_to_free_space(NULL, last_ptr);
3683
3684         if (cluster_bg != bg)
3685                 btrfs_release_block_group(cluster_bg, ffe_ctl->delalloc);
3686
3687 refill_cluster:
3688         if (ffe_ctl->loop >= LOOP_NO_EMPTY_SIZE) {
3689                 spin_unlock(&last_ptr->refill_lock);
3690                 return -ENOENT;
3691         }
3692
3693         aligned_cluster = max_t(u64,
3694                         ffe_ctl->empty_cluster + ffe_ctl->empty_size,
3695                         bg->full_stripe_len);
3696         ret = btrfs_find_space_cluster(bg, last_ptr, ffe_ctl->search_start,
3697                         ffe_ctl->num_bytes, aligned_cluster);
3698         if (ret == 0) {
3699                 /* Now pull our allocation out of this cluster */
3700                 offset = btrfs_alloc_from_cluster(bg, last_ptr,
3701                                 ffe_ctl->num_bytes, ffe_ctl->search_start,
3702                                 &ffe_ctl->max_extent_size);
3703                 if (offset) {
3704                         /* We found one, proceed */
3705                         spin_unlock(&last_ptr->refill_lock);
3706                         ffe_ctl->found_offset = offset;
3707                         trace_btrfs_reserve_extent_cluster(bg, ffe_ctl);
3708                         return 0;
3709                 }
3710         }
3711         /*
3712          * At this point we either didn't find a cluster or we weren't able to
3713          * allocate a block from our cluster.  Free the cluster we've been
3714          * trying to use, and go to the next block group.
3715          */
3716         btrfs_return_cluster_to_free_space(NULL, last_ptr);
3717         spin_unlock(&last_ptr->refill_lock);
3718         return 1;
3719 }
3720
3721 /*
3722  * Return >0 to inform caller that we find nothing
3723  * Return 0 when we found an free extent and set ffe_ctrl->found_offset
3724  */
3725 static int find_free_extent_unclustered(struct btrfs_block_group *bg,
3726                                         struct find_free_extent_ctl *ffe_ctl)
3727 {
3728         struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
3729         u64 offset;
3730
3731         /*
3732          * We are doing an unclustered allocation, set the fragmented flag so
3733          * we don't bother trying to setup a cluster again until we get more
3734          * space.
3735          */
3736         if (unlikely(last_ptr)) {
3737                 spin_lock(&last_ptr->lock);
3738                 last_ptr->fragmented = 1;
3739                 spin_unlock(&last_ptr->lock);
3740         }
3741         if (ffe_ctl->cached) {
3742                 struct btrfs_free_space_ctl *free_space_ctl;
3743
3744                 free_space_ctl = bg->free_space_ctl;
3745                 spin_lock(&free_space_ctl->tree_lock);
3746                 if (free_space_ctl->free_space <
3747                     ffe_ctl->num_bytes + ffe_ctl->empty_cluster +
3748                     ffe_ctl->empty_size) {
3749                         ffe_ctl->total_free_space = max_t(u64,
3750                                         ffe_ctl->total_free_space,
3751                                         free_space_ctl->free_space);
3752                         spin_unlock(&free_space_ctl->tree_lock);
3753                         return 1;
3754                 }
3755                 spin_unlock(&free_space_ctl->tree_lock);
3756         }
3757
3758         offset = btrfs_find_space_for_alloc(bg, ffe_ctl->search_start,
3759                         ffe_ctl->num_bytes, ffe_ctl->empty_size,
3760                         &ffe_ctl->max_extent_size);
3761         if (!offset)
3762                 return 1;
3763         ffe_ctl->found_offset = offset;
3764         return 0;
3765 }
3766
3767 static int do_allocation_clustered(struct btrfs_block_group *block_group,
3768                                    struct find_free_extent_ctl *ffe_ctl,
3769                                    struct btrfs_block_group **bg_ret)
3770 {
3771         int ret;
3772
3773         /* We want to try and use the cluster allocator, so lets look there */
3774         if (ffe_ctl->last_ptr && ffe_ctl->use_cluster) {
3775                 ret = find_free_extent_clustered(block_group, ffe_ctl, bg_ret);
3776                 if (ret >= 0)
3777                         return ret;
3778                 /* ret == -ENOENT case falls through */
3779         }
3780
3781         return find_free_extent_unclustered(block_group, ffe_ctl);
3782 }
3783
3784 /*
3785  * Tree-log block group locking
3786  * ============================
3787  *
3788  * fs_info::treelog_bg_lock protects the fs_info::treelog_bg which
3789  * indicates the starting address of a block group, which is reserved only
3790  * for tree-log metadata.
3791  *
3792  * Lock nesting
3793  * ============
3794  *
3795  * space_info::lock
3796  *   block_group::lock
3797  *     fs_info::treelog_bg_lock
3798  */
3799
3800 /*
3801  * Simple allocator for sequential-only block group. It only allows sequential
3802  * allocation. No need to play with trees. This function also reserves the
3803  * bytes as in btrfs_add_reserved_bytes.
3804  */
3805 static int do_allocation_zoned(struct btrfs_block_group *block_group,
3806                                struct find_free_extent_ctl *ffe_ctl,
3807                                struct btrfs_block_group **bg_ret)
3808 {
3809         struct btrfs_fs_info *fs_info = block_group->fs_info;
3810         struct btrfs_space_info *space_info = block_group->space_info;
3811         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3812         u64 start = block_group->start;
3813         u64 num_bytes = ffe_ctl->num_bytes;
3814         u64 avail;
3815         u64 bytenr = block_group->start;
3816         u64 log_bytenr;
3817         u64 data_reloc_bytenr;
3818         int ret = 0;
3819         bool skip = false;
3820
3821         ASSERT(btrfs_is_zoned(block_group->fs_info));
3822
3823         /*
3824          * Do not allow non-tree-log blocks in the dedicated tree-log block
3825          * group, and vice versa.
3826          */
3827         spin_lock(&fs_info->treelog_bg_lock);
3828         log_bytenr = fs_info->treelog_bg;
3829         if (log_bytenr && ((ffe_ctl->for_treelog && bytenr != log_bytenr) ||
3830                            (!ffe_ctl->for_treelog && bytenr == log_bytenr)))
3831                 skip = true;
3832         spin_unlock(&fs_info->treelog_bg_lock);
3833         if (skip)
3834                 return 1;
3835
3836         /*
3837          * Do not allow non-relocation blocks in the dedicated relocation block
3838          * group, and vice versa.
3839          */
3840         spin_lock(&fs_info->relocation_bg_lock);
3841         data_reloc_bytenr = fs_info->data_reloc_bg;
3842         if (data_reloc_bytenr &&
3843             ((ffe_ctl->for_data_reloc && bytenr != data_reloc_bytenr) ||
3844              (!ffe_ctl->for_data_reloc && bytenr == data_reloc_bytenr)))
3845                 skip = true;
3846         spin_unlock(&fs_info->relocation_bg_lock);
3847         if (skip)
3848                 return 1;
3849
3850         /* Check RO and no space case before trying to activate it */
3851         spin_lock(&block_group->lock);
3852         if (block_group->ro || btrfs_zoned_bg_is_full(block_group)) {
3853                 ret = 1;
3854                 /*
3855                  * May need to clear fs_info->{treelog,data_reloc}_bg.
3856                  * Return the error after taking the locks.
3857                  */
3858         }
3859         spin_unlock(&block_group->lock);
3860
3861         /* Metadata block group is activated at write time. */
3862         if (!ret && (block_group->flags & BTRFS_BLOCK_GROUP_DATA) &&
3863             !btrfs_zone_activate(block_group)) {
3864                 ret = 1;
3865                 /*
3866                  * May need to clear fs_info->{treelog,data_reloc}_bg.
3867                  * Return the error after taking the locks.
3868                  */
3869         }
3870
3871         spin_lock(&space_info->lock);
3872         spin_lock(&block_group->lock);
3873         spin_lock(&fs_info->treelog_bg_lock);
3874         spin_lock(&fs_info->relocation_bg_lock);
3875
3876         if (ret)
3877                 goto out;
3878
3879         ASSERT(!ffe_ctl->for_treelog ||
3880                block_group->start == fs_info->treelog_bg ||
3881                fs_info->treelog_bg == 0);
3882         ASSERT(!ffe_ctl->for_data_reloc ||
3883                block_group->start == fs_info->data_reloc_bg ||
3884                fs_info->data_reloc_bg == 0);
3885
3886         if (block_group->ro ||
3887             (!ffe_ctl->for_data_reloc &&
3888              test_bit(BLOCK_GROUP_FLAG_ZONED_DATA_RELOC, &block_group->runtime_flags))) {
3889                 ret = 1;
3890                 goto out;
3891         }
3892
3893         /*
3894          * Do not allow currently using block group to be tree-log dedicated
3895          * block group.
3896          */
3897         if (ffe_ctl->for_treelog && !fs_info->treelog_bg &&
3898             (block_group->used || block_group->reserved)) {
3899                 ret = 1;
3900                 goto out;
3901         }
3902
3903         /*
3904          * Do not allow currently used block group to be the data relocation
3905          * dedicated block group.
3906          */
3907         if (ffe_ctl->for_data_reloc && !fs_info->data_reloc_bg &&
3908             (block_group->used || block_group->reserved)) {
3909                 ret = 1;
3910                 goto out;
3911         }
3912
3913         WARN_ON_ONCE(block_group->alloc_offset > block_group->zone_capacity);
3914         avail = block_group->zone_capacity - block_group->alloc_offset;
3915         if (avail < num_bytes) {
3916                 if (ffe_ctl->max_extent_size < avail) {
3917                         /*
3918                          * With sequential allocator, free space is always
3919                          * contiguous
3920                          */
3921                         ffe_ctl->max_extent_size = avail;
3922                         ffe_ctl->total_free_space = avail;
3923                 }
3924                 ret = 1;
3925                 goto out;
3926         }
3927
3928         if (ffe_ctl->for_treelog && !fs_info->treelog_bg)
3929                 fs_info->treelog_bg = block_group->start;
3930
3931         if (ffe_ctl->for_data_reloc) {
3932                 if (!fs_info->data_reloc_bg)
3933                         fs_info->data_reloc_bg = block_group->start;
3934                 /*
3935                  * Do not allow allocations from this block group, unless it is
3936                  * for data relocation. Compared to increasing the ->ro, setting
3937                  * the ->zoned_data_reloc_ongoing flag still allows nocow
3938                  * writers to come in. See btrfs_inc_nocow_writers().
3939                  *
3940                  * We need to disable an allocation to avoid an allocation of
3941                  * regular (non-relocation data) extent. With mix of relocation
3942                  * extents and regular extents, we can dispatch WRITE commands
3943                  * (for relocation extents) and ZONE APPEND commands (for
3944                  * regular extents) at the same time to the same zone, which
3945                  * easily break the write pointer.
3946                  *
3947                  * Also, this flag avoids this block group to be zone finished.
3948                  */
3949                 set_bit(BLOCK_GROUP_FLAG_ZONED_DATA_RELOC, &block_group->runtime_flags);
3950         }
3951
3952         ffe_ctl->found_offset = start + block_group->alloc_offset;
3953         block_group->alloc_offset += num_bytes;
3954         spin_lock(&ctl->tree_lock);
3955         ctl->free_space -= num_bytes;
3956         spin_unlock(&ctl->tree_lock);
3957
3958         /*
3959          * We do not check if found_offset is aligned to stripesize. The
3960          * address is anyway rewritten when using zone append writing.
3961          */
3962
3963         ffe_ctl->search_start = ffe_ctl->found_offset;
3964
3965 out:
3966         if (ret && ffe_ctl->for_treelog)
3967                 fs_info->treelog_bg = 0;
3968         if (ret && ffe_ctl->for_data_reloc)
3969                 fs_info->data_reloc_bg = 0;
3970         spin_unlock(&fs_info->relocation_bg_lock);
3971         spin_unlock(&fs_info->treelog_bg_lock);
3972         spin_unlock(&block_group->lock);
3973         spin_unlock(&space_info->lock);
3974         return ret;
3975 }
3976
3977 static int do_allocation(struct btrfs_block_group *block_group,
3978                          struct find_free_extent_ctl *ffe_ctl,
3979                          struct btrfs_block_group **bg_ret)
3980 {
3981         switch (ffe_ctl->policy) {
3982         case BTRFS_EXTENT_ALLOC_CLUSTERED:
3983                 return do_allocation_clustered(block_group, ffe_ctl, bg_ret);
3984         case BTRFS_EXTENT_ALLOC_ZONED:
3985                 return do_allocation_zoned(block_group, ffe_ctl, bg_ret);
3986         default:
3987                 BUG();
3988         }
3989 }
3990
3991 static void release_block_group(struct btrfs_block_group *block_group,
3992                                 struct find_free_extent_ctl *ffe_ctl,
3993                                 int delalloc)
3994 {
3995         switch (ffe_ctl->policy) {
3996         case BTRFS_EXTENT_ALLOC_CLUSTERED:
3997                 ffe_ctl->retry_uncached = false;
3998                 break;
3999         case BTRFS_EXTENT_ALLOC_ZONED:
4000                 /* Nothing to do */
4001                 break;
4002         default:
4003                 BUG();
4004         }
4005
4006         BUG_ON(btrfs_bg_flags_to_raid_index(block_group->flags) !=
4007                ffe_ctl->index);
4008         btrfs_release_block_group(block_group, delalloc);
4009 }
4010
4011 static void found_extent_clustered(struct find_free_extent_ctl *ffe_ctl,
4012                                    struct btrfs_key *ins)
4013 {
4014         struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
4015
4016         if (!ffe_ctl->use_cluster && last_ptr) {
4017                 spin_lock(&last_ptr->lock);
4018                 last_ptr->window_start = ins->objectid;
4019                 spin_unlock(&last_ptr->lock);
4020         }
4021 }
4022
4023 static void found_extent(struct find_free_extent_ctl *ffe_ctl,
4024                          struct btrfs_key *ins)
4025 {
4026         switch (ffe_ctl->policy) {
4027         case BTRFS_EXTENT_ALLOC_CLUSTERED:
4028                 found_extent_clustered(ffe_ctl, ins);
4029                 break;
4030         case BTRFS_EXTENT_ALLOC_ZONED:
4031                 /* Nothing to do */
4032                 break;
4033         default:
4034                 BUG();
4035         }
4036 }
4037
4038 static int can_allocate_chunk_zoned(struct btrfs_fs_info *fs_info,
4039                                     struct find_free_extent_ctl *ffe_ctl)
4040 {
4041         /* Block group's activeness is not a requirement for METADATA block groups. */
4042         if (!(ffe_ctl->flags & BTRFS_BLOCK_GROUP_DATA))
4043                 return 0;
4044
4045         /* If we can activate new zone, just allocate a chunk and use it */
4046         if (btrfs_can_activate_zone(fs_info->fs_devices, ffe_ctl->flags))
4047                 return 0;
4048
4049         /*
4050          * We already reached the max active zones. Try to finish one block
4051          * group to make a room for a new block group. This is only possible
4052          * for a data block group because btrfs_zone_finish() may need to wait
4053          * for a running transaction which can cause a deadlock for metadata
4054          * allocation.
4055          */
4056         if (ffe_ctl->flags & BTRFS_BLOCK_GROUP_DATA) {
4057                 int ret = btrfs_zone_finish_one_bg(fs_info);
4058
4059                 if (ret == 1)
4060                         return 0;
4061                 else if (ret < 0)
4062                         return ret;
4063         }
4064
4065         /*
4066          * If we have enough free space left in an already active block group
4067          * and we can't activate any other zone now, do not allow allocating a
4068          * new chunk and let find_free_extent() retry with a smaller size.
4069          */
4070         if (ffe_ctl->max_extent_size >= ffe_ctl->min_alloc_size)
4071                 return -ENOSPC;
4072
4073         /*
4074          * Even min_alloc_size is not left in any block groups. Since we cannot
4075          * activate a new block group, allocating it may not help. Let's tell a
4076          * caller to try again and hope it progress something by writing some
4077          * parts of the region. That is only possible for data block groups,
4078          * where a part of the region can be written.
4079          */
4080         if (ffe_ctl->flags & BTRFS_BLOCK_GROUP_DATA)
4081                 return -EAGAIN;
4082
4083         /*
4084          * We cannot activate a new block group and no enough space left in any
4085          * block groups. So, allocating a new block group may not help. But,
4086          * there is nothing to do anyway, so let's go with it.
4087          */
4088         return 0;
4089 }
4090
4091 static int can_allocate_chunk(struct btrfs_fs_info *fs_info,
4092                               struct find_free_extent_ctl *ffe_ctl)
4093 {
4094         switch (ffe_ctl->policy) {
4095         case BTRFS_EXTENT_ALLOC_CLUSTERED:
4096                 return 0;
4097         case BTRFS_EXTENT_ALLOC_ZONED:
4098                 return can_allocate_chunk_zoned(fs_info, ffe_ctl);
4099         default:
4100                 BUG();
4101         }
4102 }
4103
4104 /*
4105  * Return >0 means caller needs to re-search for free extent
4106  * Return 0 means we have the needed free extent.
4107  * Return <0 means we failed to locate any free extent.
4108  */
4109 static int find_free_extent_update_loop(struct btrfs_fs_info *fs_info,
4110                                         struct btrfs_key *ins,
4111                                         struct find_free_extent_ctl *ffe_ctl,
4112                                         bool full_search)
4113 {
4114         struct btrfs_root *root = fs_info->chunk_root;
4115         int ret;
4116
4117         if ((ffe_ctl->loop == LOOP_CACHING_NOWAIT) &&
4118             ffe_ctl->have_caching_bg && !ffe_ctl->orig_have_caching_bg)
4119                 ffe_ctl->orig_have_caching_bg = true;
4120
4121         if (ins->objectid) {
4122                 found_extent(ffe_ctl, ins);
4123                 return 0;
4124         }
4125
4126         if (ffe_ctl->loop >= LOOP_CACHING_WAIT && ffe_ctl->have_caching_bg)
4127                 return 1;
4128
4129         ffe_ctl->index++;
4130         if (ffe_ctl->index < BTRFS_NR_RAID_TYPES)
4131                 return 1;
4132
4133         /* See the comments for btrfs_loop_type for an explanation of the phases. */
4134         if (ffe_ctl->loop < LOOP_NO_EMPTY_SIZE) {
4135                 ffe_ctl->index = 0;
4136                 /*
4137                  * We want to skip the LOOP_CACHING_WAIT step if we don't have
4138                  * any uncached bgs and we've already done a full search
4139                  * through.
4140                  */
4141                 if (ffe_ctl->loop == LOOP_CACHING_NOWAIT &&
4142                     (!ffe_ctl->orig_have_caching_bg && full_search))
4143                         ffe_ctl->loop++;
4144                 ffe_ctl->loop++;
4145
4146                 if (ffe_ctl->loop == LOOP_ALLOC_CHUNK) {
4147                         struct btrfs_trans_handle *trans;
4148                         int exist = 0;
4149
4150                         /* Check if allocation policy allows to create a new chunk */
4151                         ret = can_allocate_chunk(fs_info, ffe_ctl);
4152                         if (ret)
4153                                 return ret;
4154
4155                         trans = current->journal_info;
4156                         if (trans)
4157                                 exist = 1;
4158                         else
4159                                 trans = btrfs_join_transaction(root);
4160
4161                         if (IS_ERR(trans)) {
4162                                 ret = PTR_ERR(trans);
4163                                 return ret;
4164                         }
4165
4166                         ret = btrfs_chunk_alloc(trans, ffe_ctl->flags,
4167                                                 CHUNK_ALLOC_FORCE_FOR_EXTENT);
4168
4169                         /* Do not bail out on ENOSPC since we can do more. */
4170                         if (ret == -ENOSPC) {
4171                                 ret = 0;
4172                                 ffe_ctl->loop++;
4173                         }
4174                         else if (ret < 0)
4175                                 btrfs_abort_transaction(trans, ret);
4176                         else
4177                                 ret = 0;
4178                         if (!exist)
4179                                 btrfs_end_transaction(trans);
4180                         if (ret)
4181                                 return ret;
4182                 }
4183
4184                 if (ffe_ctl->loop == LOOP_NO_EMPTY_SIZE) {
4185                         if (ffe_ctl->policy != BTRFS_EXTENT_ALLOC_CLUSTERED)
4186                                 return -ENOSPC;
4187
4188                         /*
4189                          * Don't loop again if we already have no empty_size and
4190                          * no empty_cluster.
4191                          */
4192                         if (ffe_ctl->empty_size == 0 &&
4193                             ffe_ctl->empty_cluster == 0)
4194                                 return -ENOSPC;
4195                         ffe_ctl->empty_size = 0;
4196                         ffe_ctl->empty_cluster = 0;
4197                 }
4198                 return 1;
4199         }
4200         return -ENOSPC;
4201 }
4202
4203 static bool find_free_extent_check_size_class(struct find_free_extent_ctl *ffe_ctl,
4204                                               struct btrfs_block_group *bg)
4205 {
4206         if (ffe_ctl->policy == BTRFS_EXTENT_ALLOC_ZONED)
4207                 return true;
4208         if (!btrfs_block_group_should_use_size_class(bg))
4209                 return true;
4210         if (ffe_ctl->loop >= LOOP_WRONG_SIZE_CLASS)
4211                 return true;
4212         if (ffe_ctl->loop >= LOOP_UNSET_SIZE_CLASS &&
4213             bg->size_class == BTRFS_BG_SZ_NONE)
4214                 return true;
4215         return ffe_ctl->size_class == bg->size_class;
4216 }
4217
4218 static int prepare_allocation_clustered(struct btrfs_fs_info *fs_info,
4219                                         struct find_free_extent_ctl *ffe_ctl,
4220                                         struct btrfs_space_info *space_info,
4221                                         struct btrfs_key *ins)
4222 {
4223         /*
4224          * If our free space is heavily fragmented we may not be able to make
4225          * big contiguous allocations, so instead of doing the expensive search
4226          * for free space, simply return ENOSPC with our max_extent_size so we
4227          * can go ahead and search for a more manageable chunk.
4228          *
4229          * If our max_extent_size is large enough for our allocation simply
4230          * disable clustering since we will likely not be able to find enough
4231          * space to create a cluster and induce latency trying.
4232          */
4233         if (space_info->max_extent_size) {
4234                 spin_lock(&space_info->lock);
4235                 if (space_info->max_extent_size &&
4236                     ffe_ctl->num_bytes > space_info->max_extent_size) {
4237                         ins->offset = space_info->max_extent_size;
4238                         spin_unlock(&space_info->lock);
4239                         return -ENOSPC;
4240                 } else if (space_info->max_extent_size) {
4241                         ffe_ctl->use_cluster = false;
4242                 }
4243                 spin_unlock(&space_info->lock);
4244         }
4245
4246         ffe_ctl->last_ptr = fetch_cluster_info(fs_info, space_info,
4247                                                &ffe_ctl->empty_cluster);
4248         if (ffe_ctl->last_ptr) {
4249                 struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
4250
4251                 spin_lock(&last_ptr->lock);
4252                 if (last_ptr->block_group)
4253                         ffe_ctl->hint_byte = last_ptr->window_start;
4254                 if (last_ptr->fragmented) {
4255                         /*
4256                          * We still set window_start so we can keep track of the
4257                          * last place we found an allocation to try and save
4258                          * some time.
4259                          */
4260                         ffe_ctl->hint_byte = last_ptr->window_start;
4261                         ffe_ctl->use_cluster = false;
4262                 }
4263                 spin_unlock(&last_ptr->lock);
4264         }
4265
4266         return 0;
4267 }
4268
4269 static int prepare_allocation_zoned(struct btrfs_fs_info *fs_info,
4270                                     struct find_free_extent_ctl *ffe_ctl)
4271 {
4272         if (ffe_ctl->for_treelog) {
4273                 spin_lock(&fs_info->treelog_bg_lock);
4274                 if (fs_info->treelog_bg)
4275                         ffe_ctl->hint_byte = fs_info->treelog_bg;
4276                 spin_unlock(&fs_info->treelog_bg_lock);
4277         } else if (ffe_ctl->for_data_reloc) {
4278                 spin_lock(&fs_info->relocation_bg_lock);
4279                 if (fs_info->data_reloc_bg)
4280                         ffe_ctl->hint_byte = fs_info->data_reloc_bg;
4281                 spin_unlock(&fs_info->relocation_bg_lock);
4282         } else if (ffe_ctl->flags & BTRFS_BLOCK_GROUP_DATA) {
4283                 struct btrfs_block_group *block_group;
4284
4285                 spin_lock(&fs_info->zone_active_bgs_lock);
4286                 list_for_each_entry(block_group, &fs_info->zone_active_bgs, active_bg_list) {
4287                         /*
4288                          * No lock is OK here because avail is monotinically
4289                          * decreasing, and this is just a hint.
4290                          */
4291                         u64 avail = block_group->zone_capacity - block_group->alloc_offset;
4292
4293                         if (block_group_bits(block_group, ffe_ctl->flags) &&
4294                             avail >= ffe_ctl->num_bytes) {
4295                                 ffe_ctl->hint_byte = block_group->start;
4296                                 break;
4297                         }
4298                 }
4299                 spin_unlock(&fs_info->zone_active_bgs_lock);
4300         }
4301
4302         return 0;
4303 }
4304
4305 static int prepare_allocation(struct btrfs_fs_info *fs_info,
4306                               struct find_free_extent_ctl *ffe_ctl,
4307                               struct btrfs_space_info *space_info,
4308                               struct btrfs_key *ins)
4309 {
4310         switch (ffe_ctl->policy) {
4311         case BTRFS_EXTENT_ALLOC_CLUSTERED:
4312                 return prepare_allocation_clustered(fs_info, ffe_ctl,
4313                                                     space_info, ins);
4314         case BTRFS_EXTENT_ALLOC_ZONED:
4315                 return prepare_allocation_zoned(fs_info, ffe_ctl);
4316         default:
4317                 BUG();
4318         }
4319 }
4320
4321 /*
4322  * walks the btree of allocated extents and find a hole of a given size.
4323  * The key ins is changed to record the hole:
4324  * ins->objectid == start position
4325  * ins->flags = BTRFS_EXTENT_ITEM_KEY
4326  * ins->offset == the size of the hole.
4327  * Any available blocks before search_start are skipped.
4328  *
4329  * If there is no suitable free space, we will record the max size of
4330  * the free space extent currently.
4331  *
4332  * The overall logic and call chain:
4333  *
4334  * find_free_extent()
4335  * |- Iterate through all block groups
4336  * |  |- Get a valid block group
4337  * |  |- Try to do clustered allocation in that block group
4338  * |  |- Try to do unclustered allocation in that block group
4339  * |  |- Check if the result is valid
4340  * |  |  |- If valid, then exit
4341  * |  |- Jump to next block group
4342  * |
4343  * |- Push harder to find free extents
4344  *    |- If not found, re-iterate all block groups
4345  */
4346 static noinline int find_free_extent(struct btrfs_root *root,
4347                                      struct btrfs_key *ins,
4348                                      struct find_free_extent_ctl *ffe_ctl)
4349 {
4350         struct btrfs_fs_info *fs_info = root->fs_info;
4351         int ret = 0;
4352         int cache_block_group_error = 0;
4353         struct btrfs_block_group *block_group = NULL;
4354         struct btrfs_space_info *space_info;
4355         bool full_search = false;
4356
4357         WARN_ON(ffe_ctl->num_bytes < fs_info->sectorsize);
4358
4359         ffe_ctl->search_start = 0;
4360         /* For clustered allocation */
4361         ffe_ctl->empty_cluster = 0;
4362         ffe_ctl->last_ptr = NULL;
4363         ffe_ctl->use_cluster = true;
4364         ffe_ctl->have_caching_bg = false;
4365         ffe_ctl->orig_have_caching_bg = false;
4366         ffe_ctl->index = btrfs_bg_flags_to_raid_index(ffe_ctl->flags);
4367         ffe_ctl->loop = 0;
4368         ffe_ctl->retry_uncached = false;
4369         ffe_ctl->cached = 0;
4370         ffe_ctl->max_extent_size = 0;
4371         ffe_ctl->total_free_space = 0;
4372         ffe_ctl->found_offset = 0;
4373         ffe_ctl->policy = BTRFS_EXTENT_ALLOC_CLUSTERED;
4374         ffe_ctl->size_class = btrfs_calc_block_group_size_class(ffe_ctl->num_bytes);
4375
4376         if (btrfs_is_zoned(fs_info))
4377                 ffe_ctl->policy = BTRFS_EXTENT_ALLOC_ZONED;
4378
4379         ins->type = BTRFS_EXTENT_ITEM_KEY;
4380         ins->objectid = 0;
4381         ins->offset = 0;
4382
4383         trace_find_free_extent(root, ffe_ctl);
4384
4385         space_info = btrfs_find_space_info(fs_info, ffe_ctl->flags);
4386         if (!space_info) {
4387                 btrfs_err(fs_info, "No space info for %llu", ffe_ctl->flags);
4388                 return -ENOSPC;
4389         }
4390
4391         ret = prepare_allocation(fs_info, ffe_ctl, space_info, ins);
4392         if (ret < 0)
4393                 return ret;
4394
4395         ffe_ctl->search_start = max(ffe_ctl->search_start,
4396                                     first_logical_byte(fs_info));
4397         ffe_ctl->search_start = max(ffe_ctl->search_start, ffe_ctl->hint_byte);
4398         if (ffe_ctl->search_start == ffe_ctl->hint_byte) {
4399                 block_group = btrfs_lookup_block_group(fs_info,
4400                                                        ffe_ctl->search_start);
4401                 /*
4402                  * we don't want to use the block group if it doesn't match our
4403                  * allocation bits, or if its not cached.
4404                  *
4405                  * However if we are re-searching with an ideal block group
4406                  * picked out then we don't care that the block group is cached.
4407                  */
4408                 if (block_group && block_group_bits(block_group, ffe_ctl->flags) &&
4409                     block_group->cached != BTRFS_CACHE_NO) {
4410                         down_read(&space_info->groups_sem);
4411                         if (list_empty(&block_group->list) ||
4412                             block_group->ro) {
4413                                 /*
4414                                  * someone is removing this block group,
4415                                  * we can't jump into the have_block_group
4416                                  * target because our list pointers are not
4417                                  * valid
4418                                  */
4419                                 btrfs_put_block_group(block_group);
4420                                 up_read(&space_info->groups_sem);
4421                         } else {
4422                                 ffe_ctl->index = btrfs_bg_flags_to_raid_index(
4423                                                         block_group->flags);
4424                                 btrfs_lock_block_group(block_group,
4425                                                        ffe_ctl->delalloc);
4426                                 ffe_ctl->hinted = true;
4427                                 goto have_block_group;
4428                         }
4429                 } else if (block_group) {
4430                         btrfs_put_block_group(block_group);
4431                 }
4432         }
4433 search:
4434         trace_find_free_extent_search_loop(root, ffe_ctl);
4435         ffe_ctl->have_caching_bg = false;
4436         if (ffe_ctl->index == btrfs_bg_flags_to_raid_index(ffe_ctl->flags) ||
4437             ffe_ctl->index == 0)
4438                 full_search = true;
4439         down_read(&space_info->groups_sem);
4440         list_for_each_entry(block_group,
4441                             &space_info->block_groups[ffe_ctl->index], list) {
4442                 struct btrfs_block_group *bg_ret;
4443
4444                 ffe_ctl->hinted = false;
4445                 /* If the block group is read-only, we can skip it entirely. */
4446                 if (unlikely(block_group->ro)) {
4447                         if (ffe_ctl->for_treelog)
4448                                 btrfs_clear_treelog_bg(block_group);
4449                         if (ffe_ctl->for_data_reloc)
4450                                 btrfs_clear_data_reloc_bg(block_group);
4451                         continue;
4452                 }
4453
4454                 btrfs_grab_block_group(block_group, ffe_ctl->delalloc);
4455                 ffe_ctl->search_start = block_group->start;
4456
4457                 /*
4458                  * this can happen if we end up cycling through all the
4459                  * raid types, but we want to make sure we only allocate
4460                  * for the proper type.
4461                  */
4462                 if (!block_group_bits(block_group, ffe_ctl->flags)) {
4463                         u64 extra = BTRFS_BLOCK_GROUP_DUP |
4464                                 BTRFS_BLOCK_GROUP_RAID1_MASK |
4465                                 BTRFS_BLOCK_GROUP_RAID56_MASK |
4466                                 BTRFS_BLOCK_GROUP_RAID10;
4467
4468                         /*
4469                          * if they asked for extra copies and this block group
4470                          * doesn't provide them, bail.  This does allow us to
4471                          * fill raid0 from raid1.
4472                          */
4473                         if ((ffe_ctl->flags & extra) && !(block_group->flags & extra))
4474                                 goto loop;
4475
4476                         /*
4477                          * This block group has different flags than we want.
4478                          * It's possible that we have MIXED_GROUP flag but no
4479                          * block group is mixed.  Just skip such block group.
4480                          */
4481                         btrfs_release_block_group(block_group, ffe_ctl->delalloc);
4482                         continue;
4483                 }
4484
4485 have_block_group:
4486                 trace_find_free_extent_have_block_group(root, ffe_ctl, block_group);
4487                 ffe_ctl->cached = btrfs_block_group_done(block_group);
4488                 if (unlikely(!ffe_ctl->cached)) {
4489                         ffe_ctl->have_caching_bg = true;
4490                         ret = btrfs_cache_block_group(block_group, false);
4491
4492                         /*
4493                          * If we get ENOMEM here or something else we want to
4494                          * try other block groups, because it may not be fatal.
4495                          * However if we can't find anything else we need to
4496                          * save our return here so that we return the actual
4497                          * error that caused problems, not ENOSPC.
4498                          */
4499                         if (ret < 0) {
4500                                 if (!cache_block_group_error)
4501                                         cache_block_group_error = ret;
4502                                 ret = 0;
4503                                 goto loop;
4504                         }
4505                         ret = 0;
4506                 }
4507
4508                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR)) {
4509                         if (!cache_block_group_error)
4510                                 cache_block_group_error = -EIO;
4511                         goto loop;
4512                 }
4513
4514                 if (!find_free_extent_check_size_class(ffe_ctl, block_group))
4515                         goto loop;
4516
4517                 bg_ret = NULL;
4518                 ret = do_allocation(block_group, ffe_ctl, &bg_ret);
4519                 if (ret > 0)
4520                         goto loop;
4521
4522                 if (bg_ret && bg_ret != block_group) {
4523                         btrfs_release_block_group(block_group, ffe_ctl->delalloc);
4524                         block_group = bg_ret;
4525                 }
4526
4527                 /* Checks */
4528                 ffe_ctl->search_start = round_up(ffe_ctl->found_offset,
4529                                                  fs_info->stripesize);
4530
4531                 /* move on to the next group */
4532                 if (ffe_ctl->search_start + ffe_ctl->num_bytes >
4533                     block_group->start + block_group->length) {
4534                         btrfs_add_free_space_unused(block_group,
4535                                             ffe_ctl->found_offset,
4536                                             ffe_ctl->num_bytes);
4537                         goto loop;
4538                 }
4539
4540                 if (ffe_ctl->found_offset < ffe_ctl->search_start)
4541                         btrfs_add_free_space_unused(block_group,
4542                                         ffe_ctl->found_offset,
4543                                         ffe_ctl->search_start - ffe_ctl->found_offset);
4544
4545                 ret = btrfs_add_reserved_bytes(block_group, ffe_ctl->ram_bytes,
4546                                                ffe_ctl->num_bytes,
4547                                                ffe_ctl->delalloc,
4548                                                ffe_ctl->loop >= LOOP_WRONG_SIZE_CLASS);
4549                 if (ret == -EAGAIN) {
4550                         btrfs_add_free_space_unused(block_group,
4551                                         ffe_ctl->found_offset,
4552                                         ffe_ctl->num_bytes);
4553                         goto loop;
4554                 }
4555                 btrfs_inc_block_group_reservations(block_group);
4556
4557                 /* we are all good, lets return */
4558                 ins->objectid = ffe_ctl->search_start;
4559                 ins->offset = ffe_ctl->num_bytes;
4560
4561                 trace_btrfs_reserve_extent(block_group, ffe_ctl);
4562                 btrfs_release_block_group(block_group, ffe_ctl->delalloc);
4563                 break;
4564 loop:
4565                 if (!ffe_ctl->cached && ffe_ctl->loop > LOOP_CACHING_NOWAIT &&
4566                     !ffe_ctl->retry_uncached) {
4567                         ffe_ctl->retry_uncached = true;
4568                         btrfs_wait_block_group_cache_progress(block_group,
4569                                                 ffe_ctl->num_bytes +
4570                                                 ffe_ctl->empty_cluster +
4571                                                 ffe_ctl->empty_size);
4572                         goto have_block_group;
4573                 }
4574                 release_block_group(block_group, ffe_ctl, ffe_ctl->delalloc);
4575                 cond_resched();
4576         }
4577         up_read(&space_info->groups_sem);
4578
4579         ret = find_free_extent_update_loop(fs_info, ins, ffe_ctl, full_search);
4580         if (ret > 0)
4581                 goto search;
4582
4583         if (ret == -ENOSPC && !cache_block_group_error) {
4584                 /*
4585                  * Use ffe_ctl->total_free_space as fallback if we can't find
4586                  * any contiguous hole.
4587                  */
4588                 if (!ffe_ctl->max_extent_size)
4589                         ffe_ctl->max_extent_size = ffe_ctl->total_free_space;
4590                 spin_lock(&space_info->lock);
4591                 space_info->max_extent_size = ffe_ctl->max_extent_size;
4592                 spin_unlock(&space_info->lock);
4593                 ins->offset = ffe_ctl->max_extent_size;
4594         } else if (ret == -ENOSPC) {
4595                 ret = cache_block_group_error;
4596         }
4597         return ret;
4598 }
4599
4600 /*
4601  * Entry point to the extent allocator. Tries to find a hole that is at least
4602  * as big as @num_bytes.
4603  *
4604  * @root           -    The root that will contain this extent
4605  *
4606  * @ram_bytes      -    The amount of space in ram that @num_bytes take. This
4607  *                      is used for accounting purposes. This value differs
4608  *                      from @num_bytes only in the case of compressed extents.
4609  *
4610  * @num_bytes      -    Number of bytes to allocate on-disk.
4611  *
4612  * @min_alloc_size -    Indicates the minimum amount of space that the
4613  *                      allocator should try to satisfy. In some cases
4614  *                      @num_bytes may be larger than what is required and if
4615  *                      the filesystem is fragmented then allocation fails.
4616  *                      However, the presence of @min_alloc_size gives a
4617  *                      chance to try and satisfy the smaller allocation.
4618  *
4619  * @empty_size     -    A hint that you plan on doing more COW. This is the
4620  *                      size in bytes the allocator should try to find free
4621  *                      next to the block it returns.  This is just a hint and
4622  *                      may be ignored by the allocator.
4623  *
4624  * @hint_byte      -    Hint to the allocator to start searching above the byte
4625  *                      address passed. It might be ignored.
4626  *
4627  * @ins            -    This key is modified to record the found hole. It will
4628  *                      have the following values:
4629  *                      ins->objectid == start position
4630  *                      ins->flags = BTRFS_EXTENT_ITEM_KEY
4631  *                      ins->offset == the size of the hole.
4632  *
4633  * @is_data        -    Boolean flag indicating whether an extent is
4634  *                      allocated for data (true) or metadata (false)
4635  *
4636  * @delalloc       -    Boolean flag indicating whether this allocation is for
4637  *                      delalloc or not. If 'true' data_rwsem of block groups
4638  *                      is going to be acquired.
4639  *
4640  *
4641  * Returns 0 when an allocation succeeded or < 0 when an error occurred. In
4642  * case -ENOSPC is returned then @ins->offset will contain the size of the
4643  * largest available hole the allocator managed to find.
4644  */
4645 int btrfs_reserve_extent(struct btrfs_root *root, u64 ram_bytes,
4646                          u64 num_bytes, u64 min_alloc_size,
4647                          u64 empty_size, u64 hint_byte,
4648                          struct btrfs_key *ins, int is_data, int delalloc)
4649 {
4650         struct btrfs_fs_info *fs_info = root->fs_info;
4651         struct find_free_extent_ctl ffe_ctl = {};
4652         bool final_tried = num_bytes == min_alloc_size;
4653         u64 flags;
4654         int ret;
4655         bool for_treelog = (btrfs_root_id(root) == BTRFS_TREE_LOG_OBJECTID);
4656         bool for_data_reloc = (btrfs_is_data_reloc_root(root) && is_data);
4657
4658         flags = get_alloc_profile_by_root(root, is_data);
4659 again:
4660         WARN_ON(num_bytes < fs_info->sectorsize);
4661
4662         ffe_ctl.ram_bytes = ram_bytes;
4663         ffe_ctl.num_bytes = num_bytes;
4664         ffe_ctl.min_alloc_size = min_alloc_size;
4665         ffe_ctl.empty_size = empty_size;
4666         ffe_ctl.flags = flags;
4667         ffe_ctl.delalloc = delalloc;
4668         ffe_ctl.hint_byte = hint_byte;
4669         ffe_ctl.for_treelog = for_treelog;
4670         ffe_ctl.for_data_reloc = for_data_reloc;
4671
4672         ret = find_free_extent(root, ins, &ffe_ctl);
4673         if (!ret && !is_data) {
4674                 btrfs_dec_block_group_reservations(fs_info, ins->objectid);
4675         } else if (ret == -ENOSPC) {
4676                 if (!final_tried && ins->offset) {
4677                         num_bytes = min(num_bytes >> 1, ins->offset);
4678                         num_bytes = round_down(num_bytes,
4679                                                fs_info->sectorsize);
4680                         num_bytes = max(num_bytes, min_alloc_size);
4681                         ram_bytes = num_bytes;
4682                         if (num_bytes == min_alloc_size)
4683                                 final_tried = true;
4684                         goto again;
4685                 } else if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
4686                         struct btrfs_space_info *sinfo;
4687
4688                         sinfo = btrfs_find_space_info(fs_info, flags);
4689                         btrfs_err(fs_info,
4690         "allocation failed flags %llu, wanted %llu tree-log %d, relocation: %d",
4691                                   flags, num_bytes, for_treelog, for_data_reloc);
4692                         if (sinfo)
4693                                 btrfs_dump_space_info(fs_info, sinfo,
4694                                                       num_bytes, 1);
4695                 }
4696         }
4697
4698         return ret;
4699 }
4700
4701 int btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
4702                                u64 start, u64 len, int delalloc)
4703 {
4704         struct btrfs_block_group *cache;
4705
4706         cache = btrfs_lookup_block_group(fs_info, start);
4707         if (!cache) {
4708                 btrfs_err(fs_info, "Unable to find block group for %llu",
4709                           start);
4710                 return -ENOSPC;
4711         }
4712
4713         btrfs_add_free_space(cache, start, len);
4714         btrfs_free_reserved_bytes(cache, len, delalloc);
4715         trace_btrfs_reserved_extent_free(fs_info, start, len);
4716
4717         btrfs_put_block_group(cache);
4718         return 0;
4719 }
4720
4721 int btrfs_pin_reserved_extent(struct btrfs_trans_handle *trans,
4722                               const struct extent_buffer *eb)
4723 {
4724         struct btrfs_block_group *cache;
4725         int ret = 0;
4726
4727         cache = btrfs_lookup_block_group(trans->fs_info, eb->start);
4728         if (!cache) {
4729                 btrfs_err(trans->fs_info, "unable to find block group for %llu",
4730                           eb->start);
4731                 return -ENOSPC;
4732         }
4733
4734         ret = pin_down_extent(trans, cache, eb->start, eb->len, 1);
4735         btrfs_put_block_group(cache);
4736         return ret;
4737 }
4738
4739 static int alloc_reserved_extent(struct btrfs_trans_handle *trans, u64 bytenr,
4740                                  u64 num_bytes)
4741 {
4742         struct btrfs_fs_info *fs_info = trans->fs_info;
4743         int ret;
4744
4745         ret = remove_from_free_space_tree(trans, bytenr, num_bytes);
4746         if (ret)
4747                 return ret;
4748
4749         ret = btrfs_update_block_group(trans, bytenr, num_bytes, true);
4750         if (ret) {
4751                 ASSERT(!ret);
4752                 btrfs_err(fs_info, "update block group failed for %llu %llu",
4753                           bytenr, num_bytes);
4754                 return ret;
4755         }
4756
4757         trace_btrfs_reserved_extent_alloc(fs_info, bytenr, num_bytes);
4758         return 0;
4759 }
4760
4761 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
4762                                       u64 parent, u64 root_objectid,
4763                                       u64 flags, u64 owner, u64 offset,
4764                                       struct btrfs_key *ins, int ref_mod, u64 oref_root)
4765 {
4766         struct btrfs_fs_info *fs_info = trans->fs_info;
4767         struct btrfs_root *extent_root;
4768         int ret;
4769         struct btrfs_extent_item *extent_item;
4770         struct btrfs_extent_owner_ref *oref;
4771         struct btrfs_extent_inline_ref *iref;
4772         struct btrfs_path *path;
4773         struct extent_buffer *leaf;
4774         int type;
4775         u32 size;
4776         const bool simple_quota = (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_SIMPLE);
4777
4778         if (parent > 0)
4779                 type = BTRFS_SHARED_DATA_REF_KEY;
4780         else
4781                 type = BTRFS_EXTENT_DATA_REF_KEY;
4782
4783         size = sizeof(*extent_item);
4784         if (simple_quota)
4785                 size += btrfs_extent_inline_ref_size(BTRFS_EXTENT_OWNER_REF_KEY);
4786         size += btrfs_extent_inline_ref_size(type);
4787
4788         path = btrfs_alloc_path();
4789         if (!path)
4790                 return -ENOMEM;
4791
4792         extent_root = btrfs_extent_root(fs_info, ins->objectid);
4793         ret = btrfs_insert_empty_item(trans, extent_root, path, ins, size);
4794         if (ret) {
4795                 btrfs_free_path(path);
4796                 return ret;
4797         }
4798
4799         leaf = path->nodes[0];
4800         extent_item = btrfs_item_ptr(leaf, path->slots[0],
4801                                      struct btrfs_extent_item);
4802         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
4803         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
4804         btrfs_set_extent_flags(leaf, extent_item,
4805                                flags | BTRFS_EXTENT_FLAG_DATA);
4806
4807         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
4808         if (simple_quota) {
4809                 btrfs_set_extent_inline_ref_type(leaf, iref, BTRFS_EXTENT_OWNER_REF_KEY);
4810                 oref = (struct btrfs_extent_owner_ref *)(&iref->offset);
4811                 btrfs_set_extent_owner_ref_root_id(leaf, oref, oref_root);
4812                 iref = (struct btrfs_extent_inline_ref *)(oref + 1);
4813         }
4814         btrfs_set_extent_inline_ref_type(leaf, iref, type);
4815
4816         if (parent > 0) {
4817                 struct btrfs_shared_data_ref *ref;
4818                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
4819                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
4820                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
4821         } else {
4822                 struct btrfs_extent_data_ref *ref;
4823                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
4824                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
4825                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
4826                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
4827                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
4828         }
4829
4830         btrfs_mark_buffer_dirty(trans, path->nodes[0]);
4831         btrfs_free_path(path);
4832
4833         return alloc_reserved_extent(trans, ins->objectid, ins->offset);
4834 }
4835
4836 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
4837                                      struct btrfs_delayed_ref_node *node,
4838                                      struct btrfs_delayed_extent_op *extent_op)
4839 {
4840         struct btrfs_fs_info *fs_info = trans->fs_info;
4841         struct btrfs_root *extent_root;
4842         int ret;
4843         struct btrfs_extent_item *extent_item;
4844         struct btrfs_key extent_key;
4845         struct btrfs_tree_block_info *block_info;
4846         struct btrfs_extent_inline_ref *iref;
4847         struct btrfs_path *path;
4848         struct extent_buffer *leaf;
4849         u32 size = sizeof(*extent_item) + sizeof(*iref);
4850         const u64 flags = (extent_op ? extent_op->flags_to_set : 0);
4851         /* The owner of a tree block is the level. */
4852         int level = btrfs_delayed_ref_owner(node);
4853         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
4854
4855         extent_key.objectid = node->bytenr;
4856         if (skinny_metadata) {
4857                 /* The owner of a tree block is the level. */
4858                 extent_key.offset = level;
4859                 extent_key.type = BTRFS_METADATA_ITEM_KEY;
4860         } else {
4861                 extent_key.offset = node->num_bytes;
4862                 extent_key.type = BTRFS_EXTENT_ITEM_KEY;
4863                 size += sizeof(*block_info);
4864         }
4865
4866         path = btrfs_alloc_path();
4867         if (!path)
4868                 return -ENOMEM;
4869
4870         extent_root = btrfs_extent_root(fs_info, extent_key.objectid);
4871         ret = btrfs_insert_empty_item(trans, extent_root, path, &extent_key,
4872                                       size);
4873         if (ret) {
4874                 btrfs_free_path(path);
4875                 return ret;
4876         }
4877
4878         leaf = path->nodes[0];
4879         extent_item = btrfs_item_ptr(leaf, path->slots[0],
4880                                      struct btrfs_extent_item);
4881         btrfs_set_extent_refs(leaf, extent_item, 1);
4882         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
4883         btrfs_set_extent_flags(leaf, extent_item,
4884                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
4885
4886         if (skinny_metadata) {
4887                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
4888         } else {
4889                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
4890                 btrfs_set_tree_block_key(leaf, block_info, &extent_op->key);
4891                 btrfs_set_tree_block_level(leaf, block_info, level);
4892                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
4893         }
4894
4895         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY) {
4896                 btrfs_set_extent_inline_ref_type(leaf, iref,
4897                                                  BTRFS_SHARED_BLOCK_REF_KEY);
4898                 btrfs_set_extent_inline_ref_offset(leaf, iref, node->parent);
4899         } else {
4900                 btrfs_set_extent_inline_ref_type(leaf, iref,
4901                                                  BTRFS_TREE_BLOCK_REF_KEY);
4902                 btrfs_set_extent_inline_ref_offset(leaf, iref, node->ref_root);
4903         }
4904
4905         btrfs_mark_buffer_dirty(trans, leaf);
4906         btrfs_free_path(path);
4907
4908         return alloc_reserved_extent(trans, node->bytenr, fs_info->nodesize);
4909 }
4910
4911 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
4912                                      struct btrfs_root *root, u64 owner,
4913                                      u64 offset, u64 ram_bytes,
4914                                      struct btrfs_key *ins)
4915 {
4916         struct btrfs_ref generic_ref = {
4917                 .action = BTRFS_ADD_DELAYED_EXTENT,
4918                 .bytenr = ins->objectid,
4919                 .num_bytes = ins->offset,
4920                 .owning_root = btrfs_root_id(root),
4921                 .ref_root = btrfs_root_id(root),
4922         };
4923
4924         ASSERT(generic_ref.ref_root != BTRFS_TREE_LOG_OBJECTID);
4925
4926         if (btrfs_is_data_reloc_root(root) && is_fstree(root->relocation_src_root))
4927                 generic_ref.owning_root = root->relocation_src_root;
4928
4929         btrfs_init_data_ref(&generic_ref, owner, offset, 0, false);
4930         btrfs_ref_tree_mod(root->fs_info, &generic_ref);
4931
4932         return btrfs_add_delayed_data_ref(trans, &generic_ref, ram_bytes);
4933 }
4934
4935 /*
4936  * this is used by the tree logging recovery code.  It records that
4937  * an extent has been allocated and makes sure to clear the free
4938  * space cache bits as well
4939  */
4940 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
4941                                    u64 root_objectid, u64 owner, u64 offset,
4942                                    struct btrfs_key *ins)
4943 {
4944         struct btrfs_fs_info *fs_info = trans->fs_info;
4945         int ret;
4946         struct btrfs_block_group *block_group;
4947         struct btrfs_space_info *space_info;
4948         struct btrfs_squota_delta delta = {
4949                 .root = root_objectid,
4950                 .num_bytes = ins->offset,
4951                 .generation = trans->transid,
4952                 .is_data = true,
4953                 .is_inc = true,
4954         };
4955
4956         /*
4957          * Mixed block groups will exclude before processing the log so we only
4958          * need to do the exclude dance if this fs isn't mixed.
4959          */
4960         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS)) {
4961                 ret = __exclude_logged_extent(fs_info, ins->objectid,
4962                                               ins->offset);
4963                 if (ret)
4964                         return ret;
4965         }
4966
4967         block_group = btrfs_lookup_block_group(fs_info, ins->objectid);
4968         if (!block_group)
4969                 return -EINVAL;
4970
4971         space_info = block_group->space_info;
4972         spin_lock(&space_info->lock);
4973         spin_lock(&block_group->lock);
4974         space_info->bytes_reserved += ins->offset;
4975         block_group->reserved += ins->offset;
4976         spin_unlock(&block_group->lock);
4977         spin_unlock(&space_info->lock);
4978
4979         ret = alloc_reserved_file_extent(trans, 0, root_objectid, 0, owner,
4980                                          offset, ins, 1, root_objectid);
4981         if (ret)
4982                 btrfs_pin_extent(trans, ins->objectid, ins->offset, 1);
4983         ret = btrfs_record_squota_delta(fs_info, &delta);
4984         btrfs_put_block_group(block_group);
4985         return ret;
4986 }
4987
4988 #ifdef CONFIG_BTRFS_DEBUG
4989 /*
4990  * Extra safety check in case the extent tree is corrupted and extent allocator
4991  * chooses to use a tree block which is already used and locked.
4992  */
4993 static bool check_eb_lock_owner(const struct extent_buffer *eb)
4994 {
4995         if (eb->lock_owner == current->pid) {
4996                 btrfs_err_rl(eb->fs_info,
4997 "tree block %llu owner %llu already locked by pid=%d, extent tree corruption detected",
4998                              eb->start, btrfs_header_owner(eb), current->pid);
4999                 return true;
5000         }
5001         return false;
5002 }
5003 #else
5004 static bool check_eb_lock_owner(struct extent_buffer *eb)
5005 {
5006         return false;
5007 }
5008 #endif
5009
5010 static struct extent_buffer *
5011 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
5012                       u64 bytenr, int level, u64 owner,
5013                       enum btrfs_lock_nesting nest)
5014 {
5015         struct btrfs_fs_info *fs_info = root->fs_info;
5016         struct extent_buffer *buf;
5017         u64 lockdep_owner = owner;
5018
5019         buf = btrfs_find_create_tree_block(fs_info, bytenr, owner, level);
5020         if (IS_ERR(buf))
5021                 return buf;
5022
5023         if (check_eb_lock_owner(buf)) {
5024                 free_extent_buffer(buf);
5025                 return ERR_PTR(-EUCLEAN);
5026         }
5027
5028         /*
5029          * The reloc trees are just snapshots, so we need them to appear to be
5030          * just like any other fs tree WRT lockdep.
5031          *
5032          * The exception however is in replace_path() in relocation, where we
5033          * hold the lock on the original fs root and then search for the reloc
5034          * root.  At that point we need to make sure any reloc root buffers are
5035          * set to the BTRFS_TREE_RELOC_OBJECTID lockdep class in order to make
5036          * lockdep happy.
5037          */
5038         if (lockdep_owner == BTRFS_TREE_RELOC_OBJECTID &&
5039             !test_bit(BTRFS_ROOT_RESET_LOCKDEP_CLASS, &root->state))
5040                 lockdep_owner = BTRFS_FS_TREE_OBJECTID;
5041
5042         /* btrfs_clear_buffer_dirty() accesses generation field. */
5043         btrfs_set_header_generation(buf, trans->transid);
5044
5045         /*
5046          * This needs to stay, because we could allocate a freed block from an
5047          * old tree into a new tree, so we need to make sure this new block is
5048          * set to the appropriate level and owner.
5049          */
5050         btrfs_set_buffer_lockdep_class(lockdep_owner, buf, level);
5051
5052         btrfs_tree_lock_nested(buf, nest);
5053         btrfs_clear_buffer_dirty(trans, buf);
5054         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
5055         clear_bit(EXTENT_BUFFER_ZONED_ZEROOUT, &buf->bflags);
5056
5057         set_extent_buffer_uptodate(buf);
5058
5059         memzero_extent_buffer(buf, 0, sizeof(struct btrfs_header));
5060         btrfs_set_header_level(buf, level);
5061         btrfs_set_header_bytenr(buf, buf->start);
5062         btrfs_set_header_generation(buf, trans->transid);
5063         btrfs_set_header_backref_rev(buf, BTRFS_MIXED_BACKREF_REV);
5064         btrfs_set_header_owner(buf, owner);
5065         write_extent_buffer_fsid(buf, fs_info->fs_devices->metadata_uuid);
5066         write_extent_buffer_chunk_tree_uuid(buf, fs_info->chunk_tree_uuid);
5067         if (btrfs_root_id(root) == BTRFS_TREE_LOG_OBJECTID) {
5068                 buf->log_index = root->log_transid % 2;
5069                 /*
5070                  * we allow two log transactions at a time, use different
5071                  * EXTENT bit to differentiate dirty pages.
5072                  */
5073                 if (buf->log_index == 0)
5074                         set_extent_bit(&root->dirty_log_pages, buf->start,
5075                                        buf->start + buf->len - 1,
5076                                        EXTENT_DIRTY, NULL);
5077                 else
5078                         set_extent_bit(&root->dirty_log_pages, buf->start,
5079                                        buf->start + buf->len - 1,
5080                                        EXTENT_NEW, NULL);
5081         } else {
5082                 buf->log_index = -1;
5083                 set_extent_bit(&trans->transaction->dirty_pages, buf->start,
5084                                buf->start + buf->len - 1, EXTENT_DIRTY, NULL);
5085         }
5086         /* this returns a buffer locked for blocking */
5087         return buf;
5088 }
5089
5090 /*
5091  * finds a free extent and does all the dirty work required for allocation
5092  * returns the tree buffer or an ERR_PTR on error.
5093  */
5094 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
5095                                              struct btrfs_root *root,
5096                                              u64 parent, u64 root_objectid,
5097                                              const struct btrfs_disk_key *key,
5098                                              int level, u64 hint,
5099                                              u64 empty_size,
5100                                              u64 reloc_src_root,
5101                                              enum btrfs_lock_nesting nest)
5102 {
5103         struct btrfs_fs_info *fs_info = root->fs_info;
5104         struct btrfs_key ins;
5105         struct btrfs_block_rsv *block_rsv;
5106         struct extent_buffer *buf;
5107         u64 flags = 0;
5108         int ret;
5109         u32 blocksize = fs_info->nodesize;
5110         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
5111         u64 owning_root;
5112
5113 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
5114         if (btrfs_is_testing(fs_info)) {
5115                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
5116                                             level, root_objectid, nest);
5117                 if (!IS_ERR(buf))
5118                         root->alloc_bytenr += blocksize;
5119                 return buf;
5120         }
5121 #endif
5122
5123         block_rsv = btrfs_use_block_rsv(trans, root, blocksize);
5124         if (IS_ERR(block_rsv))
5125                 return ERR_CAST(block_rsv);
5126
5127         ret = btrfs_reserve_extent(root, blocksize, blocksize, blocksize,
5128                                    empty_size, hint, &ins, 0, 0);
5129         if (ret)
5130                 goto out_unuse;
5131
5132         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level,
5133                                     root_objectid, nest);
5134         if (IS_ERR(buf)) {
5135                 ret = PTR_ERR(buf);
5136                 goto out_free_reserved;
5137         }
5138         owning_root = btrfs_header_owner(buf);
5139
5140         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
5141                 if (parent == 0)
5142                         parent = ins.objectid;
5143                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
5144                 owning_root = reloc_src_root;
5145         } else
5146                 BUG_ON(parent > 0);
5147
5148         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
5149                 struct btrfs_delayed_extent_op *extent_op;
5150                 struct btrfs_ref generic_ref = {
5151                         .action = BTRFS_ADD_DELAYED_EXTENT,
5152                         .bytenr = ins.objectid,
5153                         .num_bytes = ins.offset,
5154                         .parent = parent,
5155                         .owning_root = owning_root,
5156                         .ref_root = root_objectid,
5157                 };
5158
5159                 if (!skinny_metadata || flags != 0) {
5160                         extent_op = btrfs_alloc_delayed_extent_op();
5161                         if (!extent_op) {
5162                                 ret = -ENOMEM;
5163                                 goto out_free_buf;
5164                         }
5165                         if (key)
5166                                 memcpy(&extent_op->key, key, sizeof(extent_op->key));
5167                         else
5168                                 memset(&extent_op->key, 0, sizeof(extent_op->key));
5169                         extent_op->flags_to_set = flags;
5170                         extent_op->update_key = (skinny_metadata ? false : true);
5171                         extent_op->update_flags = (flags != 0);
5172                 } else {
5173                         extent_op = NULL;
5174                 }
5175
5176                 btrfs_init_tree_ref(&generic_ref, level, btrfs_root_id(root), false);
5177                 btrfs_ref_tree_mod(fs_info, &generic_ref);
5178                 ret = btrfs_add_delayed_tree_ref(trans, &generic_ref, extent_op);
5179                 if (ret) {
5180                         btrfs_free_delayed_extent_op(extent_op);
5181                         goto out_free_buf;
5182                 }
5183         }
5184         return buf;
5185
5186 out_free_buf:
5187         btrfs_tree_unlock(buf);
5188         free_extent_buffer(buf);
5189 out_free_reserved:
5190         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 0);
5191 out_unuse:
5192         btrfs_unuse_block_rsv(fs_info, block_rsv, blocksize);
5193         return ERR_PTR(ret);
5194 }
5195
5196 struct walk_control {
5197         u64 refs[BTRFS_MAX_LEVEL];
5198         u64 flags[BTRFS_MAX_LEVEL];
5199         struct btrfs_key update_progress;
5200         struct btrfs_key drop_progress;
5201         int drop_level;
5202         int stage;
5203         int level;
5204         int shared_level;
5205         int update_ref;
5206         int keep_locks;
5207         int reada_slot;
5208         int reada_count;
5209         int restarted;
5210         /* Indicate that extent info needs to be looked up when walking the tree. */
5211         int lookup_info;
5212 };
5213
5214 /*
5215  * This is our normal stage.  We are traversing blocks the current snapshot owns
5216  * and we are dropping any of our references to any children we are able to, and
5217  * then freeing the block once we've processed all of the children.
5218  */
5219 #define DROP_REFERENCE  1
5220
5221 /*
5222  * We enter this stage when we have to walk into a child block (meaning we can't
5223  * simply drop our reference to it from our current parent node) and there are
5224  * more than one reference on it.  If we are the owner of any of the children
5225  * blocks from the current parent node then we have to do the FULL_BACKREF dance
5226  * on them in order to drop our normal ref and add the shared ref.
5227  */
5228 #define UPDATE_BACKREF  2
5229
5230 /*
5231  * Decide if we need to walk down into this node to adjust the references.
5232  *
5233  * @root:       the root we are currently deleting
5234  * @wc:         the walk control for this deletion
5235  * @eb:         the parent eb that we're currently visiting
5236  * @refs:       the number of refs for wc->level - 1
5237  * @flags:      the flags for wc->level - 1
5238  * @slot:       the slot in the eb that we're currently checking
5239  *
5240  * This is meant to be called when we're evaluating if a node we point to at
5241  * wc->level should be read and walked into, or if we can simply delete our
5242  * reference to it.  We return true if we should walk into the node, false if we
5243  * can skip it.
5244  *
5245  * We have assertions in here to make sure this is called correctly.  We assume
5246  * that sanity checking on the blocks read to this point has been done, so any
5247  * corrupted file systems must have been caught before calling this function.
5248  */
5249 static bool visit_node_for_delete(struct btrfs_root *root, struct walk_control *wc,
5250                                   struct extent_buffer *eb, u64 flags, int slot)
5251 {
5252         struct btrfs_key key;
5253         u64 generation;
5254         int level = wc->level;
5255
5256         ASSERT(level > 0);
5257         ASSERT(wc->refs[level - 1] > 0);
5258
5259         /*
5260          * The update backref stage we only want to skip if we already have
5261          * FULL_BACKREF set, otherwise we need to read.
5262          */
5263         if (wc->stage == UPDATE_BACKREF) {
5264                 if (level == 1 && flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)
5265                         return false;
5266                 return true;
5267         }
5268
5269         /*
5270          * We're the last ref on this block, we must walk into it and process
5271          * any refs it's pointing at.
5272          */
5273         if (wc->refs[level - 1] == 1)
5274                 return true;
5275
5276         /*
5277          * If we're already FULL_BACKREF then we know we can just drop our
5278          * current reference.
5279          */
5280         if (level == 1 && flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)
5281                 return false;
5282
5283         /*
5284          * This block is older than our creation generation, we can drop our
5285          * reference to it.
5286          */
5287         generation = btrfs_node_ptr_generation(eb, slot);
5288         if (!wc->update_ref || generation <= btrfs_root_origin_generation(root))
5289                 return false;
5290
5291         /*
5292          * This block was processed from a previous snapshot deletion run, we
5293          * can skip it.
5294          */
5295         btrfs_node_key_to_cpu(eb, &key, slot);
5296         if (btrfs_comp_cpu_keys(&key, &wc->update_progress) < 0)
5297                 return false;
5298
5299         /* All other cases we need to wander into the node. */
5300         return true;
5301 }
5302
5303 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
5304                                      struct btrfs_root *root,
5305                                      struct walk_control *wc,
5306                                      struct btrfs_path *path)
5307 {
5308         struct btrfs_fs_info *fs_info = root->fs_info;
5309         u64 bytenr;
5310         u64 generation;
5311         u64 refs;
5312         u64 flags;
5313         u32 nritems;
5314         struct extent_buffer *eb;
5315         int ret;
5316         int slot;
5317         int nread = 0;
5318
5319         if (path->slots[wc->level] < wc->reada_slot) {
5320                 wc->reada_count = wc->reada_count * 2 / 3;
5321                 wc->reada_count = max(wc->reada_count, 2);
5322         } else {
5323                 wc->reada_count = wc->reada_count * 3 / 2;
5324                 wc->reada_count = min_t(int, wc->reada_count,
5325                                         BTRFS_NODEPTRS_PER_BLOCK(fs_info));
5326         }
5327
5328         eb = path->nodes[wc->level];
5329         nritems = btrfs_header_nritems(eb);
5330
5331         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
5332                 if (nread >= wc->reada_count)
5333                         break;
5334
5335                 cond_resched();
5336                 bytenr = btrfs_node_blockptr(eb, slot);
5337                 generation = btrfs_node_ptr_generation(eb, slot);
5338
5339                 if (slot == path->slots[wc->level])
5340                         goto reada;
5341
5342                 if (wc->stage == UPDATE_BACKREF &&
5343                     generation <= btrfs_root_origin_generation(root))
5344                         continue;
5345
5346                 /* We don't lock the tree block, it's OK to be racy here */
5347                 ret = btrfs_lookup_extent_info(trans, fs_info, bytenr,
5348                                                wc->level - 1, 1, &refs,
5349                                                &flags, NULL);
5350                 /* We don't care about errors in readahead. */
5351                 if (ret < 0)
5352                         continue;
5353
5354                 /*
5355                  * This could be racey, it's conceivable that we raced and end
5356                  * up with a bogus refs count, if that's the case just skip, if
5357                  * we are actually corrupt we will notice when we look up
5358                  * everything again with our locks.
5359                  */
5360                 if (refs == 0)
5361                         continue;
5362
5363                 /* If we don't need to visit this node don't reada. */
5364                 if (!visit_node_for_delete(root, wc, eb, flags, slot))
5365                         continue;
5366 reada:
5367                 btrfs_readahead_node_child(eb, slot);
5368                 nread++;
5369         }
5370         wc->reada_slot = slot;
5371 }
5372
5373 /*
5374  * helper to process tree block while walking down the tree.
5375  *
5376  * when wc->stage == UPDATE_BACKREF, this function updates
5377  * back refs for pointers in the block.
5378  *
5379  * NOTE: return value 1 means we should stop walking down.
5380  */
5381 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
5382                                    struct btrfs_root *root,
5383                                    struct btrfs_path *path,
5384                                    struct walk_control *wc)
5385 {
5386         struct btrfs_fs_info *fs_info = root->fs_info;
5387         int level = wc->level;
5388         struct extent_buffer *eb = path->nodes[level];
5389         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
5390         int ret;
5391
5392         if (wc->stage == UPDATE_BACKREF && btrfs_header_owner(eb) != btrfs_root_id(root))
5393                 return 1;
5394
5395         /*
5396          * when reference count of tree block is 1, it won't increase
5397          * again. once full backref flag is set, we never clear it.
5398          */
5399         if (wc->lookup_info &&
5400             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
5401              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
5402                 ASSERT(path->locks[level]);
5403                 ret = btrfs_lookup_extent_info(trans, fs_info,
5404                                                eb->start, level, 1,
5405                                                &wc->refs[level],
5406                                                &wc->flags[level],
5407                                                NULL);
5408                 if (ret)
5409                         return ret;
5410                 if (unlikely(wc->refs[level] == 0)) {
5411                         btrfs_err(fs_info, "bytenr %llu has 0 references, expect > 0",
5412                                   eb->start);
5413                         return -EUCLEAN;
5414                 }
5415         }
5416
5417         if (wc->stage == DROP_REFERENCE) {
5418                 if (wc->refs[level] > 1)
5419                         return 1;
5420
5421                 if (path->locks[level] && !wc->keep_locks) {
5422                         btrfs_tree_unlock_rw(eb, path->locks[level]);
5423                         path->locks[level] = 0;
5424                 }
5425                 return 0;
5426         }
5427
5428         /* wc->stage == UPDATE_BACKREF */
5429         if (!(wc->flags[level] & flag)) {
5430                 ASSERT(path->locks[level]);
5431                 ret = btrfs_inc_ref(trans, root, eb, 1);
5432                 if (ret) {
5433                         btrfs_abort_transaction(trans, ret);
5434                         return ret;
5435                 }
5436                 ret = btrfs_dec_ref(trans, root, eb, 0);
5437                 if (ret) {
5438                         btrfs_abort_transaction(trans, ret);
5439                         return ret;
5440                 }
5441                 ret = btrfs_set_disk_extent_flags(trans, eb, flag);
5442                 if (ret) {
5443                         btrfs_abort_transaction(trans, ret);
5444                         return ret;
5445                 }
5446                 wc->flags[level] |= flag;
5447         }
5448
5449         /*
5450          * the block is shared by multiple trees, so it's not good to
5451          * keep the tree lock
5452          */
5453         if (path->locks[level] && level > 0) {
5454                 btrfs_tree_unlock_rw(eb, path->locks[level]);
5455                 path->locks[level] = 0;
5456         }
5457         return 0;
5458 }
5459
5460 /*
5461  * This is used to verify a ref exists for this root to deal with a bug where we
5462  * would have a drop_progress key that hadn't been updated properly.
5463  */
5464 static int check_ref_exists(struct btrfs_trans_handle *trans,
5465                             struct btrfs_root *root, u64 bytenr, u64 parent,
5466                             int level)
5467 {
5468         struct btrfs_delayed_ref_root *delayed_refs;
5469         struct btrfs_delayed_ref_head *head;
5470         struct btrfs_path *path;
5471         struct btrfs_extent_inline_ref *iref;
5472         int ret;
5473         bool exists = false;
5474
5475         path = btrfs_alloc_path();
5476         if (!path)
5477                 return -ENOMEM;
5478 again:
5479         ret = lookup_extent_backref(trans, path, &iref, bytenr,
5480                                     root->fs_info->nodesize, parent,
5481                                     btrfs_root_id(root), level, 0);
5482         if (ret != -ENOENT) {
5483                 /*
5484                  * If we get 0 then we found our reference, return 1, else
5485                  * return the error if it's not -ENOENT;
5486                  */
5487                 btrfs_free_path(path);
5488                 return (ret < 0 ) ? ret : 1;
5489         }
5490
5491         /*
5492          * We could have a delayed ref with this reference, so look it up while
5493          * we're holding the path open to make sure we don't race with the
5494          * delayed ref running.
5495          */
5496         delayed_refs = &trans->transaction->delayed_refs;
5497         spin_lock(&delayed_refs->lock);
5498         head = btrfs_find_delayed_ref_head(root->fs_info, delayed_refs, bytenr);
5499         if (!head)
5500                 goto out;
5501         if (!mutex_trylock(&head->mutex)) {
5502                 /*
5503                  * We're contended, means that the delayed ref is running, get a
5504                  * reference and wait for the ref head to be complete and then
5505                  * try again.
5506                  */
5507                 refcount_inc(&head->refs);
5508                 spin_unlock(&delayed_refs->lock);
5509
5510                 btrfs_release_path(path);
5511
5512                 mutex_lock(&head->mutex);
5513                 mutex_unlock(&head->mutex);
5514                 btrfs_put_delayed_ref_head(head);
5515                 goto again;
5516         }
5517
5518         exists = btrfs_find_delayed_tree_ref(head, root->root_key.objectid, parent);
5519         mutex_unlock(&head->mutex);
5520 out:
5521         spin_unlock(&delayed_refs->lock);
5522         btrfs_free_path(path);
5523         return exists ? 1 : 0;
5524 }
5525
5526 /*
5527  * We may not have an uptodate block, so if we are going to walk down into this
5528  * block we need to drop the lock, read it off of the disk, re-lock it and
5529  * return to continue dropping the snapshot.
5530  */
5531 static int check_next_block_uptodate(struct btrfs_trans_handle *trans,
5532                                      struct btrfs_root *root,
5533                                      struct btrfs_path *path,
5534                                      struct walk_control *wc,
5535                                      struct extent_buffer *next)
5536 {
5537         struct btrfs_tree_parent_check check = { 0 };
5538         u64 generation;
5539         int level = wc->level;
5540         int ret;
5541
5542         btrfs_assert_tree_write_locked(next);
5543
5544         generation = btrfs_node_ptr_generation(path->nodes[level], path->slots[level]);
5545
5546         if (btrfs_buffer_uptodate(next, generation, 0))
5547                 return 0;
5548
5549         check.level = level - 1;
5550         check.transid = generation;
5551         check.owner_root = btrfs_root_id(root);
5552         check.has_first_key = true;
5553         btrfs_node_key_to_cpu(path->nodes[level], &check.first_key, path->slots[level]);
5554
5555         btrfs_tree_unlock(next);
5556         if (level == 1)
5557                 reada_walk_down(trans, root, wc, path);
5558         ret = btrfs_read_extent_buffer(next, &check);
5559         if (ret) {
5560                 free_extent_buffer(next);
5561                 return ret;
5562         }
5563         btrfs_tree_lock(next);
5564         wc->lookup_info = 1;
5565         return 0;
5566 }
5567
5568 /*
5569  * If we determine that we don't have to visit wc->level - 1 then we need to
5570  * determine if we can drop our reference.
5571  *
5572  * If we are UPDATE_BACKREF then we will not, we need to update our backrefs.
5573  *
5574  * If we are DROP_REFERENCE this will figure out if we need to drop our current
5575  * reference, skipping it if we dropped it from a previous incompleted drop, or
5576  * dropping it if we still have a reference to it.
5577  */
5578 static int maybe_drop_reference(struct btrfs_trans_handle *trans, struct btrfs_root *root,
5579                                 struct btrfs_path *path, struct walk_control *wc,
5580                                 struct extent_buffer *next, u64 owner_root)
5581 {
5582         struct btrfs_ref ref = {
5583                 .action = BTRFS_DROP_DELAYED_REF,
5584                 .bytenr = next->start,
5585                 .num_bytes = root->fs_info->nodesize,
5586                 .owning_root = owner_root,
5587                 .ref_root = btrfs_root_id(root),
5588         };
5589         int level = wc->level;
5590         int ret;
5591
5592         /* We are UPDATE_BACKREF, we're not dropping anything. */
5593         if (wc->stage == UPDATE_BACKREF)
5594                 return 0;
5595
5596         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
5597                 ref.parent = path->nodes[level]->start;
5598         } else {
5599                 ASSERT(btrfs_root_id(root) == btrfs_header_owner(path->nodes[level]));
5600                 if (btrfs_root_id(root) != btrfs_header_owner(path->nodes[level])) {
5601                         btrfs_err(root->fs_info, "mismatched block owner");
5602                         return -EIO;
5603                 }
5604         }
5605
5606         /*
5607          * If we had a drop_progress we need to verify the refs are set as
5608          * expected.  If we find our ref then we know that from here on out
5609          * everything should be correct, and we can clear the
5610          * ->restarted flag.
5611          */
5612         if (wc->restarted) {
5613                 ret = check_ref_exists(trans, root, next->start, ref.parent,
5614                                        level - 1);
5615                 if (ret <= 0)
5616                         return ret;
5617                 ret = 0;
5618                 wc->restarted = 0;
5619         }
5620
5621         /*
5622          * Reloc tree doesn't contribute to qgroup numbers, and we have already
5623          * accounted them at merge time (replace_path), thus we could skip
5624          * expensive subtree trace here.
5625          */
5626         if (btrfs_root_id(root) != BTRFS_TREE_RELOC_OBJECTID &&
5627             wc->refs[level - 1] > 1) {
5628                 u64 generation = btrfs_node_ptr_generation(path->nodes[level],
5629                                                            path->slots[level]);
5630
5631                 ret = btrfs_qgroup_trace_subtree(trans, next, generation, level - 1);
5632                 if (ret) {
5633                         btrfs_err_rl(root->fs_info,
5634 "error %d accounting shared subtree, quota is out of sync, rescan required",
5635                                      ret);
5636                 }
5637         }
5638
5639         /*
5640          * We need to update the next key in our walk control so we can update
5641          * the drop_progress key accordingly.  We don't care if find_next_key
5642          * doesn't find a key because that means we're at the end and are going
5643          * to clean up now.
5644          */
5645         wc->drop_level = level;
5646         find_next_key(path, level, &wc->drop_progress);
5647
5648         btrfs_init_tree_ref(&ref, level - 1, 0, false);
5649         return btrfs_free_extent(trans, &ref);
5650 }
5651
5652 /*
5653  * helper to process tree block pointer.
5654  *
5655  * when wc->stage == DROP_REFERENCE, this function checks
5656  * reference count of the block pointed to. if the block
5657  * is shared and we need update back refs for the subtree
5658  * rooted at the block, this function changes wc->stage to
5659  * UPDATE_BACKREF. if the block is shared and there is no
5660  * need to update back, this function drops the reference
5661  * to the block.
5662  *
5663  * NOTE: return value 1 means we should stop walking down.
5664  */
5665 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
5666                                  struct btrfs_root *root,
5667                                  struct btrfs_path *path,
5668                                  struct walk_control *wc)
5669 {
5670         struct btrfs_fs_info *fs_info = root->fs_info;
5671         u64 bytenr;
5672         u64 generation;
5673         u64 owner_root = 0;
5674         struct extent_buffer *next;
5675         int level = wc->level;
5676         int ret = 0;
5677
5678         generation = btrfs_node_ptr_generation(path->nodes[level],
5679                                                path->slots[level]);
5680         /*
5681          * if the lower level block was created before the snapshot
5682          * was created, we know there is no need to update back refs
5683          * for the subtree
5684          */
5685         if (wc->stage == UPDATE_BACKREF &&
5686             generation <= btrfs_root_origin_generation(root)) {
5687                 wc->lookup_info = 1;
5688                 return 1;
5689         }
5690
5691         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
5692
5693         next = btrfs_find_create_tree_block(fs_info, bytenr, btrfs_root_id(root),
5694                                             level - 1);
5695         if (IS_ERR(next))
5696                 return PTR_ERR(next);
5697
5698         btrfs_tree_lock(next);
5699
5700         ret = btrfs_lookup_extent_info(trans, fs_info, bytenr, level - 1, 1,
5701                                        &wc->refs[level - 1],
5702                                        &wc->flags[level - 1],
5703                                        &owner_root);
5704         if (ret < 0)
5705                 goto out_unlock;
5706
5707         if (unlikely(wc->refs[level - 1] == 0)) {
5708                 btrfs_err(fs_info, "bytenr %llu has 0 references, expect > 0",
5709                           bytenr);
5710                 ret = -EUCLEAN;
5711                 goto out_unlock;
5712         }
5713         wc->lookup_info = 0;
5714
5715         /* If we don't have to walk into this node skip it. */
5716         if (!visit_node_for_delete(root, wc, path->nodes[level],
5717                                    wc->flags[level - 1], path->slots[level]))
5718                 goto skip;
5719
5720         /*
5721          * We have to walk down into this node, and if we're currently at the
5722          * DROP_REFERNCE stage and this block is shared then we need to switch
5723          * to the UPDATE_BACKREF stage in order to convert to FULL_BACKREF.
5724          */
5725         if (wc->stage == DROP_REFERENCE && wc->refs[level - 1] > 1) {
5726                 wc->stage = UPDATE_BACKREF;
5727                 wc->shared_level = level - 1;
5728         }
5729
5730         ret = check_next_block_uptodate(trans, root, path, wc, next);
5731         if (ret)
5732                 return ret;
5733
5734         level--;
5735         ASSERT(level == btrfs_header_level(next));
5736         if (level != btrfs_header_level(next)) {
5737                 btrfs_err(root->fs_info, "mismatched level");
5738                 ret = -EIO;
5739                 goto out_unlock;
5740         }
5741         path->nodes[level] = next;
5742         path->slots[level] = 0;
5743         path->locks[level] = BTRFS_WRITE_LOCK;
5744         wc->level = level;
5745         if (wc->level == 1)
5746                 wc->reada_slot = 0;
5747         return 0;
5748 skip:
5749         ret = maybe_drop_reference(trans, root, path, wc, next, owner_root);
5750         if (ret)
5751                 goto out_unlock;
5752         wc->refs[level - 1] = 0;
5753         wc->flags[level - 1] = 0;
5754         wc->lookup_info = 1;
5755         ret = 1;
5756
5757 out_unlock:
5758         btrfs_tree_unlock(next);
5759         free_extent_buffer(next);
5760
5761         return ret;
5762 }
5763
5764 /*
5765  * helper to process tree block while walking up the tree.
5766  *
5767  * when wc->stage == DROP_REFERENCE, this function drops
5768  * reference count on the block.
5769  *
5770  * when wc->stage == UPDATE_BACKREF, this function changes
5771  * wc->stage back to DROP_REFERENCE if we changed wc->stage
5772  * to UPDATE_BACKREF previously while processing the block.
5773  *
5774  * NOTE: return value 1 means we should stop walking up.
5775  */
5776 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
5777                                  struct btrfs_root *root,
5778                                  struct btrfs_path *path,
5779                                  struct walk_control *wc)
5780 {
5781         struct btrfs_fs_info *fs_info = root->fs_info;
5782         int ret = 0;
5783         int level = wc->level;
5784         struct extent_buffer *eb = path->nodes[level];
5785         u64 parent = 0;
5786
5787         if (wc->stage == UPDATE_BACKREF) {
5788                 ASSERT(wc->shared_level >= level);
5789                 if (level < wc->shared_level)
5790                         goto out;
5791
5792                 ret = find_next_key(path, level + 1, &wc->update_progress);
5793                 if (ret > 0)
5794                         wc->update_ref = 0;
5795
5796                 wc->stage = DROP_REFERENCE;
5797                 wc->shared_level = -1;
5798                 path->slots[level] = 0;
5799
5800                 /*
5801                  * check reference count again if the block isn't locked.
5802                  * we should start walking down the tree again if reference
5803                  * count is one.
5804                  */
5805                 if (!path->locks[level]) {
5806                         ASSERT(level > 0);
5807                         btrfs_tree_lock(eb);
5808                         path->locks[level] = BTRFS_WRITE_LOCK;
5809
5810                         ret = btrfs_lookup_extent_info(trans, fs_info,
5811                                                        eb->start, level, 1,
5812                                                        &wc->refs[level],
5813                                                        &wc->flags[level],
5814                                                        NULL);
5815                         if (ret < 0) {
5816                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
5817                                 path->locks[level] = 0;
5818                                 return ret;
5819                         }
5820                         if (unlikely(wc->refs[level] == 0)) {
5821                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
5822                                 btrfs_err(fs_info, "bytenr %llu has 0 references, expect > 0",
5823                                           eb->start);
5824                                 return -EUCLEAN;
5825                         }
5826                         if (wc->refs[level] == 1) {
5827                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
5828                                 path->locks[level] = 0;
5829                                 return 1;
5830                         }
5831                 }
5832         }
5833
5834         /* wc->stage == DROP_REFERENCE */
5835         ASSERT(path->locks[level] || wc->refs[level] == 1);
5836
5837         if (wc->refs[level] == 1) {
5838                 if (level == 0) {
5839                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
5840                                 ret = btrfs_dec_ref(trans, root, eb, 1);
5841                         else
5842                                 ret = btrfs_dec_ref(trans, root, eb, 0);
5843                         if (ret) {
5844                                 btrfs_abort_transaction(trans, ret);
5845                                 return ret;
5846                         }
5847                         if (is_fstree(btrfs_root_id(root))) {
5848                                 ret = btrfs_qgroup_trace_leaf_items(trans, eb);
5849                                 if (ret) {
5850                                         btrfs_err_rl(fs_info,
5851         "error %d accounting leaf items, quota is out of sync, rescan required",
5852                                              ret);
5853                                 }
5854                         }
5855                 }
5856                 /* Make block locked assertion in btrfs_clear_buffer_dirty happy. */
5857                 if (!path->locks[level]) {
5858                         btrfs_tree_lock(eb);
5859                         path->locks[level] = BTRFS_WRITE_LOCK;
5860                 }
5861                 btrfs_clear_buffer_dirty(trans, eb);
5862         }
5863
5864         if (eb == root->node) {
5865                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
5866                         parent = eb->start;
5867                 else if (btrfs_root_id(root) != btrfs_header_owner(eb))
5868                         goto owner_mismatch;
5869         } else {
5870                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
5871                         parent = path->nodes[level + 1]->start;
5872                 else if (btrfs_root_id(root) !=
5873                          btrfs_header_owner(path->nodes[level + 1]))
5874                         goto owner_mismatch;
5875         }
5876
5877         ret = btrfs_free_tree_block(trans, btrfs_root_id(root), eb, parent,
5878                                     wc->refs[level] == 1);
5879         if (ret < 0)
5880                 btrfs_abort_transaction(trans, ret);
5881 out:
5882         wc->refs[level] = 0;
5883         wc->flags[level] = 0;
5884         return ret;
5885
5886 owner_mismatch:
5887         btrfs_err_rl(fs_info, "unexpected tree owner, have %llu expect %llu",
5888                      btrfs_header_owner(eb), btrfs_root_id(root));
5889         return -EUCLEAN;
5890 }
5891
5892 /*
5893  * walk_down_tree consists of two steps.
5894  *
5895  * walk_down_proc().  Look up the reference count and reference of our current
5896  * wc->level.  At this point path->nodes[wc->level] should be populated and
5897  * uptodate, and in most cases should already be locked.  If we are in
5898  * DROP_REFERENCE and our refcount is > 1 then we've entered a shared node and
5899  * we can walk back up the tree.  If we are UPDATE_BACKREF we have to set
5900  * FULL_BACKREF on this node if it's not already set, and then do the
5901  * FULL_BACKREF conversion dance, which is to drop the root reference and add
5902  * the shared reference to all of this nodes children.
5903  *
5904  * do_walk_down().  This is where we actually start iterating on the children of
5905  * our current path->nodes[wc->level].  For DROP_REFERENCE that means dropping
5906  * our reference to the children that return false from visit_node_for_delete(),
5907  * which has various conditions where we know we can just drop our reference
5908  * without visiting the node.  For UPDATE_BACKREF we will skip any children that
5909  * visit_node_for_delete() returns false for, only walking down when necessary.
5910  * The bulk of the work for UPDATE_BACKREF occurs in the walk_up_tree() part of
5911  * snapshot deletion.
5912  */
5913 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
5914                                    struct btrfs_root *root,
5915                                    struct btrfs_path *path,
5916                                    struct walk_control *wc)
5917 {
5918         int level = wc->level;
5919         int ret = 0;
5920
5921         wc->lookup_info = 1;
5922         while (level >= 0) {
5923                 ret = walk_down_proc(trans, root, path, wc);
5924                 if (ret)
5925                         break;
5926
5927                 if (level == 0)
5928                         break;
5929
5930                 if (path->slots[level] >=
5931                     btrfs_header_nritems(path->nodes[level]))
5932                         break;
5933
5934                 ret = do_walk_down(trans, root, path, wc);
5935                 if (ret > 0) {
5936                         path->slots[level]++;
5937                         continue;
5938                 } else if (ret < 0)
5939                         break;
5940                 level = wc->level;
5941         }
5942         return (ret == 1) ? 0 : ret;
5943 }
5944
5945 /*
5946  * walk_up_tree() is responsible for making sure we visit every slot on our
5947  * current node, and if we're at the end of that node then we call
5948  * walk_up_proc() on our current node which will do one of a few things based on
5949  * our stage.
5950  *
5951  * UPDATE_BACKREF.  If we wc->level is currently less than our wc->shared_level
5952  * then we need to walk back up the tree, and then going back down into the
5953  * other slots via walk_down_tree to update any other children from our original
5954  * wc->shared_level.  Once we're at or above our wc->shared_level we can switch
5955  * back to DROP_REFERENCE, lookup the current nodes refs and flags, and carry on.
5956  *
5957  * DROP_REFERENCE. If our refs == 1 then we're going to free this tree block.
5958  * If we're level 0 then we need to btrfs_dec_ref() on all of the data extents
5959  * in our current leaf.  After that we call btrfs_free_tree_block() on the
5960  * current node and walk up to the next node to walk down the next slot.
5961  */
5962 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
5963                                  struct btrfs_root *root,
5964                                  struct btrfs_path *path,
5965                                  struct walk_control *wc, int max_level)
5966 {
5967         int level = wc->level;
5968         int ret;
5969
5970         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
5971         while (level < max_level && path->nodes[level]) {
5972                 wc->level = level;
5973                 if (path->slots[level] + 1 <
5974                     btrfs_header_nritems(path->nodes[level])) {
5975                         path->slots[level]++;
5976                         return 0;
5977                 } else {
5978                         ret = walk_up_proc(trans, root, path, wc);
5979                         if (ret > 0)
5980                                 return 0;
5981                         if (ret < 0)
5982                                 return ret;
5983
5984                         if (path->locks[level]) {
5985                                 btrfs_tree_unlock_rw(path->nodes[level],
5986                                                      path->locks[level]);
5987                                 path->locks[level] = 0;
5988                         }
5989                         free_extent_buffer(path->nodes[level]);
5990                         path->nodes[level] = NULL;
5991                         level++;
5992                 }
5993         }
5994         return 1;
5995 }
5996
5997 /*
5998  * drop a subvolume tree.
5999  *
6000  * this function traverses the tree freeing any blocks that only
6001  * referenced by the tree.
6002  *
6003  * when a shared tree block is found. this function decreases its
6004  * reference count by one. if update_ref is true, this function
6005  * also make sure backrefs for the shared block and all lower level
6006  * blocks are properly updated.
6007  *
6008  * If called with for_reloc == 0, may exit early with -EAGAIN
6009  */
6010 int btrfs_drop_snapshot(struct btrfs_root *root, int update_ref, int for_reloc)
6011 {
6012         const bool is_reloc_root = (btrfs_root_id(root) == BTRFS_TREE_RELOC_OBJECTID);
6013         struct btrfs_fs_info *fs_info = root->fs_info;
6014         struct btrfs_path *path;
6015         struct btrfs_trans_handle *trans;
6016         struct btrfs_root *tree_root = fs_info->tree_root;
6017         struct btrfs_root_item *root_item = &root->root_item;
6018         struct walk_control *wc;
6019         struct btrfs_key key;
6020         const u64 rootid = btrfs_root_id(root);
6021         int ret = 0;
6022         int level;
6023         bool root_dropped = false;
6024         bool unfinished_drop = false;
6025
6026         btrfs_debug(fs_info, "Drop subvolume %llu", btrfs_root_id(root));
6027
6028         path = btrfs_alloc_path();
6029         if (!path) {
6030                 ret = -ENOMEM;
6031                 goto out;
6032         }
6033
6034         wc = kzalloc(sizeof(*wc), GFP_NOFS);
6035         if (!wc) {
6036                 btrfs_free_path(path);
6037                 ret = -ENOMEM;
6038                 goto out;
6039         }
6040
6041         /*
6042          * Use join to avoid potential EINTR from transaction start. See
6043          * wait_reserve_ticket and the whole reservation callchain.
6044          */
6045         if (for_reloc)
6046                 trans = btrfs_join_transaction(tree_root);
6047         else
6048                 trans = btrfs_start_transaction(tree_root, 0);
6049         if (IS_ERR(trans)) {
6050                 ret = PTR_ERR(trans);
6051                 goto out_free;
6052         }
6053
6054         ret = btrfs_run_delayed_items(trans);
6055         if (ret)
6056                 goto out_end_trans;
6057
6058         /*
6059          * This will help us catch people modifying the fs tree while we're
6060          * dropping it.  It is unsafe to mess with the fs tree while it's being
6061          * dropped as we unlock the root node and parent nodes as we walk down
6062          * the tree, assuming nothing will change.  If something does change
6063          * then we'll have stale information and drop references to blocks we've
6064          * already dropped.
6065          */
6066         set_bit(BTRFS_ROOT_DELETING, &root->state);
6067         unfinished_drop = test_bit(BTRFS_ROOT_UNFINISHED_DROP, &root->state);
6068
6069         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
6070                 level = btrfs_header_level(root->node);
6071                 path->nodes[level] = btrfs_lock_root_node(root);
6072                 path->slots[level] = 0;
6073                 path->locks[level] = BTRFS_WRITE_LOCK;
6074                 memset(&wc->update_progress, 0,
6075                        sizeof(wc->update_progress));
6076         } else {
6077                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
6078                 memcpy(&wc->update_progress, &key,
6079                        sizeof(wc->update_progress));
6080
6081                 level = btrfs_root_drop_level(root_item);
6082                 BUG_ON(level == 0);
6083                 path->lowest_level = level;
6084                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6085                 path->lowest_level = 0;
6086                 if (ret < 0)
6087                         goto out_end_trans;
6088
6089                 WARN_ON(ret > 0);
6090                 ret = 0;
6091
6092                 /*
6093                  * unlock our path, this is safe because only this
6094                  * function is allowed to delete this snapshot
6095                  */
6096                 btrfs_unlock_up_safe(path, 0);
6097
6098                 level = btrfs_header_level(root->node);
6099                 while (1) {
6100                         btrfs_tree_lock(path->nodes[level]);
6101                         path->locks[level] = BTRFS_WRITE_LOCK;
6102
6103                         /*
6104                          * btrfs_lookup_extent_info() returns 0 for success,
6105                          * or < 0 for error.
6106                          */
6107                         ret = btrfs_lookup_extent_info(trans, fs_info,
6108                                                 path->nodes[level]->start,
6109                                                 level, 1, &wc->refs[level],
6110                                                 &wc->flags[level], NULL);
6111                         if (ret < 0)
6112                                 goto out_end_trans;
6113
6114                         BUG_ON(wc->refs[level] == 0);
6115
6116                         if (level == btrfs_root_drop_level(root_item))
6117                                 break;
6118
6119                         btrfs_tree_unlock(path->nodes[level]);
6120                         path->locks[level] = 0;
6121                         WARN_ON(wc->refs[level] != 1);
6122                         level--;
6123                 }
6124         }
6125
6126         wc->restarted = test_bit(BTRFS_ROOT_DEAD_TREE, &root->state);
6127         wc->level = level;
6128         wc->shared_level = -1;
6129         wc->stage = DROP_REFERENCE;
6130         wc->update_ref = update_ref;
6131         wc->keep_locks = 0;
6132         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
6133
6134         while (1) {
6135
6136                 ret = walk_down_tree(trans, root, path, wc);
6137                 if (ret < 0) {
6138                         btrfs_abort_transaction(trans, ret);
6139                         break;
6140                 }
6141
6142                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
6143                 if (ret < 0) {
6144                         btrfs_abort_transaction(trans, ret);
6145                         break;
6146                 }
6147
6148                 if (ret > 0) {
6149                         BUG_ON(wc->stage != DROP_REFERENCE);
6150                         ret = 0;
6151                         break;
6152                 }
6153
6154                 if (wc->stage == DROP_REFERENCE) {
6155                         wc->drop_level = wc->level;
6156                         btrfs_node_key_to_cpu(path->nodes[wc->drop_level],
6157                                               &wc->drop_progress,
6158                                               path->slots[wc->drop_level]);
6159                 }
6160                 btrfs_cpu_key_to_disk(&root_item->drop_progress,
6161                                       &wc->drop_progress);
6162                 btrfs_set_root_drop_level(root_item, wc->drop_level);
6163
6164                 BUG_ON(wc->level == 0);
6165                 if (btrfs_should_end_transaction(trans) ||
6166                     (!for_reloc && btrfs_need_cleaner_sleep(fs_info))) {
6167                         ret = btrfs_update_root(trans, tree_root,
6168                                                 &root->root_key,
6169                                                 root_item);
6170                         if (ret) {
6171                                 btrfs_abort_transaction(trans, ret);
6172                                 goto out_end_trans;
6173                         }
6174
6175                         if (!is_reloc_root)
6176                                 btrfs_set_last_root_drop_gen(fs_info, trans->transid);
6177
6178                         btrfs_end_transaction_throttle(trans);
6179                         if (!for_reloc && btrfs_need_cleaner_sleep(fs_info)) {
6180                                 btrfs_debug(fs_info,
6181                                             "drop snapshot early exit");
6182                                 ret = -EAGAIN;
6183                                 goto out_free;
6184                         }
6185
6186                        /*
6187                         * Use join to avoid potential EINTR from transaction
6188                         * start. See wait_reserve_ticket and the whole
6189                         * reservation callchain.
6190                         */
6191                         if (for_reloc)
6192                                 trans = btrfs_join_transaction(tree_root);
6193                         else
6194                                 trans = btrfs_start_transaction(tree_root, 0);
6195                         if (IS_ERR(trans)) {
6196                                 ret = PTR_ERR(trans);
6197                                 goto out_free;
6198                         }
6199                 }
6200         }
6201         btrfs_release_path(path);
6202         if (ret)
6203                 goto out_end_trans;
6204
6205         ret = btrfs_del_root(trans, &root->root_key);
6206         if (ret) {
6207                 btrfs_abort_transaction(trans, ret);
6208                 goto out_end_trans;
6209         }
6210
6211         if (!is_reloc_root) {
6212                 ret = btrfs_find_root(tree_root, &root->root_key, path,
6213                                       NULL, NULL);
6214                 if (ret < 0) {
6215                         btrfs_abort_transaction(trans, ret);
6216                         goto out_end_trans;
6217                 } else if (ret > 0) {
6218                         ret = 0;
6219                         /*
6220                          * If we fail to delete the orphan item this time
6221                          * around, it'll get picked up the next time.
6222                          *
6223                          * The most common failure here is just -ENOENT.
6224                          */
6225                         btrfs_del_orphan_item(trans, tree_root, btrfs_root_id(root));
6226                 }
6227         }
6228
6229         /*
6230          * This subvolume is going to be completely dropped, and won't be
6231          * recorded as dirty roots, thus pertrans meta rsv will not be freed at
6232          * commit transaction time.  So free it here manually.
6233          */
6234         btrfs_qgroup_convert_reserved_meta(root, INT_MAX);
6235         btrfs_qgroup_free_meta_all_pertrans(root);
6236
6237         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state))
6238                 btrfs_add_dropped_root(trans, root);
6239         else
6240                 btrfs_put_root(root);
6241         root_dropped = true;
6242 out_end_trans:
6243         if (!is_reloc_root)
6244                 btrfs_set_last_root_drop_gen(fs_info, trans->transid);
6245
6246         btrfs_end_transaction_throttle(trans);
6247 out_free:
6248         kfree(wc);
6249         btrfs_free_path(path);
6250 out:
6251         if (!ret && root_dropped) {
6252                 ret = btrfs_qgroup_cleanup_dropped_subvolume(fs_info, rootid);
6253                 if (ret < 0)
6254                         btrfs_warn_rl(fs_info,
6255                                       "failed to cleanup qgroup 0/%llu: %d",
6256                                       rootid, ret);
6257                 ret = 0;
6258         }
6259         /*
6260          * We were an unfinished drop root, check to see if there are any
6261          * pending, and if not clear and wake up any waiters.
6262          */
6263         if (!ret && unfinished_drop)
6264                 btrfs_maybe_wake_unfinished_drop(fs_info);
6265
6266         /*
6267          * So if we need to stop dropping the snapshot for whatever reason we
6268          * need to make sure to add it back to the dead root list so that we
6269          * keep trying to do the work later.  This also cleans up roots if we
6270          * don't have it in the radix (like when we recover after a power fail
6271          * or unmount) so we don't leak memory.
6272          */
6273         if (!for_reloc && !root_dropped)
6274                 btrfs_add_dead_root(root);
6275         return ret;
6276 }
6277
6278 /*
6279  * drop subtree rooted at tree block 'node'.
6280  *
6281  * NOTE: this function will unlock and release tree block 'node'
6282  * only used by relocation code
6283  */
6284 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
6285                         struct btrfs_root *root,
6286                         struct extent_buffer *node,
6287                         struct extent_buffer *parent)
6288 {
6289         struct btrfs_fs_info *fs_info = root->fs_info;
6290         struct btrfs_path *path;
6291         struct walk_control *wc;
6292         int level;
6293         int parent_level;
6294         int ret = 0;
6295
6296         BUG_ON(btrfs_root_id(root) != BTRFS_TREE_RELOC_OBJECTID);
6297
6298         path = btrfs_alloc_path();
6299         if (!path)
6300                 return -ENOMEM;
6301
6302         wc = kzalloc(sizeof(*wc), GFP_NOFS);
6303         if (!wc) {
6304                 btrfs_free_path(path);
6305                 return -ENOMEM;
6306         }
6307
6308         btrfs_assert_tree_write_locked(parent);
6309         parent_level = btrfs_header_level(parent);
6310         atomic_inc(&parent->refs);
6311         path->nodes[parent_level] = parent;
6312         path->slots[parent_level] = btrfs_header_nritems(parent);
6313
6314         btrfs_assert_tree_write_locked(node);
6315         level = btrfs_header_level(node);
6316         path->nodes[level] = node;
6317         path->slots[level] = 0;
6318         path->locks[level] = BTRFS_WRITE_LOCK;
6319
6320         wc->refs[parent_level] = 1;
6321         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
6322         wc->level = level;
6323         wc->shared_level = -1;
6324         wc->stage = DROP_REFERENCE;
6325         wc->update_ref = 0;
6326         wc->keep_locks = 1;
6327         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
6328
6329         while (1) {
6330                 ret = walk_down_tree(trans, root, path, wc);
6331                 if (ret < 0)
6332                         break;
6333
6334                 ret = walk_up_tree(trans, root, path, wc, parent_level);
6335                 if (ret) {
6336                         if (ret > 0)
6337                                 ret = 0;
6338                         break;
6339                 }
6340         }
6341
6342         kfree(wc);
6343         btrfs_free_path(path);
6344         return ret;
6345 }
6346
6347 /*
6348  * Unpin the extent range in an error context and don't add the space back.
6349  * Errors are not propagated further.
6350  */
6351 void btrfs_error_unpin_extent_range(struct btrfs_fs_info *fs_info, u64 start, u64 end)
6352 {
6353         unpin_extent_range(fs_info, start, end, false);
6354 }
6355
6356 /*
6357  * It used to be that old block groups would be left around forever.
6358  * Iterating over them would be enough to trim unused space.  Since we
6359  * now automatically remove them, we also need to iterate over unallocated
6360  * space.
6361  *
6362  * We don't want a transaction for this since the discard may take a
6363  * substantial amount of time.  We don't require that a transaction be
6364  * running, but we do need to take a running transaction into account
6365  * to ensure that we're not discarding chunks that were released or
6366  * allocated in the current transaction.
6367  *
6368  * Holding the chunks lock will prevent other threads from allocating
6369  * or releasing chunks, but it won't prevent a running transaction
6370  * from committing and releasing the memory that the pending chunks
6371  * list head uses.  For that, we need to take a reference to the
6372  * transaction and hold the commit root sem.  We only need to hold
6373  * it while performing the free space search since we have already
6374  * held back allocations.
6375  */
6376 static int btrfs_trim_free_extents(struct btrfs_device *device, u64 *trimmed)
6377 {
6378         u64 start = BTRFS_DEVICE_RANGE_RESERVED, len = 0, end = 0;
6379         int ret;
6380
6381         *trimmed = 0;
6382
6383         /* Discard not supported = nothing to do. */
6384         if (!bdev_max_discard_sectors(device->bdev))
6385                 return 0;
6386
6387         /* Not writable = nothing to do. */
6388         if (!test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state))
6389                 return 0;
6390
6391         /* No free space = nothing to do. */
6392         if (device->total_bytes <= device->bytes_used)
6393                 return 0;
6394
6395         ret = 0;
6396
6397         while (1) {
6398                 struct btrfs_fs_info *fs_info = device->fs_info;
6399                 u64 bytes;
6400
6401                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
6402                 if (ret)
6403                         break;
6404
6405                 find_first_clear_extent_bit(&device->alloc_state, start,
6406                                             &start, &end,
6407                                             CHUNK_TRIMMED | CHUNK_ALLOCATED);
6408
6409                 /* Check if there are any CHUNK_* bits left */
6410                 if (start > device->total_bytes) {
6411                         WARN_ON(IS_ENABLED(CONFIG_BTRFS_DEBUG));
6412                         btrfs_warn_in_rcu(fs_info,
6413 "ignoring attempt to trim beyond device size: offset %llu length %llu device %s device size %llu",
6414                                           start, end - start + 1,
6415                                           btrfs_dev_name(device),
6416                                           device->total_bytes);
6417                         mutex_unlock(&fs_info->chunk_mutex);
6418                         ret = 0;
6419                         break;
6420                 }
6421
6422                 /* Ensure we skip the reserved space on each device. */
6423                 start = max_t(u64, start, BTRFS_DEVICE_RANGE_RESERVED);
6424
6425                 /*
6426                  * If find_first_clear_extent_bit find a range that spans the
6427                  * end of the device it will set end to -1, in this case it's up
6428                  * to the caller to trim the value to the size of the device.
6429                  */
6430                 end = min(end, device->total_bytes - 1);
6431
6432                 len = end - start + 1;
6433
6434                 /* We didn't find any extents */
6435                 if (!len) {
6436                         mutex_unlock(&fs_info->chunk_mutex);
6437                         ret = 0;
6438                         break;
6439                 }
6440
6441                 ret = btrfs_issue_discard(device->bdev, start, len,
6442                                           &bytes);
6443                 if (!ret)
6444                         set_extent_bit(&device->alloc_state, start,
6445                                        start + bytes - 1, CHUNK_TRIMMED, NULL);
6446                 mutex_unlock(&fs_info->chunk_mutex);
6447
6448                 if (ret)
6449                         break;
6450
6451                 start += len;
6452                 *trimmed += bytes;
6453
6454                 if (btrfs_trim_interrupted()) {
6455                         ret = -ERESTARTSYS;
6456                         break;
6457                 }
6458
6459                 cond_resched();
6460         }
6461
6462         return ret;
6463 }
6464
6465 /*
6466  * Trim the whole filesystem by:
6467  * 1) trimming the free space in each block group
6468  * 2) trimming the unallocated space on each device
6469  *
6470  * This will also continue trimming even if a block group or device encounters
6471  * an error.  The return value will be the last error, or 0 if nothing bad
6472  * happens.
6473  */
6474 int btrfs_trim_fs(struct btrfs_fs_info *fs_info, struct fstrim_range *range)
6475 {
6476         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
6477         struct btrfs_block_group *cache = NULL;
6478         struct btrfs_device *device;
6479         u64 group_trimmed;
6480         u64 range_end = U64_MAX;
6481         u64 start;
6482         u64 end;
6483         u64 trimmed = 0;
6484         u64 bg_failed = 0;
6485         u64 dev_failed = 0;
6486         int bg_ret = 0;
6487         int dev_ret = 0;
6488         int ret = 0;
6489
6490         if (range->start == U64_MAX)
6491                 return -EINVAL;
6492
6493         /*
6494          * Check range overflow if range->len is set.
6495          * The default range->len is U64_MAX.
6496          */
6497         if (range->len != U64_MAX &&
6498             check_add_overflow(range->start, range->len, &range_end))
6499                 return -EINVAL;
6500
6501         cache = btrfs_lookup_first_block_group(fs_info, range->start);
6502         for (; cache; cache = btrfs_next_block_group(cache)) {
6503                 if (cache->start >= range_end) {
6504                         btrfs_put_block_group(cache);
6505                         break;
6506                 }
6507
6508                 start = max(range->start, cache->start);
6509                 end = min(range_end, cache->start + cache->length);
6510
6511                 if (end - start >= range->minlen) {
6512                         if (!btrfs_block_group_done(cache)) {
6513                                 ret = btrfs_cache_block_group(cache, true);
6514                                 if (ret) {
6515                                         bg_failed++;
6516                                         bg_ret = ret;
6517                                         continue;
6518                                 }
6519                         }
6520                         ret = btrfs_trim_block_group(cache,
6521                                                      &group_trimmed,
6522                                                      start,
6523                                                      end,
6524                                                      range->minlen);
6525
6526                         trimmed += group_trimmed;
6527                         if (ret) {
6528                                 bg_failed++;
6529                                 bg_ret = ret;
6530                                 continue;
6531                         }
6532                 }
6533         }
6534
6535         if (bg_failed)
6536                 btrfs_warn(fs_info,
6537                         "failed to trim %llu block group(s), last error %d",
6538                         bg_failed, bg_ret);
6539
6540         mutex_lock(&fs_devices->device_list_mutex);
6541         list_for_each_entry(device, &fs_devices->devices, dev_list) {
6542                 if (test_bit(BTRFS_DEV_STATE_MISSING, &device->dev_state))
6543                         continue;
6544
6545                 ret = btrfs_trim_free_extents(device, &group_trimmed);
6546
6547                 trimmed += group_trimmed;
6548                 if (ret) {
6549                         dev_failed++;
6550                         dev_ret = ret;
6551                         break;
6552                 }
6553         }
6554         mutex_unlock(&fs_devices->device_list_mutex);
6555
6556         if (dev_failed)
6557                 btrfs_warn(fs_info,
6558                         "failed to trim %llu device(s), last error %d",
6559                         dev_failed, dev_ret);
6560         range->len = trimmed;
6561         if (bg_ret)
6562                 return bg_ret;
6563         return dev_ret;
6564 }