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 "misc.h"
  20 #include "tree-log.h"
  21 #include "disk-io.h"
  22 #include "print-tree.h"
  23 #include "volumes.h"
  24 #include "raid56.h"
  25 #include "locking.h"
  26 #include "free-space-cache.h"
  27 #include "free-space-tree.h"
  28 #include "sysfs.h"
  29 #include "qgroup.h"
  30 #include "ref-verify.h"
  31 #include "space-info.h"
  32 #include "block-rsv.h"
  33 #include "delalloc-space.h"
  34 #include "block-group.h"
  35 #include "discard.h"
  36 #include "rcu-string.h"
  37 #include "zoned.h"
  38 #include "dev-replace.h"
  39 #include "fs.h"
  40 #include "accessors.h"
  41 #include "extent-tree.h"
  42 #include "root-tree.h"
  43 #include "file-item.h"
  44 #include "orphan.h"
  45 #include "tree-checker.h"
  46
  47 #undef SCRAMBLE_DELAYED_REFS
  48
  49
  50 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  51                                struct btrfs_delayed_ref_node *node, u64 parent,
  52                                u64 root_objectid, u64 owner_objectid,
  53                                u64 owner_offset, int refs_to_drop,
  54                                struct btrfs_delayed_extent_op *extra_op);
  55 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  56                                     struct extent_buffer *leaf,
  57                                     struct btrfs_extent_item *ei);
  58 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  59                                       u64 parent, u64 root_objectid,
  60                                       u64 flags, u64 owner, u64 offset,
  61                                       struct btrfs_key *ins, int ref_mod);
  62 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  63                                      struct btrfs_delayed_ref_node *node,
  64                                      struct btrfs_delayed_extent_op *extent_op);
  65 static int find_next_key(struct btrfs_path *path, int level,
  66                          struct btrfs_key *key);
  67
  68 static int block_group_bits(struct btrfs_block_group *cache, u64 bits)
  69 {
  70         return (cache->flags & bits) == bits;
  71 }
  72
  73 int btrfs_add_excluded_extent(struct btrfs_fs_info *fs_info,
  74                               u64 start, u64 num_bytes)
  75 {
  76         u64 end = start + num_bytes - 1;
  77         set_extent_bits(&fs_info->excluded_extents, start, end,
  78                         EXTENT_UPTODATE);
  79         return 0;
  80 }
  81
  82 void btrfs_free_excluded_extents(struct btrfs_block_group *cache)
  83 {
  84         struct btrfs_fs_info *fs_info = cache->fs_info;
  85         u64 start, end;
  86
  87         start = cache->start;
  88         end = start + cache->length - 1;
  89
  90         clear_extent_bits(&fs_info->excluded_extents, start, end,
  91                           EXTENT_UPTODATE);
  92 }
  93
  94 /* simple helper to search for an existing data extent at a given offset */
  95 int btrfs_lookup_data_extent(struct btrfs_fs_info *fs_info, u64 start, u64 len)
  96 {
  97         struct btrfs_root *root = btrfs_extent_root(fs_info, start);
  98         int ret;
  99         struct btrfs_key key;
 100         struct btrfs_path *path;
 101
 102         path = btrfs_alloc_path();
 103         if (!path)
 104                 return -ENOMEM;
 105
 106         key.objectid = start;
 107         key.offset = len;
 108         key.type = BTRFS_EXTENT_ITEM_KEY;
 109         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
 110         btrfs_free_path(path);
 111         return ret;
 112 }
 113
 114 /*
 115  * helper function to lookup reference count and flags of a tree block.
 116  *
 117  * the head node for delayed ref is used to store the sum of all the
 118  * reference count modifications queued up in the rbtree. the head
 119  * node may also store the extent flags to set. This way you can check
 120  * to see what the reference count and extent flags would be if all of
 121  * the delayed refs are not processed.
 122  */
 123 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 124                              struct btrfs_fs_info *fs_info, u64 bytenr,
 125                              u64 offset, int metadata, u64 *refs, u64 *flags)
 126 {
 127         struct btrfs_root *extent_root;
 128         struct btrfs_delayed_ref_head *head;
 129         struct btrfs_delayed_ref_root *delayed_refs;
 130         struct btrfs_path *path;
 131         struct btrfs_extent_item *ei;
 132         struct extent_buffer *leaf;
 133         struct btrfs_key key;
 134         u32 item_size;
 135         u64 num_refs;
 136         u64 extent_flags;
 137         int ret;
 138
 139         /*
 140          * If we don't have skinny metadata, don't bother doing anything
 141          * different
 142          */
 143         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA)) {
 144                 offset = fs_info->nodesize;
 145                 metadata = 0;
 146         }
 147
 148         path = btrfs_alloc_path();
 149         if (!path)
 150                 return -ENOMEM;
 151
 152         if (!trans) {
 153                 path->skip_locking = 1;
 154                 path->search_commit_root = 1;
 155         }
 156
 157 search_again:
 158         key.objectid = bytenr;
 159         key.offset = offset;
 160         if (metadata)
 161                 key.type = BTRFS_METADATA_ITEM_KEY;
 162         else
 163                 key.type = BTRFS_EXTENT_ITEM_KEY;
 164
 165         extent_root = btrfs_extent_root(fs_info, bytenr);
 166         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 167         if (ret < 0)
 168                 goto out_free;
 169
 170         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 171                 if (path->slots[0]) {
 172                         path->slots[0]--;
 173                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 174                                               path->slots[0]);
 175                         if (key.objectid == bytenr &&
 176                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 177                             key.offset == fs_info->nodesize)
 178                                 ret = 0;
 179                 }
 180         }
 181
 182         if (ret == 0) {
 183                 leaf = path->nodes[0];
 184                 item_size = btrfs_item_size(leaf, path->slots[0]);
 185                 if (item_size >= sizeof(*ei)) {
 186                         ei = btrfs_item_ptr(leaf, path->slots[0],
 187                                             struct btrfs_extent_item);
 188                         num_refs = btrfs_extent_refs(leaf, ei);
 189                         extent_flags = btrfs_extent_flags(leaf, ei);
 190                 } else {
 191                         ret = -EINVAL;
 192                         btrfs_print_v0_err(fs_info);
 193                         if (trans)
 194                                 btrfs_abort_transaction(trans, ret);
 195                         else
 196                                 btrfs_handle_fs_error(fs_info, ret, NULL);
 197
 198                         goto out_free;
 199                 }
 200
 201                 BUG_ON(num_refs == 0);
 202         } else {
 203                 num_refs = 0;
 204                 extent_flags = 0;
 205                 ret = 0;
 206         }
 207
 208         if (!trans)
 209                 goto out;
 210
 211         delayed_refs = &trans->transaction->delayed_refs;
 212         spin_lock(&delayed_refs->lock);
 213         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
 214         if (head) {
 215                 if (!mutex_trylock(&head->mutex)) {
 216                         refcount_inc(&head->refs);
 217                         spin_unlock(&delayed_refs->lock);
 218
 219                         btrfs_release_path(path);
 220
 221                         /*
 222                          * Mutex was contended, block until it's released and try
 223                          * again
 224                          */
 225                         mutex_lock(&head->mutex);
 226                         mutex_unlock(&head->mutex);
 227                         btrfs_put_delayed_ref_head(head);
 228                         goto search_again;
 229                 }
 230                 spin_lock(&head->lock);
 231                 if (head->extent_op && head->extent_op->update_flags)
 232                         extent_flags |= head->extent_op->flags_to_set;
 233                 else
 234                         BUG_ON(num_refs == 0);
 235
 236                 num_refs += head->ref_mod;
 237                 spin_unlock(&head->lock);
 238                 mutex_unlock(&head->mutex);
 239         }
 240         spin_unlock(&delayed_refs->lock);
 241 out:
 242         WARN_ON(num_refs == 0);
 243         if (refs)
 244                 *refs = num_refs;
 245         if (flags)
 246                 *flags = extent_flags;
 247 out_free:
 248         btrfs_free_path(path);
 249         return ret;
 250 }
 251
 252 /*
 253  * Back reference rules.  Back refs have three main goals:
 254  *
 255  * 1) differentiate between all holders of references to an extent so that
 256  *    when a reference is dropped we can make sure it was a valid reference
 257  *    before freeing the extent.
 258  *
 259  * 2) Provide enough information to quickly find the holders of an extent
 260  *    if we notice a given block is corrupted or bad.
 261  *
 262  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 263  *    maintenance.  This is actually the same as #2, but with a slightly
 264  *    different use case.
 265  *
 266  * There are two kinds of back refs. The implicit back refs is optimized
 267  * for pointers in non-shared tree blocks. For a given pointer in a block,
 268  * back refs of this kind provide information about the block's owner tree
 269  * and the pointer's key. These information allow us to find the block by
 270  * b-tree searching. The full back refs is for pointers in tree blocks not
 271  * referenced by their owner trees. The location of tree block is recorded
 272  * in the back refs. Actually the full back refs is generic, and can be
 273  * used in all cases the implicit back refs is used. The major shortcoming
 274  * of the full back refs is its overhead. Every time a tree block gets
 275  * COWed, we have to update back refs entry for all pointers in it.
 276  *
 277  * For a newly allocated tree block, we use implicit back refs for
 278  * pointers in it. This means most tree related operations only involve
 279  * implicit back refs. For a tree block created in old transaction, the
 280  * only way to drop a reference to it is COW it. So we can detect the
 281  * event that tree block loses its owner tree's reference and do the
 282  * back refs conversion.
 283  *
 284  * When a tree block is COWed through a tree, there are four cases:
 285  *
 286  * The reference count of the block is one and the tree is the block's
 287  * owner tree. Nothing to do in this case.
 288  *
 289  * The reference count of the block is one and the tree is not the
 290  * block's owner tree. In this case, full back refs is used for pointers
 291  * in the block. Remove these full back refs, add implicit back refs for
 292  * every pointers in the new block.
 293  *
 294  * The reference count of the block is greater than one and the tree is
 295  * the block's owner tree. In this case, implicit back refs is used for
 296  * pointers in the block. Add full back refs for every pointers in the
 297  * block, increase lower level extents' reference counts. The original
 298  * implicit back refs are entailed to the new block.
 299  *
 300  * The reference count of the block is greater than one and the tree is
 301  * not the block's owner tree. Add implicit back refs for every pointer in
 302  * the new block, increase lower level extents' reference count.
 303  *
 304  * Back Reference Key composing:
 305  *
 306  * The key objectid corresponds to the first byte in the extent,
 307  * The key type is used to differentiate between types of back refs.
 308  * There are different meanings of the key offset for different types
 309  * of back refs.
 310  *
 311  * File extents can be referenced by:
 312  *
 313  * - multiple snapshots, subvolumes, or different generations in one subvol
 314  * - different files inside a single subvolume
 315  * - different offsets inside a file (bookend extents in file.c)
 316  *
 317  * The extent ref structure for the implicit back refs has fields for:
 318  *
 319  * - Objectid of the subvolume root
 320  * - objectid of the file holding the reference
 321  * - original offset in the file
 322  * - how many bookend extents
 323  *
 324  * The key offset for the implicit back refs is hash of the first
 325  * three fields.
 326  *
 327  * The extent ref structure for the full back refs has field for:
 328  *
 329  * - number of pointers in the tree leaf
 330  *
 331  * The key offset for the implicit back refs is the first byte of
 332  * the tree leaf
 333  *
 334  * When a file extent is allocated, The implicit back refs is used.
 335  * the fields are filled in:
 336  *
 337  *     (root_key.objectid, inode objectid, offset in file, 1)
 338  *
 339  * When a file extent is removed file truncation, we find the
 340  * corresponding implicit back refs and check the following fields:
 341  *
 342  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
 343  *
 344  * Btree extents can be referenced by:
 345  *
 346  * - Different subvolumes
 347  *
 348  * Both the implicit back refs and the full back refs for tree blocks
 349  * only consist of key. The key offset for the implicit back refs is
 350  * objectid of block's owner tree. The key offset for the full back refs
 351  * is the first byte of parent block.
 352  *
 353  * When implicit back refs is used, information about the lowest key and
 354  * level of the tree block are required. These information are stored in
 355  * tree block info structure.
 356  */
 357
 358 /*
 359  * is_data == BTRFS_REF_TYPE_BLOCK, tree block type is required,
 360  * is_data == BTRFS_REF_TYPE_DATA, data type is requiried,
 361  * is_data == BTRFS_REF_TYPE_ANY, either type is OK.
 362  */
 363 int btrfs_get_extent_inline_ref_type(const struct extent_buffer *eb,
 364                                      struct btrfs_extent_inline_ref *iref,
 365                                      enum btrfs_inline_ref_type is_data)
 366 {
 367         int type = btrfs_extent_inline_ref_type(eb, iref);
 368         u64 offset = btrfs_extent_inline_ref_offset(eb, iref);
 369
 370         if (type == BTRFS_TREE_BLOCK_REF_KEY ||
 371             type == BTRFS_SHARED_BLOCK_REF_KEY ||
 372             type == BTRFS_SHARED_DATA_REF_KEY ||
 373             type == BTRFS_EXTENT_DATA_REF_KEY) {
 374                 if (is_data == BTRFS_REF_TYPE_BLOCK) {
 375                         if (type == BTRFS_TREE_BLOCK_REF_KEY)
 376                                 return type;
 377                         if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
 378                                 ASSERT(eb->fs_info);
 379                                 /*
 380                                  * Every shared one has parent tree block,
 381                                  * which must be aligned to sector size.
 382                                  */
 383                                 if (offset &&
 384                                     IS_ALIGNED(offset, eb->fs_info->sectorsize))
 385                                         return type;
 386                         }
 387                 } else if (is_data == BTRFS_REF_TYPE_DATA) {
 388                         if (type == BTRFS_EXTENT_DATA_REF_KEY)
 389                                 return type;
 390                         if (type == BTRFS_SHARED_DATA_REF_KEY) {
 391                                 ASSERT(eb->fs_info);
 392                                 /*
 393                                  * Every shared one has parent tree block,
 394                                  * which must be aligned to sector size.
 395                                  */
 396                                 if (offset &&
 397                                     IS_ALIGNED(offset, eb->fs_info->sectorsize))
 398                                         return type;
 399                         }
 400                 } else {
 401                         ASSERT(is_data == BTRFS_REF_TYPE_ANY);
 402                         return type;
 403                 }
 404         }
 405
 406         btrfs_print_leaf((struct extent_buffer *)eb);
 407         btrfs_err(eb->fs_info,
 408                   "eb %llu iref 0x%lx invalid extent inline ref type %d",
 409                   eb->start, (unsigned long)iref, type);
 410         WARN_ON(1);
 411
 412         return BTRFS_REF_TYPE_INVALID;
 413 }
 414
 415 u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
 416 {
 417         u32 high_crc = ~(u32)0;
 418         u32 low_crc = ~(u32)0;
 419         __le64 lenum;
 420
 421         lenum = cpu_to_le64(root_objectid);
 422         high_crc = btrfs_crc32c(high_crc, &lenum, sizeof(lenum));
 423         lenum = cpu_to_le64(owner);
 424         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
 425         lenum = cpu_to_le64(offset);
 426         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
 427
 428         return ((u64)high_crc << 31) ^ (u64)low_crc;
 429 }
 430
 431 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
 432                                      struct btrfs_extent_data_ref *ref)
 433 {
 434         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
 435                                     btrfs_extent_data_ref_objectid(leaf, ref),
 436                                     btrfs_extent_data_ref_offset(leaf, ref));
 437 }
 438
 439 static int match_extent_data_ref(struct extent_buffer *leaf,
 440                                  struct btrfs_extent_data_ref *ref,
 441                                  u64 root_objectid, u64 owner, u64 offset)
 442 {
 443         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
 444             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
 445             btrfs_extent_data_ref_offset(leaf, ref) != offset)
 446                 return 0;
 447         return 1;
 448 }
 449
 450 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
 451                                            struct btrfs_path *path,
 452                                            u64 bytenr, u64 parent,
 453                                            u64 root_objectid,
 454                                            u64 owner, u64 offset)
 455 {
 456         struct btrfs_root *root = btrfs_extent_root(trans->fs_info, bytenr);
 457         struct btrfs_key key;
 458         struct btrfs_extent_data_ref *ref;
 459         struct extent_buffer *leaf;
 460         u32 nritems;
 461         int ret;
 462         int recow;
 463         int err = -ENOENT;
 464
 465         key.objectid = bytenr;
 466         if (parent) {
 467                 key.type = BTRFS_SHARED_DATA_REF_KEY;
 468                 key.offset = parent;
 469         } else {
 470                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
 471                 key.offset = hash_extent_data_ref(root_objectid,
 472                                                   owner, offset);
 473         }
 474 again:
 475         recow = 0;
 476         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
 477         if (ret < 0) {
 478                 err = ret;
 479                 goto fail;
 480         }
 481
 482         if (parent) {
 483                 if (!ret)
 484                         return 0;
 485                 goto fail;
 486         }
 487
 488         leaf = path->nodes[0];
 489         nritems = btrfs_header_nritems(leaf);
 490         while (1) {
 491                 if (path->slots[0] >= nritems) {
 492                         ret = btrfs_next_leaf(root, path);
 493                         if (ret < 0)
 494                                 err = ret;
 495                         if (ret)
 496                                 goto fail;
 497
 498                         leaf = path->nodes[0];
 499                         nritems = btrfs_header_nritems(leaf);
 500                         recow = 1;
 501                 }
 502
 503                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 504                 if (key.objectid != bytenr ||
 505                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
 506                         goto fail;
 507
 508                 ref = btrfs_item_ptr(leaf, path->slots[0],
 509                                      struct btrfs_extent_data_ref);
 510
 511                 if (match_extent_data_ref(leaf, ref, root_objectid,
 512                                           owner, offset)) {
 513                         if (recow) {
 514                                 btrfs_release_path(path);
 515                                 goto again;
 516                         }
 517                         err = 0;
 518                         break;
 519                 }
 520                 path->slots[0]++;
 521         }
 522 fail:
 523         return err;
 524 }
 525
 526 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
 527                                            struct btrfs_path *path,
 528                                            u64 bytenr, u64 parent,
 529                                            u64 root_objectid, u64 owner,
 530                                            u64 offset, int refs_to_add)
 531 {
 532         struct btrfs_root *root = btrfs_extent_root(trans->fs_info, bytenr);
 533         struct btrfs_key key;
 534         struct extent_buffer *leaf;
 535         u32 size;
 536         u32 num_refs;
 537         int ret;
 538
 539         key.objectid = bytenr;
 540         if (parent) {
 541                 key.type = BTRFS_SHARED_DATA_REF_KEY;
 542                 key.offset = parent;
 543                 size = sizeof(struct btrfs_shared_data_ref);
 544         } else {
 545                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
 546                 key.offset = hash_extent_data_ref(root_objectid,
 547                                                   owner, offset);
 548                 size = sizeof(struct btrfs_extent_data_ref);
 549         }
 550
 551         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
 552         if (ret && ret != -EEXIST)
 553                 goto fail;
 554
 555         leaf = path->nodes[0];
 556         if (parent) {
 557                 struct btrfs_shared_data_ref *ref;
 558                 ref = btrfs_item_ptr(leaf, path->slots[0],
 559                                      struct btrfs_shared_data_ref);
 560                 if (ret == 0) {
 561                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
 562                 } else {
 563                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
 564                         num_refs += refs_to_add;
 565                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
 566                 }
 567         } else {
 568                 struct btrfs_extent_data_ref *ref;
 569                 while (ret == -EEXIST) {
 570                         ref = btrfs_item_ptr(leaf, path->slots[0],
 571                                              struct btrfs_extent_data_ref);
 572                         if (match_extent_data_ref(leaf, ref, root_objectid,
 573                                                   owner, offset))
 574                                 break;
 575                         btrfs_release_path(path);
 576                         key.offset++;
 577                         ret = btrfs_insert_empty_item(trans, root, path, &key,
 578                                                       size);
 579                         if (ret && ret != -EEXIST)
 580                                 goto fail;
 581
 582                         leaf = path->nodes[0];
 583                 }
 584                 ref = btrfs_item_ptr(leaf, path->slots[0],
 585                                      struct btrfs_extent_data_ref);
 586                 if (ret == 0) {
 587                         btrfs_set_extent_data_ref_root(leaf, ref,
 588                                                        root_objectid);
 589                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
 590                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
 591                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
 592                 } else {
 593                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
 594                         num_refs += refs_to_add;
 595                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
 596                 }
 597         }
 598         btrfs_mark_buffer_dirty(leaf);
 599         ret = 0;
 600 fail:
 601         btrfs_release_path(path);
 602         return ret;
 603 }
 604
 605 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
 606                                            struct btrfs_root *root,
 607                                            struct btrfs_path *path,
 608                                            int refs_to_drop)
 609 {
 610         struct btrfs_key key;
 611         struct btrfs_extent_data_ref *ref1 = NULL;
 612         struct btrfs_shared_data_ref *ref2 = NULL;
 613         struct extent_buffer *leaf;
 614         u32 num_refs = 0;
 615         int ret = 0;
 616
 617         leaf = path->nodes[0];
 618         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 619
 620         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
 621                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
 622                                       struct btrfs_extent_data_ref);
 623                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
 624         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
 625                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
 626                                       struct btrfs_shared_data_ref);
 627                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
 628         } else if (unlikely(key.type == BTRFS_EXTENT_REF_V0_KEY)) {
 629                 btrfs_print_v0_err(trans->fs_info);
 630                 btrfs_abort_transaction(trans, -EINVAL);
 631                 return -EINVAL;
 632         } else {
 633                 BUG();
 634         }
 635
 636         BUG_ON(num_refs < refs_to_drop);
 637         num_refs -= refs_to_drop;
 638
 639         if (num_refs == 0) {
 640                 ret = btrfs_del_item(trans, root, path);
 641         } else {
 642                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
 643                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
 644                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
 645                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
 646                 btrfs_mark_buffer_dirty(leaf);
 647         }
 648         return ret;
 649 }
 650
 651 static noinline u32 extent_data_ref_count(struct btrfs_path *path,
 652                                           struct btrfs_extent_inline_ref *iref)
 653 {
 654         struct btrfs_key key;
 655         struct extent_buffer *leaf;
 656         struct btrfs_extent_data_ref *ref1;
 657         struct btrfs_shared_data_ref *ref2;
 658         u32 num_refs = 0;
 659         int type;
 660
 661         leaf = path->nodes[0];
 662         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 663
 664         BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
 665         if (iref) {
 666                 /*
 667                  * If type is invalid, we should have bailed out earlier than
 668                  * this call.
 669                  */
 670                 type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
 671                 ASSERT(type != BTRFS_REF_TYPE_INVALID);
 672                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
 673                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
 674                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
 675                 } else {
 676                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
 677                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
 678                 }
 679         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
 680                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
 681                                       struct btrfs_extent_data_ref);
 682                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
 683         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
 684                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
 685                                       struct btrfs_shared_data_ref);
 686                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
 687         } else {
 688                 WARN_ON(1);
 689         }
 690         return num_refs;
 691 }
 692
 693 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
 694                                           struct btrfs_path *path,
 695                                           u64 bytenr, u64 parent,
 696                                           u64 root_objectid)
 697 {
 698         struct btrfs_root *root = btrfs_extent_root(trans->fs_info, bytenr);
 699         struct btrfs_key key;
 700         int ret;
 701
 702         key.objectid = bytenr;
 703         if (parent) {
 704                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
 705                 key.offset = parent;
 706         } else {
 707                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
 708                 key.offset = root_objectid;
 709         }
 710
 711         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
 712         if (ret > 0)
 713                 ret = -ENOENT;
 714         return ret;
 715 }
 716
 717 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
 718                                           struct btrfs_path *path,
 719                                           u64 bytenr, u64 parent,
 720                                           u64 root_objectid)
 721 {
 722         struct btrfs_root *root = btrfs_extent_root(trans->fs_info, bytenr);
 723         struct btrfs_key key;
 724         int ret;
 725
 726         key.objectid = bytenr;
 727         if (parent) {
 728                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
 729                 key.offset = parent;
 730         } else {
 731                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
 732                 key.offset = root_objectid;
 733         }
 734
 735         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
 736         btrfs_release_path(path);
 737         return ret;
 738 }
 739
 740 static inline int extent_ref_type(u64 parent, u64 owner)
 741 {
 742         int type;
 743         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
 744                 if (parent > 0)
 745                         type = BTRFS_SHARED_BLOCK_REF_KEY;
 746                 else
 747                         type = BTRFS_TREE_BLOCK_REF_KEY;
 748         } else {
 749                 if (parent > 0)
 750                         type = BTRFS_SHARED_DATA_REF_KEY;
 751                 else
 752                         type = BTRFS_EXTENT_DATA_REF_KEY;
 753         }
 754         return type;
 755 }
 756
 757 static int find_next_key(struct btrfs_path *path, int level,
 758                          struct btrfs_key *key)
 759
 760 {
 761         for (; level < BTRFS_MAX_LEVEL; level++) {
 762                 if (!path->nodes[level])
 763                         break;
 764                 if (path->slots[level] + 1 >=
 765                     btrfs_header_nritems(path->nodes[level]))
 766                         continue;
 767                 if (level == 0)
 768                         btrfs_item_key_to_cpu(path->nodes[level], key,
 769                                               path->slots[level] + 1);
 770                 else
 771                         btrfs_node_key_to_cpu(path->nodes[level], key,
 772                                               path->slots[level] + 1);
 773                 return 0;
 774         }
 775         return 1;
 776 }
 777
 778 /*
 779  * look for inline back ref. if back ref is found, *ref_ret is set
 780  * to the address of inline back ref, and 0 is returned.
 781  *
 782  * if back ref isn't found, *ref_ret is set to the address where it
 783  * should be inserted, and -ENOENT is returned.
 784  *
 785  * if insert is true and there are too many inline back refs, the path
 786  * points to the extent item, and -EAGAIN is returned.
 787  *
 788  * NOTE: inline back refs are ordered in the same way that back ref
 789  *       items in the tree are ordered.
 790  */
 791 static noinline_for_stack
 792 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
 793                                  struct btrfs_path *path,
 794                                  struct btrfs_extent_inline_ref **ref_ret,
 795                                  u64 bytenr, u64 num_bytes,
 796                                  u64 parent, u64 root_objectid,
 797                                  u64 owner, u64 offset, int insert)
 798 {
 799         struct btrfs_fs_info *fs_info = trans->fs_info;
 800         struct btrfs_root *root = btrfs_extent_root(fs_info, bytenr);
 801         struct btrfs_key key;
 802         struct extent_buffer *leaf;
 803         struct btrfs_extent_item *ei;
 804         struct btrfs_extent_inline_ref *iref;
 805         u64 flags;
 806         u64 item_size;
 807         unsigned long ptr;
 808         unsigned long end;
 809         int extra_size;
 810         int type;
 811         int want;
 812         int ret;
 813         int err = 0;
 814         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
 815         int needed;
 816
 817         key.objectid = bytenr;
 818         key.type = BTRFS_EXTENT_ITEM_KEY;
 819         key.offset = num_bytes;
 820
 821         want = extent_ref_type(parent, owner);
 822         if (insert) {
 823                 extra_size = btrfs_extent_inline_ref_size(want);
 824                 path->search_for_extension = 1;
 825                 path->keep_locks = 1;
 826         } else
 827                 extra_size = -1;
 828
 829         /*
 830          * Owner is our level, so we can just add one to get the level for the
 831          * block we are interested in.
 832          */
 833         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
 834                 key.type = BTRFS_METADATA_ITEM_KEY;
 835                 key.offset = owner;
 836         }
 837
 838 again:
 839         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
 840         if (ret < 0) {
 841                 err = ret;
 842                 goto out;
 843         }
 844
 845         /*
 846          * We may be a newly converted file system which still has the old fat
 847          * extent entries for metadata, so try and see if we have one of those.
 848          */
 849         if (ret > 0 && skinny_metadata) {
 850                 skinny_metadata = false;
 851                 if (path->slots[0]) {
 852                         path->slots[0]--;
 853                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 854                                               path->slots[0]);
 855                         if (key.objectid == bytenr &&
 856                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 857                             key.offset == num_bytes)
 858                                 ret = 0;
 859                 }
 860                 if (ret) {
 861                         key.objectid = bytenr;
 862                         key.type = BTRFS_EXTENT_ITEM_KEY;
 863                         key.offset = num_bytes;
 864                         btrfs_release_path(path);
 865                         goto again;
 866                 }
 867         }
 868
 869         if (ret && !insert) {
 870                 err = -ENOENT;
 871                 goto out;
 872         } else if (WARN_ON(ret)) {
 873                 err = -EIO;
 874                 goto out;
 875         }
 876
 877         leaf = path->nodes[0];
 878         item_size = btrfs_item_size(leaf, path->slots[0]);
 879         if (unlikely(item_size < sizeof(*ei))) {
 880                 err = -EINVAL;
 881                 btrfs_print_v0_err(fs_info);
 882                 btrfs_abort_transaction(trans, err);
 883                 goto out;
 884         }
 885
 886         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
 887         flags = btrfs_extent_flags(leaf, ei);
 888
 889         ptr = (unsigned long)(ei + 1);
 890         end = (unsigned long)ei + item_size;
 891
 892         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
 893                 ptr += sizeof(struct btrfs_tree_block_info);
 894                 BUG_ON(ptr > end);
 895         }
 896
 897         if (owner >= BTRFS_FIRST_FREE_OBJECTID)
 898                 needed = BTRFS_REF_TYPE_DATA;
 899         else
 900                 needed = BTRFS_REF_TYPE_BLOCK;
 901
 902         err = -ENOENT;
 903         while (1) {
 904                 if (ptr >= end) {
 905                         if (ptr > end) {
 906                                 err = -EUCLEAN;
 907                                 btrfs_print_leaf(path->nodes[0]);
 908                                 btrfs_crit(fs_info,
 909 "overrun extent record at slot %d while looking for inline extent for root %llu owner %llu offset %llu parent %llu",
 910                                         path->slots[0], root_objectid, owner, offset, parent);
 911                         }
 912                         break;
 913                 }
 914                 iref = (struct btrfs_extent_inline_ref *)ptr;
 915                 type = btrfs_get_extent_inline_ref_type(leaf, iref, needed);
 916                 if (type == BTRFS_REF_TYPE_INVALID) {
 917                         err = -EUCLEAN;
 918                         goto out;
 919                 }
 920
 921                 if (want < type)
 922                         break;
 923                 if (want > type) {
 924                         ptr += btrfs_extent_inline_ref_size(type);
 925                         continue;
 926                 }
 927
 928                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
 929                         struct btrfs_extent_data_ref *dref;
 930                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
 931                         if (match_extent_data_ref(leaf, dref, root_objectid,
 932                                                   owner, offset)) {
 933                                 err = 0;
 934                                 break;
 935                         }
 936                         if (hash_extent_data_ref_item(leaf, dref) <
 937                             hash_extent_data_ref(root_objectid, owner, offset))
 938                                 break;
 939                 } else {
 940                         u64 ref_offset;
 941                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
 942                         if (parent > 0) {
 943                                 if (parent == ref_offset) {
 944                                         err = 0;
 945                                         break;
 946                                 }
 947                                 if (ref_offset < parent)
 948                                         break;
 949                         } else {
 950                                 if (root_objectid == ref_offset) {
 951                                         err = 0;
 952                                         break;
 953                                 }
 954                                 if (ref_offset < root_objectid)
 955                                         break;
 956                         }
 957                 }
 958                 ptr += btrfs_extent_inline_ref_size(type);
 959         }
 960         if (err == -ENOENT && insert) {
 961                 if (item_size + extra_size >=
 962                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
 963                         err = -EAGAIN;
 964                         goto out;
 965                 }
 966                 /*
 967                  * To add new inline back ref, we have to make sure
 968                  * there is no corresponding back ref item.
 969                  * For simplicity, we just do not add new inline back
 970                  * ref if there is any kind of item for this block
 971                  */
 972                 if (find_next_key(path, 0, &key) == 0 &&
 973                     key.objectid == bytenr &&
 974                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
 975                         err = -EAGAIN;
 976                         goto out;
 977                 }
 978         }
 979         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
 980 out:
 981         if (insert) {
 982                 path->keep_locks = 0;
 983                 path->search_for_extension = 0;
 984                 btrfs_unlock_up_safe(path, 1);
 985         }
 986         return err;
 987 }
 988
 989 /*
 990  * helper to add new inline back ref
 991  */
 992 static noinline_for_stack
 993 void setup_inline_extent_backref(struct btrfs_fs_info *fs_info,
 994                                  struct btrfs_path *path,
 995                                  struct btrfs_extent_inline_ref *iref,
 996                                  u64 parent, u64 root_objectid,
 997                                  u64 owner, u64 offset, int refs_to_add,
 998                                  struct btrfs_delayed_extent_op *extent_op)
 999 {
1000         struct extent_buffer *leaf;
1001         struct btrfs_extent_item *ei;
1002         unsigned long ptr;
1003         unsigned long end;
1004         unsigned long item_offset;
1005         u64 refs;
1006         int size;
1007         int type;
1008
1009         leaf = path->nodes[0];
1010         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1011         item_offset = (unsigned long)iref - (unsigned long)ei;
1012
1013         type = extent_ref_type(parent, owner);
1014         size = btrfs_extent_inline_ref_size(type);
1015
1016         btrfs_extend_item(path, size);
1017
1018         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1019         refs = btrfs_extent_refs(leaf, ei);
1020         refs += refs_to_add;
1021         btrfs_set_extent_refs(leaf, ei, refs);
1022         if (extent_op)
1023                 __run_delayed_extent_op(extent_op, leaf, ei);
1024
1025         ptr = (unsigned long)ei + item_offset;
1026         end = (unsigned long)ei + btrfs_item_size(leaf, path->slots[0]);
1027         if (ptr < end - size)
1028                 memmove_extent_buffer(leaf, ptr + size, ptr,
1029                                       end - size - ptr);
1030
1031         iref = (struct btrfs_extent_inline_ref *)ptr;
1032         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1033         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1034                 struct btrfs_extent_data_ref *dref;
1035                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1036                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1037                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1038                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1039                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1040         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1041                 struct btrfs_shared_data_ref *sref;
1042                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1043                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1044                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1045         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1046                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1047         } else {
1048                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1049         }
1050         btrfs_mark_buffer_dirty(leaf);
1051 }
1052
1053 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1054                                  struct btrfs_path *path,
1055                                  struct btrfs_extent_inline_ref **ref_ret,
1056                                  u64 bytenr, u64 num_bytes, u64 parent,
1057                                  u64 root_objectid, u64 owner, u64 offset)
1058 {
1059         int ret;
1060
1061         ret = lookup_inline_extent_backref(trans, path, ref_ret, bytenr,
1062                                            num_bytes, parent, root_objectid,
1063                                            owner, offset, 0);
1064         if (ret != -ENOENT)
1065                 return ret;
1066
1067         btrfs_release_path(path);
1068         *ref_ret = NULL;
1069
1070         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1071                 ret = lookup_tree_block_ref(trans, path, bytenr, parent,
1072                                             root_objectid);
1073         } else {
1074                 ret = lookup_extent_data_ref(trans, path, bytenr, parent,
1075                                              root_objectid, owner, offset);
1076         }
1077         return ret;
1078 }
1079
1080 /*
1081  * helper to update/remove inline back ref
1082  */
1083 static noinline_for_stack
1084 void update_inline_extent_backref(struct btrfs_path *path,
1085                                   struct btrfs_extent_inline_ref *iref,
1086                                   int refs_to_mod,
1087                                   struct btrfs_delayed_extent_op *extent_op)
1088 {
1089         struct extent_buffer *leaf = path->nodes[0];
1090         struct btrfs_extent_item *ei;
1091         struct btrfs_extent_data_ref *dref = NULL;
1092         struct btrfs_shared_data_ref *sref = NULL;
1093         unsigned long ptr;
1094         unsigned long end;
1095         u32 item_size;
1096         int size;
1097         int type;
1098         u64 refs;
1099
1100         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1101         refs = btrfs_extent_refs(leaf, ei);
1102         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1103         refs += refs_to_mod;
1104         btrfs_set_extent_refs(leaf, ei, refs);
1105         if (extent_op)
1106                 __run_delayed_extent_op(extent_op, leaf, ei);
1107
1108         /*
1109          * If type is invalid, we should have bailed out after
1110          * lookup_inline_extent_backref().
1111          */
1112         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_ANY);
1113         ASSERT(type != BTRFS_REF_TYPE_INVALID);
1114
1115         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1116                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1117                 refs = btrfs_extent_data_ref_count(leaf, dref);
1118         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1119                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1120                 refs = btrfs_shared_data_ref_count(leaf, sref);
1121         } else {
1122                 refs = 1;
1123                 BUG_ON(refs_to_mod != -1);
1124         }
1125
1126         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1127         refs += refs_to_mod;
1128
1129         if (refs > 0) {
1130                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1131                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1132                 else
1133                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1134         } else {
1135                 size =  btrfs_extent_inline_ref_size(type);
1136                 item_size = btrfs_item_size(leaf, path->slots[0]);
1137                 ptr = (unsigned long)iref;
1138                 end = (unsigned long)ei + item_size;
1139                 if (ptr + size < end)
1140                         memmove_extent_buffer(leaf, ptr, ptr + size,
1141                                               end - ptr - size);
1142                 item_size -= size;
1143                 btrfs_truncate_item(path, item_size, 1);
1144         }
1145         btrfs_mark_buffer_dirty(leaf);
1146 }
1147
1148 static noinline_for_stack
1149 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1150                                  struct btrfs_path *path,
1151                                  u64 bytenr, u64 num_bytes, u64 parent,
1152                                  u64 root_objectid, u64 owner,
1153                                  u64 offset, int refs_to_add,
1154                                  struct btrfs_delayed_extent_op *extent_op)
1155 {
1156         struct btrfs_extent_inline_ref *iref;
1157         int ret;
1158
1159         ret = lookup_inline_extent_backref(trans, path, &iref, bytenr,
1160                                            num_bytes, parent, root_objectid,
1161                                            owner, offset, 1);
1162         if (ret == 0) {
1163                 /*
1164                  * We're adding refs to a tree block we already own, this
1165                  * should not happen at all.
1166                  */
1167                 if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1168                         btrfs_crit(trans->fs_info,
1169 "adding refs to an existing tree ref, bytenr %llu num_bytes %llu root_objectid %llu",
1170                                    bytenr, num_bytes, root_objectid);
1171                         if (IS_ENABLED(CONFIG_BTRFS_DEBUG)) {
1172                                 WARN_ON(1);
1173                                 btrfs_crit(trans->fs_info,
1174                         "path->slots[0]=%d path->nodes[0]:", path->slots[0]);
1175                                 btrfs_print_leaf(path->nodes[0]);
1176                         }
1177                         return -EUCLEAN;
1178                 }
1179                 update_inline_extent_backref(path, iref, refs_to_add, extent_op);
1180         } else if (ret == -ENOENT) {
1181                 setup_inline_extent_backref(trans->fs_info, path, iref, parent,
1182                                             root_objectid, owner, offset,
1183                                             refs_to_add, extent_op);
1184                 ret = 0;
1185         }
1186         return ret;
1187 }
1188
1189 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1190                                  struct btrfs_root *root,
1191                                  struct btrfs_path *path,
1192                                  struct btrfs_extent_inline_ref *iref,
1193                                  int refs_to_drop, int is_data)
1194 {
1195         int ret = 0;
1196
1197         BUG_ON(!is_data && refs_to_drop != 1);
1198         if (iref)
1199                 update_inline_extent_backref(path, iref, -refs_to_drop, NULL);
1200         else if (is_data)
1201                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop);
1202         else
1203                 ret = btrfs_del_item(trans, root, path);
1204         return ret;
1205 }
1206
1207 static int btrfs_issue_discard(struct block_device *bdev, u64 start, u64 len,
1208                                u64 *discarded_bytes)
1209 {
1210         int j, ret = 0;
1211         u64 bytes_left, end;
1212         u64 aligned_start = ALIGN(start, 1 << 9);
1213
1214         if (WARN_ON(start != aligned_start)) {
1215                 len -= aligned_start - start;
1216                 len = round_down(len, 1 << 9);
1217                 start = aligned_start;
1218         }
1219
1220         *discarded_bytes = 0;
1221
1222         if (!len)
1223                 return 0;
1224
1225         end = start + len;
1226         bytes_left = len;
1227
1228         /* Skip any superblocks on this device. */
1229         for (j = 0; j < BTRFS_SUPER_MIRROR_MAX; j++) {
1230                 u64 sb_start = btrfs_sb_offset(j);
1231                 u64 sb_end = sb_start + BTRFS_SUPER_INFO_SIZE;
1232                 u64 size = sb_start - start;
1233
1234                 if (!in_range(sb_start, start, bytes_left) &&
1235                     !in_range(sb_end, start, bytes_left) &&
1236                     !in_range(start, sb_start, BTRFS_SUPER_INFO_SIZE))
1237                         continue;
1238
1239                 /*
1240                  * Superblock spans beginning of range.  Adjust start and
1241                  * try again.
1242                  */
1243                 if (sb_start <= start) {
1244                         start += sb_end - start;
1245                         if (start > end) {
1246                                 bytes_left = 0;
1247                                 break;
1248                         }
1249                         bytes_left = end - start;
1250                         continue;
1251                 }
1252
1253                 if (size) {
1254                         ret = blkdev_issue_discard(bdev, start >> 9, size >> 9,
1255                                                    GFP_NOFS);
1256                         if (!ret)
1257                                 *discarded_bytes += size;
1258                         else if (ret != -EOPNOTSUPP)
1259                                 return ret;
1260                 }
1261
1262                 start = sb_end;
1263                 if (start > end) {
1264                         bytes_left = 0;
1265                         break;
1266                 }
1267                 bytes_left = end - start;
1268         }
1269
1270         if (bytes_left) {
1271                 ret = blkdev_issue_discard(bdev, start >> 9, bytes_left >> 9,
1272                                            GFP_NOFS);
1273                 if (!ret)
1274                         *discarded_bytes += bytes_left;
1275         }
1276         return ret;
1277 }
1278
1279 static int do_discard_extent(struct btrfs_discard_stripe *stripe, u64 *bytes)
1280 {
1281         struct btrfs_device *dev = stripe->dev;
1282         struct btrfs_fs_info *fs_info = dev->fs_info;
1283         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1284         u64 phys = stripe->physical;
1285         u64 len = stripe->length;
1286         u64 discarded = 0;
1287         int ret = 0;
1288
1289         /* Zone reset on a zoned filesystem */
1290         if (btrfs_can_zone_reset(dev, phys, len)) {
1291                 u64 src_disc;
1292
1293                 ret = btrfs_reset_device_zone(dev, phys, len, &discarded);
1294                 if (ret)
1295                         goto out;
1296
1297                 if (!btrfs_dev_replace_is_ongoing(dev_replace) ||
1298                     dev != dev_replace->srcdev)
1299                         goto out;
1300
1301                 src_disc = discarded;
1302
1303                 /* Send to replace target as well */
1304                 ret = btrfs_reset_device_zone(dev_replace->tgtdev, phys, len,
1305                                               &discarded);
1306                 discarded += src_disc;
1307         } else if (bdev_max_discard_sectors(stripe->dev->bdev)) {
1308                 ret = btrfs_issue_discard(dev->bdev, phys, len, &discarded);
1309         } else {
1310                 ret = 0;
1311                 *bytes = 0;
1312         }
1313
1314 out:
1315         *bytes = discarded;
1316         return ret;
1317 }
1318
1319 int btrfs_discard_extent(struct btrfs_fs_info *fs_info, u64 bytenr,
1320                          u64 num_bytes, u64 *actual_bytes)
1321 {
1322         int ret = 0;
1323         u64 discarded_bytes = 0;
1324         u64 end = bytenr + num_bytes;
1325         u64 cur = bytenr;
1326
1327         /*
1328          * Avoid races with device replace and make sure the devices in the
1329          * stripes don't go away while we are discarding.
1330          */
1331         btrfs_bio_counter_inc_blocked(fs_info);
1332         while (cur < end) {
1333                 struct btrfs_discard_stripe *stripes;
1334                 unsigned int num_stripes;
1335                 int i;
1336
1337                 num_bytes = end - cur;
1338                 stripes = btrfs_map_discard(fs_info, cur, &num_bytes, &num_stripes);
1339                 if (IS_ERR(stripes)) {
1340                         ret = PTR_ERR(stripes);
1341                         if (ret == -EOPNOTSUPP)
1342                                 ret = 0;
1343                         break;
1344                 }
1345
1346                 for (i = 0; i < num_stripes; i++) {
1347                         struct btrfs_discard_stripe *stripe = stripes + i;
1348                         u64 bytes;
1349
1350                         if (!stripe->dev->bdev) {
1351                                 ASSERT(btrfs_test_opt(fs_info, DEGRADED));
1352                                 continue;
1353                         }
1354
1355                         if (!test_bit(BTRFS_DEV_STATE_WRITEABLE,
1356                                         &stripe->dev->dev_state))
1357                                 continue;
1358
1359                         ret = do_discard_extent(stripe, &bytes);
1360                         if (ret) {
1361                                 /*
1362                                  * Keep going if discard is not supported by the
1363                                  * device.
1364                                  */
1365                                 if (ret != -EOPNOTSUPP)
1366                                         break;
1367                                 ret = 0;
1368                         } else {
1369                                 discarded_bytes += bytes;
1370                         }
1371                 }
1372                 kfree(stripes);
1373                 if (ret)
1374                         break;
1375                 cur += num_bytes;
1376         }
1377         btrfs_bio_counter_dec(fs_info);
1378         if (actual_bytes)
1379                 *actual_bytes = discarded_bytes;
1380         return ret;
1381 }
1382
1383 /* Can return -ENOMEM */
1384 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1385                          struct btrfs_ref *generic_ref)
1386 {
1387         struct btrfs_fs_info *fs_info = trans->fs_info;
1388         int ret;
1389
1390         ASSERT(generic_ref->type != BTRFS_REF_NOT_SET &&
1391                generic_ref->action);
1392         BUG_ON(generic_ref->type == BTRFS_REF_METADATA &&
1393                generic_ref->tree_ref.owning_root == BTRFS_TREE_LOG_OBJECTID);
1394
1395         if (generic_ref->type == BTRFS_REF_METADATA)
1396                 ret = btrfs_add_delayed_tree_ref(trans, generic_ref, NULL);
1397         else
1398                 ret = btrfs_add_delayed_data_ref(trans, generic_ref, 0);
1399
1400         btrfs_ref_tree_mod(fs_info, generic_ref);
1401
1402         return ret;
1403 }
1404
1405 /*
1406  * __btrfs_inc_extent_ref - insert backreference for a given extent
1407  *
1408  * The counterpart is in __btrfs_free_extent(), with examples and more details
1409  * how it works.
1410  *
1411  * @trans:          Handle of transaction
1412  *
1413  * @node:           The delayed ref node used to get the bytenr/length for
1414  *                  extent whose references are incremented.
1415  *
1416  * @parent:         If this is a shared extent (BTRFS_SHARED_DATA_REF_KEY/
1417  *                  BTRFS_SHARED_BLOCK_REF_KEY) then it holds the logical
1418  *                  bytenr of the parent block. Since new extents are always
1419  *                  created with indirect references, this will only be the case
1420  *                  when relocating a shared extent. In that case, root_objectid
1421  *                  will be BTRFS_TREE_RELOC_OBJECTID. Otherwise, parent must
1422  *                  be 0
1423  *
1424  * @root_objectid:  The id of the root where this modification has originated,
1425  *                  this can be either one of the well-known metadata trees or
1426  *                  the subvolume id which references this extent.
1427  *
1428  * @owner:          For data extents it is the inode number of the owning file.
1429  *                  For metadata extents this parameter holds the level in the
1430  *                  tree of the extent.
1431  *
1432  * @offset:         For metadata extents the offset is ignored and is currently
1433  *                  always passed as 0. For data extents it is the fileoffset
1434  *                  this extent belongs to.
1435  *
1436  * @refs_to_add     Number of references to add
1437  *
1438  * @extent_op       Pointer to a structure, holding information necessary when
1439  *                  updating a tree block's flags
1440  *
1441  */
1442 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1443                                   struct btrfs_delayed_ref_node *node,
1444                                   u64 parent, u64 root_objectid,
1445                                   u64 owner, u64 offset, int refs_to_add,
1446                                   struct btrfs_delayed_extent_op *extent_op)
1447 {
1448         struct btrfs_path *path;
1449         struct extent_buffer *leaf;
1450         struct btrfs_extent_item *item;
1451         struct btrfs_key key;
1452         u64 bytenr = node->bytenr;
1453         u64 num_bytes = node->num_bytes;
1454         u64 refs;
1455         int ret;
1456
1457         path = btrfs_alloc_path();
1458         if (!path)
1459                 return -ENOMEM;
1460
1461         /* this will setup the path even if it fails to insert the back ref */
1462         ret = insert_inline_extent_backref(trans, path, bytenr, num_bytes,
1463                                            parent, root_objectid, owner,
1464                                            offset, refs_to_add, extent_op);
1465         if ((ret < 0 && ret != -EAGAIN) || !ret)
1466                 goto out;
1467
1468         /*
1469          * Ok we had -EAGAIN which means we didn't have space to insert and
1470          * inline extent ref, so just update the reference count and add a
1471          * normal backref.
1472          */
1473         leaf = path->nodes[0];
1474         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1475         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1476         refs = btrfs_extent_refs(leaf, item);
1477         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
1478         if (extent_op)
1479                 __run_delayed_extent_op(extent_op, leaf, item);
1480
1481         btrfs_mark_buffer_dirty(leaf);
1482         btrfs_release_path(path);
1483
1484         /* now insert the actual backref */
1485         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1486                 BUG_ON(refs_to_add != 1);
1487                 ret = insert_tree_block_ref(trans, path, bytenr, parent,
1488                                             root_objectid);
1489         } else {
1490                 ret = insert_extent_data_ref(trans, path, bytenr, parent,
1491                                              root_objectid, owner, offset,
1492                                              refs_to_add);
1493         }
1494         if (ret)
1495                 btrfs_abort_transaction(trans, ret);
1496 out:
1497         btrfs_free_path(path);
1498         return ret;
1499 }
1500
1501 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
1502                                 struct btrfs_delayed_ref_node *node,
1503                                 struct btrfs_delayed_extent_op *extent_op,
1504                                 int insert_reserved)
1505 {
1506         int ret = 0;
1507         struct btrfs_delayed_data_ref *ref;
1508         struct btrfs_key ins;
1509         u64 parent = 0;
1510         u64 ref_root = 0;
1511         u64 flags = 0;
1512
1513         ins.objectid = node->bytenr;
1514         ins.offset = node->num_bytes;
1515         ins.type = BTRFS_EXTENT_ITEM_KEY;
1516
1517         ref = btrfs_delayed_node_to_data_ref(node);
1518         trace_run_delayed_data_ref(trans->fs_info, node, ref, node->action);
1519
1520         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
1521                 parent = ref->parent;
1522         ref_root = ref->root;
1523
1524         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1525                 if (extent_op)
1526                         flags |= extent_op->flags_to_set;
1527                 ret = alloc_reserved_file_extent(trans, parent, ref_root,
1528                                                  flags, ref->objectid,
1529                                                  ref->offset, &ins,
1530                                                  node->ref_mod);
1531         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1532                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
1533                                              ref->objectid, ref->offset,
1534                                              node->ref_mod, extent_op);
1535         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1536                 ret = __btrfs_free_extent(trans, node, parent,
1537                                           ref_root, ref->objectid,
1538                                           ref->offset, node->ref_mod,
1539                                           extent_op);
1540         } else {
1541                 BUG();
1542         }
1543         return ret;
1544 }
1545
1546 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
1547                                     struct extent_buffer *leaf,
1548                                     struct btrfs_extent_item *ei)
1549 {
1550         u64 flags = btrfs_extent_flags(leaf, ei);
1551         if (extent_op->update_flags) {
1552                 flags |= extent_op->flags_to_set;
1553                 btrfs_set_extent_flags(leaf, ei, flags);
1554         }
1555
1556         if (extent_op->update_key) {
1557                 struct btrfs_tree_block_info *bi;
1558                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
1559                 bi = (struct btrfs_tree_block_info *)(ei + 1);
1560                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
1561         }
1562 }
1563
1564 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
1565                                  struct btrfs_delayed_ref_head *head,
1566                                  struct btrfs_delayed_extent_op *extent_op)
1567 {
1568         struct btrfs_fs_info *fs_info = trans->fs_info;
1569         struct btrfs_root *root;
1570         struct btrfs_key key;
1571         struct btrfs_path *path;
1572         struct btrfs_extent_item *ei;
1573         struct extent_buffer *leaf;
1574         u32 item_size;
1575         int ret;
1576         int err = 0;
1577         int metadata = 1;
1578
1579         if (TRANS_ABORTED(trans))
1580                 return 0;
1581
1582         if (!btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1583                 metadata = 0;
1584
1585         path = btrfs_alloc_path();
1586         if (!path)
1587                 return -ENOMEM;
1588
1589         key.objectid = head->bytenr;
1590
1591         if (metadata) {
1592                 key.type = BTRFS_METADATA_ITEM_KEY;
1593                 key.offset = extent_op->level;
1594         } else {
1595                 key.type = BTRFS_EXTENT_ITEM_KEY;
1596                 key.offset = head->num_bytes;
1597         }
1598
1599         root = btrfs_extent_root(fs_info, key.objectid);
1600 again:
1601         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
1602         if (ret < 0) {
1603                 err = ret;
1604                 goto out;
1605         }
1606         if (ret > 0) {
1607                 if (metadata) {
1608                         if (path->slots[0] > 0) {
1609                                 path->slots[0]--;
1610                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
1611                                                       path->slots[0]);
1612                                 if (key.objectid == head->bytenr &&
1613                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
1614                                     key.offset == head->num_bytes)
1615                                         ret = 0;
1616                         }
1617                         if (ret > 0) {
1618                                 btrfs_release_path(path);
1619                                 metadata = 0;
1620
1621                                 key.objectid = head->bytenr;
1622                                 key.offset = head->num_bytes;
1623                                 key.type = BTRFS_EXTENT_ITEM_KEY;
1624                                 goto again;
1625                         }
1626                 } else {
1627                         err = -EIO;
1628                         goto out;
1629                 }
1630         }
1631
1632         leaf = path->nodes[0];
1633         item_size = btrfs_item_size(leaf, path->slots[0]);
1634
1635         if (unlikely(item_size < sizeof(*ei))) {
1636                 err = -EINVAL;
1637                 btrfs_print_v0_err(fs_info);
1638                 btrfs_abort_transaction(trans, err);
1639                 goto out;
1640         }
1641
1642         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1643         __run_delayed_extent_op(extent_op, leaf, ei);
1644
1645         btrfs_mark_buffer_dirty(leaf);
1646 out:
1647         btrfs_free_path(path);
1648         return err;
1649 }
1650
1651 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
1652                                 struct btrfs_delayed_ref_node *node,
1653                                 struct btrfs_delayed_extent_op *extent_op,
1654                                 int insert_reserved)
1655 {
1656         int ret = 0;
1657         struct btrfs_delayed_tree_ref *ref;
1658         u64 parent = 0;
1659         u64 ref_root = 0;
1660
1661         ref = btrfs_delayed_node_to_tree_ref(node);
1662         trace_run_delayed_tree_ref(trans->fs_info, node, ref, node->action);
1663
1664         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
1665                 parent = ref->parent;
1666         ref_root = ref->root;
1667
1668         if (node->ref_mod != 1) {
1669                 btrfs_err(trans->fs_info,
1670         "btree block(%llu) has %d references rather than 1: action %d ref_root %llu parent %llu",
1671                           node->bytenr, node->ref_mod, node->action, ref_root,
1672                           parent);
1673                 return -EIO;
1674         }
1675         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1676                 BUG_ON(!extent_op || !extent_op->update_flags);
1677                 ret = alloc_reserved_tree_block(trans, node, extent_op);
1678         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1679                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
1680                                              ref->level, 0, 1, extent_op);
1681         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1682                 ret = __btrfs_free_extent(trans, node, parent, ref_root,
1683                                           ref->level, 0, 1, extent_op);
1684         } else {
1685                 BUG();
1686         }
1687         return ret;
1688 }
1689
1690 /* helper function to actually process a single delayed ref entry */
1691 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
1692                                struct btrfs_delayed_ref_node *node,
1693                                struct btrfs_delayed_extent_op *extent_op,
1694                                int insert_reserved)
1695 {
1696         int ret = 0;
1697
1698         if (TRANS_ABORTED(trans)) {
1699                 if (insert_reserved)
1700                         btrfs_pin_extent(trans, node->bytenr, node->num_bytes, 1);
1701                 return 0;
1702         }
1703
1704         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
1705             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
1706                 ret = run_delayed_tree_ref(trans, node, extent_op,
1707                                            insert_reserved);
1708         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
1709                  node->type == BTRFS_SHARED_DATA_REF_KEY)
1710                 ret = run_delayed_data_ref(trans, node, extent_op,
1711                                            insert_reserved);
1712         else
1713                 BUG();
1714         if (ret && insert_reserved)
1715                 btrfs_pin_extent(trans, node->bytenr, node->num_bytes, 1);
1716         return ret;
1717 }
1718
1719 static inline struct btrfs_delayed_ref_node *
1720 select_delayed_ref(struct btrfs_delayed_ref_head *head)
1721 {
1722         struct btrfs_delayed_ref_node *ref;
1723
1724         if (RB_EMPTY_ROOT(&head->ref_tree.rb_root))
1725                 return NULL;
1726
1727         /*
1728          * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
1729          * This is to prevent a ref count from going down to zero, which deletes
1730          * the extent item from the extent tree, when there still are references
1731          * to add, which would fail because they would not find the extent item.
1732          */
1733         if (!list_empty(&head->ref_add_list))
1734                 return list_first_entry(&head->ref_add_list,
1735                                 struct btrfs_delayed_ref_node, add_list);
1736
1737         ref = rb_entry(rb_first_cached(&head->ref_tree),
1738                        struct btrfs_delayed_ref_node, ref_node);
1739         ASSERT(list_empty(&ref->add_list));
1740         return ref;
1741 }
1742
1743 static void unselect_delayed_ref_head(struct btrfs_delayed_ref_root *delayed_refs,
1744                                       struct btrfs_delayed_ref_head *head)
1745 {
1746         spin_lock(&delayed_refs->lock);
1747         head->processing = 0;
1748         delayed_refs->num_heads_ready++;
1749         spin_unlock(&delayed_refs->lock);
1750         btrfs_delayed_ref_unlock(head);
1751 }
1752
1753 static struct btrfs_delayed_extent_op *cleanup_extent_op(
1754                                 struct btrfs_delayed_ref_head *head)
1755 {
1756         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
1757
1758         if (!extent_op)
1759                 return NULL;
1760
1761         if (head->must_insert_reserved) {
1762                 head->extent_op = NULL;
1763                 btrfs_free_delayed_extent_op(extent_op);
1764                 return NULL;
1765         }
1766         return extent_op;
1767 }
1768
1769 static int run_and_cleanup_extent_op(struct btrfs_trans_handle *trans,
1770                                      struct btrfs_delayed_ref_head *head)
1771 {
1772         struct btrfs_delayed_extent_op *extent_op;
1773         int ret;
1774
1775         extent_op = cleanup_extent_op(head);
1776         if (!extent_op)
1777                 return 0;
1778         head->extent_op = NULL;
1779         spin_unlock(&head->lock);
1780         ret = run_delayed_extent_op(trans, head, extent_op);
1781         btrfs_free_delayed_extent_op(extent_op);
1782         return ret ? ret : 1;
1783 }
1784
1785 void btrfs_cleanup_ref_head_accounting(struct btrfs_fs_info *fs_info,
1786                                   struct btrfs_delayed_ref_root *delayed_refs,
1787                                   struct btrfs_delayed_ref_head *head)
1788 {
1789         int nr_items = 1;       /* Dropping this ref head update. */
1790
1791         /*
1792          * We had csum deletions accounted for in our delayed refs rsv, we need
1793          * to drop the csum leaves for this update from our delayed_refs_rsv.
1794          */
1795         if (head->total_ref_mod < 0 && head->is_data) {
1796                 spin_lock(&delayed_refs->lock);
1797                 delayed_refs->pending_csums -= head->num_bytes;
1798                 spin_unlock(&delayed_refs->lock);
1799                 nr_items += btrfs_csum_bytes_to_leaves(fs_info, head->num_bytes);
1800         }
1801
1802         btrfs_delayed_refs_rsv_release(fs_info, nr_items);
1803 }
1804
1805 static int cleanup_ref_head(struct btrfs_trans_handle *trans,
1806                             struct btrfs_delayed_ref_head *head)
1807 {
1808
1809         struct btrfs_fs_info *fs_info = trans->fs_info;
1810         struct btrfs_delayed_ref_root *delayed_refs;
1811         int ret;
1812
1813         delayed_refs = &trans->transaction->delayed_refs;
1814
1815         ret = run_and_cleanup_extent_op(trans, head);
1816         if (ret < 0) {
1817                 unselect_delayed_ref_head(delayed_refs, head);
1818                 btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
1819                 return ret;
1820         } else if (ret) {
1821                 return ret;
1822         }
1823
1824         /*
1825          * Need to drop our head ref lock and re-acquire the delayed ref lock
1826          * and then re-check to make sure nobody got added.
1827          */
1828         spin_unlock(&head->lock);
1829         spin_lock(&delayed_refs->lock);
1830         spin_lock(&head->lock);
1831         if (!RB_EMPTY_ROOT(&head->ref_tree.rb_root) || head->extent_op) {
1832                 spin_unlock(&head->lock);
1833                 spin_unlock(&delayed_refs->lock);
1834                 return 1;
1835         }
1836         btrfs_delete_ref_head(delayed_refs, head);
1837         spin_unlock(&head->lock);
1838         spin_unlock(&delayed_refs->lock);
1839
1840         if (head->must_insert_reserved) {
1841                 btrfs_pin_extent(trans, head->bytenr, head->num_bytes, 1);
1842                 if (head->is_data) {
1843                         struct btrfs_root *csum_root;
1844
1845                         csum_root = btrfs_csum_root(fs_info, head->bytenr);
1846                         ret = btrfs_del_csums(trans, csum_root, head->bytenr,
1847                                               head->num_bytes);
1848                 }
1849         }
1850
1851         btrfs_cleanup_ref_head_accounting(fs_info, delayed_refs, head);
1852
1853         trace_run_delayed_ref_head(fs_info, head, 0);
1854         btrfs_delayed_ref_unlock(head);
1855         btrfs_put_delayed_ref_head(head);
1856         return ret;
1857 }
1858
1859 static struct btrfs_delayed_ref_head *btrfs_obtain_ref_head(
1860                                         struct btrfs_trans_handle *trans)
1861 {
1862         struct btrfs_delayed_ref_root *delayed_refs =
1863                 &trans->transaction->delayed_refs;
1864         struct btrfs_delayed_ref_head *head = NULL;
1865         int ret;
1866
1867         spin_lock(&delayed_refs->lock);
1868         head = btrfs_select_ref_head(delayed_refs);
1869         if (!head) {
1870                 spin_unlock(&delayed_refs->lock);
1871                 return head;
1872         }
1873
1874         /*
1875          * Grab the lock that says we are going to process all the refs for
1876          * this head
1877          */
1878         ret = btrfs_delayed_ref_lock(delayed_refs, head);
1879         spin_unlock(&delayed_refs->lock);
1880
1881         /*
1882          * We may have dropped the spin lock to get the head mutex lock, and
1883          * that might have given someone else time to free the head.  If that's
1884          * true, it has been removed from our list and we can move on.
1885          */
1886         if (ret == -EAGAIN)
1887                 head = ERR_PTR(-EAGAIN);
1888
1889         return head;
1890 }
1891
1892 static int btrfs_run_delayed_refs_for_head(struct btrfs_trans_handle *trans,
1893                                     struct btrfs_delayed_ref_head *locked_ref,
1894                                     unsigned long *run_refs)
1895 {
1896         struct btrfs_fs_info *fs_info = trans->fs_info;
1897         struct btrfs_delayed_ref_root *delayed_refs;
1898         struct btrfs_delayed_extent_op *extent_op;
1899         struct btrfs_delayed_ref_node *ref;
1900         int must_insert_reserved = 0;
1901         int ret;
1902
1903         delayed_refs = &trans->transaction->delayed_refs;
1904
1905         lockdep_assert_held(&locked_ref->mutex);
1906         lockdep_assert_held(&locked_ref->lock);
1907
1908         while ((ref = select_delayed_ref(locked_ref))) {
1909                 if (ref->seq &&
1910                     btrfs_check_delayed_seq(fs_info, ref->seq)) {
1911                         spin_unlock(&locked_ref->lock);
1912                         unselect_delayed_ref_head(delayed_refs, locked_ref);
1913                         return -EAGAIN;
1914                 }
1915
1916                 (*run_refs)++;
1917                 ref->in_tree = 0;
1918                 rb_erase_cached(&ref->ref_node, &locked_ref->ref_tree);
1919                 RB_CLEAR_NODE(&ref->ref_node);
1920                 if (!list_empty(&ref->add_list))
1921                         list_del(&ref->add_list);
1922                 /*
1923                  * When we play the delayed ref, also correct the ref_mod on
1924                  * head
1925                  */
1926                 switch (ref->action) {
1927                 case BTRFS_ADD_DELAYED_REF:
1928                 case BTRFS_ADD_DELAYED_EXTENT:
1929                         locked_ref->ref_mod -= ref->ref_mod;
1930                         break;
1931                 case BTRFS_DROP_DELAYED_REF:
1932                         locked_ref->ref_mod += ref->ref_mod;
1933                         break;
1934                 default:
1935                         WARN_ON(1);
1936                 }
1937                 atomic_dec(&delayed_refs->num_entries);
1938
1939                 /*
1940                  * Record the must_insert_reserved flag before we drop the
1941                  * spin lock.
1942                  */
1943                 must_insert_reserved = locked_ref->must_insert_reserved;
1944                 locked_ref->must_insert_reserved = 0;
1945
1946                 extent_op = locked_ref->extent_op;
1947                 locked_ref->extent_op = NULL;
1948                 spin_unlock(&locked_ref->lock);
1949
1950                 ret = run_one_delayed_ref(trans, ref, extent_op,
1951                                           must_insert_reserved);
1952
1953                 btrfs_free_delayed_extent_op(extent_op);
1954                 if (ret) {
1955                         unselect_delayed_ref_head(delayed_refs, locked_ref);
1956                         btrfs_put_delayed_ref(ref);
1957                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d",
1958                                     ret);
1959                         return ret;
1960                 }
1961
1962                 btrfs_put_delayed_ref(ref);
1963                 cond_resched();
1964
1965                 spin_lock(&locked_ref->lock);
1966                 btrfs_merge_delayed_refs(trans, delayed_refs, locked_ref);
1967         }
1968
1969         return 0;
1970 }
1971
1972 /*
1973  * Returns 0 on success or if called with an already aborted transaction.
1974  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
1975  */
1976 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
1977                                              unsigned long nr)
1978 {
1979         struct btrfs_fs_info *fs_info = trans->fs_info;
1980         struct btrfs_delayed_ref_root *delayed_refs;
1981         struct btrfs_delayed_ref_head *locked_ref = NULL;
1982         ktime_t start = ktime_get();
1983         int ret;
1984         unsigned long count = 0;
1985         unsigned long actual_count = 0;
1986
1987         delayed_refs = &trans->transaction->delayed_refs;
1988         do {
1989                 if (!locked_ref) {
1990                         locked_ref = btrfs_obtain_ref_head(trans);
1991                         if (IS_ERR_OR_NULL(locked_ref)) {
1992                                 if (PTR_ERR(locked_ref) == -EAGAIN) {
1993                                         continue;
1994                                 } else {
1995                                         break;
1996                                 }
1997                         }
1998                         count++;
1999                 }
2000                 /*
2001                  * We need to try and merge add/drops of the same ref since we
2002                  * can run into issues with relocate dropping the implicit ref
2003                  * and then it being added back again before the drop can
2004                  * finish.  If we merged anything we need to re-loop so we can
2005                  * get a good ref.
2006                  * Or we can get node references of the same type that weren't
2007                  * merged when created due to bumps in the tree mod seq, and
2008                  * we need to merge them to prevent adding an inline extent
2009                  * backref before dropping it (triggering a BUG_ON at
2010                  * insert_inline_extent_backref()).
2011                  */
2012                 spin_lock(&locked_ref->lock);
2013                 btrfs_merge_delayed_refs(trans, delayed_refs, locked_ref);
2014
2015                 ret = btrfs_run_delayed_refs_for_head(trans, locked_ref,
2016                                                       &actual_count);
2017                 if (ret < 0 && ret != -EAGAIN) {
2018                         /*
2019                          * Error, btrfs_run_delayed_refs_for_head already
2020                          * unlocked everything so just bail out
2021                          */
2022                         return ret;
2023                 } else if (!ret) {
2024                         /*
2025                          * Success, perform the usual cleanup of a processed
2026                          * head
2027                          */
2028                         ret = cleanup_ref_head(trans, locked_ref);
2029                         if (ret > 0 ) {
2030                                 /* We dropped our lock, we need to loop. */
2031                                 ret = 0;
2032                                 continue;
2033                         } else if (ret) {
2034                                 return ret;
2035                         }
2036                 }
2037
2038                 /*
2039                  * Either success case or btrfs_run_delayed_refs_for_head
2040                  * returned -EAGAIN, meaning we need to select another head
2041                  */
2042
2043                 locked_ref = NULL;
2044                 cond_resched();
2045         } while ((nr != -1 && count < nr) || locked_ref);
2046
2047         /*
2048          * We don't want to include ref heads since we can have empty ref heads
2049          * and those will drastically skew our runtime down since we just do
2050          * accounting, no actual extent tree updates.
2051          */
2052         if (actual_count > 0) {
2053                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2054                 u64 avg;
2055
2056                 /*
2057                  * We weigh the current average higher than our current runtime
2058                  * to avoid large swings in the average.
2059                  */
2060                 spin_lock(&delayed_refs->lock);
2061                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2062                 fs_info->avg_delayed_ref_runtime = avg >> 2;    /* div by 4 */
2063                 spin_unlock(&delayed_refs->lock);
2064         }
2065         return 0;
2066 }
2067
2068 #ifdef SCRAMBLE_DELAYED_REFS
2069 /*
2070  * Normally delayed refs get processed in ascending bytenr order. This
2071  * correlates in most cases to the order added. To expose dependencies on this
2072  * order, we start to process the tree in the middle instead of the beginning
2073  */
2074 static u64 find_middle(struct rb_root *root)
2075 {
2076         struct rb_node *n = root->rb_node;
2077         struct btrfs_delayed_ref_node *entry;
2078         int alt = 1;
2079         u64 middle;
2080         u64 first = 0, last = 0;
2081
2082         n = rb_first(root);
2083         if (n) {
2084                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2085                 first = entry->bytenr;
2086         }
2087         n = rb_last(root);
2088         if (n) {
2089                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2090                 last = entry->bytenr;
2091         }
2092         n = root->rb_node;
2093
2094         while (n) {
2095                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2096                 WARN_ON(!entry->in_tree);
2097
2098                 middle = entry->bytenr;
2099
2100                 if (alt)
2101                         n = n->rb_left;
2102                 else
2103                         n = n->rb_right;
2104
2105                 alt = 1 - alt;
2106         }
2107         return middle;
2108 }
2109 #endif
2110
2111 /*
2112  * this starts processing the delayed reference count updates and
2113  * extent insertions we have queued up so far.  count can be
2114  * 0, which means to process everything in the tree at the start
2115  * of the run (but not newly added entries), or it can be some target
2116  * number you'd like to process.
2117  *
2118  * Returns 0 on success or if called with an aborted transaction
2119  * Returns <0 on error and aborts the transaction
2120  */
2121 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2122                            unsigned long count)
2123 {
2124         struct btrfs_fs_info *fs_info = trans->fs_info;
2125         struct rb_node *node;
2126         struct btrfs_delayed_ref_root *delayed_refs;
2127         struct btrfs_delayed_ref_head *head;
2128         int ret;
2129         int run_all = count == (unsigned long)-1;
2130
2131         /* We'll clean this up in btrfs_cleanup_transaction */
2132         if (TRANS_ABORTED(trans))
2133                 return 0;
2134
2135         if (test_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags))
2136                 return 0;
2137
2138         delayed_refs = &trans->transaction->delayed_refs;
2139         if (count == 0)
2140                 count = delayed_refs->num_heads_ready;
2141
2142 again:
2143 #ifdef SCRAMBLE_DELAYED_REFS
2144         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2145 #endif
2146         ret = __btrfs_run_delayed_refs(trans, count);
2147         if (ret < 0) {
2148                 btrfs_abort_transaction(trans, ret);
2149                 return ret;
2150         }
2151
2152         if (run_all) {
2153                 btrfs_create_pending_block_groups(trans);
2154
2155                 spin_lock(&delayed_refs->lock);
2156                 node = rb_first_cached(&delayed_refs->href_root);
2157                 if (!node) {
2158                         spin_unlock(&delayed_refs->lock);
2159                         goto out;
2160                 }
2161                 head = rb_entry(node, struct btrfs_delayed_ref_head,
2162                                 href_node);
2163                 refcount_inc(&head->refs);
2164                 spin_unlock(&delayed_refs->lock);
2165
2166                 /* Mutex was contended, block until it's released and retry. */
2167                 mutex_lock(&head->mutex);
2168                 mutex_unlock(&head->mutex);
2169
2170                 btrfs_put_delayed_ref_head(head);
2171                 cond_resched();
2172                 goto again;
2173         }
2174 out:
2175         return 0;
2176 }
2177
2178 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2179                                 struct extent_buffer *eb, u64 flags,
2180                                 int level)
2181 {
2182         struct btrfs_delayed_extent_op *extent_op;
2183         int ret;
2184
2185         extent_op = btrfs_alloc_delayed_extent_op();
2186         if (!extent_op)
2187                 return -ENOMEM;
2188
2189         extent_op->flags_to_set = flags;
2190         extent_op->update_flags = true;
2191         extent_op->update_key = false;
2192         extent_op->level = level;
2193
2194         ret = btrfs_add_delayed_extent_op(trans, eb->start, eb->len, extent_op);
2195         if (ret)
2196                 btrfs_free_delayed_extent_op(extent_op);
2197         return ret;
2198 }
2199
2200 static noinline int check_delayed_ref(struct btrfs_root *root,
2201                                       struct btrfs_path *path,
2202                                       u64 objectid, u64 offset, u64 bytenr)
2203 {
2204         struct btrfs_delayed_ref_head *head;
2205         struct btrfs_delayed_ref_node *ref;
2206         struct btrfs_delayed_data_ref *data_ref;
2207         struct btrfs_delayed_ref_root *delayed_refs;
2208         struct btrfs_transaction *cur_trans;
2209         struct rb_node *node;
2210         int ret = 0;
2211
2212         spin_lock(&root->fs_info->trans_lock);
2213         cur_trans = root->fs_info->running_transaction;
2214         if (cur_trans)
2215                 refcount_inc(&cur_trans->use_count);
2216         spin_unlock(&root->fs_info->trans_lock);
2217         if (!cur_trans)
2218                 return 0;
2219
2220         delayed_refs = &cur_trans->delayed_refs;
2221         spin_lock(&delayed_refs->lock);
2222         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
2223         if (!head) {
2224                 spin_unlock(&delayed_refs->lock);
2225                 btrfs_put_transaction(cur_trans);
2226                 return 0;
2227         }
2228
2229         if (!mutex_trylock(&head->mutex)) {
2230                 if (path->nowait) {
2231                         spin_unlock(&delayed_refs->lock);
2232                         btrfs_put_transaction(cur_trans);
2233                         return -EAGAIN;
2234                 }
2235
2236                 refcount_inc(&head->refs);
2237                 spin_unlock(&delayed_refs->lock);
2238
2239                 btrfs_release_path(path);
2240
2241                 /*
2242                  * Mutex was contended, block until it's released and let
2243                  * caller try again
2244                  */
2245                 mutex_lock(&head->mutex);
2246                 mutex_unlock(&head->mutex);
2247                 btrfs_put_delayed_ref_head(head);
2248                 btrfs_put_transaction(cur_trans);
2249                 return -EAGAIN;
2250         }
2251         spin_unlock(&delayed_refs->lock);
2252
2253         spin_lock(&head->lock);
2254         /*
2255          * XXX: We should replace this with a proper search function in the
2256          * future.
2257          */
2258         for (node = rb_first_cached(&head->ref_tree); node;
2259              node = rb_next(node)) {
2260                 ref = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
2261                 /* If it's a shared ref we know a cross reference exists */
2262                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
2263                         ret = 1;
2264                         break;
2265                 }
2266
2267                 data_ref = btrfs_delayed_node_to_data_ref(ref);
2268
2269                 /*
2270                  * If our ref doesn't match the one we're currently looking at
2271                  * then we have a cross reference.
2272                  */
2273                 if (data_ref->root != root->root_key.objectid ||
2274                     data_ref->objectid != objectid ||
2275                     data_ref->offset != offset) {
2276                         ret = 1;
2277                         break;
2278                 }
2279         }
2280         spin_unlock(&head->lock);
2281         mutex_unlock(&head->mutex);
2282         btrfs_put_transaction(cur_trans);
2283         return ret;
2284 }
2285
2286 static noinline int check_committed_ref(struct btrfs_root *root,
2287                                         struct btrfs_path *path,
2288                                         u64 objectid, u64 offset, u64 bytenr,
2289                                         bool strict)
2290 {
2291         struct btrfs_fs_info *fs_info = root->fs_info;
2292         struct btrfs_root *extent_root = btrfs_extent_root(fs_info, bytenr);
2293         struct extent_buffer *leaf;
2294         struct btrfs_extent_data_ref *ref;
2295         struct btrfs_extent_inline_ref *iref;
2296         struct btrfs_extent_item *ei;
2297         struct btrfs_key key;
2298         u32 item_size;
2299         int type;
2300         int ret;
2301
2302         key.objectid = bytenr;
2303         key.offset = (u64)-1;
2304         key.type = BTRFS_EXTENT_ITEM_KEY;
2305
2306         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2307         if (ret < 0)
2308                 goto out;
2309         BUG_ON(ret == 0); /* Corruption */
2310
2311         ret = -ENOENT;
2312         if (path->slots[0] == 0)
2313                 goto out;
2314
2315         path->slots[0]--;
2316         leaf = path->nodes[0];
2317         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2318
2319         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2320                 goto out;
2321
2322         ret = 1;
2323         item_size = btrfs_item_size(leaf, path->slots[0]);
2324         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2325
2326         /* If extent item has more than 1 inline ref then it's shared */
2327         if (item_size != sizeof(*ei) +
2328             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2329                 goto out;
2330
2331         /*
2332          * If extent created before last snapshot => it's shared unless the
2333          * snapshot has been deleted. Use the heuristic if strict is false.
2334          */
2335         if (!strict &&
2336             (btrfs_extent_generation(leaf, ei) <=
2337              btrfs_root_last_snapshot(&root->root_item)))
2338                 goto out;
2339
2340         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2341
2342         /* If this extent has SHARED_DATA_REF then it's shared */
2343         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
2344         if (type != BTRFS_EXTENT_DATA_REF_KEY)
2345                 goto out;
2346
2347         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2348         if (btrfs_extent_refs(leaf, ei) !=
2349             btrfs_extent_data_ref_count(leaf, ref) ||
2350             btrfs_extent_data_ref_root(leaf, ref) !=
2351             root->root_key.objectid ||
2352             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2353             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2354                 goto out;
2355
2356         ret = 0;
2357 out:
2358         return ret;
2359 }
2360
2361 int btrfs_cross_ref_exist(struct btrfs_root *root, u64 objectid, u64 offset,
2362                           u64 bytenr, bool strict, struct btrfs_path *path)
2363 {
2364         int ret;
2365
2366         do {
2367                 ret = check_committed_ref(root, path, objectid,
2368                                           offset, bytenr, strict);
2369                 if (ret && ret != -ENOENT)
2370                         goto out;
2371
2372                 ret = check_delayed_ref(root, path, objectid, offset, bytenr);
2373         } while (ret == -EAGAIN);
2374
2375 out:
2376         btrfs_release_path(path);
2377         if (btrfs_is_data_reloc_root(root))
2378                 WARN_ON(ret > 0);
2379         return ret;
2380 }
2381
2382 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2383                            struct btrfs_root *root,
2384                            struct extent_buffer *buf,
2385                            int full_backref, int inc)
2386 {
2387         struct btrfs_fs_info *fs_info = root->fs_info;
2388         u64 bytenr;
2389         u64 num_bytes;
2390         u64 parent;
2391         u64 ref_root;
2392         u32 nritems;
2393         struct btrfs_key key;
2394         struct btrfs_file_extent_item *fi;
2395         struct btrfs_ref generic_ref = { 0 };
2396         bool for_reloc = btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC);
2397         int i;
2398         int action;
2399         int level;
2400         int ret = 0;
2401
2402         if (btrfs_is_testing(fs_info))
2403                 return 0;
2404
2405         ref_root = btrfs_header_owner(buf);
2406         nritems = btrfs_header_nritems(buf);
2407         level = btrfs_header_level(buf);
2408
2409         if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state) && level == 0)
2410                 return 0;
2411
2412         if (full_backref)
2413                 parent = buf->start;
2414         else
2415                 parent = 0;
2416         if (inc)
2417                 action = BTRFS_ADD_DELAYED_REF;
2418         else
2419                 action = BTRFS_DROP_DELAYED_REF;
2420
2421         for (i = 0; i < nritems; i++) {
2422                 if (level == 0) {
2423                         btrfs_item_key_to_cpu(buf, &key, i);
2424                         if (key.type != BTRFS_EXTENT_DATA_KEY)
2425                                 continue;
2426                         fi = btrfs_item_ptr(buf, i,
2427                                             struct btrfs_file_extent_item);
2428                         if (btrfs_file_extent_type(buf, fi) ==
2429                             BTRFS_FILE_EXTENT_INLINE)
2430                                 continue;
2431                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2432                         if (bytenr == 0)
2433                                 continue;
2434
2435                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
2436                         key.offset -= btrfs_file_extent_offset(buf, fi);
2437                         btrfs_init_generic_ref(&generic_ref, action, bytenr,
2438                                                num_bytes, parent);
2439                         btrfs_init_data_ref(&generic_ref, ref_root, key.objectid,
2440                                             key.offset, root->root_key.objectid,
2441                                             for_reloc);
2442                         if (inc)
2443                                 ret = btrfs_inc_extent_ref(trans, &generic_ref);
2444                         else
2445                                 ret = btrfs_free_extent(trans, &generic_ref);
2446                         if (ret)
2447                                 goto fail;
2448                 } else {
2449                         bytenr = btrfs_node_blockptr(buf, i);
2450                         num_bytes = fs_info->nodesize;
2451                         btrfs_init_generic_ref(&generic_ref, action, bytenr,
2452                                                num_bytes, parent);
2453                         btrfs_init_tree_ref(&generic_ref, level - 1, ref_root,
2454                                             root->root_key.objectid, for_reloc);
2455                         if (inc)
2456                                 ret = btrfs_inc_extent_ref(trans, &generic_ref);
2457                         else
2458                                 ret = btrfs_free_extent(trans, &generic_ref);
2459                         if (ret)
2460                                 goto fail;
2461                 }
2462         }
2463         return 0;
2464 fail:
2465         return ret;
2466 }
2467
2468 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2469                   struct extent_buffer *buf, int full_backref)
2470 {
2471         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
2472 }
2473
2474 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2475                   struct extent_buffer *buf, int full_backref)
2476 {
2477         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
2478 }
2479
2480 static u64 get_alloc_profile_by_root(struct btrfs_root *root, int data)
2481 {
2482         struct btrfs_fs_info *fs_info = root->fs_info;
2483         u64 flags;
2484         u64 ret;
2485
2486         if (data)
2487                 flags = BTRFS_BLOCK_GROUP_DATA;
2488         else if (root == fs_info->chunk_root)
2489                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
2490         else
2491                 flags = BTRFS_BLOCK_GROUP_METADATA;
2492
2493         ret = btrfs_get_alloc_profile(fs_info, flags);
2494         return ret;
2495 }
2496
2497 static u64 first_logical_byte(struct btrfs_fs_info *fs_info)
2498 {
2499         struct rb_node *leftmost;
2500         u64 bytenr = 0;
2501
2502         read_lock(&fs_info->block_group_cache_lock);
2503         /* Get the block group with the lowest logical start address. */
2504         leftmost = rb_first_cached(&fs_info->block_group_cache_tree);
2505         if (leftmost) {
2506                 struct btrfs_block_group *bg;
2507
2508                 bg = rb_entry(leftmost, struct btrfs_block_group, cache_node);
2509                 bytenr = bg->start;
2510         }
2511         read_unlock(&fs_info->block_group_cache_lock);
2512
2513         return bytenr;
2514 }
2515
2516 static int pin_down_extent(struct btrfs_trans_handle *trans,
2517                            struct btrfs_block_group *cache,
2518                            u64 bytenr, u64 num_bytes, int reserved)
2519 {
2520         struct btrfs_fs_info *fs_info = cache->fs_info;
2521
2522         spin_lock(&cache->space_info->lock);
2523         spin_lock(&cache->lock);
2524         cache->pinned += num_bytes;
2525         btrfs_space_info_update_bytes_pinned(fs_info, cache->space_info,
2526                                              num_bytes);
2527         if (reserved) {
2528                 cache->reserved -= num_bytes;
2529                 cache->space_info->bytes_reserved -= num_bytes;
2530         }
2531         spin_unlock(&cache->lock);
2532         spin_unlock(&cache->space_info->lock);
2533
2534         set_extent_dirty(&trans->transaction->pinned_extents, bytenr,
2535                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
2536         return 0;
2537 }
2538
2539 int btrfs_pin_extent(struct btrfs_trans_handle *trans,
2540                      u64 bytenr, u64 num_bytes, int reserved)
2541 {
2542         struct btrfs_block_group *cache;
2543
2544         cache = btrfs_lookup_block_group(trans->fs_info, bytenr);
2545         BUG_ON(!cache); /* Logic error */
2546
2547         pin_down_extent(trans, cache, bytenr, num_bytes, reserved);
2548
2549         btrfs_put_block_group(cache);
2550         return 0;
2551 }
2552
2553 /*
2554  * this function must be called within transaction
2555  */
2556 int btrfs_pin_extent_for_log_replay(struct btrfs_trans_handle *trans,
2557                                     u64 bytenr, u64 num_bytes)
2558 {
2559         struct btrfs_block_group *cache;
2560         int ret;
2561
2562         cache = btrfs_lookup_block_group(trans->fs_info, bytenr);
2563         if (!cache)
2564                 return -EINVAL;
2565
2566         /*
2567          * Fully cache the free space first so that our pin removes the free space
2568          * from the cache.
2569          */
2570         ret = btrfs_cache_block_group(cache, true);
2571         if (ret)
2572                 goto out;
2573
2574         pin_down_extent(trans, cache, bytenr, num_bytes, 0);
2575
2576         /* remove us from the free space cache (if we're there at all) */
2577         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
2578 out:
2579         btrfs_put_block_group(cache);
2580         return ret;
2581 }
2582
2583 static int __exclude_logged_extent(struct btrfs_fs_info *fs_info,
2584                                    u64 start, u64 num_bytes)
2585 {
2586         int ret;
2587         struct btrfs_block_group *block_group;
2588
2589         block_group = btrfs_lookup_block_group(fs_info, start);
2590         if (!block_group)
2591                 return -EINVAL;
2592
2593         ret = btrfs_cache_block_group(block_group, true);
2594         if (ret)
2595                 goto out;
2596
2597         ret = btrfs_remove_free_space(block_group, start, num_bytes);
2598 out:
2599         btrfs_put_block_group(block_group);
2600         return ret;
2601 }
2602
2603 int btrfs_exclude_logged_extents(struct extent_buffer *eb)
2604 {
2605         struct btrfs_fs_info *fs_info = eb->fs_info;
2606         struct btrfs_file_extent_item *item;
2607         struct btrfs_key key;
2608         int found_type;
2609         int i;
2610         int ret = 0;
2611
2612         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS))
2613                 return 0;
2614
2615         for (i = 0; i < btrfs_header_nritems(eb); i++) {
2616                 btrfs_item_key_to_cpu(eb, &key, i);
2617                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2618                         continue;
2619                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
2620                 found_type = btrfs_file_extent_type(eb, item);
2621                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
2622                         continue;
2623                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
2624                         continue;
2625                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
2626                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
2627                 ret = __exclude_logged_extent(fs_info, key.objectid, key.offset);
2628                 if (ret)
2629                         break;
2630         }
2631
2632         return ret;
2633 }
2634
2635 static void
2636 btrfs_inc_block_group_reservations(struct btrfs_block_group *bg)
2637 {
2638         atomic_inc(&bg->reservations);
2639 }
2640
2641 /*
2642  * Returns the free cluster for the given space info and sets empty_cluster to
2643  * what it should be based on the mount options.
2644  */
2645 static struct btrfs_free_cluster *
2646 fetch_cluster_info(struct btrfs_fs_info *fs_info,
2647                    struct btrfs_space_info *space_info, u64 *empty_cluster)
2648 {
2649         struct btrfs_free_cluster *ret = NULL;
2650
2651         *empty_cluster = 0;
2652         if (btrfs_mixed_space_info(space_info))
2653                 return ret;
2654
2655         if (space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
2656                 ret = &fs_info->meta_alloc_cluster;
2657                 if (btrfs_test_opt(fs_info, SSD))
2658                         *empty_cluster = SZ_2M;
2659                 else
2660                         *empty_cluster = SZ_64K;
2661         } else if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) &&
2662                    btrfs_test_opt(fs_info, SSD_SPREAD)) {
2663                 *empty_cluster = SZ_2M;
2664                 ret = &fs_info->data_alloc_cluster;
2665         }
2666
2667         return ret;
2668 }
2669
2670 static int unpin_extent_range(struct btrfs_fs_info *fs_info,
2671                               u64 start, u64 end,
2672                               const bool return_free_space)
2673 {
2674         struct btrfs_block_group *cache = NULL;
2675         struct btrfs_space_info *space_info;
2676         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
2677         struct btrfs_free_cluster *cluster = NULL;
2678         u64 len;
2679         u64 total_unpinned = 0;
2680         u64 empty_cluster = 0;
2681         bool readonly;
2682
2683         while (start <= end) {
2684                 readonly = false;
2685                 if (!cache ||
2686                     start >= cache->start + cache->length) {
2687                         if (cache)
2688                                 btrfs_put_block_group(cache);
2689                         total_unpinned = 0;
2690                         cache = btrfs_lookup_block_group(fs_info, start);
2691                         BUG_ON(!cache); /* Logic error */
2692
2693                         cluster = fetch_cluster_info(fs_info,
2694                                                      cache->space_info,
2695                                                      &empty_cluster);
2696                         empty_cluster <<= 1;
2697                 }
2698
2699                 len = cache->start + cache->length - start;
2700                 len = min(len, end + 1 - start);
2701
2702                 if (return_free_space)
2703                         btrfs_add_free_space(cache, start, len);
2704
2705                 start += len;
2706                 total_unpinned += len;
2707                 space_info = cache->space_info;
2708
2709                 /*
2710                  * If this space cluster has been marked as fragmented and we've
2711                  * unpinned enough in this block group to potentially allow a
2712                  * cluster to be created inside of it go ahead and clear the
2713                  * fragmented check.
2714                  */
2715                 if (cluster && cluster->fragmented &&
2716                     total_unpinned > empty_cluster) {
2717                         spin_lock(&cluster->lock);
2718                         cluster->fragmented = 0;
2719                         spin_unlock(&cluster->lock);
2720                 }
2721
2722                 spin_lock(&space_info->lock);
2723                 spin_lock(&cache->lock);
2724                 cache->pinned -= len;
2725                 btrfs_space_info_update_bytes_pinned(fs_info, space_info, -len);
2726                 space_info->max_extent_size = 0;
2727                 if (cache->ro) {
2728                         space_info->bytes_readonly += len;
2729                         readonly = true;
2730                 } else if (btrfs_is_zoned(fs_info)) {
2731                         /* Need reset before reusing in a zoned block group */
2732                         space_info->bytes_zone_unusable += len;
2733                         readonly = true;
2734                 }
2735                 spin_unlock(&cache->lock);
2736                 if (!readonly && return_free_space &&
2737                     global_rsv->space_info == space_info) {
2738                         spin_lock(&global_rsv->lock);
2739                         if (!global_rsv->full) {
2740                                 u64 to_add = min(len, global_rsv->size -
2741                                                       global_rsv->reserved);
2742
2743                                 global_rsv->reserved += to_add;
2744                                 btrfs_space_info_update_bytes_may_use(fs_info,
2745                                                 space_info, to_add);
2746                                 if (global_rsv->reserved >= global_rsv->size)
2747                                         global_rsv->full = 1;
2748                                 len -= to_add;
2749                         }
2750                         spin_unlock(&global_rsv->lock);
2751                 }
2752                 /* Add to any tickets we may have */
2753                 if (!readonly && return_free_space && len)
2754                         btrfs_try_granting_tickets(fs_info, space_info);
2755                 spin_unlock(&space_info->lock);
2756         }
2757
2758         if (cache)
2759                 btrfs_put_block_group(cache);
2760         return 0;
2761 }
2762
2763 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans)
2764 {
2765         struct btrfs_fs_info *fs_info = trans->fs_info;
2766         struct btrfs_block_group *block_group, *tmp;
2767         struct list_head *deleted_bgs;
2768         struct extent_io_tree *unpin;
2769         u64 start;
2770         u64 end;
2771         int ret;
2772
2773         unpin = &trans->transaction->pinned_extents;
2774
2775         while (!TRANS_ABORTED(trans)) {
2776                 struct extent_state *cached_state = NULL;
2777
2778                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
2779                 ret = find_first_extent_bit(unpin, 0, &start, &end,
2780                                             EXTENT_DIRTY, &cached_state);
2781                 if (ret) {
2782                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
2783                         break;
2784                 }
2785
2786                 if (btrfs_test_opt(fs_info, DISCARD_SYNC))
2787                         ret = btrfs_discard_extent(fs_info, start,
2788                                                    end + 1 - start, NULL);
2789
2790                 clear_extent_dirty(unpin, start, end, &cached_state);
2791                 unpin_extent_range(fs_info, start, end, true);
2792                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
2793                 free_extent_state(cached_state);
2794                 cond_resched();
2795         }
2796
2797         if (btrfs_test_opt(fs_info, DISCARD_ASYNC)) {
2798                 btrfs_discard_calc_delay(&fs_info->discard_ctl);
2799                 btrfs_discard_schedule_work(&fs_info->discard_ctl, true);
2800         }
2801
2802         /*
2803          * Transaction is finished.  We don't need the lock anymore.  We
2804          * do need to clean up the block groups in case of a transaction
2805          * abort.
2806          */
2807         deleted_bgs = &trans->transaction->deleted_bgs;
2808         list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
2809                 u64 trimmed = 0;
2810
2811                 ret = -EROFS;
2812                 if (!TRANS_ABORTED(trans))
2813                         ret = btrfs_discard_extent(fs_info,
2814                                                    block_group->start,
2815                                                    block_group->length,
2816                                                    &trimmed);
2817
2818                 list_del_init(&block_group->bg_list);
2819                 btrfs_unfreeze_block_group(block_group);
2820                 btrfs_put_block_group(block_group);
2821
2822                 if (ret) {
2823                         const char *errstr = btrfs_decode_error(ret);
2824                         btrfs_warn(fs_info,
2825                            "discard failed while removing blockgroup: errno=%d %s",
2826                                    ret, errstr);
2827                 }
2828         }
2829
2830         return 0;
2831 }
2832
2833 static int do_free_extent_accounting(struct btrfs_trans_handle *trans,
2834                                      u64 bytenr, u64 num_bytes, bool is_data)
2835 {
2836         int ret;
2837
2838         if (is_data) {
2839                 struct btrfs_root *csum_root;
2840
2841                 csum_root = btrfs_csum_root(trans->fs_info, bytenr);
2842                 ret = btrfs_del_csums(trans, csum_root, bytenr, num_bytes);
2843                 if (ret) {
2844                         btrfs_abort_transaction(trans, ret);
2845                         return ret;
2846                 }
2847         }
2848
2849         ret = add_to_free_space_tree(trans, bytenr, num_bytes);
2850         if (ret) {
2851                 btrfs_abort_transaction(trans, ret);
2852                 return ret;
2853         }
2854
2855         ret = btrfs_update_block_group(trans, bytenr, num_bytes, false);
2856         if (ret)
2857                 btrfs_abort_transaction(trans, ret);
2858
2859         return ret;
2860 }
2861
2862 /*
2863  * Drop one or more refs of @node.
2864  *
2865  * 1. Locate the extent refs.
2866  *    It's either inline in EXTENT/METADATA_ITEM or in keyed SHARED_* item.
2867  *    Locate it, then reduce the refs number or remove the ref line completely.
2868  *
2869  * 2. Update the refs count in EXTENT/METADATA_ITEM
2870  *
2871  * Inline backref case:
2872  *
2873  * in extent tree we have:
2874  *
2875  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 16201 itemsize 82
2876  *              refs 2 gen 6 flags DATA
2877  *              extent data backref root FS_TREE objectid 258 offset 0 count 1
2878  *              extent data backref root FS_TREE objectid 257 offset 0 count 1
2879  *
2880  * This function gets called with:
2881  *
2882  *    node->bytenr = 13631488
2883  *    node->num_bytes = 1048576
2884  *    root_objectid = FS_TREE
2885  *    owner_objectid = 257
2886  *    owner_offset = 0
2887  *    refs_to_drop = 1
2888  *
2889  * Then we should get some like:
2890  *
2891  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 16201 itemsize 82
2892  *              refs 1 gen 6 flags DATA
2893  *              extent data backref root FS_TREE objectid 258 offset 0 count 1
2894  *
2895  * Keyed backref case:
2896  *
2897  * in extent tree we have:
2898  *
2899  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 3971 itemsize 24
2900  *              refs 754 gen 6 flags DATA
2901  *      [...]
2902  *      item 2 key (13631488 EXTENT_DATA_REF <HASH>) itemoff 3915 itemsize 28
2903  *              extent data backref root FS_TREE objectid 866 offset 0 count 1
2904  *
2905  * This function get called with:
2906  *
2907  *    node->bytenr = 13631488
2908  *    node->num_bytes = 1048576
2909  *    root_objectid = FS_TREE
2910  *    owner_objectid = 866
2911  *    owner_offset = 0
2912  *    refs_to_drop = 1
2913  *
2914  * Then we should get some like:
2915  *
2916  *      item 0 key (13631488 EXTENT_ITEM 1048576) itemoff 3971 itemsize 24
2917  *              refs 753 gen 6 flags DATA
2918  *
2919  * And that (13631488 EXTENT_DATA_REF <HASH>) gets removed.
2920  */
2921 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
2922                                struct btrfs_delayed_ref_node *node, u64 parent,
2923                                u64 root_objectid, u64 owner_objectid,
2924                                u64 owner_offset, int refs_to_drop,
2925                                struct btrfs_delayed_extent_op *extent_op)
2926 {
2927         struct btrfs_fs_info *info = trans->fs_info;
2928         struct btrfs_key key;
2929         struct btrfs_path *path;
2930         struct btrfs_root *extent_root;
2931         struct extent_buffer *leaf;
2932         struct btrfs_extent_item *ei;
2933         struct btrfs_extent_inline_ref *iref;
2934         int ret;
2935         int is_data;
2936         int extent_slot = 0;
2937         int found_extent = 0;
2938         int num_to_del = 1;
2939         u32 item_size;
2940         u64 refs;
2941         u64 bytenr = node->bytenr;
2942         u64 num_bytes = node->num_bytes;
2943         bool skinny_metadata = btrfs_fs_incompat(info, SKINNY_METADATA);
2944
2945         extent_root = btrfs_extent_root(info, bytenr);
2946         ASSERT(extent_root);
2947
2948         path = btrfs_alloc_path();
2949         if (!path)
2950                 return -ENOMEM;
2951
2952         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
2953
2954         if (!is_data && refs_to_drop != 1) {
2955                 btrfs_crit(info,
2956 "invalid refs_to_drop, dropping more than 1 refs for tree block %llu refs_to_drop %u",
2957                            node->bytenr, refs_to_drop);
2958                 ret = -EINVAL;
2959                 btrfs_abort_transaction(trans, ret);
2960                 goto out;
2961         }
2962
2963         if (is_data)
2964                 skinny_metadata = false;
2965
2966         ret = lookup_extent_backref(trans, path, &iref, bytenr, num_bytes,
2967                                     parent, root_objectid, owner_objectid,
2968                                     owner_offset);
2969         if (ret == 0) {
2970                 /*
2971                  * Either the inline backref or the SHARED_DATA_REF/
2972                  * SHARED_BLOCK_REF is found
2973                  *
2974                  * Here is a quick path to locate EXTENT/METADATA_ITEM.
2975                  * It's possible the EXTENT/METADATA_ITEM is near current slot.
2976                  */
2977                 extent_slot = path->slots[0];
2978                 while (extent_slot >= 0) {
2979                         btrfs_item_key_to_cpu(path->nodes[0], &key,
2980                                               extent_slot);
2981                         if (key.objectid != bytenr)
2982                                 break;
2983                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
2984                             key.offset == num_bytes) {
2985                                 found_extent = 1;
2986                                 break;
2987                         }
2988                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
2989                             key.offset == owner_objectid) {
2990                                 found_extent = 1;
2991                                 break;
2992                         }
2993
2994                         /* Quick path didn't find the EXTEMT/METADATA_ITEM */
2995                         if (path->slots[0] - extent_slot > 5)
2996                                 break;
2997                         extent_slot--;
2998                 }
2999
3000                 if (!found_extent) {
3001                         if (iref) {
3002                                 btrfs_crit(info,
3003 "invalid iref, no EXTENT/METADATA_ITEM found but has inline extent ref");
3004                                 btrfs_abort_transaction(trans, -EUCLEAN);
3005                                 goto err_dump;
3006                         }
3007                         /* Must be SHARED_* item, remove the backref first */
3008                         ret = remove_extent_backref(trans, extent_root, path,
3009                                                     NULL, refs_to_drop, is_data);
3010                         if (ret) {
3011                                 btrfs_abort_transaction(trans, ret);
3012                                 goto out;
3013                         }
3014                         btrfs_release_path(path);
3015
3016                         /* Slow path to locate EXTENT/METADATA_ITEM */
3017                         key.objectid = bytenr;
3018                         key.type = BTRFS_EXTENT_ITEM_KEY;
3019                         key.offset = num_bytes;
3020
3021                         if (!is_data && skinny_metadata) {
3022                                 key.type = BTRFS_METADATA_ITEM_KEY;
3023                                 key.offset = owner_objectid;
3024                         }
3025
3026                         ret = btrfs_search_slot(trans, extent_root,
3027                                                 &key, path, -1, 1);
3028                         if (ret > 0 && skinny_metadata && path->slots[0]) {
3029                                 /*
3030                                  * Couldn't find our skinny metadata item,
3031                                  * see if we have ye olde extent item.
3032                                  */
3033                                 path->slots[0]--;
3034                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
3035                                                       path->slots[0]);
3036                                 if (key.objectid == bytenr &&
3037                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
3038                                     key.offset == num_bytes)
3039                                         ret = 0;
3040                         }
3041
3042                         if (ret > 0 && skinny_metadata) {
3043                                 skinny_metadata = false;
3044                                 key.objectid = bytenr;
3045                                 key.type = BTRFS_EXTENT_ITEM_KEY;
3046                                 key.offset = num_bytes;
3047                                 btrfs_release_path(path);
3048                                 ret = btrfs_search_slot(trans, extent_root,
3049                                                         &key, path, -1, 1);
3050                         }
3051
3052                         if (ret) {
3053                                 btrfs_err(info,
3054                                           "umm, got %d back from search, was looking for %llu",
3055                                           ret, bytenr);
3056                                 if (ret > 0)
3057                                         btrfs_print_leaf(path->nodes[0]);
3058                         }
3059                         if (ret < 0) {
3060                                 btrfs_abort_transaction(trans, ret);
3061                                 goto out;
3062                         }
3063                         extent_slot = path->slots[0];
3064                 }
3065         } else if (WARN_ON(ret == -ENOENT)) {
3066                 btrfs_print_leaf(path->nodes[0]);
3067                 btrfs_err(info,
3068                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
3069                         bytenr, parent, root_objectid, owner_objectid,
3070                         owner_offset);
3071                 btrfs_abort_transaction(trans, ret);
3072                 goto out;
3073         } else {
3074                 btrfs_abort_transaction(trans, ret);
3075                 goto out;
3076         }
3077
3078         leaf = path->nodes[0];
3079         item_size = btrfs_item_size(leaf, extent_slot);
3080         if (unlikely(item_size < sizeof(*ei))) {
3081                 ret = -EINVAL;
3082                 btrfs_print_v0_err(info);
3083                 btrfs_abort_transaction(trans, ret);
3084                 goto out;
3085         }
3086         ei = btrfs_item_ptr(leaf, extent_slot,
3087                             struct btrfs_extent_item);
3088         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
3089             key.type == BTRFS_EXTENT_ITEM_KEY) {
3090                 struct btrfs_tree_block_info *bi;
3091                 if (item_size < sizeof(*ei) + sizeof(*bi)) {
3092                         btrfs_crit(info,
3093 "invalid extent item size for key (%llu, %u, %llu) owner %llu, has %u expect >= %zu",
3094                                    key.objectid, key.type, key.offset,
3095                                    owner_objectid, item_size,
3096                                    sizeof(*ei) + sizeof(*bi));
3097                         btrfs_abort_transaction(trans, -EUCLEAN);
3098                         goto err_dump;
3099                 }
3100                 bi = (struct btrfs_tree_block_info *)(ei + 1);
3101                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
3102         }
3103
3104         refs = btrfs_extent_refs(leaf, ei);
3105         if (refs < refs_to_drop) {
3106                 btrfs_crit(info,
3107                 "trying to drop %d refs but we only have %llu for bytenr %llu",
3108                           refs_to_drop, refs, bytenr);
3109                 btrfs_abort_transaction(trans, -EUCLEAN);
3110                 goto err_dump;
3111         }
3112         refs -= refs_to_drop;
3113
3114         if (refs > 0) {
3115                 if (extent_op)
3116                         __run_delayed_extent_op(extent_op, leaf, ei);
3117                 /*
3118                  * In the case of inline back ref, reference count will
3119                  * be updated by remove_extent_backref
3120                  */
3121                 if (iref) {
3122                         if (!found_extent) {
3123                                 btrfs_crit(info,
3124 "invalid iref, got inlined extent ref but no EXTENT/METADATA_ITEM found");
3125                                 btrfs_abort_transaction(trans, -EUCLEAN);
3126                                 goto err_dump;
3127                         }
3128                 } else {
3129                         btrfs_set_extent_refs(leaf, ei, refs);
3130                         btrfs_mark_buffer_dirty(leaf);
3131                 }
3132                 if (found_extent) {
3133                         ret = remove_extent_backref(trans, extent_root, path,
3134                                                     iref, refs_to_drop, is_data);
3135                         if (ret) {
3136                                 btrfs_abort_transaction(trans, ret);
3137                                 goto out;
3138                         }
3139                 }
3140         } else {
3141                 /* In this branch refs == 1 */
3142                 if (found_extent) {
3143                         if (is_data && refs_to_drop !=
3144                             extent_data_ref_count(path, iref)) {
3145                                 btrfs_crit(info,
3146                 "invalid refs_to_drop, current refs %u refs_to_drop %u",
3147                                            extent_data_ref_count(path, iref),
3148                                            refs_to_drop);
3149                                 btrfs_abort_transaction(trans, -EUCLEAN);
3150                                 goto err_dump;
3151                         }
3152                         if (iref) {
3153                                 if (path->slots[0] != extent_slot) {
3154                                         btrfs_crit(info,
3155 "invalid iref, extent item key (%llu %u %llu) doesn't have wanted iref",
3156                                                    key.objectid, key.type,
3157                                                    key.offset);
3158                                         btrfs_abort_transaction(trans, -EUCLEAN);
3159                                         goto err_dump;
3160                                 }
3161                         } else {
3162                                 /*
3163                                  * No inline ref, we must be at SHARED_* item,
3164                                  * And it's single ref, it must be:
3165                                  * |    extent_slot       ||extent_slot + 1|
3166                                  * [ EXTENT/METADATA_ITEM ][ SHARED_* ITEM ]
3167                                  */
3168                                 if (path->slots[0] != extent_slot + 1) {
3169                                         btrfs_crit(info,
3170         "invalid SHARED_* item, previous item is not EXTENT/METADATA_ITEM");
3171                                         btrfs_abort_transaction(trans, -EUCLEAN);
3172                                         goto err_dump;
3173                                 }
3174                                 path->slots[0] = extent_slot;
3175                                 num_to_del = 2;
3176                         }
3177                 }
3178
3179                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
3180                                       num_to_del);
3181                 if (ret) {
3182                         btrfs_abort_transaction(trans, ret);
3183                         goto out;
3184                 }
3185                 btrfs_release_path(path);
3186
3187                 ret = do_free_extent_accounting(trans, bytenr, num_bytes, is_data);
3188         }
3189         btrfs_release_path(path);
3190
3191 out:
3192         btrfs_free_path(path);
3193         return ret;
3194 err_dump:
3195         /*
3196          * Leaf dump can take up a lot of log buffer, so we only do full leaf
3197          * dump for debug build.
3198          */
3199         if (IS_ENABLED(CONFIG_BTRFS_DEBUG)) {
3200                 btrfs_crit(info, "path->slots[0]=%d extent_slot=%d",
3201                            path->slots[0], extent_slot);
3202                 btrfs_print_leaf(path->nodes[0]);
3203         }
3204
3205         btrfs_free_path(path);
3206         return -EUCLEAN;
3207 }
3208
3209 /*
3210  * when we free an block, it is possible (and likely) that we free the last
3211  * delayed ref for that extent as well.  This searches the delayed ref tree for
3212  * a given extent, and if there are no other delayed refs to be processed, it
3213  * removes it from the tree.
3214  */
3215 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
3216                                       u64 bytenr)
3217 {
3218         struct btrfs_delayed_ref_head *head;
3219         struct btrfs_delayed_ref_root *delayed_refs;
3220         int ret = 0;
3221
3222         delayed_refs = &trans->transaction->delayed_refs;
3223         spin_lock(&delayed_refs->lock);
3224         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
3225         if (!head)
3226                 goto out_delayed_unlock;
3227
3228         spin_lock(&head->lock);
3229         if (!RB_EMPTY_ROOT(&head->ref_tree.rb_root))
3230                 goto out;
3231
3232         if (cleanup_extent_op(head) != NULL)
3233                 goto out;
3234
3235         /*
3236          * waiting for the lock here would deadlock.  If someone else has it
3237          * locked they are already in the process of dropping it anyway
3238          */
3239         if (!mutex_trylock(&head->mutex))
3240                 goto out;
3241
3242         btrfs_delete_ref_head(delayed_refs, head);
3243         head->processing = 0;
3244
3245         spin_unlock(&head->lock);
3246         spin_unlock(&delayed_refs->lock);
3247
3248         BUG_ON(head->extent_op);
3249         if (head->must_insert_reserved)
3250                 ret = 1;
3251
3252         btrfs_cleanup_ref_head_accounting(trans->fs_info, delayed_refs, head);
3253         mutex_unlock(&head->mutex);
3254         btrfs_put_delayed_ref_head(head);
3255         return ret;
3256 out:
3257         spin_unlock(&head->lock);
3258
3259 out_delayed_unlock:
3260         spin_unlock(&delayed_refs->lock);
3261         return 0;
3262 }
3263
3264 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
3265                            u64 root_id,
3266                            struct extent_buffer *buf,
3267                            u64 parent, int last_ref)
3268 {
3269         struct btrfs_fs_info *fs_info = trans->fs_info;
3270         struct btrfs_ref generic_ref = { 0 };
3271         int ret;
3272
3273         btrfs_init_generic_ref(&generic_ref, BTRFS_DROP_DELAYED_REF,
3274                                buf->start, buf->len, parent);
3275         btrfs_init_tree_ref(&generic_ref, btrfs_header_level(buf),
3276                             root_id, 0, false);
3277
3278         if (root_id != BTRFS_TREE_LOG_OBJECTID) {
3279                 btrfs_ref_tree_mod(fs_info, &generic_ref);
3280                 ret = btrfs_add_delayed_tree_ref(trans, &generic_ref, NULL);
3281                 BUG_ON(ret); /* -ENOMEM */
3282         }
3283
3284         if (last_ref && btrfs_header_generation(buf) == trans->transid) {
3285                 struct btrfs_block_group *cache;
3286                 bool must_pin = false;
3287
3288                 if (root_id != BTRFS_TREE_LOG_OBJECTID) {
3289                         ret = check_ref_cleanup(trans, buf->start);
3290                         if (!ret) {
3291                                 btrfs_redirty_list_add(trans->transaction, buf);
3292                                 goto out;
3293                         }
3294                 }
3295
3296                 cache = btrfs_lookup_block_group(fs_info, buf->start);
3297
3298                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
3299                         pin_down_extent(trans, cache, buf->start, buf->len, 1);
3300                         btrfs_put_block_group(cache);
3301                         goto out;
3302                 }
3303
3304                 /*
3305                  * If there are tree mod log users we may have recorded mod log
3306                  * operations for this node.  If we re-allocate this node we
3307                  * could replay operations on this node that happened when it
3308                  * existed in a completely different root.  For example if it
3309                  * was part of root A, then was reallocated to root B, and we
3310                  * are doing a btrfs_old_search_slot(root b), we could replay
3311                  * operations that happened when the block was part of root A,
3312                  * giving us an inconsistent view of the btree.
3313                  *
3314                  * We are safe from races here because at this point no other
3315                  * node or root points to this extent buffer, so if after this
3316                  * check a new tree mod log user joins we will not have an
3317                  * existing log of operations on this node that we have to
3318                  * contend with.
3319                  */
3320                 if (test_bit(BTRFS_FS_TREE_MOD_LOG_USERS, &fs_info->flags))
3321                         must_pin = true;
3322
3323                 if (must_pin || btrfs_is_zoned(fs_info)) {
3324                         btrfs_redirty_list_add(trans->transaction, buf);
3325                         pin_down_extent(trans, cache, buf->start, buf->len, 1);
3326                         btrfs_put_block_group(cache);
3327                         goto out;
3328                 }
3329
3330                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
3331
3332                 btrfs_add_free_space(cache, buf->start, buf->len);
3333                 btrfs_free_reserved_bytes(cache, buf->len, 0);
3334                 btrfs_put_block_group(cache);
3335                 trace_btrfs_reserved_extent_free(fs_info, buf->start, buf->len);
3336         }
3337 out:
3338         if (last_ref) {
3339                 /*
3340                  * Deleting the buffer, clear the corrupt flag since it doesn't
3341                  * matter anymore.
3342                  */
3343                 clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
3344         }
3345 }
3346
3347 /* Can return -ENOMEM */
3348 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_ref *ref)
3349 {
3350         struct btrfs_fs_info *fs_info = trans->fs_info;
3351         int ret;
3352
3353         if (btrfs_is_testing(fs_info))
3354                 return 0;
3355
3356         /*
3357          * tree log blocks never actually go into the extent allocation
3358          * tree, just update pinning info and exit early.
3359          */
3360         if ((ref->type == BTRFS_REF_METADATA &&
3361              ref->tree_ref.owning_root == BTRFS_TREE_LOG_OBJECTID) ||
3362             (ref->type == BTRFS_REF_DATA &&
3363              ref->data_ref.owning_root == BTRFS_TREE_LOG_OBJECTID)) {
3364                 /* unlocks the pinned mutex */
3365                 btrfs_pin_extent(trans, ref->bytenr, ref->len, 1);
3366                 ret = 0;
3367         } else if (ref->type == BTRFS_REF_METADATA) {
3368                 ret = btrfs_add_delayed_tree_ref(trans, ref, NULL);
3369         } else {
3370                 ret = btrfs_add_delayed_data_ref(trans, ref, 0);
3371         }
3372
3373         if (!((ref->type == BTRFS_REF_METADATA &&
3374                ref->tree_ref.owning_root == BTRFS_TREE_LOG_OBJECTID) ||
3375               (ref->type == BTRFS_REF_DATA &&
3376                ref->data_ref.owning_root == BTRFS_TREE_LOG_OBJECTID)))
3377                 btrfs_ref_tree_mod(fs_info, ref);
3378
3379         return ret;
3380 }
3381
3382 enum btrfs_loop_type {
3383         LOOP_CACHING_NOWAIT,
3384         LOOP_CACHING_WAIT,
3385         LOOP_ALLOC_CHUNK,
3386         LOOP_NO_EMPTY_SIZE,
3387 };
3388
3389 static inline void
3390 btrfs_lock_block_group(struct btrfs_block_group *cache,
3391                        int delalloc)
3392 {
3393         if (delalloc)
3394                 down_read(&cache->data_rwsem);
3395 }
3396
3397 static inline void btrfs_grab_block_group(struct btrfs_block_group *cache,
3398                        int delalloc)
3399 {
3400         btrfs_get_block_group(cache);
3401         if (delalloc)
3402                 down_read(&cache->data_rwsem);
3403 }
3404
3405 static struct btrfs_block_group *btrfs_lock_cluster(
3406                    struct btrfs_block_group *block_group,
3407                    struct btrfs_free_cluster *cluster,
3408                    int delalloc)
3409         __acquires(&cluster->refill_lock)
3410 {
3411         struct btrfs_block_group *used_bg = NULL;
3412
3413         spin_lock(&cluster->refill_lock);
3414         while (1) {
3415                 used_bg = cluster->block_group;
3416                 if (!used_bg)
3417                         return NULL;
3418
3419                 if (used_bg == block_group)
3420                         return used_bg;
3421
3422                 btrfs_get_block_group(used_bg);
3423
3424                 if (!delalloc)
3425                         return used_bg;
3426
3427                 if (down_read_trylock(&used_bg->data_rwsem))
3428                         return used_bg;
3429
3430                 spin_unlock(&cluster->refill_lock);
3431
3432                 /* We should only have one-level nested. */
3433                 down_read_nested(&used_bg->data_rwsem, SINGLE_DEPTH_NESTING);
3434
3435                 spin_lock(&cluster->refill_lock);
3436                 if (used_bg == cluster->block_group)
3437                         return used_bg;
3438
3439                 up_read(&used_bg->data_rwsem);
3440                 btrfs_put_block_group(used_bg);
3441         }
3442 }
3443
3444 static inline void
3445 btrfs_release_block_group(struct btrfs_block_group *cache,
3446                          int delalloc)
3447 {
3448         if (delalloc)
3449                 up_read(&cache->data_rwsem);
3450         btrfs_put_block_group(cache);
3451 }
3452
3453 enum btrfs_extent_allocation_policy {
3454         BTRFS_EXTENT_ALLOC_CLUSTERED,
3455         BTRFS_EXTENT_ALLOC_ZONED,
3456 };
3457
3458 /*
3459  * Structure used internally for find_free_extent() function.  Wraps needed
3460  * parameters.
3461  */
3462 struct find_free_extent_ctl {
3463         /* Basic allocation info */
3464         u64 ram_bytes;
3465         u64 num_bytes;
3466         u64 min_alloc_size;
3467         u64 empty_size;
3468         u64 flags;
3469         int delalloc;
3470
3471         /* Where to start the search inside the bg */
3472         u64 search_start;
3473
3474         /* For clustered allocation */
3475         u64 empty_cluster;
3476         struct btrfs_free_cluster *last_ptr;
3477         bool use_cluster;
3478
3479         bool have_caching_bg;
3480         bool orig_have_caching_bg;
3481
3482         /* Allocation is called for tree-log */
3483         bool for_treelog;
3484
3485         /* Allocation is called for data relocation */
3486         bool for_data_reloc;
3487
3488         /* RAID index, converted from flags */
3489         int index;
3490
3491         /*
3492          * Current loop number, check find_free_extent_update_loop() for details
3493          */
3494         int loop;
3495
3496         /*
3497          * Whether we're refilling a cluster, if true we need to re-search
3498          * current block group but don't try to refill the cluster again.
3499          */
3500         bool retry_clustered;
3501
3502         /*
3503          * Whether we're updating free space cache, if true we need to re-search
3504          * current block group but don't try updating free space cache again.
3505          */
3506         bool retry_unclustered;
3507
3508         /* If current block group is cached */
3509         int cached;
3510
3511         /* Max contiguous hole found */
3512         u64 max_extent_size;
3513
3514         /* Total free space from free space cache, not always contiguous */
3515         u64 total_free_space;
3516
3517         /* Found result */
3518         u64 found_offset;
3519
3520         /* Hint where to start looking for an empty space */
3521         u64 hint_byte;
3522
3523         /* Allocation policy */
3524         enum btrfs_extent_allocation_policy policy;
3525 };
3526
3527
3528 /*
3529  * Helper function for find_free_extent().
3530  *
3531  * Return -ENOENT to inform caller that we need fallback to unclustered mode.
3532  * Return -EAGAIN to inform caller that we need to re-search this block group
3533  * Return >0 to inform caller that we find nothing
3534  * Return 0 means we have found a location and set ffe_ctl->found_offset.
3535  */
3536 static int find_free_extent_clustered(struct btrfs_block_group *bg,
3537                                       struct find_free_extent_ctl *ffe_ctl,
3538                                       struct btrfs_block_group **cluster_bg_ret)
3539 {
3540         struct btrfs_block_group *cluster_bg;
3541         struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
3542         u64 aligned_cluster;
3543         u64 offset;
3544         int ret;
3545
3546         cluster_bg = btrfs_lock_cluster(bg, last_ptr, ffe_ctl->delalloc);
3547         if (!cluster_bg)
3548                 goto refill_cluster;
3549         if (cluster_bg != bg && (cluster_bg->ro ||
3550             !block_group_bits(cluster_bg, ffe_ctl->flags)))
3551                 goto release_cluster;
3552
3553         offset = btrfs_alloc_from_cluster(cluster_bg, last_ptr,
3554                         ffe_ctl->num_bytes, cluster_bg->start,
3555                         &ffe_ctl->max_extent_size);
3556         if (offset) {
3557                 /* We have a block, we're done */
3558                 spin_unlock(&last_ptr->refill_lock);
3559                 trace_btrfs_reserve_extent_cluster(cluster_bg,
3560                                 ffe_ctl->search_start, ffe_ctl->num_bytes);
3561                 *cluster_bg_ret = cluster_bg;
3562                 ffe_ctl->found_offset = offset;
3563                 return 0;
3564         }
3565         WARN_ON(last_ptr->block_group != cluster_bg);
3566
3567 release_cluster:
3568         /*
3569          * If we are on LOOP_NO_EMPTY_SIZE, we can't set up a new clusters, so
3570          * lets just skip it and let the allocator find whatever block it can
3571          * find. If we reach this point, we will have tried the cluster
3572          * allocator plenty of times and not have found anything, so we are
3573          * likely way too fragmented for the clustering stuff to find anything.
3574          *
3575          * However, if the cluster is taken from the current block group,
3576          * release the cluster first, so that we stand a better chance of
3577          * succeeding in the unclustered allocation.
3578          */
3579         if (ffe_ctl->loop >= LOOP_NO_EMPTY_SIZE && cluster_bg != bg) {
3580                 spin_unlock(&last_ptr->refill_lock);
3581                 btrfs_release_block_group(cluster_bg, ffe_ctl->delalloc);
3582                 return -ENOENT;
3583         }
3584
3585         /* This cluster didn't work out, free it and start over */
3586         btrfs_return_cluster_to_free_space(NULL, last_ptr);
3587
3588         if (cluster_bg != bg)
3589                 btrfs_release_block_group(cluster_bg, ffe_ctl->delalloc);
3590
3591 refill_cluster:
3592         if (ffe_ctl->loop >= LOOP_NO_EMPTY_SIZE) {
3593                 spin_unlock(&last_ptr->refill_lock);
3594                 return -ENOENT;
3595         }
3596
3597         aligned_cluster = max_t(u64,
3598                         ffe_ctl->empty_cluster + ffe_ctl->empty_size,
3599                         bg->full_stripe_len);
3600         ret = btrfs_find_space_cluster(bg, last_ptr, ffe_ctl->search_start,
3601                         ffe_ctl->num_bytes, aligned_cluster);
3602         if (ret == 0) {
3603                 /* Now pull our allocation out of this cluster */
3604                 offset = btrfs_alloc_from_cluster(bg, last_ptr,
3605                                 ffe_ctl->num_bytes, ffe_ctl->search_start,
3606                                 &ffe_ctl->max_extent_size);
3607                 if (offset) {
3608                         /* We found one, proceed */
3609                         spin_unlock(&last_ptr->refill_lock);
3610                         trace_btrfs_reserve_extent_cluster(bg,
3611                                         ffe_ctl->search_start,
3612                                         ffe_ctl->num_bytes);
3613                         ffe_ctl->found_offset = offset;
3614                         return 0;
3615                 }
3616         } else if (!ffe_ctl->cached && ffe_ctl->loop > LOOP_CACHING_NOWAIT &&
3617                    !ffe_ctl->retry_clustered) {
3618                 spin_unlock(&last_ptr->refill_lock);
3619
3620                 ffe_ctl->retry_clustered = true;
3621                 btrfs_wait_block_group_cache_progress(bg, ffe_ctl->num_bytes +
3622                                 ffe_ctl->empty_cluster + ffe_ctl->empty_size);
3623                 return -EAGAIN;
3624         }
3625         /*
3626          * At this point we either didn't find a cluster or we weren't able to
3627          * allocate a block from our cluster.  Free the cluster we've been
3628          * trying to use, and go to the next block group.
3629          */
3630         btrfs_return_cluster_to_free_space(NULL, last_ptr);
3631         spin_unlock(&last_ptr->refill_lock);
3632         return 1;
3633 }
3634
3635 /*
3636  * Return >0 to inform caller that we find nothing
3637  * Return 0 when we found an free extent and set ffe_ctrl->found_offset
3638  * Return -EAGAIN to inform caller that we need to re-search this block group
3639  */
3640 static int find_free_extent_unclustered(struct btrfs_block_group *bg,
3641                                         struct find_free_extent_ctl *ffe_ctl)
3642 {
3643         struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
3644         u64 offset;
3645
3646         /*
3647          * We are doing an unclustered allocation, set the fragmented flag so
3648          * we don't bother trying to setup a cluster again until we get more
3649          * space.
3650          */
3651         if (unlikely(last_ptr)) {
3652                 spin_lock(&last_ptr->lock);
3653                 last_ptr->fragmented = 1;
3654                 spin_unlock(&last_ptr->lock);
3655         }
3656         if (ffe_ctl->cached) {
3657                 struct btrfs_free_space_ctl *free_space_ctl;
3658
3659                 free_space_ctl = bg->free_space_ctl;
3660                 spin_lock(&free_space_ctl->tree_lock);
3661                 if (free_space_ctl->free_space <
3662                     ffe_ctl->num_bytes + ffe_ctl->empty_cluster +
3663                     ffe_ctl->empty_size) {
3664                         ffe_ctl->total_free_space = max_t(u64,
3665                                         ffe_ctl->total_free_space,
3666                                         free_space_ctl->free_space);
3667                         spin_unlock(&free_space_ctl->tree_lock);
3668                         return 1;
3669                 }
3670                 spin_unlock(&free_space_ctl->tree_lock);
3671         }
3672
3673         offset = btrfs_find_space_for_alloc(bg, ffe_ctl->search_start,
3674                         ffe_ctl->num_bytes, ffe_ctl->empty_size,
3675                         &ffe_ctl->max_extent_size);
3676
3677         /*
3678          * If we didn't find a chunk, and we haven't failed on this block group
3679          * before, and this block group is in the middle of caching and we are
3680          * ok with waiting, then go ahead and wait for progress to be made, and
3681          * set @retry_unclustered to true.
3682          *
3683          * If @retry_unclustered is true then we've already waited on this
3684          * block group once and should move on to the next block group.
3685          */
3686         if (!offset && !ffe_ctl->retry_unclustered && !ffe_ctl->cached &&
3687             ffe_ctl->loop > LOOP_CACHING_NOWAIT) {
3688                 btrfs_wait_block_group_cache_progress(bg, ffe_ctl->num_bytes +
3689                                                       ffe_ctl->empty_size);
3690                 ffe_ctl->retry_unclustered = true;
3691                 return -EAGAIN;
3692         } else if (!offset) {
3693                 return 1;
3694         }
3695         ffe_ctl->found_offset = offset;
3696         return 0;
3697 }
3698
3699 static int do_allocation_clustered(struct btrfs_block_group *block_group,
3700                                    struct find_free_extent_ctl *ffe_ctl,
3701                                    struct btrfs_block_group **bg_ret)
3702 {
3703         int ret;
3704
3705         /* We want to try and use the cluster allocator, so lets look there */
3706         if (ffe_ctl->last_ptr && ffe_ctl->use_cluster) {
3707                 ret = find_free_extent_clustered(block_group, ffe_ctl, bg_ret);
3708                 if (ret >= 0 || ret == -EAGAIN)
3709                         return ret;
3710                 /* ret == -ENOENT case falls through */
3711         }
3712
3713         return find_free_extent_unclustered(block_group, ffe_ctl);
3714 }
3715
3716 /*
3717  * Tree-log block group locking
3718  * ============================
3719  *
3720  * fs_info::treelog_bg_lock protects the fs_info::treelog_bg which
3721  * indicates the starting address of a block group, which is reserved only
3722  * for tree-log metadata.
3723  *
3724  * Lock nesting
3725  * ============
3726  *
3727  * space_info::lock
3728  *   block_group::lock
3729  *     fs_info::treelog_bg_lock
3730  */
3731
3732 /*
3733  * Simple allocator for sequential-only block group. It only allows sequential
3734  * allocation. No need to play with trees. This function also reserves the
3735  * bytes as in btrfs_add_reserved_bytes.
3736  */
3737 static int do_allocation_zoned(struct btrfs_block_group *block_group,
3738                                struct find_free_extent_ctl *ffe_ctl,
3739                                struct btrfs_block_group **bg_ret)
3740 {
3741         struct btrfs_fs_info *fs_info = block_group->fs_info;
3742         struct btrfs_space_info *space_info = block_group->space_info;
3743         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3744         u64 start = block_group->start;
3745         u64 num_bytes = ffe_ctl->num_bytes;
3746         u64 avail;
3747         u64 bytenr = block_group->start;
3748         u64 log_bytenr;
3749         u64 data_reloc_bytenr;
3750         int ret = 0;
3751         bool skip = false;
3752
3753         ASSERT(btrfs_is_zoned(block_group->fs_info));
3754
3755         /*
3756          * Do not allow non-tree-log blocks in the dedicated tree-log block
3757          * group, and vice versa.
3758          */
3759         spin_lock(&fs_info->treelog_bg_lock);
3760         log_bytenr = fs_info->treelog_bg;
3761         if (log_bytenr && ((ffe_ctl->for_treelog && bytenr != log_bytenr) ||
3762                            (!ffe_ctl->for_treelog && bytenr == log_bytenr)))
3763                 skip = true;
3764         spin_unlock(&fs_info->treelog_bg_lock);
3765         if (skip)
3766                 return 1;
3767
3768         /*
3769          * Do not allow non-relocation blocks in the dedicated relocation block
3770          * group, and vice versa.
3771          */
3772         spin_lock(&fs_info->relocation_bg_lock);
3773         data_reloc_bytenr = fs_info->data_reloc_bg;
3774         if (data_reloc_bytenr &&
3775             ((ffe_ctl->for_data_reloc && bytenr != data_reloc_bytenr) ||
3776              (!ffe_ctl->for_data_reloc && bytenr == data_reloc_bytenr)))
3777                 skip = true;
3778         spin_unlock(&fs_info->relocation_bg_lock);
3779         if (skip)
3780                 return 1;
3781
3782         /* Check RO and no space case before trying to activate it */
3783         spin_lock(&block_group->lock);
3784         if (block_group->ro || btrfs_zoned_bg_is_full(block_group)) {
3785                 ret = 1;
3786                 /*
3787                  * May need to clear fs_info->{treelog,data_reloc}_bg.
3788                  * Return the error after taking the locks.
3789                  */
3790         }
3791         spin_unlock(&block_group->lock);
3792
3793         if (!ret && !btrfs_zone_activate(block_group)) {
3794                 ret = 1;
3795                 /*
3796                  * May need to clear fs_info->{treelog,data_reloc}_bg.
3797                  * Return the error after taking the locks.
3798                  */
3799         }
3800
3801         spin_lock(&space_info->lock);
3802         spin_lock(&block_group->lock);
3803         spin_lock(&fs_info->treelog_bg_lock);
3804         spin_lock(&fs_info->relocation_bg_lock);
3805
3806         if (ret)
3807                 goto out;
3808
3809         ASSERT(!ffe_ctl->for_treelog ||
3810                block_group->start == fs_info->treelog_bg ||
3811                fs_info->treelog_bg == 0);
3812         ASSERT(!ffe_ctl->for_data_reloc ||
3813                block_group->start == fs_info->data_reloc_bg ||
3814                fs_info->data_reloc_bg == 0);
3815
3816         if (block_group->ro ||
3817             test_bit(BLOCK_GROUP_FLAG_ZONED_DATA_RELOC, &block_group->runtime_flags)) {
3818                 ret = 1;
3819                 goto out;
3820         }
3821
3822         /*
3823          * Do not allow currently using block group to be tree-log dedicated
3824          * block group.
3825          */
3826         if (ffe_ctl->for_treelog && !fs_info->treelog_bg &&
3827             (block_group->used || block_group->reserved)) {
3828                 ret = 1;
3829                 goto out;
3830         }
3831
3832         /*
3833          * Do not allow currently used block group to be the data relocation
3834          * dedicated block group.
3835          */
3836         if (ffe_ctl->for_data_reloc && !fs_info->data_reloc_bg &&
3837             (block_group->used || block_group->reserved)) {
3838                 ret = 1;
3839                 goto out;
3840         }
3841
3842         WARN_ON_ONCE(block_group->alloc_offset > block_group->zone_capacity);
3843         avail = block_group->zone_capacity - block_group->alloc_offset;
3844         if (avail < num_bytes) {
3845                 if (ffe_ctl->max_extent_size < avail) {
3846                         /*
3847                          * With sequential allocator, free space is always
3848                          * contiguous
3849                          */
3850                         ffe_ctl->max_extent_size = avail;
3851                         ffe_ctl->total_free_space = avail;
3852                 }
3853                 ret = 1;
3854                 goto out;
3855         }
3856
3857         if (ffe_ctl->for_treelog && !fs_info->treelog_bg)
3858                 fs_info->treelog_bg = block_group->start;
3859
3860         if (ffe_ctl->for_data_reloc && !fs_info->data_reloc_bg)
3861                 fs_info->data_reloc_bg = block_group->start;
3862
3863         ffe_ctl->found_offset = start + block_group->alloc_offset;
3864         block_group->alloc_offset += num_bytes;
3865         spin_lock(&ctl->tree_lock);
3866         ctl->free_space -= num_bytes;
3867         spin_unlock(&ctl->tree_lock);
3868
3869         /*
3870          * We do not check if found_offset is aligned to stripesize. The
3871          * address is anyway rewritten when using zone append writing.
3872          */
3873
3874         ffe_ctl->search_start = ffe_ctl->found_offset;
3875
3876 out:
3877         if (ret && ffe_ctl->for_treelog)
3878                 fs_info->treelog_bg = 0;
3879         if (ret && ffe_ctl->for_data_reloc &&
3880             fs_info->data_reloc_bg == block_group->start) {
3881                 /*
3882                  * Do not allow further allocations from this block group.
3883                  * Compared to increasing the ->ro, setting the
3884                  * ->zoned_data_reloc_ongoing flag still allows nocow
3885                  *  writers to come in. See btrfs_inc_nocow_writers().
3886                  *
3887                  * We need to disable an allocation to avoid an allocation of
3888                  * regular (non-relocation data) extent. With mix of relocation
3889                  * extents and regular extents, we can dispatch WRITE commands
3890                  * (for relocation extents) and ZONE APPEND commands (for
3891                  * regular extents) at the same time to the same zone, which
3892                  * easily break the write pointer.
3893                  */
3894                 set_bit(BLOCK_GROUP_FLAG_ZONED_DATA_RELOC, &block_group->runtime_flags);
3895                 fs_info->data_reloc_bg = 0;
3896         }
3897         spin_unlock(&fs_info->relocation_bg_lock);
3898         spin_unlock(&fs_info->treelog_bg_lock);
3899         spin_unlock(&block_group->lock);
3900         spin_unlock(&space_info->lock);
3901         return ret;
3902 }
3903
3904 static int do_allocation(struct btrfs_block_group *block_group,
3905                          struct find_free_extent_ctl *ffe_ctl,
3906                          struct btrfs_block_group **bg_ret)
3907 {
3908         switch (ffe_ctl->policy) {
3909         case BTRFS_EXTENT_ALLOC_CLUSTERED:
3910                 return do_allocation_clustered(block_group, ffe_ctl, bg_ret);
3911         case BTRFS_EXTENT_ALLOC_ZONED:
3912                 return do_allocation_zoned(block_group, ffe_ctl, bg_ret);
3913         default:
3914                 BUG();
3915         }
3916 }
3917
3918 static void release_block_group(struct btrfs_block_group *block_group,
3919                                 struct find_free_extent_ctl *ffe_ctl,
3920                                 int delalloc)
3921 {
3922         switch (ffe_ctl->policy) {
3923         case BTRFS_EXTENT_ALLOC_CLUSTERED:
3924                 ffe_ctl->retry_clustered = false;
3925                 ffe_ctl->retry_unclustered = false;
3926                 break;
3927         case BTRFS_EXTENT_ALLOC_ZONED:
3928                 /* Nothing to do */
3929                 break;
3930         default:
3931                 BUG();
3932         }
3933
3934         BUG_ON(btrfs_bg_flags_to_raid_index(block_group->flags) !=
3935                ffe_ctl->index);
3936         btrfs_release_block_group(block_group, delalloc);
3937 }
3938
3939 static void found_extent_clustered(struct find_free_extent_ctl *ffe_ctl,
3940                                    struct btrfs_key *ins)
3941 {
3942         struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
3943
3944         if (!ffe_ctl->use_cluster && last_ptr) {
3945                 spin_lock(&last_ptr->lock);
3946                 last_ptr->window_start = ins->objectid;
3947                 spin_unlock(&last_ptr->lock);
3948         }
3949 }
3950
3951 static void found_extent(struct find_free_extent_ctl *ffe_ctl,
3952                          struct btrfs_key *ins)
3953 {
3954         switch (ffe_ctl->policy) {
3955         case BTRFS_EXTENT_ALLOC_CLUSTERED:
3956                 found_extent_clustered(ffe_ctl, ins);
3957                 break;
3958         case BTRFS_EXTENT_ALLOC_ZONED:
3959                 /* Nothing to do */
3960                 break;
3961         default:
3962                 BUG();
3963         }
3964 }
3965
3966 static int can_allocate_chunk_zoned(struct btrfs_fs_info *fs_info,
3967                                     struct find_free_extent_ctl *ffe_ctl)
3968 {
3969         /* If we can activate new zone, just allocate a chunk and use it */
3970         if (btrfs_can_activate_zone(fs_info->fs_devices, ffe_ctl->flags))
3971                 return 0;
3972
3973         /*
3974          * We already reached the max active zones. Try to finish one block
3975          * group to make a room for a new block group. This is only possible
3976          * for a data block group because btrfs_zone_finish() may need to wait
3977          * for a running transaction which can cause a deadlock for metadata
3978          * allocation.
3979          */
3980         if (ffe_ctl->flags & BTRFS_BLOCK_GROUP_DATA) {
3981                 int ret = btrfs_zone_finish_one_bg(fs_info);
3982
3983                 if (ret == 1)
3984                         return 0;
3985                 else if (ret < 0)
3986                         return ret;
3987         }
3988
3989         /*
3990          * If we have enough free space left in an already active block group
3991          * and we can't activate any other zone now, do not allow allocating a
3992          * new chunk and let find_free_extent() retry with a smaller size.
3993          */
3994         if (ffe_ctl->max_extent_size >= ffe_ctl->min_alloc_size)
3995                 return -ENOSPC;
3996
3997         /*
3998          * Even min_alloc_size is not left in any block groups. Since we cannot
3999          * activate a new block group, allocating it may not help. Let's tell a
4000          * caller to try again and hope it progress something by writing some
4001          * parts of the region. That is only possible for data block groups,
4002          * where a part of the region can be written.
4003          */
4004         if (ffe_ctl->flags & BTRFS_BLOCK_GROUP_DATA)
4005                 return -EAGAIN;
4006
4007         /*
4008          * We cannot activate a new block group and no enough space left in any
4009          * block groups. So, allocating a new block group may not help. But,
4010          * there is nothing to do anyway, so let's go with it.
4011          */
4012         return 0;
4013 }
4014
4015 static int can_allocate_chunk(struct btrfs_fs_info *fs_info,
4016                               struct find_free_extent_ctl *ffe_ctl)
4017 {
4018         switch (ffe_ctl->policy) {
4019         case BTRFS_EXTENT_ALLOC_CLUSTERED:
4020                 return 0;
4021         case BTRFS_EXTENT_ALLOC_ZONED:
4022                 return can_allocate_chunk_zoned(fs_info, ffe_ctl);
4023         default:
4024                 BUG();
4025         }
4026 }
4027
4028 static int chunk_allocation_failed(struct find_free_extent_ctl *ffe_ctl)
4029 {
4030         switch (ffe_ctl->policy) {
4031         case BTRFS_EXTENT_ALLOC_CLUSTERED:
4032                 /*
4033                  * If we can't allocate a new chunk we've already looped through
4034                  * at least once, move on to the NO_EMPTY_SIZE case.
4035                  */
4036                 ffe_ctl->loop = LOOP_NO_EMPTY_SIZE;
4037                 return 0;
4038         case BTRFS_EXTENT_ALLOC_ZONED:
4039                 /* Give up here */
4040                 return -ENOSPC;
4041         default:
4042                 BUG();
4043         }
4044 }
4045
4046 /*
4047  * Return >0 means caller needs to re-search for free extent
4048  * Return 0 means we have the needed free extent.
4049  * Return <0 means we failed to locate any free extent.
4050  */
4051 static int find_free_extent_update_loop(struct btrfs_fs_info *fs_info,
4052                                         struct btrfs_key *ins,
4053                                         struct find_free_extent_ctl *ffe_ctl,
4054                                         bool full_search)
4055 {
4056         struct btrfs_root *root = fs_info->chunk_root;
4057         int ret;
4058
4059         if ((ffe_ctl->loop == LOOP_CACHING_NOWAIT) &&
4060             ffe_ctl->have_caching_bg && !ffe_ctl->orig_have_caching_bg)
4061                 ffe_ctl->orig_have_caching_bg = true;
4062
4063         if (ins->objectid) {
4064                 found_extent(ffe_ctl, ins);
4065                 return 0;
4066         }
4067
4068         if (ffe_ctl->loop >= LOOP_CACHING_WAIT && ffe_ctl->have_caching_bg)
4069                 return 1;
4070
4071         ffe_ctl->index++;
4072         if (ffe_ctl->index < BTRFS_NR_RAID_TYPES)
4073                 return 1;
4074
4075         /*
4076          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
4077          *                      caching kthreads as we move along
4078          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
4079          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
4080          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
4081          *                     again
4082          */
4083         if (ffe_ctl->loop < LOOP_NO_EMPTY_SIZE) {
4084                 ffe_ctl->index = 0;
4085                 if (ffe_ctl->loop == LOOP_CACHING_NOWAIT) {
4086                         /*
4087                          * We want to skip the LOOP_CACHING_WAIT step if we
4088                          * don't have any uncached bgs and we've already done a
4089                          * full search through.
4090                          */
4091                         if (ffe_ctl->orig_have_caching_bg || !full_search)
4092                                 ffe_ctl->loop = LOOP_CACHING_WAIT;
4093                         else
4094                                 ffe_ctl->loop = LOOP_ALLOC_CHUNK;
4095                 } else {
4096                         ffe_ctl->loop++;
4097                 }
4098
4099                 if (ffe_ctl->loop == LOOP_ALLOC_CHUNK) {
4100                         struct btrfs_trans_handle *trans;
4101                         int exist = 0;
4102
4103                         /*Check if allocation policy allows to create a new chunk */
4104                         ret = can_allocate_chunk(fs_info, ffe_ctl);
4105                         if (ret)
4106                                 return ret;
4107
4108                         trans = current->journal_info;
4109                         if (trans)
4110                                 exist = 1;
4111                         else
4112                                 trans = btrfs_join_transaction(root);
4113
4114                         if (IS_ERR(trans)) {
4115                                 ret = PTR_ERR(trans);
4116                                 return ret;
4117                         }
4118
4119                         ret = btrfs_chunk_alloc(trans, ffe_ctl->flags,
4120                                                 CHUNK_ALLOC_FORCE_FOR_EXTENT);
4121
4122                         /* Do not bail out on ENOSPC since we can do more. */
4123                         if (ret == -ENOSPC)
4124                                 ret = chunk_allocation_failed(ffe_ctl);
4125                         else if (ret < 0)
4126                                 btrfs_abort_transaction(trans, ret);
4127                         else
4128                                 ret = 0;
4129                         if (!exist)
4130                                 btrfs_end_transaction(trans);
4131                         if (ret)
4132                                 return ret;
4133                 }
4134
4135                 if (ffe_ctl->loop == LOOP_NO_EMPTY_SIZE) {
4136                         if (ffe_ctl->policy != BTRFS_EXTENT_ALLOC_CLUSTERED)
4137                                 return -ENOSPC;
4138
4139                         /*
4140                          * Don't loop again if we already have no empty_size and
4141                          * no empty_cluster.
4142                          */
4143                         if (ffe_ctl->empty_size == 0 &&
4144                             ffe_ctl->empty_cluster == 0)
4145                                 return -ENOSPC;
4146                         ffe_ctl->empty_size = 0;
4147                         ffe_ctl->empty_cluster = 0;
4148                 }
4149                 return 1;
4150         }
4151         return -ENOSPC;
4152 }
4153
4154 static int prepare_allocation_clustered(struct btrfs_fs_info *fs_info,
4155                                         struct find_free_extent_ctl *ffe_ctl,
4156                                         struct btrfs_space_info *space_info,
4157                                         struct btrfs_key *ins)
4158 {
4159         /*
4160          * If our free space is heavily fragmented we may not be able to make
4161          * big contiguous allocations, so instead of doing the expensive search
4162          * for free space, simply return ENOSPC with our max_extent_size so we
4163          * can go ahead and search for a more manageable chunk.
4164          *
4165          * If our max_extent_size is large enough for our allocation simply
4166          * disable clustering since we will likely not be able to find enough
4167          * space to create a cluster and induce latency trying.
4168          */
4169         if (space_info->max_extent_size) {
4170                 spin_lock(&space_info->lock);
4171                 if (space_info->max_extent_size &&
4172                     ffe_ctl->num_bytes > space_info->max_extent_size) {
4173                         ins->offset = space_info->max_extent_size;
4174                         spin_unlock(&space_info->lock);
4175                         return -ENOSPC;
4176                 } else if (space_info->max_extent_size) {
4177                         ffe_ctl->use_cluster = false;
4178                 }
4179                 spin_unlock(&space_info->lock);
4180         }
4181
4182         ffe_ctl->last_ptr = fetch_cluster_info(fs_info, space_info,
4183                                                &ffe_ctl->empty_cluster);
4184         if (ffe_ctl->last_ptr) {
4185                 struct btrfs_free_cluster *last_ptr = ffe_ctl->last_ptr;
4186
4187                 spin_lock(&last_ptr->lock);
4188                 if (last_ptr->block_group)
4189                         ffe_ctl->hint_byte = last_ptr->window_start;
4190                 if (last_ptr->fragmented) {
4191                         /*
4192                          * We still set window_start so we can keep track of the
4193                          * last place we found an allocation to try and save
4194                          * some time.
4195                          */
4196                         ffe_ctl->hint_byte = last_ptr->window_start;
4197                         ffe_ctl->use_cluster = false;
4198                 }
4199                 spin_unlock(&last_ptr->lock);
4200         }
4201
4202         return 0;
4203 }
4204
4205 static int prepare_allocation(struct btrfs_fs_info *fs_info,
4206                               struct find_free_extent_ctl *ffe_ctl,
4207                               struct btrfs_space_info *space_info,
4208                               struct btrfs_key *ins)
4209 {
4210         switch (ffe_ctl->policy) {
4211         case BTRFS_EXTENT_ALLOC_CLUSTERED:
4212                 return prepare_allocation_clustered(fs_info, ffe_ctl,
4213                                                     space_info, ins);
4214         case BTRFS_EXTENT_ALLOC_ZONED:
4215                 if (ffe_ctl->for_treelog) {
4216                         spin_lock(&fs_info->treelog_bg_lock);
4217                         if (fs_info->treelog_bg)
4218                                 ffe_ctl->hint_byte = fs_info->treelog_bg;
4219                         spin_unlock(&fs_info->treelog_bg_lock);
4220                 }
4221                 if (ffe_ctl->for_data_reloc) {
4222                         spin_lock(&fs_info->relocation_bg_lock);
4223                         if (fs_info->data_reloc_bg)
4224                                 ffe_ctl->hint_byte = fs_info->data_reloc_bg;
4225                         spin_unlock(&fs_info->relocation_bg_lock);
4226                 }
4227                 return 0;
4228         default:
4229                 BUG();
4230         }
4231 }
4232
4233 /*
4234  * walks the btree of allocated extents and find a hole of a given size.
4235  * The key ins is changed to record the hole:
4236  * ins->objectid == start position
4237  * ins->flags = BTRFS_EXTENT_ITEM_KEY
4238  * ins->offset == the size of the hole.
4239  * Any available blocks before search_start are skipped.
4240  *
4241  * If there is no suitable free space, we will record the max size of
4242  * the free space extent currently.
4243  *
4244  * The overall logic and call chain:
4245  *
4246  * find_free_extent()
4247  * |- Iterate through all block groups
4248  * |  |- Get a valid block group
4249  * |  |- Try to do clustered allocation in that block group
4250  * |  |- Try to do unclustered allocation in that block group
4251  * |  |- Check if the result is valid
4252  * |  |  |- If valid, then exit
4253  * |  |- Jump to next block group
4254  * |
4255  * |- Push harder to find free extents
4256  *    |- If not found, re-iterate all block groups
4257  */
4258 static noinline int find_free_extent(struct btrfs_root *root,
4259                                      struct btrfs_key *ins,
4260                                      struct find_free_extent_ctl *ffe_ctl)
4261 {
4262         struct btrfs_fs_info *fs_info = root->fs_info;
4263         int ret = 0;
4264         int cache_block_group_error = 0;
4265         struct btrfs_block_group *block_group = NULL;
4266         struct btrfs_space_info *space_info;
4267         bool full_search = false;
4268
4269         WARN_ON(ffe_ctl->num_bytes < fs_info->sectorsize);
4270
4271         ffe_ctl->search_start = 0;
4272         /* For clustered allocation */
4273         ffe_ctl->empty_cluster = 0;
4274         ffe_ctl->last_ptr = NULL;
4275         ffe_ctl->use_cluster = true;
4276         ffe_ctl->have_caching_bg = false;
4277         ffe_ctl->orig_have_caching_bg = false;
4278         ffe_ctl->index = btrfs_bg_flags_to_raid_index(ffe_ctl->flags);
4279         ffe_ctl->loop = 0;
4280         /* For clustered allocation */
4281         ffe_ctl->retry_clustered = false;
4282         ffe_ctl->retry_unclustered = false;
4283         ffe_ctl->cached = 0;
4284         ffe_ctl->max_extent_size = 0;
4285         ffe_ctl->total_free_space = 0;
4286         ffe_ctl->found_offset = 0;
4287         ffe_ctl->policy = BTRFS_EXTENT_ALLOC_CLUSTERED;
4288
4289         if (btrfs_is_zoned(fs_info))
4290                 ffe_ctl->policy = BTRFS_EXTENT_ALLOC_ZONED;
4291
4292         ins->type = BTRFS_EXTENT_ITEM_KEY;
4293         ins->objectid = 0;
4294         ins->offset = 0;
4295
4296         trace_find_free_extent(root, ffe_ctl->num_bytes, ffe_ctl->empty_size,
4297                                ffe_ctl->flags);
4298
4299         space_info = btrfs_find_space_info(fs_info, ffe_ctl->flags);
4300         if (!space_info) {
4301                 btrfs_err(fs_info, "No space info for %llu", ffe_ctl->flags);
4302                 return -ENOSPC;
4303         }
4304
4305         ret = prepare_allocation(fs_info, ffe_ctl, space_info, ins);
4306         if (ret < 0)
4307                 return ret;
4308
4309         ffe_ctl->search_start = max(ffe_ctl->search_start,
4310                                     first_logical_byte(fs_info));
4311         ffe_ctl->search_start = max(ffe_ctl->search_start, ffe_ctl->hint_byte);
4312         if (ffe_ctl->search_start == ffe_ctl->hint_byte) {
4313                 block_group = btrfs_lookup_block_group(fs_info,
4314                                                        ffe_ctl->search_start);
4315                 /*
4316                  * we don't want to use the block group if it doesn't match our
4317                  * allocation bits, or if its not cached.
4318                  *
4319                  * However if we are re-searching with an ideal block group
4320                  * picked out then we don't care that the block group is cached.
4321                  */
4322                 if (block_group && block_group_bits(block_group, ffe_ctl->flags) &&
4323                     block_group->cached != BTRFS_CACHE_NO) {
4324                         down_read(&space_info->groups_sem);
4325                         if (list_empty(&block_group->list) ||
4326                             block_group->ro) {
4327                                 /*
4328                                  * someone is removing this block group,
4329                                  * we can't jump into the have_block_group
4330                                  * target because our list pointers are not
4331                                  * valid
4332                                  */
4333                                 btrfs_put_block_group(block_group);
4334                                 up_read(&space_info->groups_sem);
4335                         } else {
4336                                 ffe_ctl->index = btrfs_bg_flags_to_raid_index(
4337                                                         block_group->flags);
4338                                 btrfs_lock_block_group(block_group,
4339                                                        ffe_ctl->delalloc);
4340                                 goto have_block_group;
4341                         }
4342                 } else if (block_group) {
4343                         btrfs_put_block_group(block_group);
4344                 }
4345         }
4346 search:
4347         ffe_ctl->have_caching_bg = false;
4348         if (ffe_ctl->index == btrfs_bg_flags_to_raid_index(ffe_ctl->flags) ||
4349             ffe_ctl->index == 0)
4350                 full_search = true;
4351         down_read(&space_info->groups_sem);
4352         list_for_each_entry(block_group,
4353                             &space_info->block_groups[ffe_ctl->index], list) {
4354                 struct btrfs_block_group *bg_ret;
4355
4356                 /* If the block group is read-only, we can skip it entirely. */
4357                 if (unlikely(block_group->ro)) {
4358                         if (ffe_ctl->for_treelog)
4359                                 btrfs_clear_treelog_bg(block_group);
4360                         if (ffe_ctl->for_data_reloc)
4361                                 btrfs_clear_data_reloc_bg(block_group);
4362                         continue;
4363                 }
4364
4365                 btrfs_grab_block_group(block_group, ffe_ctl->delalloc);
4366                 ffe_ctl->search_start = block_group->start;
4367
4368                 /*
4369                  * this can happen if we end up cycling through all the
4370                  * raid types, but we want to make sure we only allocate
4371                  * for the proper type.
4372                  */
4373                 if (!block_group_bits(block_group, ffe_ctl->flags)) {
4374                         u64 extra = BTRFS_BLOCK_GROUP_DUP |
4375                                 BTRFS_BLOCK_GROUP_RAID1_MASK |
4376                                 BTRFS_BLOCK_GROUP_RAID56_MASK |
4377                                 BTRFS_BLOCK_GROUP_RAID10;
4378
4379                         /*
4380                          * if they asked for extra copies and this block group
4381                          * doesn't provide them, bail.  This does allow us to
4382                          * fill raid0 from raid1.
4383                          */
4384                         if ((ffe_ctl->flags & extra) && !(block_group->flags & extra))
4385                                 goto loop;
4386
4387                         /*
4388                          * This block group has different flags than we want.
4389                          * It's possible that we have MIXED_GROUP flag but no
4390                          * block group is mixed.  Just skip such block group.
4391                          */
4392                         btrfs_release_block_group(block_group, ffe_ctl->delalloc);
4393                         continue;
4394                 }
4395
4396 have_block_group:
4397                 ffe_ctl->cached = btrfs_block_group_done(block_group);
4398                 if (unlikely(!ffe_ctl->cached)) {
4399                         ffe_ctl->have_caching_bg = true;
4400                         ret = btrfs_cache_block_group(block_group, false);
4401
4402                         /*
4403                          * If we get ENOMEM here or something else we want to
4404                          * try other block groups, because it may not be fatal.
4405                          * However if we can't find anything else we need to
4406                          * save our return here so that we return the actual
4407                          * error that caused problems, not ENOSPC.
4408                          */
4409                         if (ret < 0) {
4410                                 if (!cache_block_group_error)
4411                                         cache_block_group_error = ret;
4412                                 ret = 0;
4413                                 goto loop;
4414                         }
4415                         ret = 0;
4416                 }
4417
4418                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
4419                         goto loop;
4420
4421                 bg_ret = NULL;
4422                 ret = do_allocation(block_group, ffe_ctl, &bg_ret);
4423                 if (ret == 0) {
4424                         if (bg_ret && bg_ret != block_group) {
4425                                 btrfs_release_block_group(block_group,
4426                                                           ffe_ctl->delalloc);
4427                                 block_group = bg_ret;
4428                         }
4429                 } else if (ret == -EAGAIN) {
4430                         goto have_block_group;
4431                 } else if (ret > 0) {
4432                         goto loop;
4433                 }
4434
4435                 /* Checks */
4436                 ffe_ctl->search_start = round_up(ffe_ctl->found_offset,
4437                                                  fs_info->stripesize);
4438
4439                 /* move on to the next group */
4440                 if (ffe_ctl->search_start + ffe_ctl->num_bytes >
4441                     block_group->start + block_group->length) {
4442                         btrfs_add_free_space_unused(block_group,
4443                                             ffe_ctl->found_offset,
4444                                             ffe_ctl->num_bytes);
4445                         goto loop;
4446                 }
4447
4448                 if (ffe_ctl->found_offset < ffe_ctl->search_start)
4449                         btrfs_add_free_space_unused(block_group,
4450                                         ffe_ctl->found_offset,
4451                                         ffe_ctl->search_start - ffe_ctl->found_offset);
4452
4453                 ret = btrfs_add_reserved_bytes(block_group, ffe_ctl->ram_bytes,
4454                                                ffe_ctl->num_bytes,
4455                                                ffe_ctl->delalloc);
4456                 if (ret == -EAGAIN) {
4457                         btrfs_add_free_space_unused(block_group,
4458                                         ffe_ctl->found_offset,
4459                                         ffe_ctl->num_bytes);
4460                         goto loop;
4461                 }
4462                 btrfs_inc_block_group_reservations(block_group);
4463
4464                 /* we are all good, lets return */
4465                 ins->objectid = ffe_ctl->search_start;
4466                 ins->offset = ffe_ctl->num_bytes;
4467
4468                 trace_btrfs_reserve_extent(block_group, ffe_ctl->search_start,
4469                                            ffe_ctl->num_bytes);
4470                 btrfs_release_block_group(block_group, ffe_ctl->delalloc);
4471                 break;
4472 loop:
4473                 release_block_group(block_group, ffe_ctl, ffe_ctl->delalloc);
4474                 cond_resched();
4475         }
4476         up_read(&space_info->groups_sem);
4477
4478         ret = find_free_extent_update_loop(fs_info, ins, ffe_ctl, full_search);
4479         if (ret > 0)
4480                 goto search;
4481
4482         if (ret == -ENOSPC && !cache_block_group_error) {
4483                 /*
4484                  * Use ffe_ctl->total_free_space as fallback if we can't find
4485                  * any contiguous hole.
4486                  */
4487                 if (!ffe_ctl->max_extent_size)
4488                         ffe_ctl->max_extent_size = ffe_ctl->total_free_space;
4489                 spin_lock(&space_info->lock);
4490                 space_info->max_extent_size = ffe_ctl->max_extent_size;
4491                 spin_unlock(&space_info->lock);
4492                 ins->offset = ffe_ctl->max_extent_size;
4493         } else if (ret == -ENOSPC) {
4494                 ret = cache_block_group_error;
4495         }
4496         return ret;
4497 }
4498
4499 /*
4500  * btrfs_reserve_extent - entry point to the extent allocator. Tries to find a
4501  *                        hole that is at least as big as @num_bytes.
4502  *
4503  * @root           -    The root that will contain this extent
4504  *
4505  * @ram_bytes      -    The amount of space in ram that @num_bytes take. This
4506  *                      is used for accounting purposes. This value differs
4507  *                      from @num_bytes only in the case of compressed extents.
4508  *
4509  * @num_bytes      -    Number of bytes to allocate on-disk.
4510  *
4511  * @min_alloc_size -    Indicates the minimum amount of space that the
4512  *                      allocator should try to satisfy. In some cases
4513  *                      @num_bytes may be larger than what is required and if
4514  *                      the filesystem is fragmented then allocation fails.
4515  *                      However, the presence of @min_alloc_size gives a
4516  *                      chance to try and satisfy the smaller allocation.
4517  *
4518  * @empty_size     -    A hint that you plan on doing more COW. This is the
4519  *                      size in bytes the allocator should try to find free
4520  *                      next to the block it returns.  This is just a hint and
4521  *                      may be ignored by the allocator.
4522  *
4523  * @hint_byte      -    Hint to the allocator to start searching above the byte
4524  *                      address passed. It might be ignored.
4525  *
4526  * @ins            -    This key is modified to record the found hole. It will
4527  *                      have the following values:
4528  *                      ins->objectid == start position
4529  *                      ins->flags = BTRFS_EXTENT_ITEM_KEY
4530  *                      ins->offset == the size of the hole.
4531  *
4532  * @is_data        -    Boolean flag indicating whether an extent is
4533  *                      allocated for data (true) or metadata (false)
4534  *
4535  * @delalloc       -    Boolean flag indicating whether this allocation is for
4536  *                      delalloc or not. If 'true' data_rwsem of block groups
4537  *                      is going to be acquired.
4538  *
4539  *
4540  * Returns 0 when an allocation succeeded or < 0 when an error occurred. In
4541  * case -ENOSPC is returned then @ins->offset will contain the size of the
4542  * largest available hole the allocator managed to find.
4543  */
4544 int btrfs_reserve_extent(struct btrfs_root *root, u64 ram_bytes,
4545                          u64 num_bytes, u64 min_alloc_size,
4546                          u64 empty_size, u64 hint_byte,
4547                          struct btrfs_key *ins, int is_data, int delalloc)
4548 {
4549         struct btrfs_fs_info *fs_info = root->fs_info;
4550         struct find_free_extent_ctl ffe_ctl = {};
4551         bool final_tried = num_bytes == min_alloc_size;
4552         u64 flags;
4553         int ret;
4554         bool for_treelog = (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
4555         bool for_data_reloc = (btrfs_is_data_reloc_root(root) && is_data);
4556
4557         flags = get_alloc_profile_by_root(root, is_data);
4558 again:
4559         WARN_ON(num_bytes < fs_info->sectorsize);
4560
4561         ffe_ctl.ram_bytes = ram_bytes;
4562         ffe_ctl.num_bytes = num_bytes;
4563         ffe_ctl.min_alloc_size = min_alloc_size;
4564         ffe_ctl.empty_size = empty_size;
4565         ffe_ctl.flags = flags;
4566         ffe_ctl.delalloc = delalloc;
4567         ffe_ctl.hint_byte = hint_byte;
4568         ffe_ctl.for_treelog = for_treelog;
4569         ffe_ctl.for_data_reloc = for_data_reloc;
4570
4571         ret = find_free_extent(root, ins, &ffe_ctl);
4572         if (!ret && !is_data) {
4573                 btrfs_dec_block_group_reservations(fs_info, ins->objectid);
4574         } else if (ret == -ENOSPC) {
4575                 if (!final_tried && ins->offset) {
4576                         num_bytes = min(num_bytes >> 1, ins->offset);
4577                         num_bytes = round_down(num_bytes,
4578                                                fs_info->sectorsize);
4579                         num_bytes = max(num_bytes, min_alloc_size);
4580                         ram_bytes = num_bytes;
4581                         if (num_bytes == min_alloc_size)
4582                                 final_tried = true;
4583                         goto again;
4584                 } else if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
4585                         struct btrfs_space_info *sinfo;
4586
4587                         sinfo = btrfs_find_space_info(fs_info, flags);
4588                         btrfs_err(fs_info,
4589         "allocation failed flags %llu, wanted %llu tree-log %d, relocation: %d",
4590                                   flags, num_bytes, for_treelog, for_data_reloc);
4591                         if (sinfo)
4592                                 btrfs_dump_space_info(fs_info, sinfo,
4593                                                       num_bytes, 1);
4594                 }
4595         }
4596
4597         return ret;
4598 }
4599
4600 int btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
4601                                u64 start, u64 len, int delalloc)
4602 {
4603         struct btrfs_block_group *cache;
4604
4605         cache = btrfs_lookup_block_group(fs_info, start);
4606         if (!cache) {
4607                 btrfs_err(fs_info, "Unable to find block group for %llu",
4608                           start);
4609                 return -ENOSPC;
4610         }
4611
4612         btrfs_add_free_space(cache, start, len);
4613         btrfs_free_reserved_bytes(cache, len, delalloc);
4614         trace_btrfs_reserved_extent_free(fs_info, start, len);
4615
4616         btrfs_put_block_group(cache);
4617         return 0;
4618 }
4619
4620 int btrfs_pin_reserved_extent(struct btrfs_trans_handle *trans, u64 start,
4621                               u64 len)
4622 {
4623         struct btrfs_block_group *cache;
4624         int ret = 0;
4625
4626         cache = btrfs_lookup_block_group(trans->fs_info, start);
4627         if (!cache) {
4628                 btrfs_err(trans->fs_info, "unable to find block group for %llu",
4629                           start);
4630                 return -ENOSPC;
4631         }
4632
4633         ret = pin_down_extent(trans, cache, start, len, 1);
4634         btrfs_put_block_group(cache);
4635         return ret;
4636 }
4637
4638 static int alloc_reserved_extent(struct btrfs_trans_handle *trans, u64 bytenr,
4639                                  u64 num_bytes)
4640 {
4641         struct btrfs_fs_info *fs_info = trans->fs_info;
4642         int ret;
4643
4644         ret = remove_from_free_space_tree(trans, bytenr, num_bytes);
4645         if (ret)
4646                 return ret;
4647
4648         ret = btrfs_update_block_group(trans, bytenr, num_bytes, true);
4649         if (ret) {
4650                 ASSERT(!ret);
4651                 btrfs_err(fs_info, "update block group failed for %llu %llu",
4652                           bytenr, num_bytes);
4653                 return ret;
4654         }
4655
4656         trace_btrfs_reserved_extent_alloc(fs_info, bytenr, num_bytes);
4657         return 0;
4658 }
4659
4660 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
4661                                       u64 parent, u64 root_objectid,
4662                                       u64 flags, u64 owner, u64 offset,
4663                                       struct btrfs_key *ins, int ref_mod)
4664 {
4665         struct btrfs_fs_info *fs_info = trans->fs_info;
4666         struct btrfs_root *extent_root;
4667         int ret;
4668         struct btrfs_extent_item *extent_item;
4669         struct btrfs_extent_inline_ref *iref;
4670         struct btrfs_path *path;
4671         struct extent_buffer *leaf;
4672         int type;
4673         u32 size;
4674
4675         if (parent > 0)
4676                 type = BTRFS_SHARED_DATA_REF_KEY;
4677         else
4678                 type = BTRFS_EXTENT_DATA_REF_KEY;
4679
4680         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
4681
4682         path = btrfs_alloc_path();
4683         if (!path)
4684                 return -ENOMEM;
4685
4686         extent_root = btrfs_extent_root(fs_info, ins->objectid);
4687         ret = btrfs_insert_empty_item(trans, extent_root, path, ins, size);
4688         if (ret) {
4689                 btrfs_free_path(path);
4690                 return ret;
4691         }
4692
4693         leaf = path->nodes[0];
4694         extent_item = btrfs_item_ptr(leaf, path->slots[0],
4695                                      struct btrfs_extent_item);
4696         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
4697         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
4698         btrfs_set_extent_flags(leaf, extent_item,
4699                                flags | BTRFS_EXTENT_FLAG_DATA);
4700
4701         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
4702         btrfs_set_extent_inline_ref_type(leaf, iref, type);
4703         if (parent > 0) {
4704                 struct btrfs_shared_data_ref *ref;
4705                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
4706                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
4707                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
4708         } else {
4709                 struct btrfs_extent_data_ref *ref;
4710                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
4711                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
4712                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
4713                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
4714                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
4715         }
4716
4717         btrfs_mark_buffer_dirty(path->nodes[0]);
4718         btrfs_free_path(path);
4719
4720         return alloc_reserved_extent(trans, ins->objectid, ins->offset);
4721 }
4722
4723 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
4724                                      struct btrfs_delayed_ref_node *node,
4725                                      struct btrfs_delayed_extent_op *extent_op)
4726 {
4727         struct btrfs_fs_info *fs_info = trans->fs_info;
4728         struct btrfs_root *extent_root;
4729         int ret;
4730         struct btrfs_extent_item *extent_item;
4731         struct btrfs_key extent_key;
4732         struct btrfs_tree_block_info *block_info;
4733         struct btrfs_extent_inline_ref *iref;
4734         struct btrfs_path *path;
4735         struct extent_buffer *leaf;
4736         struct btrfs_delayed_tree_ref *ref;
4737         u32 size = sizeof(*extent_item) + sizeof(*iref);
4738         u64 flags = extent_op->flags_to_set;
4739         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
4740
4741         ref = btrfs_delayed_node_to_tree_ref(node);
4742
4743         extent_key.objectid = node->bytenr;
4744         if (skinny_metadata) {
4745                 extent_key.offset = ref->level;
4746                 extent_key.type = BTRFS_METADATA_ITEM_KEY;
4747         } else {
4748                 extent_key.offset = node->num_bytes;
4749                 extent_key.type = BTRFS_EXTENT_ITEM_KEY;
4750                 size += sizeof(*block_info);
4751         }
4752
4753         path = btrfs_alloc_path();
4754         if (!path)
4755                 return -ENOMEM;
4756
4757         extent_root = btrfs_extent_root(fs_info, extent_key.objectid);
4758         ret = btrfs_insert_empty_item(trans, extent_root, path, &extent_key,
4759                                       size);
4760         if (ret) {
4761                 btrfs_free_path(path);
4762                 return ret;
4763         }
4764
4765         leaf = path->nodes[0];
4766         extent_item = btrfs_item_ptr(leaf, path->slots[0],
4767                                      struct btrfs_extent_item);
4768         btrfs_set_extent_refs(leaf, extent_item, 1);
4769         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
4770         btrfs_set_extent_flags(leaf, extent_item,
4771                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
4772
4773         if (skinny_metadata) {
4774                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
4775         } else {
4776                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
4777                 btrfs_set_tree_block_key(leaf, block_info, &extent_op->key);
4778                 btrfs_set_tree_block_level(leaf, block_info, ref->level);
4779                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
4780         }
4781
4782         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY) {
4783                 btrfs_set_extent_inline_ref_type(leaf, iref,
4784                                                  BTRFS_SHARED_BLOCK_REF_KEY);
4785                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->parent);
4786         } else {
4787                 btrfs_set_extent_inline_ref_type(leaf, iref,
4788                                                  BTRFS_TREE_BLOCK_REF_KEY);
4789                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->root);
4790         }
4791
4792         btrfs_mark_buffer_dirty(leaf);
4793         btrfs_free_path(path);
4794
4795         return alloc_reserved_extent(trans, node->bytenr, fs_info->nodesize);
4796 }
4797
4798 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
4799                                      struct btrfs_root *root, u64 owner,
4800                                      u64 offset, u64 ram_bytes,
4801                                      struct btrfs_key *ins)
4802 {
4803         struct btrfs_ref generic_ref = { 0 };
4804
4805         BUG_ON(root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
4806
4807         btrfs_init_generic_ref(&generic_ref, BTRFS_ADD_DELAYED_EXTENT,
4808                                ins->objectid, ins->offset, 0);
4809         btrfs_init_data_ref(&generic_ref, root->root_key.objectid, owner,
4810                             offset, 0, false);
4811         btrfs_ref_tree_mod(root->fs_info, &generic_ref);
4812
4813         return btrfs_add_delayed_data_ref(trans, &generic_ref, ram_bytes);
4814 }
4815
4816 /*
4817  * this is used by the tree logging recovery code.  It records that
4818  * an extent has been allocated and makes sure to clear the free
4819  * space cache bits as well
4820  */
4821 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
4822                                    u64 root_objectid, u64 owner, u64 offset,
4823                                    struct btrfs_key *ins)
4824 {
4825         struct btrfs_fs_info *fs_info = trans->fs_info;
4826         int ret;
4827         struct btrfs_block_group *block_group;
4828         struct btrfs_space_info *space_info;
4829
4830         /*
4831          * Mixed block groups will exclude before processing the log so we only
4832          * need to do the exclude dance if this fs isn't mixed.
4833          */
4834         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS)) {
4835                 ret = __exclude_logged_extent(fs_info, ins->objectid,
4836                                               ins->offset);
4837                 if (ret)
4838                         return ret;
4839         }
4840
4841         block_group = btrfs_lookup_block_group(fs_info, ins->objectid);
4842         if (!block_group)
4843                 return -EINVAL;
4844
4845         space_info = block_group->space_info;
4846         spin_lock(&space_info->lock);
4847         spin_lock(&block_group->lock);
4848         space_info->bytes_reserved += ins->offset;
4849         block_group->reserved += ins->offset;
4850         spin_unlock(&block_group->lock);
4851         spin_unlock(&space_info->lock);
4852
4853         ret = alloc_reserved_file_extent(trans, 0, root_objectid, 0, owner,
4854                                          offset, ins, 1);
4855         if (ret)
4856                 btrfs_pin_extent(trans, ins->objectid, ins->offset, 1);
4857         btrfs_put_block_group(block_group);
4858         return ret;
4859 }
4860
4861 static struct extent_buffer *
4862 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
4863                       u64 bytenr, int level, u64 owner,
4864                       enum btrfs_lock_nesting nest)
4865 {
4866         struct btrfs_fs_info *fs_info = root->fs_info;
4867         struct extent_buffer *buf;
4868         u64 lockdep_owner = owner;
4869
4870         buf = btrfs_find_create_tree_block(fs_info, bytenr, owner, level);
4871         if (IS_ERR(buf))
4872                 return buf;
4873
4874         /*
4875          * Extra safety check in case the extent tree is corrupted and extent
4876          * allocator chooses to use a tree block which is already used and
4877          * locked.
4878          */
4879         if (buf->lock_owner == current->pid) {
4880                 btrfs_err_rl(fs_info,
4881 "tree block %llu owner %llu already locked by pid=%d, extent tree corruption detected",
4882                         buf->start, btrfs_header_owner(buf), current->pid);
4883                 free_extent_buffer(buf);
4884                 return ERR_PTR(-EUCLEAN);
4885         }
4886
4887         /*
4888          * The reloc trees are just snapshots, so we need them to appear to be
4889          * just like any other fs tree WRT lockdep.
4890          *
4891          * The exception however is in replace_path() in relocation, where we
4892          * hold the lock on the original fs root and then search for the reloc
4893          * root.  At that point we need to make sure any reloc root buffers are
4894          * set to the BTRFS_TREE_RELOC_OBJECTID lockdep class in order to make
4895          * lockdep happy.
4896          */
4897         if (lockdep_owner == BTRFS_TREE_RELOC_OBJECTID &&
4898             !test_bit(BTRFS_ROOT_RESET_LOCKDEP_CLASS, &root->state))
4899                 lockdep_owner = BTRFS_FS_TREE_OBJECTID;
4900
4901         /* btrfs_clean_tree_block() accesses generation field. */
4902         btrfs_set_header_generation(buf, trans->transid);
4903
4904         /*
4905          * This needs to stay, because we could allocate a freed block from an
4906          * old tree into a new tree, so we need to make sure this new block is
4907          * set to the appropriate level and owner.
4908          */
4909         btrfs_set_buffer_lockdep_class(lockdep_owner, buf, level);
4910
4911         __btrfs_tree_lock(buf, nest);
4912         btrfs_clean_tree_block(buf);
4913         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
4914         clear_bit(EXTENT_BUFFER_NO_CHECK, &buf->bflags);
4915
4916         set_extent_buffer_uptodate(buf);
4917
4918         memzero_extent_buffer(buf, 0, sizeof(struct btrfs_header));
4919         btrfs_set_header_level(buf, level);
4920         btrfs_set_header_bytenr(buf, buf->start);
4921         btrfs_set_header_generation(buf, trans->transid);
4922         btrfs_set_header_backref_rev(buf, BTRFS_MIXED_BACKREF_REV);
4923         btrfs_set_header_owner(buf, owner);
4924         write_extent_buffer_fsid(buf, fs_info->fs_devices->metadata_uuid);
4925         write_extent_buffer_chunk_tree_uuid(buf, fs_info->chunk_tree_uuid);
4926         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
4927                 buf->log_index = root->log_transid % 2;
4928                 /*
4929                  * we allow two log transactions at a time, use different
4930                  * EXTENT bit to differentiate dirty pages.
4931                  */
4932                 if (buf->log_index == 0)
4933                         set_extent_dirty(&root->dirty_log_pages, buf->start,
4934                                         buf->start + buf->len - 1, GFP_NOFS);
4935                 else
4936                         set_extent_new(&root->dirty_log_pages, buf->start,
4937                                         buf->start + buf->len - 1);
4938         } else {
4939                 buf->log_index = -1;
4940                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
4941                          buf->start + buf->len - 1, GFP_NOFS);
4942         }
4943         /* this returns a buffer locked for blocking */
4944         return buf;
4945 }
4946
4947 /*
4948  * finds a free extent and does all the dirty work required for allocation
4949  * returns the tree buffer or an ERR_PTR on error.
4950  */
4951 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
4952                                              struct btrfs_root *root,
4953                                              u64 parent, u64 root_objectid,
4954                                              const struct btrfs_disk_key *key,
4955                                              int level, u64 hint,
4956                                              u64 empty_size,
4957                                              enum btrfs_lock_nesting nest)
4958 {
4959         struct btrfs_fs_info *fs_info = root->fs_info;
4960         struct btrfs_key ins;
4961         struct btrfs_block_rsv *block_rsv;
4962         struct extent_buffer *buf;
4963         struct btrfs_delayed_extent_op *extent_op;
4964         struct btrfs_ref generic_ref = { 0 };
4965         u64 flags = 0;
4966         int ret;
4967         u32 blocksize = fs_info->nodesize;
4968         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
4969
4970 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4971         if (btrfs_is_testing(fs_info)) {
4972                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
4973                                             level, root_objectid, nest);
4974                 if (!IS_ERR(buf))
4975                         root->alloc_bytenr += blocksize;
4976                 return buf;
4977         }
4978 #endif
4979
4980         block_rsv = btrfs_use_block_rsv(trans, root, blocksize);
4981         if (IS_ERR(block_rsv))
4982                 return ERR_CAST(block_rsv);
4983
4984         ret = btrfs_reserve_extent(root, blocksize, blocksize, blocksize,
4985                                    empty_size, hint, &ins, 0, 0);
4986         if (ret)
4987                 goto out_unuse;
4988
4989         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level,
4990                                     root_objectid, nest);
4991         if (IS_ERR(buf)) {
4992                 ret = PTR_ERR(buf);
4993                 goto out_free_reserved;
4994         }
4995
4996         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
4997                 if (parent == 0)
4998                         parent = ins.objectid;
4999                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
5000         } else
5001                 BUG_ON(parent > 0);
5002
5003         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
5004                 extent_op = btrfs_alloc_delayed_extent_op();
5005                 if (!extent_op) {
5006                         ret = -ENOMEM;
5007                         goto out_free_buf;
5008                 }
5009                 if (key)
5010                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
5011                 else
5012                         memset(&extent_op->key, 0, sizeof(extent_op->key));
5013                 extent_op->flags_to_set = flags;
5014                 extent_op->update_key = skinny_metadata ? false : true;
5015                 extent_op->update_flags = true;
5016                 extent_op->level = level;
5017
5018                 btrfs_init_generic_ref(&generic_ref, BTRFS_ADD_DELAYED_EXTENT,
5019                                        ins.objectid, ins.offset, parent);
5020                 btrfs_init_tree_ref(&generic_ref, level, root_objectid,
5021                                     root->root_key.objectid, false);
5022                 btrfs_ref_tree_mod(fs_info, &generic_ref);
5023                 ret = btrfs_add_delayed_tree_ref(trans, &generic_ref, extent_op);
5024                 if (ret)
5025                         goto out_free_delayed;
5026         }
5027         return buf;
5028
5029 out_free_delayed:
5030         btrfs_free_delayed_extent_op(extent_op);
5031 out_free_buf:
5032         btrfs_tree_unlock(buf);
5033         free_extent_buffer(buf);
5034 out_free_reserved:
5035         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 0);
5036 out_unuse:
5037         btrfs_unuse_block_rsv(fs_info, block_rsv, blocksize);
5038         return ERR_PTR(ret);
5039 }
5040
5041 struct walk_control {
5042         u64 refs[BTRFS_MAX_LEVEL];
5043         u64 flags[BTRFS_MAX_LEVEL];
5044         struct btrfs_key update_progress;
5045         struct btrfs_key drop_progress;
5046         int drop_level;
5047         int stage;
5048         int level;
5049         int shared_level;
5050         int update_ref;
5051         int keep_locks;
5052         int reada_slot;
5053         int reada_count;
5054         int restarted;
5055 };
5056
5057 #define DROP_REFERENCE  1
5058 #define UPDATE_BACKREF  2
5059
5060 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
5061                                      struct btrfs_root *root,
5062                                      struct walk_control *wc,
5063                                      struct btrfs_path *path)
5064 {
5065         struct btrfs_fs_info *fs_info = root->fs_info;
5066         u64 bytenr;
5067         u64 generation;
5068         u64 refs;
5069         u64 flags;
5070         u32 nritems;
5071         struct btrfs_key key;
5072         struct extent_buffer *eb;
5073         int ret;
5074         int slot;
5075         int nread = 0;
5076
5077         if (path->slots[wc->level] < wc->reada_slot) {
5078                 wc->reada_count = wc->reada_count * 2 / 3;
5079                 wc->reada_count = max(wc->reada_count, 2);
5080         } else {
5081                 wc->reada_count = wc->reada_count * 3 / 2;
5082                 wc->reada_count = min_t(int, wc->reada_count,
5083                                         BTRFS_NODEPTRS_PER_BLOCK(fs_info));
5084         }
5085
5086         eb = path->nodes[wc->level];
5087         nritems = btrfs_header_nritems(eb);
5088
5089         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
5090                 if (nread >= wc->reada_count)
5091                         break;
5092
5093                 cond_resched();
5094                 bytenr = btrfs_node_blockptr(eb, slot);
5095                 generation = btrfs_node_ptr_generation(eb, slot);
5096
5097                 if (slot == path->slots[wc->level])
5098                         goto reada;
5099
5100                 if (wc->stage == UPDATE_BACKREF &&
5101                     generation <= root->root_key.offset)
5102                         continue;
5103
5104                 /* We don't lock the tree block, it's OK to be racy here */
5105                 ret = btrfs_lookup_extent_info(trans, fs_info, bytenr,
5106                                                wc->level - 1, 1, &refs,
5107                                                &flags);
5108                 /* We don't care about errors in readahead. */
5109                 if (ret < 0)
5110                         continue;
5111                 BUG_ON(refs == 0);
5112
5113                 if (wc->stage == DROP_REFERENCE) {
5114                         if (refs == 1)
5115                                 goto reada;
5116
5117                         if (wc->level == 1 &&
5118                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5119                                 continue;
5120                         if (!wc->update_ref ||
5121                             generation <= root->root_key.offset)
5122                                 continue;
5123                         btrfs_node_key_to_cpu(eb, &key, slot);
5124                         ret = btrfs_comp_cpu_keys(&key,
5125                                                   &wc->update_progress);
5126                         if (ret < 0)
5127                                 continue;
5128                 } else {
5129                         if (wc->level == 1 &&
5130                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5131                                 continue;
5132                 }
5133 reada:
5134                 btrfs_readahead_node_child(eb, slot);
5135                 nread++;
5136         }
5137         wc->reada_slot = slot;
5138 }
5139
5140 /*
5141  * helper to process tree block while walking down the tree.
5142  *
5143  * when wc->stage == UPDATE_BACKREF, this function updates
5144  * back refs for pointers in the block.
5145  *
5146  * NOTE: return value 1 means we should stop walking down.
5147  */
5148 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
5149                                    struct btrfs_root *root,
5150                                    struct btrfs_path *path,
5151                                    struct walk_control *wc, int lookup_info)
5152 {
5153         struct btrfs_fs_info *fs_info = root->fs_info;
5154         int level = wc->level;
5155         struct extent_buffer *eb = path->nodes[level];
5156         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
5157         int ret;
5158
5159         if (wc->stage == UPDATE_BACKREF &&
5160             btrfs_header_owner(eb) != root->root_key.objectid)
5161                 return 1;
5162
5163         /*
5164          * when reference count of tree block is 1, it won't increase
5165          * again. once full backref flag is set, we never clear it.
5166          */
5167         if (lookup_info &&
5168             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
5169              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
5170                 BUG_ON(!path->locks[level]);
5171                 ret = btrfs_lookup_extent_info(trans, fs_info,
5172                                                eb->start, level, 1,
5173                                                &wc->refs[level],
5174                                                &wc->flags[level]);
5175                 BUG_ON(ret == -ENOMEM);
5176                 if (ret)
5177                         return ret;
5178                 BUG_ON(wc->refs[level] == 0);
5179         }
5180
5181         if (wc->stage == DROP_REFERENCE) {
5182                 if (wc->refs[level] > 1)
5183                         return 1;
5184
5185                 if (path->locks[level] && !wc->keep_locks) {
5186                         btrfs_tree_unlock_rw(eb, path->locks[level]);
5187                         path->locks[level] = 0;
5188                 }
5189                 return 0;
5190         }
5191
5192         /* wc->stage == UPDATE_BACKREF */
5193         if (!(wc->flags[level] & flag)) {
5194                 BUG_ON(!path->locks[level]);
5195                 ret = btrfs_inc_ref(trans, root, eb, 1);
5196                 BUG_ON(ret); /* -ENOMEM */
5197                 ret = btrfs_dec_ref(trans, root, eb, 0);
5198                 BUG_ON(ret); /* -ENOMEM */
5199                 ret = btrfs_set_disk_extent_flags(trans, eb, flag,
5200                                                   btrfs_header_level(eb));
5201                 BUG_ON(ret); /* -ENOMEM */
5202                 wc->flags[level] |= flag;
5203         }
5204
5205         /*
5206          * the block is shared by multiple trees, so it's not good to
5207          * keep the tree lock
5208          */
5209         if (path->locks[level] && level > 0) {
5210                 btrfs_tree_unlock_rw(eb, path->locks[level]);
5211                 path->locks[level] = 0;
5212         }
5213         return 0;
5214 }
5215
5216 /*
5217  * This is used to verify a ref exists for this root to deal with a bug where we
5218  * would have a drop_progress key that hadn't been updated properly.
5219  */
5220 static int check_ref_exists(struct btrfs_trans_handle *trans,
5221                             struct btrfs_root *root, u64 bytenr, u64 parent,
5222                             int level)
5223 {
5224         struct btrfs_path *path;
5225         struct btrfs_extent_inline_ref *iref;
5226         int ret;
5227
5228         path = btrfs_alloc_path();
5229         if (!path)
5230                 return -ENOMEM;
5231
5232         ret = lookup_extent_backref(trans, path, &iref, bytenr,
5233                                     root->fs_info->nodesize, parent,
5234                                     root->root_key.objectid, level, 0);
5235         btrfs_free_path(path);
5236         if (ret == -ENOENT)
5237                 return 0;
5238         if (ret < 0)
5239                 return ret;
5240         return 1;
5241 }
5242
5243 /*
5244  * helper to process tree block pointer.
5245  *
5246  * when wc->stage == DROP_REFERENCE, this function checks
5247  * reference count of the block pointed to. if the block
5248  * is shared and we need update back refs for the subtree
5249  * rooted at the block, this function changes wc->stage to
5250  * UPDATE_BACKREF. if the block is shared and there is no
5251  * need to update back, this function drops the reference
5252  * to the block.
5253  *
5254  * NOTE: return value 1 means we should stop walking down.
5255  */
5256 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
5257                                  struct btrfs_root *root,
5258                                  struct btrfs_path *path,
5259                                  struct walk_control *wc, int *lookup_info)
5260 {
5261         struct btrfs_fs_info *fs_info = root->fs_info;
5262         u64 bytenr;
5263         u64 generation;
5264         u64 parent;
5265         struct btrfs_tree_parent_check check = { 0 };
5266         struct btrfs_key key;
5267         struct btrfs_ref ref = { 0 };
5268         struct extent_buffer *next;
5269         int level = wc->level;
5270         int reada = 0;
5271         int ret = 0;
5272         bool need_account = false;
5273
5274         generation = btrfs_node_ptr_generation(path->nodes[level],
5275                                                path->slots[level]);
5276         /*
5277          * if the lower level block was created before the snapshot
5278          * was created, we know there is no need to update back refs
5279          * for the subtree
5280          */
5281         if (wc->stage == UPDATE_BACKREF &&
5282             generation <= root->root_key.offset) {
5283                 *lookup_info = 1;
5284                 return 1;
5285         }
5286
5287         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
5288
5289         check.level = level - 1;
5290         check.transid = generation;
5291         check.owner_root = root->root_key.objectid;
5292         check.has_first_key = true;
5293         btrfs_node_key_to_cpu(path->nodes[level], &check.first_key,
5294                               path->slots[level]);
5295
5296         next = find_extent_buffer(fs_info, bytenr);
5297         if (!next) {
5298                 next = btrfs_find_create_tree_block(fs_info, bytenr,
5299                                 root->root_key.objectid, level - 1);
5300                 if (IS_ERR(next))
5301                         return PTR_ERR(next);
5302                 reada = 1;
5303         }
5304         btrfs_tree_lock(next);
5305
5306         ret = btrfs_lookup_extent_info(trans, fs_info, bytenr, level - 1, 1,
5307                                        &wc->refs[level - 1],
5308                                        &wc->flags[level - 1]);
5309         if (ret < 0)
5310                 goto out_unlock;
5311
5312         if (unlikely(wc->refs[level - 1] == 0)) {
5313                 btrfs_err(fs_info, "Missing references.");
5314                 ret = -EIO;
5315                 goto out_unlock;
5316         }
5317         *lookup_info = 0;
5318
5319         if (wc->stage == DROP_REFERENCE) {
5320                 if (wc->refs[level - 1] > 1) {
5321                         need_account = true;
5322                         if (level == 1 &&
5323                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5324                                 goto skip;
5325
5326                         if (!wc->update_ref ||
5327                             generation <= root->root_key.offset)
5328                                 goto skip;
5329
5330                         btrfs_node_key_to_cpu(path->nodes[level], &key,
5331                                               path->slots[level]);
5332                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
5333                         if (ret < 0)
5334                                 goto skip;
5335
5336                         wc->stage = UPDATE_BACKREF;
5337                         wc->shared_level = level - 1;
5338                 }
5339         } else {
5340                 if (level == 1 &&
5341                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5342                         goto skip;
5343         }
5344
5345         if (!btrfs_buffer_uptodate(next, generation, 0)) {
5346                 btrfs_tree_unlock(next);
5347                 free_extent_buffer(next);
5348                 next = NULL;
5349                 *lookup_info = 1;
5350         }
5351
5352         if (!next) {
5353                 if (reada && level == 1)
5354                         reada_walk_down(trans, root, wc, path);
5355                 next = read_tree_block(fs_info, bytenr, &check);
5356                 if (IS_ERR(next)) {
5357                         return PTR_ERR(next);
5358                 } else if (!extent_buffer_uptodate(next)) {
5359                         free_extent_buffer(next);
5360                         return -EIO;
5361                 }
5362                 btrfs_tree_lock(next);
5363         }
5364
5365         level--;
5366         ASSERT(level == btrfs_header_level(next));
5367         if (level != btrfs_header_level(next)) {
5368                 btrfs_err(root->fs_info, "mismatched level");
5369                 ret = -EIO;
5370                 goto out_unlock;
5371         }
5372         path->nodes[level] = next;
5373         path->slots[level] = 0;
5374         path->locks[level] = BTRFS_WRITE_LOCK;
5375         wc->level = level;
5376         if (wc->level == 1)
5377                 wc->reada_slot = 0;
5378         return 0;
5379 skip:
5380         wc->refs[level - 1] = 0;
5381         wc->flags[level - 1] = 0;
5382         if (wc->stage == DROP_REFERENCE) {
5383                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
5384                         parent = path->nodes[level]->start;
5385                 } else {
5386                         ASSERT(root->root_key.objectid ==
5387                                btrfs_header_owner(path->nodes[level]));
5388                         if (root->root_key.objectid !=
5389                             btrfs_header_owner(path->nodes[level])) {
5390                                 btrfs_err(root->fs_info,
5391                                                 "mismatched block owner");
5392                                 ret = -EIO;
5393                                 goto out_unlock;
5394                         }
5395                         parent = 0;
5396                 }
5397
5398                 /*
5399                  * If we had a drop_progress we need to verify the refs are set
5400                  * as expected.  If we find our ref then we know that from here
5401                  * on out everything should be correct, and we can clear the
5402                  * ->restarted flag.
5403                  */
5404                 if (wc->restarted) {
5405                         ret = check_ref_exists(trans, root, bytenr, parent,
5406                                                level - 1);
5407                         if (ret < 0)
5408                                 goto out_unlock;
5409                         if (ret == 0)
5410                                 goto no_delete;
5411                         ret = 0;
5412                         wc->restarted = 0;
5413                 }
5414
5415                 /*
5416                  * Reloc tree doesn't contribute to qgroup numbers, and we have
5417                  * already accounted them at merge time (replace_path),
5418                  * thus we could skip expensive subtree trace here.
5419                  */
5420                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
5421                     need_account) {
5422                         ret = btrfs_qgroup_trace_subtree(trans, next,
5423                                                          generation, level - 1);
5424                         if (ret) {
5425                                 btrfs_err_rl(fs_info,
5426                                              "Error %d accounting shared subtree. Quota is out of sync, rescan required.",
5427                                              ret);
5428                         }
5429                 }
5430
5431                 /*
5432                  * We need to update the next key in our walk control so we can
5433                  * update the drop_progress key accordingly.  We don't care if
5434                  * find_next_key doesn't find a key because that means we're at
5435                  * the end and are going to clean up now.
5436                  */
5437                 wc->drop_level = level;
5438                 find_next_key(path, level, &wc->drop_progress);
5439
5440                 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
5441                                        fs_info->nodesize, parent);
5442                 btrfs_init_tree_ref(&ref, level - 1, root->root_key.objectid,
5443                                     0, false);
5444                 ret = btrfs_free_extent(trans, &ref);
5445                 if (ret)
5446                         goto out_unlock;
5447         }
5448 no_delete:
5449         *lookup_info = 1;
5450         ret = 1;
5451
5452 out_unlock:
5453         btrfs_tree_unlock(next);
5454         free_extent_buffer(next);
5455
5456         return ret;
5457 }
5458
5459 /*
5460  * helper to process tree block while walking up the tree.
5461  *
5462  * when wc->stage == DROP_REFERENCE, this function drops
5463  * reference count on the block.
5464  *
5465  * when wc->stage == UPDATE_BACKREF, this function changes
5466  * wc->stage back to DROP_REFERENCE if we changed wc->stage
5467  * to UPDATE_BACKREF previously while processing the block.
5468  *
5469  * NOTE: return value 1 means we should stop walking up.
5470  */
5471 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
5472                                  struct btrfs_root *root,
5473                                  struct btrfs_path *path,
5474                                  struct walk_control *wc)
5475 {
5476         struct btrfs_fs_info *fs_info = root->fs_info;
5477         int ret;
5478         int level = wc->level;
5479         struct extent_buffer *eb = path->nodes[level];
5480         u64 parent = 0;
5481
5482         if (wc->stage == UPDATE_BACKREF) {
5483                 BUG_ON(wc->shared_level < level);
5484                 if (level < wc->shared_level)
5485                         goto out;
5486
5487                 ret = find_next_key(path, level + 1, &wc->update_progress);
5488                 if (ret > 0)
5489                         wc->update_ref = 0;
5490
5491                 wc->stage = DROP_REFERENCE;
5492                 wc->shared_level = -1;
5493                 path->slots[level] = 0;
5494
5495                 /*
5496                  * check reference count again if the block isn't locked.
5497                  * we should start walking down the tree again if reference
5498                  * count is one.
5499                  */
5500                 if (!path->locks[level]) {
5501                         BUG_ON(level == 0);
5502                         btrfs_tree_lock(eb);
5503                         path->locks[level] = BTRFS_WRITE_LOCK;
5504
5505                         ret = btrfs_lookup_extent_info(trans, fs_info,
5506                                                        eb->start, level, 1,
5507                                                        &wc->refs[level],
5508                                                        &wc->flags[level]);
5509                         if (ret < 0) {
5510                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
5511                                 path->locks[level] = 0;
5512                                 return ret;
5513                         }
5514                         BUG_ON(wc->refs[level] == 0);
5515                         if (wc->refs[level] == 1) {
5516                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
5517                                 path->locks[level] = 0;
5518                                 return 1;
5519                         }
5520                 }
5521         }
5522
5523         /* wc->stage == DROP_REFERENCE */
5524         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
5525
5526         if (wc->refs[level] == 1) {
5527                 if (level == 0) {
5528                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
5529                                 ret = btrfs_dec_ref(trans, root, eb, 1);
5530                         else
5531                                 ret = btrfs_dec_ref(trans, root, eb, 0);
5532                         BUG_ON(ret); /* -ENOMEM */
5533                         if (is_fstree(root->root_key.objectid)) {
5534                                 ret = btrfs_qgroup_trace_leaf_items(trans, eb);
5535                                 if (ret) {
5536                                         btrfs_err_rl(fs_info,
5537         "error %d accounting leaf items, quota is out of sync, rescan required",
5538                                              ret);
5539                                 }
5540                         }
5541                 }
5542                 /* make block locked assertion in btrfs_clean_tree_block happy */
5543                 if (!path->locks[level] &&
5544                     btrfs_header_generation(eb) == trans->transid) {
5545                         btrfs_tree_lock(eb);
5546                         path->locks[level] = BTRFS_WRITE_LOCK;
5547                 }
5548                 btrfs_clean_tree_block(eb);
5549         }
5550
5551         if (eb == root->node) {
5552                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
5553                         parent = eb->start;
5554                 else if (root->root_key.objectid != btrfs_header_owner(eb))
5555                         goto owner_mismatch;
5556         } else {
5557                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
5558                         parent = path->nodes[level + 1]->start;
5559                 else if (root->root_key.objectid !=
5560                          btrfs_header_owner(path->nodes[level + 1]))
5561                         goto owner_mismatch;
5562         }
5563
5564         btrfs_free_tree_block(trans, btrfs_root_id(root), eb, parent,
5565                               wc->refs[level] == 1);
5566 out:
5567         wc->refs[level] = 0;
5568         wc->flags[level] = 0;
5569         return 0;
5570
5571 owner_mismatch:
5572         btrfs_err_rl(fs_info, "unexpected tree owner, have %llu expect %llu",
5573                      btrfs_header_owner(eb), root->root_key.objectid);
5574         return -EUCLEAN;
5575 }
5576
5577 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
5578                                    struct btrfs_root *root,
5579                                    struct btrfs_path *path,
5580                                    struct walk_control *wc)
5581 {
5582         int level = wc->level;
5583         int lookup_info = 1;
5584         int ret;
5585
5586         while (level >= 0) {
5587                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
5588                 if (ret > 0)
5589                         break;
5590
5591                 if (level == 0)
5592                         break;
5593
5594                 if (path->slots[level] >=
5595                     btrfs_header_nritems(path->nodes[level]))
5596                         break;
5597
5598                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
5599                 if (ret > 0) {
5600                         path->slots[level]++;
5601                         continue;
5602                 } else if (ret < 0)
5603                         return ret;
5604                 level = wc->level;
5605         }
5606         return 0;
5607 }
5608
5609 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
5610                                  struct btrfs_root *root,
5611                                  struct btrfs_path *path,
5612                                  struct walk_control *wc, int max_level)
5613 {
5614         int level = wc->level;
5615         int ret;
5616
5617         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
5618         while (level < max_level && path->nodes[level]) {
5619                 wc->level = level;
5620                 if (path->slots[level] + 1 <
5621                     btrfs_header_nritems(path->nodes[level])) {
5622                         path->slots[level]++;
5623                         return 0;
5624                 } else {
5625                         ret = walk_up_proc(trans, root, path, wc);
5626                         if (ret > 0)
5627                                 return 0;
5628                         if (ret < 0)
5629                                 return ret;
5630
5631                         if (path->locks[level]) {
5632                                 btrfs_tree_unlock_rw(path->nodes[level],
5633                                                      path->locks[level]);
5634                                 path->locks[level] = 0;
5635                         }
5636                         free_extent_buffer(path->nodes[level]);
5637                         path->nodes[level] = NULL;
5638                         level++;
5639                 }
5640         }
5641         return 1;
5642 }
5643
5644 /*
5645  * drop a subvolume tree.
5646  *
5647  * this function traverses the tree freeing any blocks that only
5648  * referenced by the tree.
5649  *
5650  * when a shared tree block is found. this function decreases its
5651  * reference count by one. if update_ref is true, this function
5652  * also make sure backrefs for the shared block and all lower level
5653  * blocks are properly updated.
5654  *
5655  * If called with for_reloc == 0, may exit early with -EAGAIN
5656  */
5657 int btrfs_drop_snapshot(struct btrfs_root *root, int update_ref, int for_reloc)
5658 {
5659         const bool is_reloc_root = (root->root_key.objectid ==
5660                                     BTRFS_TREE_RELOC_OBJECTID);
5661         struct btrfs_fs_info *fs_info = root->fs_info;
5662         struct btrfs_path *path;
5663         struct btrfs_trans_handle *trans;
5664         struct btrfs_root *tree_root = fs_info->tree_root;
5665         struct btrfs_root_item *root_item = &root->root_item;
5666         struct walk_control *wc;
5667         struct btrfs_key key;
5668         int err = 0;
5669         int ret;
5670         int level;
5671         bool root_dropped = false;
5672         bool unfinished_drop = false;
5673
5674         btrfs_debug(fs_info, "Drop subvolume %llu", root->root_key.objectid);
5675
5676         path = btrfs_alloc_path();
5677         if (!path) {
5678                 err = -ENOMEM;
5679                 goto out;
5680         }
5681
5682         wc = kzalloc(sizeof(*wc), GFP_NOFS);
5683         if (!wc) {
5684                 btrfs_free_path(path);
5685                 err = -ENOMEM;
5686                 goto out;
5687         }
5688
5689         /*
5690          * Use join to avoid potential EINTR from transaction start. See
5691          * wait_reserve_ticket and the whole reservation callchain.
5692          */
5693         if (for_reloc)
5694                 trans = btrfs_join_transaction(tree_root);
5695         else
5696                 trans = btrfs_start_transaction(tree_root, 0);
5697         if (IS_ERR(trans)) {
5698                 err = PTR_ERR(trans);
5699                 goto out_free;
5700         }
5701
5702         err = btrfs_run_delayed_items(trans);
5703         if (err)
5704                 goto out_end_trans;
5705
5706         /*
5707          * This will help us catch people modifying the fs tree while we're
5708          * dropping it.  It is unsafe to mess with the fs tree while it's being
5709          * dropped as we unlock the root node and parent nodes as we walk down
5710          * the tree, assuming nothing will change.  If something does change
5711          * then we'll have stale information and drop references to blocks we've
5712          * already dropped.
5713          */
5714         set_bit(BTRFS_ROOT_DELETING, &root->state);
5715         unfinished_drop = test_bit(BTRFS_ROOT_UNFINISHED_DROP, &root->state);
5716
5717         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
5718                 level = btrfs_header_level(root->node);
5719                 path->nodes[level] = btrfs_lock_root_node(root);
5720                 path->slots[level] = 0;
5721                 path->locks[level] = BTRFS_WRITE_LOCK;
5722                 memset(&wc->update_progress, 0,
5723                        sizeof(wc->update_progress));
5724         } else {
5725                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
5726                 memcpy(&wc->update_progress, &key,
5727                        sizeof(wc->update_progress));
5728
5729                 level = btrfs_root_drop_level(root_item);
5730                 BUG_ON(level == 0);
5731                 path->lowest_level = level;
5732                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5733                 path->lowest_level = 0;
5734                 if (ret < 0) {
5735                         err = ret;
5736                         goto out_end_trans;
5737                 }
5738                 WARN_ON(ret > 0);
5739
5740                 /*
5741                  * unlock our path, this is safe because only this
5742                  * function is allowed to delete this snapshot
5743                  */
5744                 btrfs_unlock_up_safe(path, 0);
5745
5746                 level = btrfs_header_level(root->node);
5747                 while (1) {
5748                         btrfs_tree_lock(path->nodes[level]);
5749                         path->locks[level] = BTRFS_WRITE_LOCK;
5750
5751                         ret = btrfs_lookup_extent_info(trans, fs_info,
5752                                                 path->nodes[level]->start,
5753                                                 level, 1, &wc->refs[level],
5754                                                 &wc->flags[level]);
5755                         if (ret < 0) {
5756                                 err = ret;
5757                                 goto out_end_trans;
5758                         }
5759                         BUG_ON(wc->refs[level] == 0);
5760
5761                         if (level == btrfs_root_drop_level(root_item))
5762                                 break;
5763
5764                         btrfs_tree_unlock(path->nodes[level]);
5765                         path->locks[level] = 0;
5766                         WARN_ON(wc->refs[level] != 1);
5767                         level--;
5768                 }
5769         }
5770
5771         wc->restarted = test_bit(BTRFS_ROOT_DEAD_TREE, &root->state);
5772         wc->level = level;
5773         wc->shared_level = -1;
5774         wc->stage = DROP_REFERENCE;
5775         wc->update_ref = update_ref;
5776         wc->keep_locks = 0;
5777         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
5778
5779         while (1) {
5780
5781                 ret = walk_down_tree(trans, root, path, wc);
5782                 if (ret < 0) {
5783                         err = ret;
5784                         break;
5785                 }
5786
5787                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
5788                 if (ret < 0) {
5789                         err = ret;
5790                         break;
5791                 }
5792
5793                 if (ret > 0) {
5794                         BUG_ON(wc->stage != DROP_REFERENCE);
5795                         break;
5796                 }
5797
5798                 if (wc->stage == DROP_REFERENCE) {
5799                         wc->drop_level = wc->level;
5800                         btrfs_node_key_to_cpu(path->nodes[wc->drop_level],
5801                                               &wc->drop_progress,
5802                                               path->slots[wc->drop_level]);
5803                 }
5804                 btrfs_cpu_key_to_disk(&root_item->drop_progress,
5805                                       &wc->drop_progress);
5806                 btrfs_set_root_drop_level(root_item, wc->drop_level);
5807
5808                 BUG_ON(wc->level == 0);
5809                 if (btrfs_should_end_transaction(trans) ||
5810                     (!for_reloc && btrfs_need_cleaner_sleep(fs_info))) {
5811                         ret = btrfs_update_root(trans, tree_root,
5812                                                 &root->root_key,
5813                                                 root_item);
5814                         if (ret) {
5815                                 btrfs_abort_transaction(trans, ret);
5816                                 err = ret;
5817                                 goto out_end_trans;
5818                         }
5819
5820                         if (!is_reloc_root)
5821                                 btrfs_set_last_root_drop_gen(fs_info, trans->transid);
5822
5823                         btrfs_end_transaction_throttle(trans);
5824                         if (!for_reloc && btrfs_need_cleaner_sleep(fs_info)) {
5825                                 btrfs_debug(fs_info,
5826                                             "drop snapshot early exit");
5827                                 err = -EAGAIN;
5828                                 goto out_free;
5829                         }
5830
5831                        /*
5832                         * Use join to avoid potential EINTR from transaction
5833                         * start. See wait_reserve_ticket and the whole
5834                         * reservation callchain.
5835                         */
5836                         if (for_reloc)
5837                                 trans = btrfs_join_transaction(tree_root);
5838                         else
5839                                 trans = btrfs_start_transaction(tree_root, 0);
5840                         if (IS_ERR(trans)) {
5841                                 err = PTR_ERR(trans);
5842                                 goto out_free;
5843                         }
5844                 }
5845         }
5846         btrfs_release_path(path);
5847         if (err)
5848                 goto out_end_trans;
5849
5850         ret = btrfs_del_root(trans, &root->root_key);
5851         if (ret) {
5852                 btrfs_abort_transaction(trans, ret);
5853                 err = ret;
5854                 goto out_end_trans;
5855         }
5856
5857         if (!is_reloc_root) {
5858                 ret = btrfs_find_root(tree_root, &root->root_key, path,
5859                                       NULL, NULL);
5860                 if (ret < 0) {
5861                         btrfs_abort_transaction(trans, ret);
5862                         err = ret;
5863                         goto out_end_trans;
5864                 } else if (ret > 0) {
5865                         /* if we fail to delete the orphan item this time
5866                          * around, it'll get picked up the next time.
5867                          *
5868                          * The most common failure here is just -ENOENT.
5869                          */
5870                         btrfs_del_orphan_item(trans, tree_root,
5871                                               root->root_key.objectid);
5872                 }
5873         }
5874
5875         /*
5876          * This subvolume is going to be completely dropped, and won't be
5877          * recorded as dirty roots, thus pertrans meta rsv will not be freed at
5878          * commit transaction time.  So free it here manually.
5879          */
5880         btrfs_qgroup_convert_reserved_meta(root, INT_MAX);
5881         btrfs_qgroup_free_meta_all_pertrans(root);
5882
5883         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state))
5884                 btrfs_add_dropped_root(trans, root);
5885         else
5886                 btrfs_put_root(root);
5887         root_dropped = true;
5888 out_end_trans:
5889         if (!is_reloc_root)
5890                 btrfs_set_last_root_drop_gen(fs_info, trans->transid);
5891
5892         btrfs_end_transaction_throttle(trans);
5893 out_free:
5894         kfree(wc);
5895         btrfs_free_path(path);
5896 out:
5897         /*
5898          * We were an unfinished drop root, check to see if there are any
5899          * pending, and if not clear and wake up any waiters.
5900          */
5901         if (!err && unfinished_drop)
5902                 btrfs_maybe_wake_unfinished_drop(fs_info);
5903
5904         /*
5905          * So if we need to stop dropping the snapshot for whatever reason we
5906          * need to make sure to add it back to the dead root list so that we
5907          * keep trying to do the work later.  This also cleans up roots if we
5908          * don't have it in the radix (like when we recover after a power fail
5909          * or unmount) so we don't leak memory.
5910          */
5911         if (!for_reloc && !root_dropped)
5912                 btrfs_add_dead_root(root);
5913         return err;
5914 }
5915
5916 /*
5917  * drop subtree rooted at tree block 'node'.
5918  *
5919  * NOTE: this function will unlock and release tree block 'node'
5920  * only used by relocation code
5921  */
5922 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
5923                         struct btrfs_root *root,
5924                         struct extent_buffer *node,
5925                         struct extent_buffer *parent)
5926 {
5927         struct btrfs_fs_info *fs_info = root->fs_info;
5928         struct btrfs_path *path;
5929         struct walk_control *wc;
5930         int level;
5931         int parent_level;
5932         int ret = 0;
5933         int wret;
5934
5935         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
5936
5937         path = btrfs_alloc_path();
5938         if (!path)
5939                 return -ENOMEM;
5940
5941         wc = kzalloc(sizeof(*wc), GFP_NOFS);
5942         if (!wc) {
5943                 btrfs_free_path(path);
5944                 return -ENOMEM;
5945         }
5946
5947         btrfs_assert_tree_write_locked(parent);
5948         parent_level = btrfs_header_level(parent);
5949         atomic_inc(&parent->refs);
5950         path->nodes[parent_level] = parent;
5951         path->slots[parent_level] = btrfs_header_nritems(parent);
5952
5953         btrfs_assert_tree_write_locked(node);
5954         level = btrfs_header_level(node);
5955         path->nodes[level] = node;
5956         path->slots[level] = 0;
5957         path->locks[level] = BTRFS_WRITE_LOCK;
5958
5959         wc->refs[parent_level] = 1;
5960         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
5961         wc->level = level;
5962         wc->shared_level = -1;
5963         wc->stage = DROP_REFERENCE;
5964         wc->update_ref = 0;
5965         wc->keep_locks = 1;
5966         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
5967
5968         while (1) {
5969                 wret = walk_down_tree(trans, root, path, wc);
5970                 if (wret < 0) {
5971                         ret = wret;
5972                         break;
5973                 }
5974
5975                 wret = walk_up_tree(trans, root, path, wc, parent_level);
5976                 if (wret < 0)
5977                         ret = wret;
5978                 if (wret != 0)
5979                         break;
5980         }
5981
5982         kfree(wc);
5983         btrfs_free_path(path);
5984         return ret;
5985 }
5986
5987 int btrfs_error_unpin_extent_range(struct btrfs_fs_info *fs_info,
5988                                    u64 start, u64 end)
5989 {
5990         return unpin_extent_range(fs_info, start, end, false);
5991 }
5992
5993 /*
5994  * It used to be that old block groups would be left around forever.
5995  * Iterating over them would be enough to trim unused space.  Since we
5996  * now automatically remove them, we also need to iterate over unallocated
5997  * space.
5998  *
5999  * We don't want a transaction for this since the discard may take a
6000  * substantial amount of time.  We don't require that a transaction be
6001  * running, but we do need to take a running transaction into account
6002  * to ensure that we're not discarding chunks that were released or
6003  * allocated in the current transaction.
6004  *
6005  * Holding the chunks lock will prevent other threads from allocating
6006  * or releasing chunks, but it won't prevent a running transaction
6007  * from committing and releasing the memory that the pending chunks
6008  * list head uses.  For that, we need to take a reference to the
6009  * transaction and hold the commit root sem.  We only need to hold
6010  * it while performing the free space search since we have already
6011  * held back allocations.
6012  */
6013 static int btrfs_trim_free_extents(struct btrfs_device *device, u64 *trimmed)
6014 {
6015         u64 start = BTRFS_DEVICE_RANGE_RESERVED, len = 0, end = 0;
6016         int ret;
6017
6018         *trimmed = 0;
6019
6020         /* Discard not supported = nothing to do. */
6021         if (!bdev_max_discard_sectors(device->bdev))
6022                 return 0;
6023
6024         /* Not writable = nothing to do. */
6025         if (!test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state))
6026                 return 0;
6027
6028         /* No free space = nothing to do. */
6029         if (device->total_bytes <= device->bytes_used)
6030                 return 0;
6031
6032         ret = 0;
6033
6034         while (1) {
6035                 struct btrfs_fs_info *fs_info = device->fs_info;
6036                 u64 bytes;
6037
6038                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
6039                 if (ret)
6040                         break;
6041
6042                 find_first_clear_extent_bit(&device->alloc_state, start,
6043                                             &start, &end,
6044                                             CHUNK_TRIMMED | CHUNK_ALLOCATED);
6045
6046                 /* Check if there are any CHUNK_* bits left */
6047                 if (start > device->total_bytes) {
6048                         WARN_ON(IS_ENABLED(CONFIG_BTRFS_DEBUG));
6049                         btrfs_warn_in_rcu(fs_info,
6050 "ignoring attempt to trim beyond device size: offset %llu length %llu device %s device size %llu",
6051                                           start, end - start + 1,
6052                                           btrfs_dev_name(device),
6053                                           device->total_bytes);
6054                         mutex_unlock(&fs_info->chunk_mutex);
6055                         ret = 0;
6056                         break;
6057                 }
6058
6059                 /* Ensure we skip the reserved space on each device. */
6060                 start = max_t(u64, start, BTRFS_DEVICE_RANGE_RESERVED);
6061
6062                 /*
6063                  * If find_first_clear_extent_bit find a range that spans the
6064                  * end of the device it will set end to -1, in this case it's up
6065                  * to the caller to trim the value to the size of the device.
6066                  */
6067                 end = min(end, device->total_bytes - 1);
6068
6069                 len = end - start + 1;
6070
6071                 /* We didn't find any extents */
6072                 if (!len) {
6073                         mutex_unlock(&fs_info->chunk_mutex);
6074                         ret = 0;
6075                         break;
6076                 }
6077
6078                 ret = btrfs_issue_discard(device->bdev, start, len,
6079                                           &bytes);
6080                 if (!ret)
6081                         set_extent_bits(&device->alloc_state, start,
6082                                         start + bytes - 1,
6083                                         CHUNK_TRIMMED);
6084                 mutex_unlock(&fs_info->chunk_mutex);
6085
6086                 if (ret)
6087                         break;
6088
6089                 start += len;
6090                 *trimmed += bytes;
6091
6092                 if (fatal_signal_pending(current)) {
6093                         ret = -ERESTARTSYS;
6094                         break;
6095                 }
6096
6097                 cond_resched();
6098         }
6099
6100         return ret;
6101 }
6102
6103 /*
6104  * Trim the whole filesystem by:
6105  * 1) trimming the free space in each block group
6106  * 2) trimming the unallocated space on each device
6107  *
6108  * This will also continue trimming even if a block group or device encounters
6109  * an error.  The return value will be the last error, or 0 if nothing bad
6110  * happens.
6111  */
6112 int btrfs_trim_fs(struct btrfs_fs_info *fs_info, struct fstrim_range *range)
6113 {
6114         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
6115         struct btrfs_block_group *cache = NULL;
6116         struct btrfs_device *device;
6117         u64 group_trimmed;
6118         u64 range_end = U64_MAX;
6119         u64 start;
6120         u64 end;
6121         u64 trimmed = 0;
6122         u64 bg_failed = 0;
6123         u64 dev_failed = 0;
6124         int bg_ret = 0;
6125         int dev_ret = 0;
6126         int ret = 0;
6127
6128         if (range->start == U64_MAX)
6129                 return -EINVAL;
6130
6131         /*
6132          * Check range overflow if range->len is set.
6133          * The default range->len is U64_MAX.
6134          */
6135         if (range->len != U64_MAX &&
6136             check_add_overflow(range->start, range->len, &range_end))
6137                 return -EINVAL;
6138
6139         cache = btrfs_lookup_first_block_group(fs_info, range->start);
6140         for (; cache; cache = btrfs_next_block_group(cache)) {
6141                 if (cache->start >= range_end) {
6142                         btrfs_put_block_group(cache);
6143                         break;
6144                 }
6145
6146                 start = max(range->start, cache->start);
6147                 end = min(range_end, cache->start + cache->length);
6148
6149                 if (end - start >= range->minlen) {
6150                         if (!btrfs_block_group_done(cache)) {
6151                                 ret = btrfs_cache_block_group(cache, true);
6152                                 if (ret) {
6153                                         bg_failed++;
6154                                         bg_ret = ret;
6155                                         continue;
6156                                 }
6157                         }
6158                         ret = btrfs_trim_block_group(cache,
6159                                                      &group_trimmed,
6160                                                      start,
6161                                                      end,
6162                                                      range->minlen);
6163
6164                         trimmed += group_trimmed;
6165                         if (ret) {
6166                                 bg_failed++;
6167                                 bg_ret = ret;
6168                                 continue;
6169                         }
6170                 }
6171         }
6172
6173         if (bg_failed)
6174                 btrfs_warn(fs_info,
6175                         "failed to trim %llu block group(s), last error %d",
6176                         bg_failed, bg_ret);
6177
6178         mutex_lock(&fs_devices->device_list_mutex);
6179         list_for_each_entry(device, &fs_devices->devices, dev_list) {
6180                 if (test_bit(BTRFS_DEV_STATE_MISSING, &device->dev_state))
6181                         continue;
6182
6183                 ret = btrfs_trim_free_extents(device, &group_trimmed);
6184                 if (ret) {
6185                         dev_failed++;
6186                         dev_ret = ret;
6187                         break;
6188                 }
6189
6190                 trimmed += group_trimmed;
6191         }
6192         mutex_unlock(&fs_devices->device_list_mutex);
6193
6194         if (dev_failed)
6195                 btrfs_warn(fs_info,
6196                         "failed to trim %llu device(s), last error %d",
6197                         dev_failed, dev_ret);
6198         range->len = trimmed;
6199         if (bg_ret)
6200                 return bg_ret;
6201         return dev_ret;
6202 }